Research on click models usually focuses on developing effective approaches to reduce biases in user clicks. However, one of the major drawbacks of existing click models is the lack of scalability. In this work, we tackle the scalability of Expectation-Maximization (EM)-based click models by introducing ParClick, a new parallel algorithm designed by following the Partitioning-Communication-Aggregation-Mapping (PCAM) method. To this end, we first provide a generic formulation of EM-based click models. Then, we design an efficient parallel version of this generic click model following the PCAM approach: we partition user click logs and model parameters into separate tasks, analyze communication among them, and aggregate these tasks to reduce communication overhead. Finally, we provide a scalable, parallel implementation of the proposed design, which maps well on a multi-core machine. Our experiments on the Yandex relevance prediction dataset show that ParClick scales well when increasing the amount of training data and computational resources. In particular, ParClick is 24.7 times faster to train with 40 million search sessions and 40 threads compared to the standard sequential version of the Click Chain Model (CCM) without any degradation in effectiveness.

In counterfactual learning to rank (CLTR) user interactions are used as a source of supervision. Since user interactions come with bias, an important focus of research in this field lies in developing methods to correct for the bias of interactions. Inverse propensity scoring (IPS) is a popular method suitable for correcting position bias. Affine correction (AC) is a generalization of IPS that corrects for position bias and trust bias. IPS and AC provably remove bias, conditioned on an accurate estimation of the bias parameters. Estimating the bias parameters, in turn, requires an accurate estimation of the relevance probabilities. This cyclic dependency introduces practical limitations in terms of sensitivity, convergence and efficiency.

As the size and complexity of models and datasets grow, so does the need for communication-efficient variants of stochastic gradient descent that can be deployed to perform parallel model training. One popular communication-compression method for data-parallel SGD is QSGD (Alistarh et al., 2017), which quantizes and encodes gradients to reduce communication costs. The baseline variant of QSGD provides strong theoretical guarantees, however, for practical purposes, the authors proposed a heuristic variant which we call QSGDinf, which demonstrated impressive empirical gains for distributed training of large neural networks. In this paper, we build on this work to propose a new gradient quantization scheme, and show that it has both stronger theoretical guarantees than QSGD, and matches and exceeds the empirical performance of the QSGDinf heuristic and of other compression methods.

Unbiased counterfactual learning to rank (CLTR) requires click propensities to compensate for the difference between user clicks and true relevance of search results via inverse propensity scoring (IPS). Current propensity estimation methods assume that user click behavior follows the position-based click model (PBM) and estimate click propensities based on this assumption. However, in reality, user clicks often follow the cascade model (CM), where users scan search results from top to bottom and where each next click depends on the previous one. In this cascade scenario, PBM-based estimates of propensities are not accurate, which, in turn, hurts CLTR performance. In this paper, we propose a propensity estimation method for the cascade scenario, called cascade model-based inverse propensity scoring (CM-IPS). We show that CM-IPS keeps CLTR performance close to the full-information performance in case the user clicks follow the CM, while PBM-based CLTR has a significant gap towards the full-information. The opposite is true if the user clicks follow PBM instead of the CM. Finally, we suggest a way to select between CM- and PBM-based propensity estimation methods based on historical user clicks.

The top- N recommendation problem has been studied extensively. Item-based collaborative filtering recommendation algorithms show promising results for the problem. They predict a user’s preferences by estimating similarities between a target and user-rated items. Top- N recommendation remains a challenging task in scenarios where there is a lack of preference history for new items. Feature-based Similarity Models (FSMs) address this particular problem by extending item-based collaborative filtering by estimating similarity functions of item features. The quality of the estimated similarity function determines the accuracy of the recommendation. However, existing FSMs only estimate global similarity functions; i.e., they estimate using preference information across all users. Moreover, the estimated similarity functions are linear ; hence, they may fail to capture the complex structure underlying item features. In this article, we propose to improve FSMs by estimating local similarity functions, where each function is estimated for a subset of like-minded users. To capture global preference patterns, we extend the global similarity function from linear to nonlinear, based on the effectiveness of variational autoencoders. We propose a Bayesian generative model, called the Local Variational Feature-based Similarity Model, to encapsulate local and global similarity functions. We present a variational Expectation Minimization algorithm for efficient approximate inference. Extensive experiments on a large number of real-world datasets demonstrate the effectiveness of our proposed model.

Contextual bandit problems are a natural fit for many information retrieval tasks, such as learning to rank, text classification, recommendation, and so on. However, existing learning methods for contextual bandit problems have one of two drawbacks: They either do not explore the space of all possible document rankings (i.e., actions) and, thus, may miss the optimal ranking, or they present suboptimal rankings to a user and, thus, may harm the user experience. We introduce a new learning method for contextual bandit problems, Safe Exploration Algorithm (SEA), which overcomes the above drawbacks. SEA starts by using a baseline (or production) ranking system (i.e., policy), which does not harm the user experience and, thus, is safe to execute but has suboptimal performance and, thus, needs to be improved. Then SEA uses counterfactual learning to learn a new policy based on the behavior of the baseline policy. SEA also uses high-confidence off-policy evaluation to estimate the performance of the newly learned policy. Once the performance of the newly learned policy is at least as good as the performance of the baseline policy, SEA starts using the new policy to execute new actions, allowing it to actively explore favorable regions of the action space. This way, SEA never performs worse than the baseline policy and, thus, does not harm the user experience, while still exploring the action space and, thus, being able to find an optimal policy. Our experiments using text classification and document retrieval confirm the above by comparing SEA (and a boundless variant called BSEA) to online and offline learning methods for contextual bandit problems.

Online ranker evaluation is one of the key challenges in information retrieval. Although the preferences of rankers can be inferred by interleaving methods, the problem of how to effectively choose the ranker pair that generates the interleaved list without degrading the user experience too much is still challenging. On the one hand, if two rankers have not been compared enough, the inferred preference can be noisy and inaccurate. On the other hand, if two rankers are compared too many times, the interleaving process inevitably hurts the user experience too much. This dilemma is known as the exploration versus exploitation tradeoff. It is captured by the K -armed dueling bandit problem, which is a variant of the K -armed bandit problem, where the feedback comes in the form of pairwise preferences. Today’s deployed search systems can evaluate a large number of rankers concurrently, and scaling effectively in the presence of numerous rankers is a critical aspect of K -armed dueling bandit problems. In this article, we focus on solving the large-scale online ranker evaluation problem under the so-called Condorcet assumption, where there exists an optimal ranker that is preferred to all other rankers. We propose Merge Double Thompson Sampling (MergeDTS), which first utilizes a divide-and-conquer strategy that localizes the comparisons carried out by the algorithm to small batches of rankers, and then employs Thompson Sampling to reduce the comparisons between suboptimal rankers inside these small batches. The effectiveness (regret) and efficiency (time complexity) of MergeDTS are extensively evaluated using examples from the domain of online evaluation for web search. Our main finding is that for large-scale Condorcet ranker evaluation problems, MergeDTS outperforms the state-of-the-art dueling bandit algorithms.

We consider the novel problem of evaluating a recommendation policy offline in environments where the reward signal is non-stationary. Non-stationarity appears in many Information Retrieval (IR) applications such as recommendation and advertising, but its effect on off-policy evaluation has not been studied at all. We are the first to address this issue. First, we analyze standard off-policy estimators in non-stationary environments and show both theoretically and experimentally that their bias grows with time. Then, we propose new off-policy estimators with moving averages and show that their bias is independent of time and can be bounded. Furthermore, we provide a method to trade-off bias and variance in a principled way to get an off-policy estimator that works well in both non-stationary and stationary environments. We experiment on publicly available recommendation datasets and show that our newly proposed moving average estimators accurately capture changes in non-stationary environments, while standard off-policy estimators fail to do so.

The visual appearance of a webpage carries valuable information about the page's quality and can be used to improve the performance of learning to rank (LTR). We introduce the Visual learning TO Rank (ViTOR) model that integrates state-of-the-art visual features extraction methods: (i) transfer learning from a pre-trained image classification model, and (ii) synthetic saliency heat maps generated from webpage snapshots. Since there is currently no public dataset for the task of LTR with visual features, we also introduce and release the ViTOR dataset, containing visually rich and diverse webpages. The ViTOR dataset consists of visual snapshots, non-visual features and relevance judgments for ClueWeb12 webpages and TREC Web Track queries. We experiment with the proposed ViTOR model on the ViTOR dataset and show that it significantly improves the performance of LTR with visual features.

We show that click models trained with suboptimal hyperparameters suffer from the issue of bad calibration. This means that their predicted click probabilities do not agree with the observed proportions of clicks in the held-out data. To repair this discrepancy, we adapt a non-parametric calibration method called isotonic regression. Our experimental results show that isotonic regression significantly improves click models trained with suboptimal hyperparameters in terms of perplexity, and that it makes click models less sensitive to the choice of hyperparameters. Interestingly, the relative ranking of existing click models in terms of their predictive performance changes depending on whether or not their predictions are calibrated. Therefore, we advocate that calibration becomes a mandatory part of the click model evaluation protocol.

Getting a better understanding of user behavior is important for advancing information retrieval systems. Existing work focuses on modeling and predicting single interaction events, such as clicks. In this paper, we for the first time focus on modeling and predicting sequences of interaction events. And in particular, sequences of clicks. We formulate the problem of click sequence prediction and propose a click sequence model (CSM) that aims to predict the order in which a user will interact with search engine results. CSM is based on a neural network that follows the encoder-decoder architecture. The encoder computes contextual embeddings of the results. The decoder predicts the sequence of positions of the clicked results. It uses an attention mechanism to extract necessary information about the results at each timestep. We optimize the parameters of CSM by maximizing the likelihood of observed click sequences. We test the effectiveness of CSM on three new tasks: (i) predicting click sequences, (ii) predicting the number of clicks, and (iii) predicting whether or not a user will interact with the results in the order these results are presented on a search engine result page (SERP). Also, we show that CSM achieves state-of-the-art results on a standard click prediction task, where the goal is to predict an unordered set of results a user will click on.

We address the task of fusing ranked lists of documents that are retrieved in response to a query. Past work on this task of rank aggregation often assumes that documents in the lists being fused are independent and that only the documents that are ranked high in many lists are likely to be relevant to a given topic. We propose manifold learning aggregation approaches, ManX and v-ManX, that build on the cluster hypothesis and exploit inter-document similarity information. ManX regularizes document fusion scores, so that documents that appear to be similar within a manifold, receive similar scores, whereas v-ManX first generates virtual adversarial documents and then regularizes the fusion scores of both original and virtual adversarial documents. Since aggregation methods built on the cluster hypothesis are computationally expensive, we adopt an optimization method that uses the top-k documents as anchors and considerably reduces the computational complexity of manifold-based methods, resulting in two efficient aggregation approaches, a-ManX and a-v-ManX. We assess the proposed approaches experimentally and show that they significantly outperform the state-of-the-art aggregation approaches, while a-ManX and a-v-ManX run faster than ManX, v-ManX, respectively.

Click models allow us to interpret user click behavior in search interactions and to remove various types of bias from user clicks. Existing studies on click models consider a static scenario where user click behavior does not change over time. We show empirically that click models deteriorate over time if retraining is avoided. We then adapt online expectation-maximization (EM) techniques to efficiently incorporate new click/skip observations into a trained click model. Our instantiation of Online EM for click models is orders of magnitude more efficient than retraining the model from scratch using standard EM, while loosing little in quality. To deal with outdated click information, we propose a variant of online EM called EM with Forgetting, which surpasses the performance of complete retraining while being as efficient as Online EM.

In web search, information about times between user actions has been shown to be a good indicator of users' satisfaction with the search results. Existing work uses the mean values of the observed times, or fits probability distributions to the observed times. This implies a context-independence assumption that the time elapsed between a pair of user actions does not depend on the context, in which the first action takes place. We validate this assumption using logs of a commercial web search engine and discover that it does not always hold. For between 37% to 80% of query-result pairs, depending on the number of observations, the distributions of click dwell times have statistically significant differences in query sessions for which a given result (i) is the first item to be clicked and (ii) is not the first. To account for this context bias effect, we propose a context-aware time model (CATM). The CATM allows us (i) to predict times between user actions in contexts, in which these actions were not observed, and (ii) to compute context-independent estimates of the times by predicting them in predefined contexts. Our experimental results show that the CATM provides better means than existing methods to predict and interpret times between user actions.

In this tutorial we give an overview of click models for web search. We show how the framework of probabilistic graphical models helps to explain user behavior, build new evaluation metrics and perform simulations. The tutorial discusses foundational aspects alongside experimental details and applications, with live demos and discussions of publicly available resources.

Understanding user browsing behavior in web search is key to improving web search effectiveness. Many click models have been proposed to explain or predict user clicks on search engine results. They are based on the probabilistic graphical model (PGM) framework, in which user behavior is represented as a sequence of observable and hidden events. The PGM framework provides a mathematically solid way to reason about a set of events given some information about other events. But the structure of the dependencies between the events has to be set manually. Different click models use different hand-crafted sets of dependencies. We propose an alternative based on the idea of distributed representations: to represent the user's information need and the information available to the user with a vector state. The components of the vector state are learned to represent concepts that are useful for modeling user behavior. And user behavior is modeled as a sequence of vector states associated with a query session: the vector state is initialized with a query, and then iteratively updated based on information about interactions with the search engine results. This approach allows us to directly understand user browsing behavior from click-through data, i.e., without the need for a predefined set of rules as is customary for PGM-based click models. We illustrate our approach using a set of neural click models. Our experimental results show that the neural click model that uses the same training data as traditional PGM-based click models, has better performance on the click prediction task (i.e., predicting user click on search engine results) and the relevance prediction task (i.e., ranking documents by their relevance to a query). An analysis of the best performing neural click model shows that it learns similar concepts to those used in traditional click models, and that it also learns other concepts that cannot be designed manually.

This tutorial concerns with more advanced and more recent topics in the area of click models. Here, we discuss recent developments in the area with a particular focus on applications of click models. The tutorial features a guest talk and a live demo where participants have a chance to build their own advanced click model.

In this introductory tutorial we give an overview of click models for web search. We show how the framework of probabilistic graphical models help to explain user behavior, build new evaluation metrics and perform simulations. The tutorial is augmented with a live demo where participants have a chance to implement a click model and to test it on a publicly available dataset.

Today's web search systems present users with heterogeneous information coming from sources of different types, also known as verticals. Evaluating such systems is an important but complex task, which is still far from being solved. In this paper we examine the hypothesis that the use of models that capture user search behavior on heterogeneous result pages helps to improve the quality of offline metrics. We propose two vertical-aware metrics based on user click models for federated search and evaluate them using query logs of the Yandex search engine. We show that depending on the type of vertical, the proposed metrics have higher correlation with online user behavior than other state-of-the-art techniques.

In a distributed retrieval setup, resource selection is the problem of identifying and ranking relevant sources of information for a given user’s query. For better usage of existing resource-selection techniques, it is desirable to know what the fundamental differences between them are and in what settings one is superior to others. However, little is understood still about the actual behavior of resource-selection methods. In this work, we focus on small-document approaches to resource selection that rank and select sources based on the ranking of their documents. We pose a number of research questions and approach them by three types of analyses. First, we present existing small-document techniques in a unified framework and analyze them theoretically. Second, we propose using a qualitative analysis to study the behavior of different small-document approaches. Third, we present a novel experimental methodology to evaluate small-document techniques and to validate the results of the qualitative analysis. This way, we answer the posed research questions and provide insights about small-document methods in general and about each technique in particular.

We give a fresh look into score normalization for merging result-lists, isolating the problem from other components. We focus on three of the simplest, practical, and widely-used linear methods which do not require any training data, i.e. MinMax, Sum, and Z-Score. We provide theoretical arguments on why and when the methods work, and evaluate them experimentally. We find that MinMax is the most robust under many circumstances, and that Sum is - in contrast to previous literature - the worst. Based on the insights gained, we propose another three simple methods which work as good or better than the baselines.

Distributed Information Retrieval (DIR), also known as Federated Search, integrates multiple searchable collections and provides direct access to them through a unified interface [3]. This is done by a centralized broker, that receives user queries, forwards them to appropriate collections and returns merged results to users.