The elephant in the room: getting value from Big Data

Big Data, and its 4 Vs – volume, velocity, variety, and veracity – have been at the forefront of societal, scientific and engineering discourse. Arguably the most important 5th V, value, is not talked about as much. How can we make sure that our data is not just big, but also valuable? WebDB 2015, the premier workshop on Web and Databases, focuses on this important topic this year. To set the stage, we have interviewed several prominent members of the data management community, soliciting their opinions on how we can ensure that data is not just available in quantity, but also in quality.

We interviewed Serge Abiteboul (INRIA Saclay & ENS Cachan), Oren Etzioni (Allen Institute for Artificial Intelligence), Divesh Srivastava (AT&T Labs-Research) with Luna Dong (Google Inc.), and Gerhard Weikum (Max Planck Institute for Informatics). We asked them about their motivation for doing research in the area of data quality, their current work, and their view on the future of the field.

Serge Abiteboul is a Senior Researcher INRIA Saclay, and an affiliated professor at Ecole Normale Supérieure de Cachan. He obtained his Ph.D. from the University of Southern California, and a State Doctoral Thesis from the University of Paris-Sud. He was a Lecturer at the École Polytechnique and Visiting Professor at Stanford and Oxford University. He has been Chair Professor at Collège de France in 2011-12 and Francqui Chair Professor at Namur University in 2012-2013. He co-founded the company Xyleme in 2000. Serge Abiteboul has received the ACM SIGMOD Innovation Award in 1998, the EADS Award from the French Academy of sciences in 2007; the Milner Award from the Royal Society in 2013; and a European Research Council Fellowship (2008-2013). He became a member of the French Academy of Sciences in 2008, and a member the Academy of Europe in 2011. He is a member of the Conseil national du numérique. His research work focuses mainly on data, information and knowledge management, particularly on the Web.

What is your motivation for doing research on the value of Big Data?

My experience is that it is getting easier and easier to get data but if you are not careful all you get is garbage. So quality is extremely important, never over-valued and certainly relevant. For instance, with some students we crawled the French Web. If you crawl naively, it turns out that very rapidly all the URLs you try to load are wrong, meaning they do not correspond to real pages, or they return pages without real content. You need to use something such as PageRank to focus your resources on relevant pages.

So then what is your current work for finding the equivalent of “relevant pages” in Big Data?

I am working on personal information where very often, the difficulty is to get the proper knowledge and, for instance, align correctly entities from different sources. My long-term goal also working for instance with Amélie Marian is the construction of a Personal Knowledge Base that gathers all the knowledge someone can get about his/her life. For each one of us, such knowledge has enormous potential value, but for the moment it lives in different silos and we cannot get this value.

This line of work is not purely technical, but involves societal issues as well. We are living in a world where companies and governments have loads of data on us and we don’t even know what they have and how they are using it. Personal Information Management is an attempt to rebalance the situation, and make personal data more easily accessible to the individuals. I have a paper on Personal Information Management Systems, talking about just that, to appear in CACM (with Benjamin André and Daniel Kaplan).

And what is your view of the killer app of Big Data?

Relational databases was a big technical success in the 1970s-80s. Recommendation of information was a big one in the 1990s-2000s, from PageRank to social recommendation. After data, after information, the next big technical success is going to be “knowledge”, say in the 2010s-20s :). It is not an easy sell because knowledge management has often been disappointing – not delivering on its promises. By knowledge management, I mean systems capable of acquiring knowledge at a large scale, reasoning with this knowledge, exchanging knowledge in a distributed manner. I mean techniques such as that used at Berkeley with Bud or at INRIA with Webdamlog. To build such systems, beyond scale and distribution, we have to solve quality issues: the knowledge is going to be imprecise, possibly missing, with inconsistencies. I see knowledge management as the next killer app!

Oren Etzioni is Chief Executive Officer of the Allen Institute for Artificial Intelligence. He has been a Professor at the University of Washington’s Computer Science department starting in 1991, receiving several awards including GeekWire’s Geek of the Year (2013), the Robert Engelmore Memorial Award (2007), the IJCAI Distinguished Paper Award (2005), AAAI Fellow (2003), and a National Young Investigator Award (1993). He was also the founder or co-founder of several companies including Farecast (sold to Microsoft in 2008) and Decide (sold to eBay in 2013), and the author of over 100 technical papers that have garnered over 22,000 citations. The goal of Oren’s research is to solve fundamental problems in AI, particularly the automatic learning of knowledge from text. Oren received his Ph.D. from Carnegie Mellon University in 1991, and his B.A. from Harvard in 1986.

Oren, how did you get started in your work on Big Data?

I like to say that I’ve been working on Big Data from the early days when it was only “small data”. Our 2003 KDD paper on predictive pricing started with a data set with 12K data points. By the time Farecast was sold to Microsoft, in 2008, we were approaching a trillion labeled data points. Big price data was the essence of Farecast’s predictive model, and had the elegant property that it was “self labeling”. That is, if we can label the airfare on a flight from Seattle to Boston with either a “buy now” or “wait” label—all we have to do is monitor the price movement over time to determine the appropriate label. 20/20 hindsight allows us to produce labels automatically. But for Farecast, and other applications of Big Data, the labeled data points are only part of the story. Background knowledge, reasoning, and more sophisticated semantic models are necessary to take predictive accuracy to the next level.

So what is the AI2 working on to bring us to this next level?

Beginning in January 1, 2014 we launched the Allen Institute for AI, a research center dedicated to leveraging modern data mining, text mining, and more in order to make progress on fundamental AI questions, and to develop high-impact AI applications.
One key project is Semantic Scholar, which utilizes big data over millions of academic papers to revolutionize the process of homing in on relevant papers and important citations. We are developing information extraction methods that map PDF files to key attributes including the problem in the paper, the methods used, the data sets employed, and the results reported. See this link for more information, and to be notified when Semantic Scholar launches as a free service later in 2015. Beyond text, we find that figures are an important source of information in academic papers and have begun a research program to extract figures from the papers and analyze them. The first results in this research program (including open-source software for extracting figures) are available here.

And thinking ahead, what would be the killer application that you have in mind for Big Data?

Ideas like “background knowledge” and “common-sense reasoning” are investigated in AI whereas Big Data and data mining has developed into its own vibrant community. Over the next 10 years, I see the potential for these communities to re-engage with the goal of producing methods that are still scalable, but require less manual engineering and “human intelligence” to work. The killer application would be a Big Data application that easily adapts to a new domain, and that doesn’t make egregious errors because it has “more intelligence”.
More concretely, we are looking at systems like Semantic Scholar, which will operate over a graph of more than 100M papers linked by citations to each other, as an application that will drive exciting research in AI and Big Data methods coming together to make literature search, which scientists and doctors do daily, more efficient than ever.

Divesh Srivastava is the head of the Database Research Department at AT&T Labs-Research. He received his Ph.D. from the University of Wisconsin, Madison, and his B.Tech from the Indian Institute of Technology, Bombay. He is an ACM fellow, on the board of trustees of the VLDB Endowment, the managing editor of the Proceedings of the VLDB Endowment (PVLDB), and an associate editor of the ACM Transactions on Database Systems. His research interests and publications span a variety of topics in data management.

Xin Luna Dong is a senior research scientist at Google. She works on enriching and cleaning knowledge for the Google Knowledge Graph. Her research interest includes data integration, data cleaning, and knowledge management. Prior to joining Google, she worked for AT&T Labs – Research and received her Ph.D. in Computer Science and Engineering at the University of Washington. She is the co-chair for WAIM’15 and has served as an area chair for SIGMOD’15, ICDE’13, and CIKM’11. She won the best-demo award in SIGMOD’05.

Divesh and Luna, you have been working on several aspects of Big Data Value. What attracts you to this topic?

Value, the 5th V of big data, is arguably the promise of the big data era. The choices of what data to collect and integrate, what analyses to perform, and what data-driven decisions to make, are driven by their perceived value – to society, to organizations, and to individuals. It is worth noting that while value and quality of big data may be correlated, they are conceptually different. For example, one can have high quality data about the names of all the countries in North America, but this list of names may not have much perceived value. In contrast, even relatively incomplete data about the shopping habits of people can be quite valuable to online advertisers.

It should not be surprising that early efforts to extract value from big data have focused on integrating and extracting knowledge from the low-hanging fruit of “head” data – data about popular entities, in the current world, from large sources. This is true both in industry (often heavily relying on manual curation) and in academia (often as underlying assumptions of the proposed integration techniques). However, focusing exclusively on head data leaves behind a considerable volume of “tail” data, including data about less popular entities, in less popular verticals, about non-current (historical) facts, from smaller sources, in languages other than English, and so on. While each data item in the “long tail” may provide only little value, the total value present in the long tail can be substantial, possibly even exceeding the total value that can be extracted solely from head data. This is akin to shops making a big profit from a large number of specialty items, each sold in a small quantity, in addition to the profit made by selling large quantities of a few popular items.

We issue a call to arms – “leave no valuable data behind” in our quest to extract significant value from big data.

What is your recent work in this quest?

Our work in this area focuses on the acquisition, integration, and knowledge extraction from big data. More recently, we have been considering a variety of ideas, including looking at collaboratively edited databases, news stories, and “local” information, where multiple perspectives and timeliness can be even more important than guaranteeing extremely high accuracy (e.g., 99% accuracy requirement for Google’s Knowledge Graph).

We started this body of work a few years ago with the Solomon project for data fusion, to make wise decisions about finding the truth when faced with conflicting information from multiple sources. We quickly realized the importance of copy detection between sources of structured data to solve this problem, and developed techniques that iteratively perform copy detection, source trustworthiness evaluation, and truth discovery. The Knowledge Vault (KV) project and the Sonya project naturally extend the Solomon project to address the challenge of web-scale data. They focus on knowledge fusion, finding truthfulness of extracted knowledge from web-scale data (see here), and building probabilistic knowledge bases, in the presence of source errors and extraction errors, with the latter dominating (see here). The Sonya project in addition measures knowledge-based trust, determining the trustworthiness of web sources based on the correctness of the facts they provide.

Big data often has a temporal dimension, reflecting the dynamic nature of the real-world, with evolving entities, relationships and stories. Over the years we have worked on many big data integration problems dealing with evolving data. For example, our work on temporal record linkage addressed the challenging problem of entity resolution over time, which has to deal with evolution of entities wherein their attribute values can change over time, as well as the possibility that different entities are more likely to share similar attribute values over time. We have also looked at quality issues in collaboratively edited databases, with some recent work on automatically identifying fine-grained controversies over time in Wikipedia articles (upcoming paper in ICDE 2015).

More recently, we have been working on the novel topic of data source management, which is of increasing interest because of the proliferation of a large number of data sources in almost every domain of interest. Our initial research on this topic involves assessing the evolving quality of data sources, and enabling the discovery of valuable sources to integrate before actually performing the integration (see here and here).

Finally, we make a shameless plug for our new book “Big Data Integration” that should be published very soon, which we hope will serve as a starting point for interested readers to pursue additional work on this exciting topic.

And where do you think will research head tomorrow?

In keeping with our theme of “no valuable data left behind”, we think that effectively collecting, integrating, and using tail data is a challenging research direction for the big data community. There are many interesting questions that need to be answered. How should one acquire, integrate, and extract knowledge on tail entities, and for tail verticals, when there may not be many data sources providing relevant data? How can one understand the quality and value of tail data sources? How can such sources be used without compromising on value, even if the data are not of extremely high quality? How does one integrate historical data, including entities that evolve over time, and enable the exploration of the history of web data sources? In addition to freshness, what additional metrics are relevant to capturing quality over time? How does one deal with sources that provide data about future events? How can one integrate data across multiple languages and cultures? Answering these challenging questions will keep our community busy for many years to come.

Gerhard Weikum is a scientific director at the Max Planck Institute for Informatics in Saarbruecken, Germany, where he is leading the department on databases and information systems. He co-authored a comprehensive textbook on transactional systems, received the VLDB 10-Year Award for his work on automatic DB tuning, and is one of the creators of the YAGO knowledge base. Gerhard is an ACM Fellow, a member of several scientific academies in Germany and Europe, and a recipient of a Google Focused Research Award, an ACM SIGMOD Contributions Award, and an ERC Synergy Grant.

What is your motivation for doing research in the area of Big Data Value?

Big Data is the New Oil! This often heard metaphor refers to the world’s most precious raw asset — of this century and of the previous century. However, raw oil does not power any engines or contribute to high-tech materials. Oil needs to be cleaned, refined, and put in an application context to gain its true value. The same holds for Big Data. The raw data itself does not hold any value, unless it is processed in analytical tasks from which humans or downstream applications can derive insight. Here is where data quality comes into play, and in a crucial role.

Some applications may do well with huge amounts of inaccurate or partly erroneous data, but truly mission-critical applications would often prefer less data of higher accuracy and correctness. This Veracity dimension of the data is widely underestimated. In many applications, the workflows for Big Data analytics include major efforts on data cleaning, to eliminate or correct spurious data. Often, a substantial amount of manual data curation is unavoidable and incurs a major cost fraction.

OK, I see. Then what is your recent work in the area of oil refinery?

Much of the research in my group at the Max Planck Institute could actually be cast under this alternative – and much cooler – metaphor: Big Text Data is the New Chocolate!

We believe that many applications would enormously gain from tapping unstructured text data, like news, product reviews in social media, discussion forums, customer requests, and more. Chocolate is a lot more sophisticated and tasteful than oil — and so is natural-language text. Text is full of finesse, vagueness and ambiguities, and so could at best be seen as Uncertain Big Data. A major goal of our research is to automatically understand and enrich text data in terms of entities and relationships and this way enable its use in analytic tasks — on par with structured big data.

We have developed versatile and robust methods for discovering mentions of named entities in text documents, like news articles or posts in social media, and disambiguating them onto entities in a knowledge base or entity catalog. The AIDA software is freely available as open source code. These methods allow us to group documents by entities, entity pairs or entity categories, and compute aggregates on these groups. Our STICS demonstrator shows some of the capabilities for semantic search and analytics. We can further combine this with the detection and canonicalization of text phrases that denote relations between entities, and we can capture special kinds of text expressions that bear sentiments (like/dislike/support/oppose/doubt/etc.) or other important information.

Having nailed down the entities, we can obtain additional structured data from entity-indexed data and knowledge bases to further enrich our text documents and document groups. All this enables a wealth of co-occurrence-based analytics for comparisons, trends, outliers, and more. Obviously, for lifting unstructured text data to this value-added level, the Veracity of mapping names and phrases into entities and relations is decisive.

For example, when performing a political opinion analysis about the Ukrainian politician and former boxer Klitschko, one needs to be careful about not confusing him with his brother who is actively boxing. A news text like “former box champion Klitschko is now the mayor of Kiev” needs to be distinguished from “box champion Klitschko visited his brother in Kiev”. Conversely, a recent text like “the mayor of Kiev met with the German chancellor” should also count towards the politician Vitali Klitschko although it does not explicitly mention his name.

Why New Chocolate?

Well, in the Aztec Empire, cocoa beans were so valuable that they were used as currency! Moreover, cocoa contains chemicals that trigger the production of the neurotransmitter Serotonin in our brains – a happiness substance! Yes, you may have to eat thousands of chocolate bars before you experience any notable kicks, but for the sake of the principle: chocolate is so much richer and creates so much more happiness than oil.

Thanks for this culinary encouragement :-) What do you think will be the future of the field?

Data quality is more than the quest for Veracity. Even if we could ensure that the database has only fully correct and accurate data points, there are other quality dimensions that often create major problems: incompleteness, bias and staleness are three aspects of paramount importance.

No data or knowledge base can ever be perfectly complete, but how do we know which parts we do not know? For a simple example, consider an entertainment music database, where we have captured a song and ten different cover versions of it. How can we tell that there really are only ten covers of that song? If there are more, how can we rule out that our choice of having these ten in the database is not biased in any way – for example, reflecting only Western culture versions and ignoring Asian covers? Tapping into text sources, in the New Chocolate sense, can help completing the data, but is also prone to “reporting bias”. The possibility that some of the data is stale, to different degrees, makes the situation even more complex.

Finally, add the Variety dimension on top of all this — not a single database but many independent data and text sources with different levels of incompleteness, bias, and staleness. Assessing the overall quality that such heterogeneous and diverse data provides for a given analytic task is a grand challenge. Ideally, we would like to understand how the quality of that data affects the quality of the insight we derive from it. If we consider data cleaning measures, what costs do we need to pay to achieve which improvements in data quality and analytic-output quality? I believe these are pressing research issues; their complexity will keep the field busy for the coming years.


Do you agree with Serge Abiteboul that knowledge management will be the killer app of Big Data? Do you share Gerhard Weikum’s opinion that Big Text Data is the New Chocolate? Do you have ideas on how to achieve the “no valuable data left behind” mantra that Divesh Srivastava and Luna Dong evoke? Does your work marry the domains of AI and Big Data, as Oren Etzioni proposes? We would be delighted to hear your opinion and your latest contribution in this field! This year’s WebDB workshop, which will be co-located with ACM SIGMOD, will provide a premier venue for discussing issues of big data quality. Its theme is “Freshness, Correctness, Quality of Information and Knowledge on the Web”. This theme encompasses a wide range of research directions, from focused crawling and time-aware search and ranking, to information extraction and data integration, to management, alignment, curation, and integration of structured knowledge, and to information corroboration and provenance. However, papers on all aspects of the Web and databases are solicited. We are looking forward to your submissions, and to interesting discussions about whether the future will bring us not just big data, but also good data!

Blogger Profile: julia Julia Stoyanovich is an Assistant Professor of Computer Science at Drexel University. She was previously a postdoctoral researcher and a CIFellow at the University of Pennsylvania. Julia holds M.S. and Ph.D. degrees in Computer Science from Columbia University, and a B.S. in Computer Science and in Mathematics and Statistics from the University of Massachusetts at Amherst. After receiving her B.S. Julia went on to work for two start-ups and one real company in New York City, where she interacted with, and was puzzled by, a variety of massive datasets. Julia’s research focuses on developing novel information discovery approaches for large datasets in presence of rich semantic and statistical structure. Her work has been supported by the NSF and by Google.

Blogger Profile: FabianSuchanek_HiRes Fabian M. Suchanek is an associate professor at the Telecom ParisTech University in Paris. Fabian developed inter alia the YAGO-Ontology, one of the largest public knowledge bases on the Semantic Web, which earned him a honorable mention of the SIGMOD dissertation award. His interests include information extraction, automated reasoning, and knowledge bases. Fabian has published around 40 scientific articles, among others at ISWC, VLDB, SIGMOD, WWW, CIKM, ICDE, and SIGIR, and his work has been cited more than 3500 times.

Graph data management

amolGraph data management has seen a resurgence in recent years, because of an increasing realization that querying and reasoning about the structure of the interconnections between entities can lead to interesting and deep insights into a variety of phenomena. The application domains where graph or network analytics are regularly applied include social media, finance, communication networks, biological networks, and many others. Despite much work on the topic, graph data management is still a nascent topic with many open questions. At the same time, I feel that the research in the database community is fragmented and somewhat disconnected from application domains, and many important questions are not being investigated in our community. This blog post is an attempt to summarize some of my thoughts on this topic, and what exciting and important research problems I think are still open.

At its simplest, graph data management is about managing, querying, and analyzing a set of entities (nodes) and interconnections (edges) between them, both of which may have attributes associated with them. Although much of the research has focused on homogeneous graphs, most real-world graphs are heterogeneous, and the entities and the edges can usually be grouped into a small number of well-defined classes.

Graph processing tasks can be broadly divided into a few categories. (1) First, we may to want execute standard SQL queries, especially aggregations, by treating the node and edge classes as relations. (2) Second, we may have queries focused on the interconnection structure and its properties; examples include subgraph pattern matching (and variants), keyword proximity search, reachability queries, counting or aggregating over patterns (e.g., triangle/motif counting), grouping nodes based on their interconnection structures, path queries, and others. (3) Third, there is usually a need to execute basic or advanced graph algorithms on the graphs or their subgraphs, e.g., bipartite matching, spanning trees, network flow, shortest paths, traversals, finding cliques or dense subgraphs, graph bisection/partitioning, etc. (4) Fourth, there are “network science” or “graph mining” tasks where the goal is to understand the interconnection network, build predictive models for it, and/or identify interesting events or different types of structures; examples of such tasks include community detection, centrality analysis, influence propagation, ego-centric analysis, modeling evolution over time, link prediction, frequent subgraph mining, and many others [New10]. There is much research still being done on developing new such techniques; however, there is also increasing interest in applying the more mature techniques to very large graphs and doing so in real-time. (5) Finally, many general-purpose machine learning and optimization algorithms (e.g., logistic regression, stochastic gradient descent, ADMM) can be cast as graph processing tasks in appropriately constructed graphs, allowing us to solve problems like topic modeling, recommendations, matrix factorization, etc., on very large inputs [Low12].

Prior work on graph data management could itself be roughly divided into work on specialized graph databases and on large-scale graph analytics, which have largely evolved separately from each other; the former has considered end-to-end data management issues including storage representations, transactions, and query languages, whereas the latter work has typically focused on processing specific tasks or types of tasks over large volumes of data. I will discuss those separately, focusing on whether we need “new” systems for graph data management and on open problems.

Graph Databases and Querying

The first question I am usually asked when I mention graph databases is whether we really need a separate database system for graphs, or whether relational databases suffice. Personally I believe that graph databases provide a significant value for a large class of applications and will emerge as another vertical.

If the goal is to support some simple graph algorithms or graph queries on data that is stored in an RDBMS, then it is often possible to do those using SQL and user-defined functions and aggregations. However, for more complex queries, a specialized graph database engine is likely to be much more user-friendly and likely to provide significant performance advantages. Many of the queries listed above either cannot be mapped to SQL (e.g., flexible subgraph pattern matching, keyword proximity search) or the equivalent SQL is complex and hard to understand or debug. An abstraction layer that converts queries from a graph query language to SQL could address some of these shortcomings, but that will likely only cover a small fraction of the queries mentioned above. More importantly, graph databases provide efficient programmatic access to the graph, allowing one to write arbitrary algorithms against them if needed. Since there is usually a need to execute some graph algorithms or network science tasks in the application domains discussed above, that feature alone makes a graph database very appealing. Most graph data models also support flexible schemas — although an orthogonal issue, new deployments may choose a graph database for that reason.

Whether a specialized graph database provides significant performance advantages over RDBMSs for the functionality common to both is somewhat less clear. For many graph queries, the equivalent SQL, if one exists, can involve many joins and unions and it is unlikely the RDBMS query optimizer could optimize those queries well (especially given the higher use of self-joins). It may also be difficult to choose among different ways to map a graph query into SQL. Queries that require recursion (e.g., reachability) are difficult to execute in a relational database, but are natural for graph databases. Graph databases can also employ specific optimizations geared towards graph queries and traversals. For example, graph databases typically store all the edges for a node with the node to avoid joins, and such denormalization can significantly help with traversal queries, especially queries that traverse multiple types of edges simultaneously (e.g., for subgraph pattern matching). Replicating node information with neighbors can reduce the number of cache misses and distributed traversals for most graph queries (at the expense of increased storage and update costs). Similarly, min cut-based graph partitioning techniques help in reducing the number of distributed queries or transactions, and similar optimizations can be effective in multi-core environments as well. On the other hand, there is less work on query optimization in graph databases, and for simple queries (especially simple subgraph pattern matching queries), the query optimizer in relational databases may make better decisions than many of today’s graph databases.

I think exploring such optimizations and understanding the tradeoffs better are rich topics for further research. For example, how the graph is laid out, both in persistent storage and in memory, can have a significant impact on the performance, especially in multi-core environments. We also need to better understand the common access patterns that are induced by different types of queries or tasks, and the impact of different storage representations on the performance of those access patterns. Another key challenge for graph querying is developing a practical query language. There is much theoretical work on this problem [Woo12] and several languages are currently used in practice, including SPARQL, Cypher, Gremlin, and Datalog. Among those, SPARQL and Cypher are based primarily on subgraph pattern matching and can handle a limited set of queries, whereas Gremlin is a somewhat low-level language and may not be easy to optimize. Datalog (used in LogicBlox and Datomic) perhaps strikes the best balance, but is not as user-friendly for graph querying and may need standardization of some of the advanced constructs, especially aggregates.

Unfortunately I see little work on these problems, and on end-to-end graph databases in general, in our community. There are quite a few graph data management systems being actively built in the industry, including Neo4j, Titan, OrientDB, Datomic, DEX, to name a few, where these issues are being explored. Much of the work in our community, on the other hand, is more narrowly focused on developing search algorithms and indexing techniques for specific types of queries; while that work has resulted in many innovative techniques, for wide applicability and impact, it is also important to understand how those fit into a general-purpose graph data management system.

Large-scale Graph Analytics

Unlike the above scenario, the case of new systems for graph analysis tasks, broadly defined to include graph algorithms and network science and graph mining tasks, is more persuasive. Batch analytics systems like relational data warehouses and MapReduce-based systems are not a good fit for graph analytics as is. From the usability perspective, it is not natural to write graph analysis tasks using those programming frameworks. Some graph programming models (e.g., the vertex-centric programming model [Mal10]) can be supported in a relational database through use of UDFs and UDAs [Jin14]; however it is not clear if the richer programming frameworks (discussed below) can also be efficiently supported. Further, many graph analysis tasks are inherently iterative, with many iterations and very little work per vertex per iteration, and thus the overheads of
those systems may start dominating and may be hard to amortize away.

A more critical question, in my opinion, is whether the popular vertex-centric programming model is really a good model for graph analytics. To briefly recap, in this model, users write vertex-level compute programs, that are then executed iteratively by the framework in either a bulk synchronous fashion or asynchronous fashion using message passing or shared memory. This model is well-suited for some graph processing tasks like computing PageRank or connected components, and also for several distributed machine learning and optimization tasks that can be mapped to message passing algorithms in appropriately constructed graphs [Low12]. Originally introduced in this context in Google’s Pregel system [Mal10], several graph analytics systems are built around this model (e.g., Giraph, Hama, GraphLab, PowerGraph, GRACE, GPS, GraphX).

However, most graph analysis tasks (e.g., the popular modularity optimization algorithm for community detection, betweenness centralities) or graph algorithms (e.g., matching, partitioning) cannot be written using the vertex-centric programming model while permitting efficient execution. The model limits the compute program’s access to a single vertex’s state and so the overall computation needs to be decomposed into smaller local tasks that can be (largely) independently executed; it is not clear how to do this for most of the computations discussed above, without requiring a large number of iterations. Even local neighborhood-centric analysis tasks (e.g., counting motifs, identifying social circles, computing local clustering coefficients) are inefficient to execute using this model; one could execute such a task by constructing multi-hop neighborhoods in each node’s local state by exchanging neighbor lists, but the memory required to hold that state can quickly make it infeasible [Qua14]. I believe these limitations are the main reason why most of the papers about this model focus on a small set of tasks like PageRank, and also why we don’t see broad adoption of this model for graph analysis tasks, unlike the MapReduce framework, which was very quickly and widely adopted.

Some of the alternative, and more expressive, programming models proposed in recent years include distributed Datalog-based framework used by Socialite [Seo13], data-centric programming models of Ligra [Shu13] and Galois [Ngu13], Green-Marl DSL [Hon12], and NScale framework from our recent work [Qua14]. Unlike the vertex-centric frameworks, however, distributed data-parallel execution is not straightforward for these frameworks, and investigating the trade-offs between expressiveness, ability to parallelize the computations, and ease-of-use remains a crucial challenge. The equivalence between graph analytics and matrix operations [Mat13], and whether that leads to better graph analysis systems, also need to be explored in more depth.

On a related note, the need for distributed execution of graph processing tasks is often taken as a given. However, graphs with 10′s to 100′s of billions of edges can be loaded onto a single powerful machine today, depending on the amount of information per node that needs to be processed (Ligra reports experiments on a graph with 12.9 billion edges with 256GB memory [Shu13]; in our recent work, we were able to execute a large number of streaming aggregate queries over a graph with 300 million edges on a 64GB machine [Mon14a]). Given the difficulty in distributing many graph analysis/querying tasks, it may be better to investigate approaches that eliminate the need to execute any single query or task in a distributed fashion (e.g., through aggressive compression, careful encoding of adjacency lists, or careful staging to disk or SSD (a la GraphChi)), while parallelizing independent queries/tasks across different machines.

Other Open questions

Despite much work, there are many important and hard problems that remain open in graph data management, in addition to the ones discussed above; more challenges are likely to come up as graph querying and analytics are broadly adopted.

Need for a Benchmark: Given the complex tradeoffs, many of the questions discussed above would be hard to answer without some representative workloads and benchmarks, especially because the performance of a system may be quite sensitive to the skew in the degree distribution and the graph diameter. Some of issues, e.g., storage representation, have been studied in depth in the context of RDF triple-stores, but the benchmarks established there appear to focus on scale and do not feature sufficient variety in the queries. A benchmark covering a variety of graph analysis tasks would also help significantly towards evaluating and comparing the expressive power and the performance of different frameworks and systems. Benchmarks would also help reconcile some of the recent conflicting empirical comparisons, and would help shed light on specific design decisions that impact performance significantly.

Temporal and real-time analytics: Most real-world graphs are highly dynamic in nature and often generate large volumes of data at a very rapid rate. Temporal analytics or querying over historical traces can lead to deeper insights into various phenomena, especially those related to evolution or change. One of the key challenges here is how to store the historical trace compactly while still enabling efficient execution of point queries and global or neighborhood-centric analysis tasks [Khu13]. Key differences from temporal databases, a topic that has seen much work, appear to be the scale of data, focus on distributed and in-memory environments, and the need to support global analysis tasks (which usually require loading entire historical snapshots into memory). Similarly real-time querying and analytics, especially anomaly detection, present several unique challenges not encountered in relational data stream processing [Mon14b].

Graph extraction: Another interesting, and practically important, question is how to efficiently extract a graph, or a collection of graphs, from non-graph data stores. Most graph analytics systems assume that the graph is provided explicitly. However, in many cases, the graph may have to be constructed by joining and combining information spread across a set of relations or files or key-value stores. A general framework that allows one to specify what graph or graphs need to be constructed for analysis, and how they map to the data stored in the persistent data stores would significantly simplify the end-to-end process of graph analytics. Even if the data is stored in a graph data store, often we only need to load a set of subgraphs of that graph for further analysis [Qua14], and similar framework would be needed to specify what subgraphs are to be extracted.

The above list is naturally somewhat skewed towards the problems we are working on in our group at Maryland, and towards developing general-purpose graph data management systems. In addition, there is also much work that needs to be done in the application areas like social media, finance, cybersecurity, etc.; in developing graph analytics techniques that lead to meaningful insights in those domains; in understanding what types of query workloads are typical; and in handling those over large volumes of data.


[Hon12] Green-Marl: A DSL for Easy and Efficient Graph Analysis; S. Hong, H. Chafi, E. Sedlar, and K. Olukotun; ASPLOS 2012

[Jin14] Vertexica: Your Relational Friend for Graph Analytics!; A. Jindal, P. Rawlani, E. Wu, S. Madden, A. Deshpande, M. Stonebraker; VLDB 2014

[Khu13] Efficient Snapshot Retrieval over Historical Graph Data; U. Khurana, A. Deshpande; ICDE 2013.

[Low12] Distributed GraphLab: a framework for machine learning and data mining in the cloud; Low et al.; VLDB 2012.

[Mal10] Pregel: a system for large-scale graph processing; Malewicz et al.; SIGMOD 2010

[Mat13] Standards for Graph Algorithm Primitives; Mattson et al.; HPEC 2013

[Mon14a] EAGr: Supporting Continuous Ego-centric Aggregate Queries over Large Dynamic Graphs; J. Mondal, A. Deshpande; SIGMOD 2014

[Mon14b] Stream Querying and Reasoning on Social Data; J. Mondal, A. Deshpande; ESONAM 2014

[New10] Networks: An Introduction; M. Newman 2010.

[Ngu13] A lightweight infrastructure for graph analytics; D. Nguyen, A. Lenharth, K. Pingali; SOSP 2013

[Qua14] NScale: Neighborhood-centric Large-Scale Graph Analytics in the Cloud; A. Quamar, A. Deshpande, J. Lin; arXiv:1405.1499, 2014

[Seo13] Distributed socialite: a datalog-based language for large-scale graph analysis; J. Seo, J. Park, J. Shin, M. Lam; VLDB 2013

[Shu13] Ligra: a lightweight graph processing framework for shared memory; J. Shun, G. Blelloch; PPoPP 2013

[Woo12] Query languages for graph databases; P. Wood; ACM SIGMOD Record 2012.

Blogger Profile: Amol Deshpande is an Associate Professor in the Department of Computer Science at the University of Maryland with a joint appointment in the University of Maryland Institute for Advanced Computer Studies (UMIACS). He received his Ph.D. from University of California at Berkeley in 2004. His research interests include uncertain data management, graph analytics, adaptive query processing, data streams, and sensor networks.

Why your friends have more to be thankful for


As Thanksgiving approaches, it may feel like everyone else has so much more to be thankful for. Just check your Facebook, Twitter or Instagram: your friends seem to dine at finer restaurants, take more exotic vacations, and attend more exciting parties. Research suggests this is not simply a matter of perception, but a mathematical fact (for most of us anyway). This unsettling observation is rooted in the friendship paradox, which states that “on average, your friends are more popular than you are”. This means that if you ask a random person who her friends are, the average number of friends her friends have is likely to be larger than the number of friends she has. Friendship paradox holds online too: 98% of Twittter users follow others who have larger followings, on average. Unless you are Lady Gaga, most of your followers are also more popular than you are.

The friendship paradox is not merely a mathematical curiosity, but has useful applications in disease monitoring and trend prediction. Researchers used it to spot flu outbreaks on a college campus in their early stages devise efficient strategies to predict trending topics on Twitter weeks before they became popular. Similarly, if you arrive in an African village with only five Ebola vaccines, the best strategy is not to vaccinate five random people, but ask those people who their friends are and vaccinate five of these friends. Due to the friendship paradox, the friends are likely to be more central, both in the Twitterverse and in the village, and thus more likely get sickened early by the virus or to tweet about topics that later become popular.

Although it sounds strange, the friendship paradox has a simple mathematical explanation. People are diverse: most of us have a few dozen friends, and then there is Lady Gaga. This rare outlier skews the average friend count of many people, putting them in the paradox regime. Mathematicians advise using the median when dealing with distributions that include such extremely large values. The median is the half-way point: half of the numbers in the distribution lie below the median and half above. Unlike the average, it is not easily skewed by a few extremely large numbers. The median is used, for example, to report the income of US households, where extreme fortunes of the top 1% of the population skew the average household income.

Remarkably, my group at University of Southern California has shown that friendship paradox still holds for the median. In other words, most of your friends have more friends than you do, not on average, but most! We showed that over 95% of Twitter users have fewer followers than most of the people they follow, or most of the people who follow them. Stranger still, the paradox holds not only for popularity, but for other personal attributes. As an example, consider how frequently a user posts status updates on Twitter. There is a paradox for that: most of the people you follow post more status updates than you do. Similarly, most of the people you follow receive more novel and diverse information than you do. Also, most of the people you follow receive information that ends up spreading much farther than what you see in your stream.

The friendship paradox helps explain why you are not as cool or interesting as your friends. Extraordinary people are likely to be better socially connected and have more friends than the more ordinary people like you and I. Extraordinary people are also likely to be more active and post more frequently about their extraordinary experiences. This is all it takes to skew our perceptions of the quality of our lives relative to those of our friends. So, if you feel that your friends have more to be grateful for, at least in this you are not alone.


Blogger Profile:
Kristina Lerman is a Project Leader at the Information Sciences Institute and holds a joint appointment as a Research Associate Professor in the USC Viterbi School of Engineering’s Computer Science Department. Her research focuses on applying network- and machine learning-based methods to problems in social computing.

Big Data Is Opening the Door To Revolutions: Databases Should Be Next


Most professional fields, whether in business or academia, rely on data and have done so for centuries. In the digital age and with the emergence of Big Data, this dependency is growing dramatically – perhaps out of proportion to its current value given the concepts, tools, and techniques presently available. For example, how do you tell if the results of data-intensive analysis are correct and reliable and not weak or even spurious? Most data-intensive disciplines have statistical measures that attempt to calculate meaning or truth. Efficacy quantifies the strength of a relationship within a system, such as biology or business. For example, when researchers investigate a new drug, they compare its effectiveness to a placebo, using statistics to determine whether the drug worked. This approach, where data selection and processing is predicated on complex models rather than simple comparison, is a far cry from select-project-join queries.

Efficacy is the capacity to produce a desired result or effect. In medicine, it is the ability of an intervention or drug to produce an outcome. P-values have been a conventional empirical metric of efficacy for 100 years.

Moreover, the underlying data in these fields is complex, uncertain, and multimodal. Despite a large body of research data management for science applications, there has been little adoption of relational techniques in the science disciplines. In this post, we examine two challenges. First, modeling data around domain-specific efficacy rather than set theory. Second, support for ensembles of data models to enable many perspectives on a single data set.

The big picture is compelling. Since the late 1980’s one of us estimated in papers and keynotes that databases contain less than 10% of the world’s data and dropping fast as non-database data growth exploded. A corresponding fraction of the world’s applications – data and computation – are amenable to traditional databases. Modelling the 90% opens the door for the database community to the requirements of the rest of the world’s data and a new, vastly larger generation of database research and technology. This calls for a shift in our community commensurate with the profound changes introduced by Big Data.

Efficacy first, then efficiency

Since meaning and truth are relative to a system, efficacy measures are of accuracy, correctness, precision, and significance with respect to a context. That we can compute an answer efficiently – at lightning speed over massive data sets – is entirely irrelevant or even harmful if we cannot demonstrate that the answer is meaningful or at least approximately right in a given context. As fields develop and complexity increases, efficacy measures become increasingly sophisticated, refined, and debated. For example, p-values – the gold standard of empirical efficacy – have been questioned for decades, especially under the pressure of increasing irreproducibility in science. The same is true for precision and recall in information retrieval. In fact, since most fields that depend on data involve uncertainty, measures of efficacy are being questioned everywhere, with the notable exception of data management.

Big Data, broadly construed, is inherently multidisciplinary but often lacks the efficacy measures of its constituent disciplines – statistics, machine learning, empiricism, data mining, information retrieval, among others – let alone those of application domains such as finance, biology, clinical studies, high-energy physics, drug discovery, and astrophysics. One reason for this is that efficacy measures that have been developed in the small data world, based on statistics and other fields, do not necessarily hold true over massive data sets [3][5]. Efficacy in this context is an important, open, and rich research challenge. The value and success of data-intensive discovery (Big Data) depends on achieving adequate means of evaluating the efficacy of its results. A notable exception is the Baylor-Watson result [7] that focused first on efficacy, i.e., modeling, that then contributed to efficiency. But efficacy is one aspect of a larger challenge – modeling.

Relational data is the servant of the data model and the query. It was right to constrain data when we had a well-defined model. And we could always get the model right – right?

As data management evolved it distinguished itself from information retrieval by not requiring efficacy measures since databases were bounded, discrete, and complied with well-defined models, e.g., schemas. In contrast, information retrieval (and later machine learning) searched data sets for complex correlations rather than rigidly defined predicates. Finding relationships like “select all pairs where their covariance is greater than x” are inherently iterative and compute-intensive. In contrast, the contents of a database either match a query or they do not – black or white. No need for estimating accuracy, confidence, or probabilities. Relational data is the servant of the data model and the query. This permitted massive performance improvements that led to the widespread adoption of databases in applications for which schemas made sense. If the data did not comply with the schema and the query within that, then the data was erroneous by definition and should be rejected or corrected. It was right to constrain data when we had a well-defined model. And we could always get the model right – right? Many courageous researchers over the past 50 years have studied this problem, including probabilistic databases and fuzzy logic, (and more [1]) but none has seen widespread adoption. Why?

While the non-database world – life sciences, high-energy physics, astrophysics, finance – opened the door to Big Data and its possibilities, the data management world is aspiring to take ownership of their infrastructure – the storage, management, manipulation, querying, and searching of massive datasets. Currently much of this work is done in an ad-hoc manner using tools like R and Python. What is required for a more general solution? The non-database world is driven by applications – solving problems with real-world constraints – achieving efficacy within the models and definitions of their domain – often with 400 or 500 years of history.

In contrast, the database landscape is predominantly concerned with efficiency and has not dealt head on with efficacy yet. Some of these issues have been addressed in the database context in terms of specific models, languages, and design, but seldom have those concerns impacted the core database infrastructure, let alone gained adoption. Perhaps database researchers focused only on application domains that are well-behaved. While efficacy is a critical requirement – possibly the most critical requirement – in domains that make extensive use of data, it is part of the broader requirement for modelling unmet by database systems.

For more than a decade physics, astrophysics, photonics, biology, indeed most physical sciences as well as statistics and machine learning have made the modest assumption that multiple perspectives may be more valuable than a single model.

Data Models for the 90%

The database community, like many others, perhaps has not fully internalized the paradigm shift from small to big data. Big Data – or data per se – does not create the change nor is itself the change. Big Data opens the door to a revolution in thinking. One aspect involves data-driven methods. A profound shift involves viewing phenomena from multiple perspectives simultaneously.

A significant aspect of this shift is that every Big Data activity (small data activities also, but with less impact) requires measures of efficacy for each perspective or model. This is not simply owing to the reframing of corresponding principles from empirical science, but also to the multiple meanings of data, each of which requires mechanisms for addressing efficacy.

If the data management community is about to provide solutions for this nascent challenge, then it will need to deal with efficacy. This essentially has to do with modelling, a chaotic and ad-hoc database topic that has been largely unsuccessful, again measured by adoption. The relational model has dominated databases for over 40 years largely owing to efficiency. The database community knows how to optimize anything expressed relationally. While the relational model has proven to be amazingly general, its adoption has been limited in many domains, especially the sciences.

A related limitation of the database world is the assumption of a single perspective, e.g., a single version of truth, one schema per database even with multiple views. For more than a decade physics, astrophysics, photonics, biology, indeed most physical sciences as well as statistics and machine learning have made the modest assumption that multiple perspectives may be more valuable than a single model.

In [8] the author argued that science undergoes paradigm shifts only when there are rival theories about the fundamentals of a discipline. It is his position that rival paradigms are incommensurable using entirely different concepts and evaluation metrics from one another. One such example was the wave and particle theories of light. Each has entirely different models and measures of efficacy. Understanding the big picture necessitates finding consistencies and anomalies in both theories.

Ensemble models are one approach to addressing this challenge. Let’s consider an example in evolutionary biology where researchers use a collection of models to learn about how the human genome has changed over time. In [3] the authors identified positive examples of natural selection in recent human populations. Their discoveries have two parts: the affected gene’s location and its (improved) mutation. By composing many signals of natural selection, the authors increase the resolution of their genomic map by up to 100x. This research computes genetic signals at many levels, from clustering genes that are likely to be inherited together to looking at the high-level geographic distribution of different mutations. In present database modeling, the former might be represented as a graph database, whereas the latter is more likely to fall into the purview of geospatial databases. How can we bring them together? Perhaps neither of these models is designed for computing how effective different genetic variations are at producing advantageous traits. This pattern repeats itself in meteorology, physics, and a myriad of other domains that mathematically model large, dynamic systems.

Stepping into the void of uncertainty, unboundedness, ensemble models, and open-ended model exploration is far harder and scarier. We call it Computing Reality

Ensemble models pose substantial challenges to the data management community. How do you simultaneously store, manage, query, and update this variety of models, applying to a single dataset with many, possibly conflicting schemas? Database folks may first be concerned about doing this efficiently. Nope – wrong question. The first step is to understand the problem, to ask the right questions, to get the model correct and only then to make it efficient. How do you support ensemble models and their requirements including efficacy?

This may be why application domains that use massive data sets have grown their own data management tools, such as Hadoop, ADAM, Wikidata, and Scientific Data Management Systems, let alone a plethora of such tools in most physical science communities that the database community has never heard of. It’s not just that their data does not fit the relational model; databases do not support ensemble models, efficacy, or many of the fundamental concepts used to understand data. Why would any application domain (e.g., physical sciences, clinical studies, drug discovery) or discipline (e.g., information retrieval, machine learning, statistics) want to partner with an infrastructure technology that did not support its basic principles?

The database community has developed amazing technology that has changed the world. Since the early 1990’s it has extended its models to non-relational models such for networks, text, graphs, arrays, and many more. But efficacy is not just an issue of expressing eScience applications relationally, as UDFs or in R, but modeling and computing hypotheses under the complex contexts defined by domain experts, none of which map easily to set theory or other discrete mathematics. Stepping into the void of uncertainty, unboundedness, ensemble models, and open-ended model exploration is far harder and scarier. We call it Computing Reality [1][6].

Big Data is opening the door to a paradigm shift in many human endeavors. Machine learning was first through the door with real, albeit preliminary, results and it is already on to the next generation with deep learning [3]. Analytics and other domains are riding the wave of machine learning. The database community is heading for the door now, but it will be challenging. We first have to understand the problem and get the requirements right. To paraphrase Ron Fagin, we need to focus on asking the right questions. The rest may be a breeze but efficacy before efficiency!

So not only are we leaving the relational world that was dominated one model or a class of discrete models, but we are leaving the world of a single model for each dataset and embarking on a journey into a world of ensemble models of including probabilistic, fuzzy, and even potentially the richest model of them all, a model-free approach that enables us to listen to the data. All at scale. This seems scary to us but also just what we need.

Are we crazy, naive? Isn’t it our mission to dig in this data goldmine, to contribute to accelerating scientific discovery? What do you think? We are all ears.

[1] Brodie, Michael and Jennie Duggan. What versus Why. Towards Computing Reality, April 17, 2014, published April 22, 2014

[2] Duggan, Jennie and Michael L. Brodie, Hephaestus: Virtual Experiments for Data-Intensive Science, In CIDR 2015 (to appear)

[3] Gomes, Lee. Machine-Learning Maestro Michael Jordan on the Delusions of Big Data and Other Huge Engineering Efforts, IEEE Spectrum, 20 Oct 2014

[4] Grossman, Sharon R., et al. “A composite of multiple signals distinguishes causal variants in regions of positive selection.” Science 327.5967 (2010): 883-886.

[5] National Research Council. Frontiers in Massive Data Analysis. Washington, DC: The National Academies Press, 2013

[6] Piatetsky, Gregory, Interview: Michael Brodie, leading database researcher, industry leader, thinker. SIGKDD Explor. Newsl. 16, 1 (September 2014), 57-63. DOI=10.1145/2674026.2674035

[7] Scott Spangler, Angela D. Wilkins, Benjamin J. Bachman, Meena Nagarajan, Tajhal Dayaram, Peter Haas, Sam Regenbogen, Curtis R. Pickering, Austin Comer, Jeffrey N. Myers, Ioana Stanoi, Linda Kato, Ana Lelescu, Jacques J. Labrie, Neha Parikh, Andreas Martin Lisewski, Lawrence Donehower, Ying Chen, and Olivier Lichtarge. 2014. Automated hypothesis generation based on mining scientific literature. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining (KDD ’14). ACM, New York, NY, USA, 1877-1886. DOI=10.1145/2623330.2623667

[8] Kuhn, Thomas S. The structure of scientific revolutions. University of Chicago press, 2012.

Bloggers’ Profiles:
Dr. Brodie has over 40 years experience in research and industrial practice in databases, distributed systems, integration, artificial intelligence, and multi-disciplinary problem solving. He is concerned with the Big Picture aspects of information ecosystems including business, economic, social, application, and technical. Dr. Brodie is a Research Scientist, Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology; advises startups; serves on Advisory Boards of national and international research organizations; and is an adjunct professor at the National University of Ireland, Galway. For over 20 years he served as Chief Scientist of IT, Verizon, a Fortune 20 company, responsible for advanced technologies, architectures, and methodologies for Information Technology strategies and for guiding industrial scale deployments of emergent technologies. His current research and applied interests include Big Data, Data Science, and data curation at scale and a related start up He has served on several National Academy of Science committees. Dr. Brodie holds a PhD in Databases from the University of Toronto 
and a Doctor of Science (honoris causa) from the National University of Ireland.

Jennie Duggan is a postdoctoral associate at MIT CSAIL working with Michael Stonebraker and an adjunct assistant professor at Northwestern University. She received her Ph.D. from Brown University in 2012 under the supervision of Ugur Cetintemel. Her research interests include scientific data management, database workload modeling, and cloud computing. She is especially focused on making data-driven science more accessible and scalable.

Quality Experiments Crowdsourced


Crowdsourcing is a powerful paradigm that democratizes the creation, collection, analysis, curation, and dissemination of data, contributed by millions of individuals, who are called workers. Early crowdsourcing platforms in the context of citizen reporting and citizen sciences were dedicated to specific tasks. Today, many other platforms that expect workers with different expertise exist. Among them, we can find Turkit, Mob4hire, uTest, Freelancer, eLance, oDesk, Guru, Topcoder, Trada, 99design, Innocentive, CloudCrowd, and CloudFlower. The most popular one is Amazon Mechanical Turk (AMT) that is generic and allows any requester to post virtually any kind of task in which any worker can participate via self-appointment.

Tasks can be as simple as binary requests such as recognizing a landmark in a city, or more sophisticated ones such as meta-data enrichment (picture and video tagging in Flickr or YouTube, food description in Open Food Facts), opinion solicitation (restaurants reviews), collaborative intelligence (city map alignment in, identifying stars in GalaxyZoo, character recognition in Recaptcha), and knowledge-intensive crowdsourcing with tasks such as collaborative editing (Wikipedia) and idea generation. The database community has been offering its own platforms (e.g., Qurk on top of AMT and Deco for query processing) and today, scientists are resorting to crowdsourcing for a particular kind of task, that is, quality experiments to validate their findings.

There are several challenges one faces in crowd-based quality experiments. Some are inherent to using a crowd. For example, worker filtering to make sure only qualified ones participate, has to be done by task requesters. Those qualification tests are often difficult to design and do not always lead to flawless experiments. Data collection to build worker profiles is also a challenge (e.g., workers are asked to rate at least 30 movies to build a profile). Other challenges depend on the actual platform. The availability of a large worker pool and the ability to easily set up virtually any task, make AMT the platform of choice for most research experiments. In fact, AMT is used by most researchers in the DB community, and tasks range from answering simple binary questions such as determining whether a tweet expresses a positive or a negative sentiment, to more sophisticated tasks that require some domain knowledge such as collaborative journalism. However, the closed nature of AMT (and of all crowdsourcing platforms we know of today) forces task requesters to find workarounds outside the system in order to select the right pool of workers and assign to them the right set of tasks (AMT is often used as a hiring platform only and workers are redirected to another website). Moreover, monetary compensation, as is the case in AMT, is not always the best choice for incentivizing workers (think open-source contributions) and tends to attract more cheaters when the reward is too high. To address that, post-quality control is enforced via majority voting or by monitoring task completion time. The bottom line is that in a volatile environment, enforcing some level of correctness is difficult, time-consuming and costly.

I have used AMT in several qualitative analyses. The first experiment was in 2009 [1] to evaluate travel itineraries extracted from Flickr against carefully handcrafted ones. It was a large-scale experiment “to validate the wisdom of the crowd”. About 450 workers participated in the experiment and evaluated itineraries in 5 popular and geographically distributed cities (San Francisco, New York City, London, Paris and Barcelona). Half of the workers were involved in an independent study that evaluated the usefulness of our itineraries, the points of interests they contained and the landmark visit and transit times. The other half participated in a comparative study where our itineraries were put side-by-side with handcrafted ones and were preferred in a large majority of the cases. It worked so well that the Wall Street Journal, a prominent newspaper in the US, talked about it [2]. In this experiment, special attention was dedicated to the design of qualification tests to filter out workers. We had to think carefully about a test for which answers are not readily available with simple Web search. Below is one of the tests we used.


The second time we used AMT [3] relied on input from 100 workers to evaluate how well different rating aggregation functions (least misery function, majority voting, pairwise disagreement, and global average) to compute movie recommendations from user groups with different characteristics (small/large groups, similar/dissimilar groups). We found several insightful and sometimes surprising results including that least misery, i.e., optimizing for the lowest rating in a group, performed better than averaging ratings and is the best aggregation function for small groups formed by similar users. Pairwise disagreement on the other hand, did very well with large groups of dissimilar users. In the first round of that experiment, we had inconsistent findings. We then looked carefully in AMT logs and realized that there were cheaters whose task completion time was unrealistic.

The latest work we used AMT for was crowdsourcing the validation of articles produced collaboratively by a crowd to another crowd [4]. Using the crowd to validate the crowd is in fact a common practice and has shown to work well in many applications such as sentence translation by non-experts. Participating workers were instructed to join a group and together write a short summary of current world events on some topic. The qualification test included answering factual questions on those events. Of course, cheaters could use their favorite search engine to find answers to the qualification test.

I hope that the examples above illustrate both the advantage of having a generic platform that enables reaching out to a large worker pool but also showcases the challenges of using crowds for qualitative analysis. Using crowds bears similarities to running surveys and to using cohorts in medical drug trials. While there are practices on how to select such cohorts and ensure result quality (e.g., controlled/test groups, independently chosen subjects, double blind tests in medical trials), none exist for running a crowd-based quality experiments.

With my colleagues Beatrice Valeri and Shady El Bassuoni, we examined 4395 papers published in SIGMOD, PVLDB, ICDE, EDBT, ICDM, KDD, ICWSM, RecSys, and Hypertext between 2010 and 2014 (2014 statistics are partial) and found that 60 of them used crowd-based quality experiments. Among those, all of them use AMT or CrowdFlower, a layer on top of AMT. Most tasks are about collecting golden data used in evaluation and among those, labeling data is the most common. 48 out of the 60 papers have at least one author in the USA. European authors contributed to a total of 14 papers followed by Israel (6 papers), China (4 papers), Qatar and Canada (3 papers each), Singapore and Japan (2 papers each). As shown below, the number of publications that crowdsource experiments has been increasing with a worker base ranging from around 20 to 500 per experiment.



Most papers use workers’ acceptance rate (recorded in AMT based on previous tasks completed by those workers) and only a few use a thorough qualification test. Those tests are followed by a post-processing step where majority voting, manual checking (by involving another crowd), task completion time, and answer justification, are used to check workers’ input. There goes design efficiency.



What is our opportunity today?
Crowd-based quality experiments in the database community are recent. In addition, research in crowdsourcing is gaining traction in our community. What could we do to make a difference? We need a crowd, a platform and some wisdom. Lab experiments have historically been conducted with a small number of colleagues and students who are physically present in a lab. We need to get them online. We also need a platform with returning workers that can be profiled, a platform where we can conduct crowd-based quality experiments but also test our own research on crowdsourcing: worker profiling strategies, our task-to-worker assignment algorithms. That would reduce the burden on all of us when reporting quality experiments on papers since it would be enough to point to results summarized on a single website (double-blind reviewing could be enforced). Such a platform would need to be generic in order to allow any kind of task and any worker to register, open in order to deploy our own crowd-based algorithms and advance the state of research in crowdsourcing, and free!

I do not expect our community to suddenly change the course of things. I myself still run experiments on AMT and will continue to do it. Of course, there are questions related to where to start and how to maintain the platform. Crowd4U ( is a good place to start. It’s an all-academic generic platform that is being developed at the Univ. of Tsukuba and that is open to the rest of the academic world. Here in Grenoble, we ran a campaign on campus where we advertised crowdsourcing and gave out cookies and drinks (not totally free!) and passers-by performed over 1600 tasks to label tweets in less than 3 hours.

You can start your own platform or join Crowd4U but keep it open! In [4], we argue that human factors such as worker skills, worker availability, expected wage, and ability to work together, could be accounted for to make better use of the crowd. Such parameters can only be tested in an open and generic crowdsourcing system where task assignment and skill learning algorithms can be deployed for virtually any task.

So let’s take this opportunity, have some wisdom, and build and nurture our own crowdsourcing platform(s).

[1] Munmun De Choudhury, Moran Feldman, Sihem Amer-Yahia, Nadav Golbandi, Ronny Lempel, Cong Yu: Automatic construction of travel itineraries using social breadcrumbs. HT 2010: 35-44


[3] Senjuti Basu Roy, Sihem Amer-Yahia, Ashish Chawla, Gautam Das, Cong Yu: Space efficiency in group recommendation. VLDB J. 19(6): 877-900 (2010)

[4] Senjuti Basu Roy, Ioanna Lykourentzou, Saravanan Thirumuruganathan, Sihem Amer-Yahia, Gautam Das: Optimization in Knowledge-Intensive Crowdsourcing. CoRR abs/1401.1302 (2014)

Blogger’s Profile:
Sihem Amer-Yahia is a 1st class CNRS (Centre National de Recherche Scientifique) Research Director at LIG (Laboratoire d’Informatique de Grenoble) in France. Sihem heads the SLIDE group (ScaLable Information Discovery and Exploitation) that sits at the intersection of large-scale data management and Web data analytics with an emphasis on the social and the semantic Web. Until July 2012, Sihem was Principal Scientist at the Qatar Computing Research Institute (QCRI) where she led a group in Social Computing and worked with local Universities on student mentoring and with Al Jazeera Online on news traffic analytics. From 2006 to 2011, she was Senior Scientist at Yahoo! Research and worked on revisiting relevance models and scalable top-k processing algorithms for Delicious, Yahoo! Travel and Personals, and Flickr. Before that, she spent 7 years at at&t Labs in New Jersey, working on XML query optimization and XML full-text search in conjunction with the W3C. Sihem has served on the SIGMOD Executive Board, is a member of the VLDB and the EDBT Endowments. She serves on the editorial boards of ACM TODS, the VLDB Journal and the Information Systems Journal. She was track chair of SIGIR 2013 and of PVLDB 2013. She is PC chair of EDBT 2014 and will be PC chair of BDA 2015 (French DB conference) and PC co-chair of SIGMOD Industrial 2015. Sihem received her Ph.D. in Computer Science from Paris-Orsay and INRIA in 1999, and her engineering degree from ESI/INI, Algeria.