Off-policy learning is a framework for optimizing policies without deploying them, using data collected by another policy. In recommender systems, this is especially challenging due to the imbalance in logged data: some items are recommended and thus logged more frequently than others. This is further perpetuated when recommending a list of items, as the action space is combinatorial. To address this challenge, we study pessimistic off-policy optimization for learning to rank. The key idea is to compute lower confidence bounds on parameters of click models and then return the list with the highest pessimistic estimate of its value. This approach is computationally efficient, and we analyze it. We study its Bayesian and frequentist variants and overcome the limitation of unknown prior by incorporating empirical Bayes. To show the empirical effectiveness of our approach, we compare it to off-policy optimizers that use inverse propensity scores or neglect uncertainty. Our approach outperforms all baselines and is both robust and general.
Unbiased learning to rank (ULTR) aims to train an unbiased ranking model from biased user click logs. Most of the current ULTR methods are based on the examination hypothesis (EH), which assumes that the click probability can be factorized into two scalar functions, one related to ranking features and the other related to bias factors. Unfortunately, the interactions among features, bias factors and clicks are complicated in practice, and usually cannot be factorized in this independent way. Fitting click data with EH could lead to model misspecification and bring the approximation error. In this paper, we propose a vector-based EH and formulate the click probability as a dot product of two vector functions. This solution is complete due to its universality in fitting arbitrary click functions. Based on it, we propose a novel model named Vectorization to adaptively learn the relevance embeddings and sort documents by projecting embeddings onto a base vector. Extensive experiments show that our method significantly outperforms the state-of-the-art ULTR methods on complex real clicks as well as simple simulated clicks.
Interpretable Learning to Rank (LtR) is an emerging field within the research area of explainable AI, aiming at developing intelligible and accurate predictive models. While most of the previous research efforts focus on creating post-hoc explanations, in this paper we investigate how to train effective and intrinsically-interpretable ranking models. Developing these models is particularly challenging and it also requires finding a trade-off between ranking quality and model complexity. State-of-the-art rankers, made of either large ensembles of trees or several neural layers, exploit in fact an unlimited number of feature interactions making them black boxes. Previous approaches on intrinsically-interpretable ranking models address this issue by avoiding interactions between features thus paying a significant performance drop with respect to full-complexity models. Conversely, ILMART, our novel and interpretable LtR solution based on LambdaMART, is able to train effective and intelligible models by exploiting a limited and controlled number of pairwise feature interactions. Exhaustive and reproducible experiments conducted on three publicly-available LtR datasets show that ILMART outperforms the current state-of-the-art solution for interpretable ranking of a large margin with a gain of nDCG of up to 8%.
Crowd image is arguably one of the most laborious data to annotate. In this paper, we devote to reduce the massive demand of densely labeled crowd data, and propose a novel weakly-supervised setting, in which we leverage the binary ranking of two images with high-contrast crowd counts as training guidance. To enable training under this new setting, we convert the crowd count regression problem to a ranking potential prediction problem. In particular, we tailor a Siamese Ranking Network that predicts the potential scores of two images indicating the ordering of the counts. Hence, the ultimate goal is to assign appropriate potentials for all the crowd images to ensure their orderings obey the ranking labels. On the other hand, potentials reveal the relative crowd sizes but cannot yield an exact crowd count. We resolve this problem by introducing "anchors" during the inference stage. Concretely, anchors are a few images with count labels used for referencing the corresponding counts from potential scores by a simple linear mapping function. We conduct extensive experiments to study various combinations of supervision, and we show that the proposed method outperforms existing weakly-supervised methods without additional labeling effort by a large margin.
While China has become the biggest online market in the world with around 1 billion internet users, Baidu runs the world largest Chinese search engine serving more than hundreds of millions of daily active users and responding billions queries per day. To handle the diverse query requests from users at web-scale, Baidu has done tremendous efforts in understanding users' queries, retrieve relevant contents from a pool of trillions of webpages, and rank the most relevant webpages on the top of results. Among these components used in Baidu search, learning to rank (LTR) plays a critical role and we need to timely label an extremely large number of queries together with relevant webpages to train and update the online LTR models. To reduce the costs and time consumption of queries/webpages labeling, we study the problem of Activ Learning to Rank (active LTR) that selects unlabeled queries for annotation and training in this work. Specifically, we first investigate the criterion -- Ranking Entropy (RE) characterizing the entropy of relevant webpages under a query produced by a sequence of online LTR models updated by different checkpoints, using a Query-By-Committee (QBC) method. Then, we explore a new criterion namely Prediction Variances (PV) that measures the variance of prediction results for all relevant webpages under a query. Our empirical studies find that RE may favor low-frequency queries from the pool for labeling while PV prioritizing high-frequency queries more. Finally, we combine these two complementary criteria as the sample selection strategies for active learning. Extensive experiments with comparisons to baseline algorithms show that the proposed approach could train LTR models achieving higher Discounted Cumulative Gain (i.e., the relative improvement {\Delta}DCG4=1.38%) with the same budgeted labeling efforts.
Decision forest (decision tree ensemble) is one of the most popular machine learning algorithms. To use large models on big data, like document scoring with learning-to-rank models, we need to evaluate these models efficiently. In this paper, we explore MatrixNet, the ancestor of the popular CatBoost library. Both libraries use the SSE instruction set for scoring on CPU. This paper investigates the opportunities given by the AVX instruction set to evaluate models more efficiently. We achieved 35% speedup on the binarization stage (nodes conditions comparison), and 20% speedup on the trees apply stage on the ranking model.
The Plackett-Luce (PL) model is ubiquitous in learning-to-rank (LTR) because it provides a useful and intuitive probabilistic model for sampling ranked lists. Counterfactual offline evaluation and optimization of ranking metrics are pivotal for using LTR methods in production. When adopting the PL model as a ranking policy, both tasks require the computation of expectations with respect to the model. These are usually approximated via Monte-Carlo (MC) sampling, since the combinatorial scaling in the number of items to be ranked makes their analytical computation intractable. Despite recent advances in improving the computational efficiency of the sampling process via the Gumbel top-k trick, the MC estimates can suffer from high variance. We develop a novel approach to producing more sample-efficient estimators of expectations in the PL model by combining the Gumbel top-k trick with quasi-Monte Carlo (QMC) sampling, a well-established technique for variance reduction. We illustrate our findings both theoretically and empirically using real-world recommendation data from Amazon Music and the Yahoo learning-to-rank challenge.
Learning-to-rank, a machine learning technique widely used in information retrieval, has recently been applied to the problem of ligand-based virtual screening, to accelerate the early stages of new drug development. Ranking prediction models learn based on ordinal relationships, making them suitable for integrating assay data from various environments. Existing studies of rank prediction in compound screening have generally used a learning-to-rank method called RankSVM. However, they have not been compared with or validated against the gradient boosting decision tree (GBDT)-based learning-to-rank methods that have gained popularity recently. Furthermore, although the ranking metric called Normalized Discounted Cumulative Gain (NDCG) is widely used in information retrieval, it only determines whether the predictions are better than those of other models. In other words, NDCG is incapable of recognizing when a prediction model produces worse than random results. Nevertheless, NDCG is still used in the performance evaluation of compound screening using learning-to-rank. This study used the GBDT model with ranking loss functions, called lambdarank and lambdaloss, for ligand-based virtual screening; results were compared with existing RankSVM methods and GBDT models using regression. We also proposed a new ranking metric, Normalized Enrichment Discounted Cumulative Gain (NEDCG), which aims to properly evaluate the goodness of ranking predictions. Results showed that the GBDT model with learning-to-rank outperformed existing regression methods using GBDT and RankSVM on diverse datasets. Moreover, NEDCG showed that predictions by regression were comparable to random predictions in multi-assay, multi-family datasets, demonstrating its usefulness for a more direct assessment of compound screening performance.
There are many contexts in which dyadic data are present. Social networks are a well-known example. In these contexts, pairs of elements are linked building a network that reflects interactions. Explaining why these relationships are established is essential to obtain transparency, an increasingly important notion. These explanations are often presented using text, thanks to the spread of the natural language understanding tasks. Our aim is to represent and explain pairs established by any agent (e.g., a recommender system or a paid promotion mechanism), so that text-based personalisation is taken into account. We have focused on the TripAdvisor platform, considering the applicability to other dyadic data contexts. The items are a subset of users and restaurants and the interactions the reviews posted by these users. We propose the PTER (Personalised TExt-based Reviews) model. We predict, from the available reviews for a given restaurant, those that fit to the specific user interactions. PTER leverages the BERT (Bidirectional Encoders Representations from Transformers) transformer-encoder model. We customised a deep neural network following the feature-based approach, presenting a LTR (Learning To Rank) downstream task. We carried out several comparisons of our proposal with a random baseline and other models of the state of the art, following the EXTRA (EXplanaTion RAnking) benchmark. Our method outperforms other collaborative filtering proposals.
There are several measures for fairness in ranking, based on different underlying assumptions and perspectives. PL optimization with the REINFORCE algorithm can be used for optimizing black-box objective functions over permutations. In particular, it can be used for optimizing fairness measures. However, though effective for queries with a moderate number of repeating sessions, PL optimization has room for improvement for queries with a small number of repeating sessions. In this paper, we present a novel way of representing permutation distributions, based on the notion of permutation graphs. Similar to PL, our distribution representation, called PPG, can be used for black-box optimization of fairness. Different from PL, where pointwise logits are used as the distribution parameters, in PPG pairwise inversion probabilities together with a reference permutation construct the distribution. As such, the reference permutation can be set to the best sampled permutation regarding the objective function, making PPG suitable for both deterministic and stochastic rankings. Our experiments show that PPG, while comparable to PL for larger session repetitions (i.e., stochastic ranking), improves over PL for optimizing fairness metrics for queries with one session (i.e., deterministic ranking). Additionally, when accurate utility estimations are available, e.g., in tabular models, the performance of PPG in fairness optimization is significantly boosted compared to lower quality utility estimations from a learning to rank model, leading to a large performance gap with PL. Finally, the pairwise probabilities make it possible to impose pairwise constraints such as "item $d_1$ should always be ranked higher than item $d_2$." Such constraints can be used to simultaneously optimize the fairness metric and control another objective such as ranking performance.
Movie Recommender System is widely applied in commercial environments such as NetFlix and Tubi. Classic recommender models utilize technologies such as collaborative filtering, learning to rank, matrix factorization and deep learning models to achieve lower marketing expenses and higher revenues. However, audience of movies have different ratings of the same movie in different contexts. Important movie watching contexts include audience mood, location, weather, etc. Tobe able to take advantage of contextual information is of great benefit to recommender builders. However, popular techniques such as tensor factorization consumes an impractical amount of storage, which greatly reduces its feasibility in real world environment. In this paper, we take advantage of the MatMat framework, which factorizes matrices by matrix fitting to build a context-aware movie recommender system that is superior to classic matrix factorization and comparable in the fairness metric.
While large-scale pre-trained language models like BERT have advanced the state-of-the-art in IR, its application in query performance prediction (QPP) is so far based on pointwise modeling of individual queries. Meanwhile, recent studies suggest that the cross-attention modeling of a group of documents can effectively boost performances for both learning-to-rank algorithms and BERT-based re-ranking. To this end, a BERT-based groupwise QPP model is proposed, in which the ranking contexts of a list of queries are jointly modeled to predict the relative performance of individual queries. Extensive experiments on three standard TREC collections showcase effectiveness of our approach. Our code is available at https://github.com/VerdureChen/Group-QPP.
Conventional methods for query autocompletion aim to predict which completed query a user will select from a list. A shortcoming of this approach is that users often do not know which query will provide the best retrieval performance on the current information retrieval system, meaning that any query autocompletion methods trained to mimic user behavior can lead to suboptimal query suggestions. To overcome this limitation, we propose a new approach that explicitly optimizes the query suggestions for downstream retrieval performance. We formulate this as a problem of ranking a set of rankings, where each query suggestion is represented by the downstream item ranking it produces. We then present a learning method that ranks query suggestions by the quality of their item rankings. The algorithm is based on a counterfactual learning approach that is able to leverage feedback on the items (e.g., clicks, purchases) to evaluate query suggestions through an unbiased estimator, thus avoiding the assumption that users write or select optimal queries. We establish theoretical support for the proposed approach and provide learning-theoretic guarantees. We also present empirical results on publicly available datasets, and demonstrate real-world applicability using data from an online shopping store.
Plackett-Luce gradient estimation enables the optimization of stochastic ranking models within feasible time constraints through sampling techniques. Unfortunately, the computational complexity of existing methods does not scale well with the length of the rankings, i.e. the ranking cutoff, nor with the item collection size. In this paper, we introduce the novel PL-Rank-3 algorithm that performs unbiased gradient estimation with a computational complexity comparable to the best sorting algorithms. As a result, our novel learning-to-rank method is applicable in any scenario where standard sorting is feasible in reasonable time. Our experimental results indicate large gains in the time required for optimization, without any loss in performance. For the field, our contribution could potentially allow state-of-the-art learning-to-rank methods to be applied to much larger scales than previously feasible.
In this perspective paper we study the effect of non independent and identically distributed (non-IID) data on federated online learning to rank (FOLTR) and chart directions for future work in this new and largely unexplored research area of Information Retrieval. In the FOLTR process, clients participate in a federation to jointly create an effective ranker from the implicit click signal originating in each client, without the need to share data (documents, queries, clicks). A well-known factor that affects the performance of federated learning systems, and that poses serious challenges to these approaches, is that there may be some type of bias in the way data is distributed across clients. While FOLTR systems are on their own rights a type of federated learning system, the presence and effect of non-IID data in FOLTR has not been studied. To this aim, we first enumerate possible data distribution settings that may showcase data bias across clients and thus give rise to the non-IID problem. Then, we study the impact of each setting on the performance of the current state-of-the-art FOLTR approach, the Federated Pairwise Differentiable Gradient Descent (FPDGD), and we highlight which data distributions may pose a problem for FOLTR methods. We also explore how common approaches proposed in the federated learning literature address non-IID issues in FOLTR. This allows us to unveil new research gaps that, we argue, future research in FOLTR should consider. This is an important contribution to the current state of FOLTR field because, for FOLTR systems to be deployed, the factors affecting their performance, including the impact of non-IID data, need to be thoroughly understood.
In practical multi-criterion decision-making, it is cumbersome if a decision maker (DM) is asked to choose among a set of trade-off alternatives covering the whole Pareto-optimal front. This is a paradox in conventional evolutionary multi-objective optimization (EMO) that always aim to achieve a well balance between convergence and diversity. In essence, the ultimate goal of multi-objective optimization is to help a decision maker (DM) identify solution(s) of interest (SOI) achieving satisfactory trade-offs among multiple conflicting criteria. Bearing this in mind, this paper develops a framework for designing preference-based EMO algorithms to find SOI in an interactive manner. Its core idea is to involve human in the loop of EMO. After every several iterations, the DM is invited to elicit her feedback with regard to a couple of incumbent candidates. By collecting such information, her preference is progressively learned by a learning-to-rank neural network and then applied to guide the baseline EMO algorithm. Note that this framework is so general that any existing EMO algorithm can be applied in a plug-in manner. Experiments on $48$ benchmark test problems with up to 10 objectives fully demonstrate the effectiveness of our proposed algorithms for finding SOI.
Unbiased learning to rank has been proposed to alleviate the biases in the search ranking, making it possible to train ranking models with user interaction data. In real applications, search engines are designed to display only the most relevant k documents from the retrieved candidate set. The rest candidates are discarded. As a consequence, position bias and sample selection bias usually occur simultaneously. Existing unbiased learning to rank approaches either focus on one type of bias (e.g., position bias) or mitigate the position bias and sample selection bias with separate components, overlooking their associations. In this study, we first analyze the mechanisms and associations of position bias and sample selection bias from the viewpoint of a causal graph. Based on the analysis, we propose Causal Likelihood Decomposition (CLD), a unified approach to simultaneously mitigating these two biases in top-k learning to rank. By decomposing the log-likelihood of the biased data as an unbiased term that only related to relevance, plus other terms related to biases, CLD successfully detaches the relevance from position bias and sample selection bias. An unbiased ranking model can be obtained from the unbiased term, via maximizing the whole likelihood. An extension to the pairwise neural ranking is also developed. Advantages of CLD include theoretical soundness and a unified framework for pointwise and pairwise unbiased top-k learning to rank. Extensive experimental results verified that CLD, including its pairwise neural extension, outperformed the baselines by mitigating both the position bias and the sample selection bias. Empirical studies also showed that CLD is robust to the variation of bias severity and the click noise.
Nowadays, state-of-the-art learning-to-rank methods are based on gradient-boosted decision trees (GBDT). The most well-known algorithm is LambdaMART which was proposed more than a decade ago. Recently, several other GBDT-based ranking algorithms were proposed. In this paper, we thoroughly analyze these methods in a unified setup. In particular, we address the following questions. Is direct optimization of a smoothed ranking loss preferable over optimizing a convex surrogate? How to properly construct and smooth surrogate ranking losses? To address these questions, we compare LambdaMART with YetiRank and StochasticRank methods and their modifications. We also propose a simple improvement of the YetiRank approach that allows for optimizing specific ranking loss functions. As a result, we gain insights into learning-to-rank techniques and obtain a new state-of-the-art algorithm.
Learning from a label distribution has achieved promising results on ordinal regression tasks such as facial age and head pose estimation wherein, the concept of adaptive label distribution learning (ALDL) has drawn lots of attention recently for its superiority in theory. However, compared with the methods assuming fixed form label distribution, ALDL methods have not achieved better performance. We argue that existing ALDL algorithms do not fully exploit the intrinsic properties of ordinal regression. In this paper, we emphatically summarize that learning an adaptive label distribution on ordinal regression tasks should follow three principles. First, the probability corresponding to the ground-truth should be the highest in label distribution. Second, the probabilities of neighboring labels should decrease with the increase of distance away from the ground-truth, i.e., the distribution is unimodal. Third, the label distribution should vary with samples changing, and even be distinct for different instances with the same label, due to the different levels of difficulty and ambiguity. Under the premise of these principles, we propose a novel loss function for fully adaptive label distribution learning, namely unimodal-concentrated loss. Specifically, the unimodal loss derived from the learning to rank strategy constrains the distribution to be unimodal. Furthermore, the estimation error and the variance of the predicted distribution for a specific sample are integrated into the proposed concentrated loss to make the predicted distribution maximize at the ground-truth and vary according to the predicting uncertainty. Extensive experimental results on typical ordinal regression tasks including age and head pose estimation, show the superiority of our proposed unimodal-concentrated loss compared with existing loss functions.
Clicks on rankings suffer from position-bias: generally items on lower ranks are less likely to be examined - and thus clicked - by users, in spite of their actual preferences between items. The prevalent approach to unbiased click-based learning-to-rank (LTR) is based on counterfactual inverse-propensity-scoring (IPS) estimation. In contrast with general reinforcement learning, counterfactual doubly-robust (DR) estimation has not been applied to click-based LTR in previous literature. In this paper, we introduce a novel DR estimator that is the first DR approach specifically designed for position-bias. The difficulty with position-bias is that the treatment - user examination - is not directly observable in click data. As a solution, our estimator uses the expected treatment per rank, instead of the actual treatment that existing DR estimators use. Our novel DR estimator has more robust unbiasedness conditions than the existing IPS approach, and in addition, provides enormous decreases in variance: our experimental results indicate it requires several orders of magnitude fewer datapoints to converge at optimal performance. For the unbiased LTR field, our DR estimator contributes both increases in state-of-the-art performance and the most robust theoretical guarantees of all known LTR estimators.