Making ideal decisions as a product leader in a web-facing company is incredibly challenging. Beyond navigating the ambiguity of customer satisfaction and achieving business goals, leaders must also ensure their products and services remain relevant, desirable, and profitable. Data and experimentation are crucial for testing product hypotheses and informing decisions. Online controlled experiments, such as A/B testing, can provide highly reliable data to support decisions. However, these experiments can be time-consuming and costly, particularly when assessing impacts on key business metrics like retention or long-term value.

Listeners of long-form talk-audio content, such as podcast episodes, often find it challenging to understand the overall structure and locate relevant sections. A practical solution is to divide episodes into chapters--semantically coherent segments labeled with titles and timestamps. Since most episodes on our platform at Spotify currently lack creator-provided chapters, automating the creation of chapters is essential. Scaling the chapterization of podcast episodes presents unique challenges. First, episodes tend to be less structured than written texts, featuring spontaneous discussions with nuanced transitions. Second, the transcripts are usually lengthy, averaging about 16,000 tokens, which necessitates efficient processing that can preserve context. To address these challenges, we introduce PODTILE, a fine-tuned encoder-decoder transformer to segment conversational data. The model simultaneously generates chapter transitions and titles for the input transcript. To preserve context, each input text is augmented with global context, including the episode's title, description, and previous chapter titles. In our intrinsic evaluation, PODTILE achieved an 11% improvement in ROUGE score over the strongest baseline. Additionally, we provide insights into the practical benefits of auto-generated chapters for listeners navigating episode content. Our findings indicate that auto-generated chapters serve as a useful tool for engaging with less popular podcasts. Finally, we present empirical evidence that using chapter titles can enhance effectiveness of sparse retrieval in search tasks.

Short- and long-term outcomes of an algorithm often differ, with damaging downstream effects. A known example is a click-bait algorithm, which may increase short-term clicks but damage long-term user engagement. A possible solution to estimate the long-term outcome is to run an online experiment or A/B test for the potential algorithms, but it takes months or even longer to observe the long-term outcomes of interest, making the algorithm selection process unacceptably slow. This work thus studies the problem of feasibly yet accurately estimating the long-term outcome of an algorithm using only historical and short-term experiment data. Existing approaches to this problem either need a restrictive assumption about the short-term outcomes called surrogacy or cannot effectively use short-term outcomes, which is inefficient. Therefore, we propose a new framework called Long-term Off-Policy Evaluation (LOPE), which is based on reward function decomposition. LOPE works under a more relaxed assumption than surrogacy and effectively leverages short-term rewards to substantially reduce the variance. Synthetic experiments show that LOPE outperforms existing approaches particularly when surrogacy is severely violated and the long-term reward is noisy. In addition, real-world experiments on large-scale A/B test data collected on a music streaming platform show that LOPE can estimate the long-term outcome of actual algorithms more accurately than existing feasible methods.

Current practice for evaluating recommender systems typically focuses on point estimates of user-oriented effectiveness metrics or business metrics, sometimes combined with additional metrics for considerations such as diversity and novelty. In this article, we argue for the need for researchers and practitioners to attend more closely to various distributions that arise from a recommender system (or other information access system) and the sources of uncertainty that lead to these distributions. One immediate implication of our argument is that both researchers and practitioners must report and examine more thoroughly the distribution of utility between and within different stakeholder groups. However, distributions of various forms arise in many more aspects of the recommender systems experimental process, and distributional thinking has substantial ramifications for how we design, evaluate, and present recommender systems evaluation and research results. Leveraging and emphasizing distributions in the evaluation of recommender systems is a necessary step to ensure that the systems provide appropriate and equitably distributed benefit to the people they affect.

Search systems in online content platforms are typically biased toward a minority of highly consumed items, reflecting the most common user behavior of navigating toward content that is already familiar and popular. Query suggestions are a powerful tool to support query formulation and to encourage exploratory search and content discovery. However, classic approaches for query suggestions typically rely either on semantic similarity, which lacks diversity and does not reflect user searching behavior, or on a collaborative similarity measure mined from search logs, which suffers from data sparsity and is biased by highly popular queries. In this work, we argue that the task of query suggestion can be modelled as a link prediction task on a heterogeneous graph including queries and documents, enabling Graph Learning methods to effectively generate query suggestions encompassing both semantic and collaborative information. We perform an offline evaluation on an internal Spotify dataset of search logs and on two public datasets, showing that node2vec leads to an accurate and diversified set of results, especially on the large scale real-world data. We then describe the implementation in an instant search scenario and discuss a set of additional challenges tied to the specific production environment. Finally, we report the results of a large scale A/B test involving millions of users and prove that node2vec query suggestions lead to an increase in online metrics such as coverage (+1.42% shown search results pages with suggestions) and engagement (+1.21% clicks), with a specifically notable boost in the number of clicks on exploratory search queries (+9.37%).

Calibration in recommender systems has recently gained significant attention. In the recommended list of items, calibration ensures that the various (past) areas of interest of a user are reflected with their corresponding proportions. For instance, if a user has watched, say, 80 romance movies and 20 action movies, then it is reasonable to expect the recommended list of movies to be comprised of about 80% romance and 20% action movies as well. Calibration is particularly important given that optimizing towards accuracy often leads to the user's minority interests being dominated by their main interests, or by a few overall popular items, in the recommendations they receive. In this paper, we propose a novel approach based on the max flow problem for generating calibrated recommendations. In a series of experiments using two publicly available datasets, we demonstrate the superior performance of our proposed approach compared to the state-of-the-art in generating relevant and calibrated recommendation lists.

The AudioCHI 2022 workshop focusses on human engagement with spoken material in search settings, including live stream audio and collections. Spoken material comes in many forms, including for example: factual or entertaining (or both!), timely or of historical interest, local or global, single speaker or conversations. Users engage with spoken material for a variety of reasons, including entertainment, current affairs, education, and research. While there has been considerable previous work studying spoken document retrieval or more generally spoken content retrieval, AudioCHI 2022 is the first meeting to explore user engagement with audio content, including discussing: (i) how content analysis might establish verbal and non-verbal features for rich content representations, and (ii) and use cases and human factors in interaction with spoken audio content, and their interaction with more established topics relating to spoken content retrieval. The workshop brings together researchers in spoken content retrieval with expertise on human computer interaction in information access to examine opportunities and challenges for advancing technologies for search and interaction with spoken content.

Recommender systems play a key role in helping users find their favorite music to play among an often extremely large catalog of items on online streaming services. To correctly identify users' interests, recommendation algorithms rely on past user behavior and feedback to aim at learning users' preferences through the logged interactions. User modeling is a fundamental part of this large-scale system as it enables the model to learn an optimal representation for each user. For instance, in music recommendation, the focus of this paper, users' interests at any time is shaped by their general preferences for music as well as their recent or momentary interests in a particular type of music. In this paper, we present a novel approach for learning user representation based on general and slow-changing user interests as well as fast-moving current preferences. We propose a variational autoencoder-based model that takes fast and slow-moving features and learns an optimal user representation. Our model, which we call FS-VAE, consists of sequential and non-sequential encoders to capture patterns in user-item interactions and learn users' representations. We evaluate FS-VAE on a real-world music streaming dataset. Our experimental results show a clear improvement in learning optimal representations compared to state-of-the-art baselines on the next item recommendation task. We also demonstrate how each of the model components, slow input feature, and fast ones play a role in achieving the best results in next item prediction and learning users' representations.

Online controlled experiments, in which different variants of a product are compared based on an Overall Evaluation Criterion (OEC), have emerged as a gold standard for decision making in online services. It is vital that the OEC is aligned with the overall goal of stakeholders for effective decision making. However, this is a challenge when the overall goal is not immediately observable. For instance, we might want to understand the effect of deploying a feature on long-term retention, where the outcome (retention) is not observable at the end of an A/B test.

Recommender systems are modulating what billions of people are exposed to on a daily basis. Typically, these systems are optimized for user engagement signals such as clicks, streams, likes, or a weighted combination of such sets. Despite the pervasiveness of this practice, little research has been done to explore the downstream impacts of optimization choice on users, creators and the ecosystem they are offered in. We used a platform that caters recommendations to millions of people and show in practice what you optimize for can have a large impact on the content users are exposed to, as well as what they end up consuming.

Podcasts are a popular medium for rapid dissemination of information, entertainment, and casual conversations. Content aggregators are taking an increased interest in recommending podcasts to listeners to help them build larger audiences. With many podcasts released every day, many podcasts that would be of interest to listeners remain underserved by these recommendation systems. In this paper, we study variables related to podcast appeal to listeners selected at random in a large online study, in a production setting, involving more than five million recommendations. We present the results of two observational studies, which suggests that underserved podcast have the potential to grow their audiences. To mitigate the rich-get-richer effect, we propose leveraging semantic information, via means of knowledge graphs, to recommend underserved podcasts to listeners. Finally, we conduct empirical experiments that show our method is effective at recommending underserved podcasts, in comparison to baseline methods that rely on listening behavior.

Over recent years, podcasts have emerged as a novel medium for sharing and broadcasting information over the Internet. Audio streaming platforms originally designed for music content, such as Amazon Music, Pandora, and Spotify, have reported a rapid growth, with millions of users consuming podcasts every day. With podcasts emerging as a new medium for consuming information, the need to develop information access systems that enable efficient and effective discovery from a heterogeneous collection of music and podcasts is more important than ever. However, information access in such domains still remains understudied. In this work, we conduct a large-scale log analysis to study and compare podcast and music search behavior on Spotify, a major audio streaming platform. Our findings suggest that there exist fundamental differences in user behavior while searching for podcasts compared to music. Specifically, we identify the need to improve podcast search performance. We propose a simple yet effective transformer-based neural instant search model that retrieves items from a heterogeneous collection of music and podcast content. Our model takes advantage of multi-task learning to optimize for a ranking objective in addition to a query intent type identification objective. Our experiments on large-scale search logs show that the proposed model significantly outperforms strong baselines for both podcast and music queries.

Rapidly growing online podcast archives contain diverse content on a wide range of topics. These archives form an important resource for entertainment and professional use, but their value can only be realized if users can rapidly and reliably locate content of interest. Search for relevant content can be based on metadata provided by content creators, but also on transcripts of the spoken content itself. Excavating relevant content from deep within these audio streams for diverse types of information needs requires varying the approach to systems prototyping. We describe a set of diverse podcast information needs and different approaches to assessing retrieved content for relevance. We use these information needs in an investigation of the utility and effectiveness of these information sources. Based on our analysis, we recommend approaches for indexing and retrieving podcast content for ad hoc search.

Podcasts are spoken documents across a wide-range of genres and styles, with growing listenership across the world, and a rapidly lowering barrier to entry for both listeners and creators. The great strides in search and recommendation in research and industry have yet to see impact in the podcast space, where recommendations are still largely driven by word of mouth. In this perspective paper, we highlight the many differences between podcasts and other media, and discuss our perspective on challenges and future research directions in the domain of podcast information access.

There is increasing attention to evaluating the fairness of search system ranking decisions. These metrics often consider the membership of items to particular groups, often identified using protected attributes such as gender or ethnicity. To date, these metrics typically assume the availability and completeness of protected attribute labels of items. However, the protected attributes of individuals are rarely present, limiting the application of fair ranking metrics in large scale systems. In order to address this problem, we propose a sampling strategy and estimation technique for four fair ranking metrics. We formulate a robust and unbiased estimator which can operate even with very limited number of labeled items. We evaluate our approach using both simulated and real world data. Our experimental results demonstrate that our method can estimate this family of fair ranking metrics and provides a robust, reliable alternative to exhaustive or random data annotation.

As conversational agents like Siri and Alexa gain in popularity and use, conversation is becoming a more and more important mode of interaction for search. Conversational search shares some features with traditional search, but differs in some important respects: conversational search systems are less likely to return ranked lists of results (a SERP), more likely to involve iterated interactions, and more likely to feature longer, well-formed user queries in the form of natural language questions. Because of these differences, traditional methods for search evaluation (such as the Cranfield paradigm) do not translate easily to conversational search. In this work, we propose a framework for offline evaluation of conversational search, which includes a methodology for creating test collections with relevance judgments, an evaluation measure based on a user interaction model, and an approach to collecting user interaction data to train the model. The framework is based on the idea of “subtopics”, often used to model novelty and diversity in search and recommendation, and the user model is similar to the geometric browsing model introduced by RBP and used in ERR. As far as we know, this is the first work to combine these ideas into a comprehensive framework for offline evaluation of conversational search.

We introduce the concept of expected exposure as the average attention ranked items receive from users over repeated samples of the same query. Furthermore, we advocate for the adoption of the principle of equal expected exposure: given a fixed information need, no item should receive more or less expected exposure than any other item of the same relevance grade. We argue that this principle is desirable for many retrieval objectives and scenarios, including topical diversity and fair ranking. %Leveraging user models from existing retrieval metrics, we propose a general evaluation methodology based on expected exposure and draw connections to related metrics in information retrieval evaluation. Importantly, this methodology relaxes classic information retrieval assumptions, allowing a system, in response to a query, to produce a distribution over rankings instead of a single fixed ranking. We study the behavior of the expected exposure metric and stochastic rankers across a variety of information access conditions, including ad hoc retrieval and recommendation. %We believe that measuring and optimizing expected exposure metrics using randomization opens a new area for retrieval algorithm development and progress.

Users of music streaming, video streaming, news recommendation, and e-commerce services often engage with content in a sequential manner. Providing and evaluating good sequences of recommendations is therefore a central problem for these services. Prior reweighting-based counterfactual evaluation methods either suffer from high variance or make strong independence assumptions about rewards. We propose a new counterfactual estimator that allows for sequential interactions in the rewards with lower variance in an asymptotically unbiased manner. Our method uses graphical assumptions about the causal relationships of the slate to reweight the rewards in the logging policy in a way that approximates the expected sum of rewards under the target policy. Extensive experiments in simulation and on a live recommender system show that our approach outperforms existing methods in terms of bias and data efficiency for the sequential track recommendations problem.

The tutorial focuses on two major themes of recent advances in recommender systems: Part A: Recommendations in a Marketplace: Multi-sided marketplaces are steadily emerging as valuable ecosystems in many applications (e.g. Amazon, AirBnb, Uber), wherein the platforms have customers not only on the demand side (e.g. users), but also on the supply side (e.g. retailer). This tutorial focuses on designing search & recommendation frameworks that power such multi-stakeholder platforms. We discuss multi-objective ranking/recommendation techniques, discuss different ways in which stakeholders specify their objectives, highlight user specific characteristics (e.g. user receptivity) which could be leveraged when developing joint optimization modules and finally present a number of real world case-studies of such multi-stakeholder platforms.

Recommender systems offer great opportunity not only for users to discover new content, but also for the providers of that content to find new audience, followers, and fans. Users often come to a recommender system with certain expectations about what it will recommend to them, and a recommender system that is optimized for creating opportunities for content creators may provide recommendations that are very different from what a user is expecting. We hypothesize that some users’ expectations have a much wider range of acceptability than others, and users with more ”receptivity” to subversion of their expectations are likely to accept such divergence in the recommended content. Understanding users’ responses to such recommendations is vital to platforms that need to serve multiple stakeholders. In this work we investigate logged behavioral responses of users of an audio streaming platform to recommendations that deviate from their expectation, or “divergent” recommendations. We present three classes of listener response to divergent recommendations that can be identified in interaction logs with the aim of predicting which users can be targeted for future divergent recommendations. We derive a number of user characteristics based on user’s music consumption which we think are predictive of user’s receptivity, train models to predict receptivity of these users, and run a live A/B test to validate our approach by correlating with engagement.

Information retrieval (IR) ranking models in production systems continually evolve in response to user feedback, insights from research, and new developments. Rather than investing all engineering resources to produce a single challenger to the existing system, a commercial provider might choose to explore multiple new ranking models simultaneously. However, even small changes to a complex model can have unintended consequences. In particular, the per-topic effectiveness profile is likely to change, and even when an overall improvement is achieved, gains are rarely observed for every query, introducing the risk that some users or queries may be negatively impacted by the new model if deployed into production.

Recommender systems are increasingly used to predict and serve content that aligns with user taste, yet the task of matching new users with relevant content remains a challenge. We consider podcasting to be an emerging medium with rapid growth in adoption, and discuss challenges that arise when applying traditional recommendation approaches to address the cold-start problem. Using music consumption behavior, we examine two main techniques in inferring Spotify users preferences over more than 200k podcasts. Our results show significant improvements in consumption of up to 50% for both offline and online experiments. We provide extensive analysis on model performance and examine the degree to which music data as an input source introduces bias in recommendations.

Over the past decade, podcasts have been one of the fastest growing online streaming media. Many online audio streaming platforms such as Pandora, Spotify, etc. that traditionally focused on music content have started to incorporate services related to podcasts. Although incorporating new media types such as podcasts has created tremendous opportunities for these streaming platforms to expand their content offering, it also introduces new challenges. Since the functional use of podcasts and music may largely overlap for many people, the two types of content may compete with one another for the finite amount of time that users may allocate for audio streaming. As a result, incorporating podcast listening may influence and change the way users have originally consumed music. Adopting quasi-experimental techniques, the current study assesses the causal influence of adding a new class of content on user listening behavior by using large scale observational data collected from a widely used audio streaming platform. Our results demonstrate that podcast and music consumption compete slightly but do not replace one another – users open another time window to listen to podcasts. In addition, users who have added podcasts to their music listening demonstrate significantly different consumption habits for podcasts vs. music in terms of the streaming time, duration and frequency. Taking all the differences as input features to a machine learning model, we demonstrate that a podcast listening session is predictable at the start of a new listening session. Our study provides a novel contribution for online audio streaming and consumption services to understand their potential consumers and to best support their current users with an improved recommendation system.

In recent years, two sided marketplaces have emerged as viable business models in many real world applications (e.g. Uber, AirBnb), wherein the platforms have customers not only on the demand side (e.g. users), but also on the supply side (e.g. drivers, hosts). Such multi-sided marketplace involves interaction between multiple stakeholders among which there are different individuals with assorted needs. While traditional recommender systems focused specifically towards increasing consumer satisfaction by providing relevant content to consumers, two-sided marketplaces face an interesting problem of optimizing their models for supplier preferences, and visibility. In this tutorial, we consider a number of research problems which need to be address when developing a recommendation framework powering a multi-stakeholder marketplace, and provides audience with a profound introduction to this upcoming area and presents directions of further research. Tutorial material available at: https://rishabhmehrotra.github.io/recs-in-marketplace/

Evaluating algorithmic recommendations is an important, but difficult, problem. Evaluations conducted offline using data collected from user interactions with an online system often suffer from biases arising from the user interface or the recommendation engine. Online evaluation (A/B testing) can more easily address problems of bias, but depending on setting can be time-consuming and incur risk of negatively impacting the user experience, not to mention that it is generally more difficult when access to a large user base is not taken as granted. A compromise based on \em counterfactual analysis is to present some subset of online users with recommendation results that have been randomized or otherwise manipulated, log their interactions, and then use those to de-bias offline evaluations on historical data. However, previous work does not offer clear conclusions on how well such methods correlate with and are able to predict the results of online A/B tests. Understanding this is crucial to widespread adoption of new offline evaluation techniques in recommender systems. In this work we present a comparison of offline and online evaluation results for a particular recommendation problem: recommending playlists of tracks to a user looking for music. We describe two different ways to think about de-biasing offline collections for more accurate evaluation. Our results show that, contrary to much of the previous work on this topic, properly-conducted offline experiments do correlate well to A/B test results, and moreover that we can expect an offline evaluation to identify the best candidate systems for online testing with high probability.

Recently, there has been considerable interest in the use of historical logged user interaction data—queries and clicks—for evaluation of search systems in the context of counterfactual analysis [8,10]. Recent approaches attempt to de-bias the historical log data by conducting randomization experiments and modeling the bias in user behavior. Thus far, the focus has been on addressing bias that arises due to the position of the document being clicked (position-bias) or sparsity of clicks on certain query-document pairs (selection-bias). However, there is another source of bias that could arise: the bias due to the context in which a document was presented to the user. The propensity of the user clicking on a document depends not only on its position but also on many other contextual factors. In this work, we show that the existing counterfactual estimators fail to capture one type of bias, specifically, the effect on click-through rates due to the relevance of documents ranked above. Further, we propose a modification to the existing estimator that takes into account this bias. We rely on full result randomization that allows us to control for the click context at various ranks; we demonstrate the effectiveness of our methods in evaluating retrieval system through experiments on a simulation setup that is designed to cover a wide variety of scenarios.

Evaluation is a fundamental part of a recommendation system. Evaluation typically takes one of three forms: (1) smaller lab studies with real users; (2) batch tests with offline collections, judgements, and measures; (3) large-scale controlled experiments (e.g. A/B tests) looking at implicit feedback. But it is rare for the first to inform and influence the latter two; in particular, implicit feedback metrics often have to be continuously revised and updated as assumptions are found to be poorly supported.

We investigate the use of logged user interaction data---queries and clicks---for offline evaluation of new search systems in the context of counterfactual analysis. The challenge of evaluating a new ranker against log data collected from a static production ranker is that new rankers may retrieve documents that have never been seen in the logs before, and thus lack any logged feedback from users. Additionally, the ranker itself could bias user actions such that even documents that have been seen in the logs would have exhibited different interaction patterns had they been retrieved and ranked by the new ranker. We present a methodology for incrementally logging interactions on previously-unseen documents for use in computation of an unbiased estimator of a new ranker's effectiveness. Our method is very lightly invasive with respect to the production ranker results to insure against users becoming dissatisfied if the new ranker is poor. We demonstrate how well our methods work in a simulation environment designed to be challenging for such methods to argue that they are likely to work in a wide variety of scenarios.

Users interact with search engine result pages in various ways, including their clicks, cursor movements, and page scrolls. Researchers model such user interaction behavior in order to understand users and improve search result presentation. In this paper we propose nine different user click models that take various real life search behavior into account. Our main assumption is that users base their clicks primarily on features of document titles, URLs, and snippets over the entire results page, as well as the queries submitted to the search engine and the documents seen previously in their session history. We evaluate our click models by their effectiveness at predicting actual user clicks using standard classification evaluation measures precision, recall, and area under the ROC curve (AUC). We show that incorporating information about the entire results page gives nearly 40% improvement in precision, and including information about the session history along with it increases precision by up to 43%.

We analyze 5,792 IR conference papers published over 20 years to investigate how researchers have used and are using statistical significance testing in their experiments

The past 20 years have seen a great improvement in the rigor of information retrieval experimentation, due primarily to two factors: high-quality, public, portable test collections such as those produced by TREC (the Text REtrieval Conference), and the increased practice of statistical hypothesis testing to determine whether measured improvements can be ascribed to something other than random chance. Together these create a very useful standard for reviewers, program committees, and journal editors; work in information retrieval (IR) increasingly cannot be published unless it has been evaluated using a well-constructed test collection and shown to produce a statistically significant improvement over a good baseline. But, as the saying goes, any tool sharp enough to be useful is also sharp enough to be dangerous. Statistical tests of significance are widely misunderstood. Most researchers and developers treat them as a "black box": evaluation results go in and a p-value comes out. But because significance is such an important factor in determining what research directions to explore and what is published, using p-values obtained without thought can have consequences for everyone doing research in IR. Ioannidis has argued that the main consequence in the biomedical sciences is that most published research findings are false; could that be the case in IR as well?

Information Retrieval (IR) research has traditionally focused on serving the best results for a single query - so-called ad hoc retrieval. However, users typically search iteratively, refining and reformulating their queries during a session. A key challenge in the study of this interaction is the creation of suitable evaluation resources to assess the effectiveness of IR systems over sessions. This paper describes the TREC Session Track, which ran from 2010 through to 2014, which focussed on forming test collections that included various forms of implicit feedback. We describe the test collections; a brief analysis of the differences between datasets over the years; and the evaluation results that demonstrate that the use of user session data significantly improved effectiveness.

Data fusion has been shown to be a simple and effective way to improve retrieval results. Most existing data fusion methods combine ranked lists from different retrieval functions for a single given query-but in most real search settings, the diversity of retrieval functions required to achieve good fusion performance is not available. This paper presents a method for data fusion based on combining ranked lists from different queries that users could have entered for their information need, keeping the retrieval function fixed. We argue that if we can obtain a set of "possible queries" for an information need, we can achieve high effectiveness by fusing the rankings over the possible queries. In order to demonstrate effectiveness, we present experimental results on 5 different datasets covering tasks such as ad-hoc search, novelty and diversity search, and search in the presence of implicit user feedback. Our results show strong performances for our method, it is competitive with state-of-the-art methods on the same datasets, and in some cases outperforms them.

Similarity measures have been used widely in information retrieval research. Most research has been done on query-document or document-document similarity without much attention to the user's perception of similarity in the context of the information need. In this study, we collect user preference judgements of web document similarity in order to investigate: (1) the correlation between similarity measures and users' perception of similarity, (2) the correlation between the web document features plus document-query features and users' similarity judgements. We analyze the performance of various similarity methods at predicting user preferences, in both unsupervised and supervised settings. We show that a supervised approach using many features is able to predict user preferences close to the level of agreement between users, and moreover achieve a 15% improvement in AUC over an unsupervised approach.

Different users may be attempting to satisfy different information needs while providing the same query to a search engine. Addressing that issue is addressing Novelty and Diversity in information retrieval. Novelty and Diversity search task models the task wherein users are interested in seeing more and more documents that are not only relevant, but also cover more aspects (or subtopics) related to the topic of interest. This is in contrast with the traditional IR task where topical relevance is the only factor in evaluating search results. In this paper, we conduct a user study where users are asked to give a preference between one of two documents B and C given a query and also given that they have already seen a document A. We then test a total of ten hypotheses pertaining to the relationship between the "comprehensiveness" of documents (i.e. the number of subtopics a document is relevant to) and real users' preference judgments. Our results show that users are inclined to prefer documents with higher comprehensiveness, even when the prior document A already covers more aspects than the two documents being compared, and even when the least preferred has a higher relevance grade. In fact, users are inclined to prefer documents with higher overall aspect-coverage even in cases where B and C are relevant to the same number of novel subtopics.

Reusable test collections allow researchers to rapidly test different algorithms to find the one that works "best". But because of randomness in the topic sample, or in relevance judgments, or in interactions among system components, extreme results can be seen entirely due to chance, particularly when a collection becomes very popular. We argue that the best known published effectiveness on any given collection could be measured as much as 20% higher than its "true" intrinsic effectiveness, and that there are many other systems with lower measured effectiveness that could have substantially higher intrinsic effectiveness.

The past 20 years have seen a great improvement in the rigor of information retrieval experimentation, due primarily to two factors: high-quality, public, portable test collections such as those produced by TREC (the Text REtrieval Con- ference [2]), and the increased practice of statistical hypothesis testing to determine whether measured improvements can be ascribed to something other than random chance. Together these create a very useful standard for reviewers, program committees, and journal editors; work in information retrieval (IR) increasingly cannot be published unless it has been evaluated using a well-constructed test collection and shown to produce a statistically significant improvement over a good baseline.

Novel and diverse document ranking is an effective strategy that involves reducing redundancy in a ranked list to maximize the amount of novel and relevant information available to users. Evaluation for novelty and diversity typically involves an assessor judging each document for relevance against a set of pre-identified subtopics, which may be disambiguations of the query, facets of an information need, or nuggets of information. Alternately, when expressing a \emph{preference} for document A or document B, users may implicitly take subtopics into account, but may also take into account other factors such as recency, readability, length, and so on, each of which may have more or less importance depending on user. A \emph{user profile} contains information about the extent to which each factor, including subtopic relevance, plays a role in the user's preference for one document over another. A preference-based evaluation can then take this user profile information into account to better model utility to the space of users.

The notion of relevance differs between assessors, thus giving rise to assessor disagreement. Although assessor disagreement has been frequently observed, the factors leading to disagreement are still an open problem. In this paper we study the relationship between assessor disagreement and various topic independent factors such as readability and cohesiveness. We build a logistic model using reading level and other simple document features to predict assessor disagreement and rank documents by decreasing probability of disagreement. We compare the predictive power of these document-level features with that of a meta-search feature that aggregates a document's ranking across multiple retrieval runs. Our features are shown to be on a par with the meta-search feature, without requiring a large and diverse set of retrieval runs to calculate. Surprisingly, however, we find that the reading level features are negatively correlated with disagreement, suggesting that they are detecting some other aspect of document content.

In this paper, we present a medical record search system which is useful for identifying cohorts required in clinical studies. In particular, we propose a query-adaptive weighting method that can dynamically aggregate and score evidence in multiple medical reports (from different hospital departments or from different tests within the same department) of a patient. Furthermore, we explore several informative features for learning our retrieval model.

Click logs present a wealth of evidence about how users interact with a search system. This evidence has been used for many things: learning rankings, personalizing, evaluating effectiveness, and more. But it is almost always distilled into point estimates of feature or parameter values, ignoring what may be the most salient feature of users---their variability. No two users interact with a system in exactly the same way, and even a single user may interact with results for the same query differently depending on information need, mood, time of day, and a host of other factors. We present a Bayesian approach to using logs to compute posterior distributions for probabilistic models of user interactions. Since they are distributions rather than point estimates, they naturally capture variability in the population. We show how to cluster posterior distributions to discover patterns of user interactions in logs, and discuss how to use the clusters to evaluate search engines according to a user model. Because the approach is Bayesian, our methods can be applied to very large logs (such as those possessed by Web search engines) as well as very small (such as those found in almost any other setting).

Assessors are well known to disagree frequently on the relevance of documents to a topic, but the factors leading to assessor disagreement are still poorly understood. In this paper, we examine the relationship between the rank at which a document is returned by a set of retrieval systems and the likelihood that a second assessor will disagree with the relevance assessment of the initial assessor, and find that there is a strong and consistent correlation between the two. We adopt a metarank method of summarizing a document's rank across multiple runs, and propose a logistic regression predictive model of second assessor disagreement given metarank and initially-assessed relevance. The consistency of the model parameters across different topics, assessor pairs, and collections is considered. The model gives comparatively accurate predictions of absolute system scores, but less consistent predictions of relative scores than a simpler rank-insensitive model. We demonstrate that the logistic regression model is robust to using sampled, rather than exhaustive, dual assessment. We demonstrate the use of the sampled predictive model to incorporate assessor disagreement into tests of statistical significance.

The goal of the tutorial is to provide attendees with a comprehensive overview of the latest advances in the development of information retrieval evaluation measures and discuss the current challenges in the area. A number of topics are covered, including background in traditional evaluation paradigm and traditional evaluation measures, evaluation measures based on user models, advanced models of user interaction with search engines, measures based on these models, measures for novelty and diversity, and session-based measures.

There has been considerable interest in incorporating diversity in search results to account for redundancy and the space of possible user needs. Most work on this problem is based on subtopics: diversity rankers score documents against a set of hypothesized subtopics, and diversity rankings are evaluated by assigning a value to each ranked document based on the number of novel (and redundant) subtopics it is relevant to. This can be seen as modeling a user who is always interested in seeing more novel subtopics, with progressively decreasing interest in seeing the same subtopic multiple times. We put this model to test: if it is correct, then users, when given a choice, should prefer to see a document that has more value to the evaluation. We formulate some specific hypotheses from this model and test them with actual users in a novel preference-based design in which users express a preference for document A or document B given document C. We argue that while the user study shows the subtopic model is good, there are many other factors apart from novelty and redundancy that may be influencing user preferences. From this, we introduce a new framework to construct an ideal diversity ranking using only preference judgments, with no explicit subtopic judgments whatsoever.

Recently, researchers have shown interest in the use of preference judgments for evaluation in IR literature. Although preference judgments have several advantages over absolute judgment, one of the major disadvantages is that the number of judgments needed increases polynomially as the number of documents in the pool increases. We propose a novel method using PageRank to minimize the number of judgments required to evaluate systems using preference judgments. We test the proposed hypotheses using the TREC 2004 to 2006 Terabyte dataset to show that it is possible to reduce the evaluation cost considerably. Further, we study the susceptibility of the methods due to assessor errors.

High-quality reusable test collections and formal statistical hypothesis testing together support a rigorous experimental environment for information retrieval research. But as Armstrong et al. [2009b] recently argued, global analysis of experiments suggests that there has actually been little real improvement in ad hoc retrieval effectiveness over time. We investigate this phenomenon in the context of simultaneous testing of many hypotheses using a fixed set of data. We argue that the most common approaches to significance testing ignore a great deal of information about the world. Taking into account even a fairly small amount of this information can lead to very different conclusions about systems than those that have appeared in published literature. We demonstrate how to model a set of IR experiments for analysis both mathematically and practically, and show that doing so can cause p -values from statistical hypothesis tests to increase by orders of magnitude. This has major consequences on the interpretation of experimental results using reusable test collections: it is very difficult to conclude that anything is significant once we have modeled many of the sources of randomness in experimental design and analysis.

Information retrieval effectiveness evaluation typically takes one of two forms: batch experiments based on static test collections, or lab studies measuring actual users interacting with a system. Test collection experiments are sometimes viewed as introducing too many simplifying assumptions to accurately predict the usefulness of a system to its users. As a result, there is great interest in creating test collections and measures that better model user behavior. One line of research involves developing measures that include a parameterized user model; choosing a parameter value simulates a particular type of user. We propose that these measures offer an opportunity to more accurately simulate the variance due to user behavior, and thus to analyze system effectiveness to a simulated user population. We introduce a Bayesian procedure for producing sampling distributions from click data, and show how to use statistical tools to quantify the effects of variance due to parameter selection.

There is great interest in producing effectiveness measures that model user behavior in order to better model the utility of a system to its users. These measures are often formulated as a sum over the product of a discount function of ranks and a gain function mapping relevance assessments to numeric utility values. We develop a conceptual framework for analyzing such effectiveness measures based on classifying members of this broad family of measures into four distinct families, each of which reflects a different notion of system utility. Within this framework we can hypothesize about the properties that such a measure should have and test those hypotheses against user and system data. Along the way we present a collection of novel results about specific measures and relationships between them.

The standard system-based evaluation paradigm has focused on assessing the performance of retrieval systems in serving the best results for a single query. Real users, however, often begin an interaction with a search engine with a sufficiently under-specified query that they will need to reformulate before they find what they are looking for. In this work we consider the problem of evaluating retrieval systems over test collections of multi-query sessions. We propose two families of measures: a model-free family that makes no assumption about the user's behavior over a session, and a model-based family with a simple model of user interactions over the session. In both cases we generalize traditional evaluation metrics such as average precision to multi-query session evaluation. We demonstrate the behavior of the proposed metrics by using the new TREC 2010 Session track collection and simulations over the TREC-9 Query track collection.

Portable, reusable test collections are a vital part of research and development in information retrieval. Reusability is difficult to assess, however. The standard approach— simulating judgment collection when groups of systems are held out, then evaluating those held-out systems—only works when there is a large set of relevance judgments to draw on during the simulation. As test collections adapt to larger and larger corpora, it becomes less and less likely that there will be sufficient judgments for such simulation experiments. Thus we propose a methodology for information retrieval experimentation that collects evidence for or against the reusability of a test collection while judgments are being made. Using this methodology along with the appropriate statistical analyses, researchers will be able to estimate the reusability of their test collections while building them and implement “course corrections ” if the collection does not seem to be achieving desired levels of reusability. We show the robustness of our design to inherent sources of variance, and provide a description of an actual implementation of the framework for creating a large test collection.

Search corpora are growing larger and larger: over the last 10 years, the IR research community has moved from the several hundred thousand documents on the TREC disks to the tens of millions of U.S. government web pages of GOV2 to the one billion general-interest web pages in the new ClueWeb09 collection. But traditional means of acquiring relevance judgments and evaluating - e.g. pooling documents to calculate average precision - do not seem to scale well to these new large collections. They require substantially more cost in human assessments for the same reliability in evaluation; if the additional cost goes over the assessing budget, errors in evaluation are inevitable.

Recent efforts in test collection building have focused on scaling back the number of necessary relevance judgments and then scaling up the number of search topics. Since the largest source of variation in a Cranfield-style experiment comes from the topics, this is a reasonable approach. However, as topic set sizes grow, and researchers look to crowdsourcing and Amazon's Mechanical Turk to collect relevance judgments, we are faced with issues of quality control. This paper examines the robustness of the TREC Million Query track methods when some assessors make significant and systematic errors. We find that while averages are robust, assessor errors can have a large effect on system rankings.

A set of words is often insufficient to express a user's information need. In order to account for various information needs associated with a query, diversification seems to be a reasonable strategy. By diversifying the result set, we increase the probability of results being relevant to the user's information needs when the given query is ambiguous. A diverse result set must contain a set of documents that cover various subtopics for a given query. We propose a graph based method which exploits the link structure of the web to return a ranked list that provides complete coverage for a query. Our method not only provides diversity to the results set, but also avoids excessive redundancy. Moreover, the probability of relevance of a document is conditioned on the documents that appear before it in the result list. We show the effectiveness of our method by comparing it with a query-likelihood model as the baseline.

While test collection construction is a time-consuming and expensive process, the true cost is amortized by reusing the collection over hundreds or thousands of experiments. Some of these experiments may involve systems that retrieve documents not judged during the initial construction phase, and some of these systems may be "hard" to evaluate: depending on which judgments are missing and which judged documents were retrieved, the experimenter's confidence in an evaluation could potentially be very low. We propose two methods for quantifying the reusability of a test collection for evaluating new systems. The proposed methods provide simple yet highly effective tests for determining whether an existing set of judgments is useful for evaluating a new system. Empirical evaluations using TREC datasets confirm the usefulness of our proposed reusability measures. In particular, we show that our methods can reliably estimate confidence intervals that are indicative of collection reusability.

Traditional models of information retrieval assume documents are independently relevant. But when the goal is retrieving diverse or novel information about a topic, retrieval models need to capture dependencies between documents. Such tasks require alternative evaluation and optimization methods that operate on different types of relevance judgments. We define faceted topic retrieval as a particular novelty-driven task with the goal of finding a set of documents that cover the different facets of an information need. A faceted topic retrieval system must be able to cover as many facets as possible with the smallest number of documents. We introduce two novel models for faceted topic retrieval, one based on pruning a set of retrieved documents and one based on retrieving sets of documents through direct optimization of evaluation measures. We compare the performance of our models to MMR and the probabilistic model due to Zhai et al. on a set of 60 topics annotated with facets, showing that our models are competitive.

Rank correlation statistics are useful for determining whether a there is a correspondence between two measurements, particularly when the measures themselves are of less interest than their relative ordering. Kendall's - in particular has found use in Information Retrieval as a "meta-evaluation" measure: it has been used to compare evaluation measures, evaluate system rankings, and evaluate predicted performance. In the meta-evaluation domain, however, correlations between systems confound relationships between measurements, practically guaranteeing a positive and significant estimate of - regardless of any actual correlation between the measurements. We introduce an alternative measure of distance between rankings that corrects this by explicitly accounting for correlations between systems over a sample of topics, and moreover has a probabilistic interpretation for use in a test of statistical significance. We validate our measure with theory, simulated data, and experiment.

Research has shown that little practical difference exists between the randomization, Student's paired t, and bootstrap tests of statistical significance for TREC ad-hoc retrieval experiments with 50 topics. We compared these three tests on runs with topic sizes down to 10 topics. We found that these tests show increasing disagreement as the number of topics decreases. At smaller numbers of topics, the randomization test tended to produce smaller p-values than the t-test for p-values less than 0.1. The bootstrap exhibited a systematic bias towards p-values strictly less than the t-test with this bias increasing as the number of topics decreased. We recommend the use of the randomization test although the t-test appears to be suitable even when the number of topics is small.

There has been recent interest in collecting user or assessor preferences, rather than absolute judgments of relevance, for the evaluation or learning of ranking algorithms. Since measures like precision, recall, and DCG are defined over absolute judgments, evaluation over preferences will require new evaluation measures that explicitly model them. We describe a class of such measures and compare absolute and preference measures over a large TREC collection.

Information retrieval evaluation has typically been performed over several dozen queries, each judged to near-completeness. There has been a great deal of recent work on evaluation over much smaller judgment sets: how to select the best set of documents to judge and how to estimate evaluation measures when few judgments are available. In light of this, it should be possible to evaluate over many more queries without much more total judging effort. The Million Query Track at TREC 2007 used two document selection algorithms to acquire relevance judgments for more than 1,800 queries. We present results of the track, along with deeper analysis: investigating tradeoffs between the number of queries and number of judgments shows that, up to a point, evaluation over more queries with fewer judgments is more cost-effective and as reliable as fewer queries with more judgments. Total assessor effort can be reduced by 95% with no appreciable increase in evaluation errors.

Semiautomatic evaluation of retrieval systems using document similarities.

Information retrieval experimentation generally proceeds in a cycle of development, evaluation, and hypothesis testing. Ideally, the evaluation and testing phases should be short and easy, so as to maximize the amount of time spent in development. There has been recent work on reducing the amount of assessor effort needed to evaluate retrieval sys-tems, but it has not, for the most part, investigated the ef-fects of these methods on tests of significance. In this work, we explore in detail the effects of reduced sets of judgments on the sign test. We demonstrate both analytically and em-pirically the relationship between the power of the test, the number of topics evaluated, and the number of judgments available. Using these relationships, we can determine the number of topics and judgments needed for the least-cost but highest-confidence significance evaluation. Specifically, testing pairwise significance over 192 topics with fewer than 5 judgments for each is as good as testing significance over 25 topics with an average of 166 judgments for each—85% less effort producing no additional errors.

Information retrieval (IR) researchers commonly use three tests of statistical significance: the Student's paired t-test, the Wilcoxon signed rank test, and the sign test. Other researchers have previously proposed using both the bootstrap and Fisher's randomization (permutation) test as non-parametric significance tests for IR but these tests have seen little use. For each of these five tests, we took the ad-hoc retrieval runs submitted to TRECs 3 and 5-8, and for each pair of runs, we measured the statistical significance of the difference in their mean average precision. We discovered that there is little practical difference between the randomization, bootstrap, and t tests. Both the Wilcoxon and sign test have a poor ability to detect significance and have the potential to lead to false detections of significance. The Wilcoxon and sign tests are simplified variants of the randomization test and their use should be discontinued for measuring the significance of a difference between means.

Low-cost methods for acquiring relevance judgments can be a boon to researchers who need to evaluate new retrieval tasks or topics but do not have the resources to make thousands of judgments. While these judgments are very useful for a one-time evaluation, it is not clear that they can be trusted when re-used to evaluate new systems. In this work, we formally define what it means for judgments to be reusable: the confidence in an evaluation of new systems can be accurately assessed from an existing set of relevance judgments. We then present a method for augmenting a set of relevance judgments with relevance estimates that require no additional assessor effort. Using this method practically guarantees reusability: with as few as five judgments per topic taken from only two systems, we can reliably evaluate a larger set of ten systems. Even the smallest sets of judgments can be useful for evaluation of new systems.

Accurate estimation of information retrieval evaluation metrics such as average precision require large sets of relevance judgments. Building sets large enough for evaluation of real-world implementations is at best inefficient, at worst infeasible. In this work we link evaluation with test collection construction to gain an understanding of the minimal judging effort that must be done to have high confidence in the outcome of an evaluation. A new way of looking at average precision leads to a natural algorithm for selecting documents to judge and allows us to estimate the degree of confidence by defining a distribution over possible document judgments. A study with annotators shows that this method can be used by a small group of researchers to rank a set of systems in under three hours with 95% confidence.

We introduce a novel approach to combining rankings from multiple retrieval systems. We use a logistic regression model or an SVM to learn a ranking from pairwise document preferences. Our approach requires no training data or relevance scores, and outperforms a popular voting algorithm.

Corpora and topics are readily available for information retrieval research. Relevance judgments, which are necessary for system evaluation, are expensive; the cost of obtaining them prohibits in-house evaluation of retrieval systems on new corpora or new topics. We present an algorithm for cheaply constructing sets of relevance judgments. Our method intelligently selects documents to be judged and decides when to stop in such a way that with very little work there can be a high degree of confidence in the result of the evaluation. We demonstrate the algorithm's effectiveness by showing that it produces small sets of relevance judgments that reliably discriminate between two systems. The algorithm can be used to incrementally design retrieval systems by simultaneously comparing sets of systems. The number of additional judgments needed after each incremental design change decreases at a rate reciprocal to the number of systems being compared. To demonstrate the effectiveness of our method, we evaluate TREC ad hoc submissions, showing that with 95% fewer relevance judgments we can reach a Kendall's tau rank correlation of at least 0.9.

We describe a user study that examined the relationship between the quality of an Information Retrieval system and the effectiveness of its users in performing a task. The task involves finding answer facets of questions pertaining to a collection of newswire documents over a six month period. We artificially created sets of ranked lists at increasing levels of quality by blending the output of a state-of-the-art retrieval system with truth data created by annotators. Subjects performed the task by using these ranked lists to guide their labeling of answer passages in the retrieved articles. We found that as system accuracy improves, subject time on task and error rate decrease, and the rate of finding new correct answers increases. There is a large intermediary region in which the utility difference is not significant; our results suggest that there is some threshold of accuracy for this task beyond which user utility improves rapidly, but more experiments are needed to examine the area around that threshold closely.