While personalization increases the utility of recommender systems, it also brings the issue of filter bubbles. E.g., if the system keeps exposing and recommending the items that the user is interested in, it may also make the user feel bored and less satisfied. Existing work studies filter bubbles in static recommendation, where the effect of overexposure is hard to capture. In contrast, we believe it is more meaningful to study the issue in interactive recommendation and optimize long-term user satisfaction. Nevertheless, it is unrealistic to train the model online due to the high cost. As such, we have to leverage offline training data and disentangle the causal effect on user satisfaction. To achieve this goal, we propose a counterfactual interactive recommender system (CIRS) that augments offline reinforcement learning (offline RL) with causal inference. The basic idea is to first learn a causal user model on historical data to capture the overexposure effect of items on user satisfaction. It then uses the learned causal user model to help the planning of the RL policy. To conduct evaluation offline, we innovatively create an authentic RL environment (KuaiEnv) based on a real-world fully observed user rating dataset. The experiments show the effectiveness of CIRS in bursting filter bubbles and achieving long-term success in interactive recommendation. The implementation of CIRS is available via https://github.com/chongminggao/CIRS-codes.
Learning discriminative representation from the complex spatio-temporal dynamic space is essential for video recognition. On top of those stylized spatio-temporal computational units, further refining the learnt feature with axial contexts is demonstrated to be promising in achieving this goal. However, previous works generally focus on utilizing a single kind of contexts to calibrate entire feature channels and could hardly apply to deal with diverse video activities. The problem can be tackled by using pair-wise spatio-temporal attentions to recompute feature response with cross-axis contexts at the expense of heavy computations. In this paper, we propose an efficient feature refinement method that decomposes the feature channels into several groups and separately refines them with different axial contexts in parallel. We refer this lightweight feature calibration as group contextualization (GC). Specifically, we design a family of efficient element-wise calibrators, i.e., ECal-G/S/T/L, where their axial contexts are information dynamics aggregated from other axes either globally or locally, to contextualize feature channel groups. The GC module can be densely plugged into each residual layer of the off-the-shelf video networks. With little computational overhead, consistent improvement is observed when plugging in GC on different networks. By utilizing calibrators to embed feature with four different kinds of contexts in parallel, the learnt representation is expected to be more resilient to diverse types of activities. On videos with rich temporal variations, empirically GC can boost the performance of 2D-CNN (e.g., TSN and TSM) to a level comparable to the state-of-the-art video networks. Code is available at https://github.com/haoyanbin918/Group-Contextualization.
Recommender systems are usually developed and evaluated on the historical user-item logs. However, most offline recommendation datasets are highly sparse and contain various biases, which hampers the evaluation of recommendation policies. Existing efforts aim to improve the data quality by collecting users' preferences on randomly selected items (e.g., Yahoo! and Coat). However, they still suffer from the high variance issue caused by the sparsely observed data. To fundamentally solve the problem, we present KuaiRec, a fully-observed dataset collected from the social video-sharing mobile App, Kuaishou. The feedback of 1,411 users on almost all of the 3,327 videos is explicitly observed. To the best of our knowledge, this is the first real-world fully-observed dataset with millions of user-item interactions in recommendation. To demonstrate the advantage of KuaiRec, we leverage it to explore the key questions in evaluating conversational recommender systems. The experimental results show that two factors in traditional partially-observed data -- the data density and the exposure bias -- greatly affect the evaluation results. This entails the significance of our fully-observed data in researching many directions in recommender systems, e.g., the unbiased recommendation, interactive/conversational recommendation, and evaluation. We release the dataset and the pipeline implementation for evaluation at https://chongminggao.github.io/KuaiRec/.
A good personalized product search (PPS) system should not only focus on retrieving relevant products, but also consider user personalized preference. Recent work on PPS mainly adopts the representation learning paradigm, e.g., learning representations for each entity (including user, product and query) from historical user behaviors (aka. user-product-query interactions). However, we argue that existing methods do not sufficiently exploit the crucial collaborative signal, which is latent in historical interactions to reveal the affinity between the entities. Collaborative signal is quite helpful for generating high-quality representation, exploiting which would benefit the representation learning of one node from its connected nodes. To tackle this limitation, in this work, we propose a new model IHGNN for personalized product search. IHGNN resorts to a hypergraph constructed from the historical user-product-query interactions, which could completely preserve ternary relations and express collaborative signal based on the topological structure. On this basis, we develop a specific interactive hypergraph neural network to explicitly encode the structure information (i.e., collaborative signal) into the embedding process. It collects the information from the hypergraph neighbors and explicitly models neighbor feature interaction to enhance the representation of the target entity. Extensive experiments on three real-world datasets validate the superiority of our proposal over the state-of-the-arts.
Explainability of graph neural networks (GNNs) aims to answer "Why the GNN made a certain prediction?", which is crucial to interpret the model prediction. The feature attribution framework distributes a GNN's prediction to its input features (e.g., edges), identifying an influential subgraph as the explanation. When evaluating the explanation (i.e., subgraph importance), a standard way is to audit the model prediction based on the subgraph solely. However, we argue that a distribution shift exists between the full graph and the subgraph, causing the out-of-distribution problem. Furthermore, with an in-depth causal analysis, we find the OOD effect acts as the confounder, which brings spurious associations between the subgraph importance and model prediction, making the evaluation less reliable. In this work, we propose Deconfounded Subgraph Evaluation (DSE) which assesses the causal effect of an explanatory subgraph on the model prediction. While the distribution shift is generally intractable, we employ the front-door adjustment and introduce a surrogate variable of the subgraphs. Specifically, we devise a generative model to generate the plausible surrogates that conform to the data distribution, thus approaching the unbiased estimation of subgraph importance. Empirical results demonstrate the effectiveness of DSE in terms of explanation fidelity.
Intrinsic interpretability of graph neural networks (GNNs) is to find a small subset of the input graph's features -- rationale -- which guides the model prediction. Unfortunately, the leading rationalization models often rely on data biases, especially shortcut features, to compose rationales and make predictions without probing the critical and causal patterns. Moreover, such data biases easily change outside the training distribution. As a result, these models suffer from a huge drop in interpretability and predictive performance on out-of-distribution data. In this work, we propose a new strategy of discovering invariant rationale (DIR) to construct intrinsically interpretable GNNs. It conducts interventions on the training distribution to create multiple interventional distributions. Then it approaches the causal rationales that are invariant across different distributions while filtering out the spurious patterns that are unstable. Experiments on both synthetic and real-world datasets validate the superiority of our DIR in terms of interpretability and generalization ability on graph classification over the leading baselines. Code and datasets are available at https://github.com/Wuyxin/DIR-GNN.
Learning objectives of recommender models remain largely unexplored. Most methods routinely adopt either pointwise or pairwise loss to train the model parameters, while rarely pay attention to softmax loss due to the high computational cost. Sampled softmax loss emerges as an efficient substitute for softmax loss. Its special case, InfoNCE loss, has been widely used in self-supervised learning and exhibited remarkable performance for contrastive learning. Nonetheless, limited studies use sampled softmax loss as the learning objective to train the recommender. Worse still, none of them explore its properties and answer "Does sampled softmax loss suit for item recommendation?" and "What are the conceptual advantages of sampled softmax loss, as compared with the prevalent losses?", to the best of our knowledge. In this work, we aim to better understand sampled softmax loss for item recommendation. Specifically, we first theoretically reveal three model-agnostic advantages: (1) mitigating popularity bias, which is beneficial to long-tail recommendation; (2) mining hard negative samples, which offers informative gradients to optimize model parameters; and (3) maximizing the ranking metric, which facilitates top-K performance. Moreover, we probe the model-specific characteristics on the top of various recommenders. Experimental results suggest that sampled softmax loss is more friendly to history and graph-based recommenders (e.g., SVD++ and LightGCN), but performs poorly for ID-based models (e.g., MF). We ascribe this to its shortcoming in learning representation magnitude, making the combination with the models that are also incapable of adjusting representation magnitude learn poor representations. In contrast, the history- and graph-based models, which naturally adjust representation magnitude according to node degree, are able to compensate for the shortcoming of sampled softmax loss.
Learning powerful representations is one central theme of graph neural networks (GNNs). It requires refining the critical information from the input graph, instead of the trivial patterns, to enrich the representations. Towards this end, graph attention and pooling methods prevail. They mostly follow the paradigm of "learning to attend". It maximizes the mutual information between the attended subgraph and the ground-truth label. However, this training paradigm is prone to capture the spurious correlations between the trivial subgraph and the label. Such spurious correlations are beneficial to in-distribution (ID) test evaluations, but cause poor generalization in the out-of-distribution (OOD) test data. In this work, we revisit the GNN modeling from the causal perspective. On the top of our causal assumption, the trivial information serves as a confounder between the critical information and the label, which opens a backdoor path between them and makes them spuriously correlated. Hence, we present a new paradigm of deconfounded training (DTP) that better mitigates the confounding effect and latches on the critical information, to enhance the representation and generalization ability. Specifically, we adopt the attention modules to disentangle the critical subgraph and trivial subgraph. Then we make each critical subgraph fairly interact with diverse trivial subgraphs to achieve a stable prediction. It allows GNNs to capture a more reliable subgraph whose relation with the label is robust across different distributions. We conduct extensive experiments on synthetic and real-world datasets to demonstrate the effectiveness.
The ubiquity of implicit feedback makes it indispensable for building recommender systems. However, it does not actually reflect the actual satisfaction of users. For example, in E-commerce, a large portion of clicks do not translate to purchases, and many purchases end up with negative reviews. As such, it is of importance to account for the inevitable noises in implicit feedback. However, little work on recommendation has taken the noisy nature of implicit feedback into consideration. In this work, we explore the central theme of denoising implicit feedback for recommender learning, including training and inference. By observing the process of normal recommender training, we find that noisy feedback typically has large loss values in the early stages. Inspired by this observation, we propose a new training strategy named Adaptive Denoising Training (ADT), which adaptively prunes the noisy interactions by two paradigms (i.e., Truncated Loss and Reweighted Loss). Furthermore, we consider extra feedback (e.g., rating) as auxiliary signal and propose three strategies to incorporate extra feedback into ADT: finetuning, warm-up training, and colliding inference. We instantiate the two paradigms on the widely used binary cross-entropy loss and test them on three representative recommender models. Extensive experiments on three benchmarks demonstrate that ADT significantly improves the quality of recommendation over normal training without using extra feedback. Besides, the proposed three strategies for using extra feedback largely enhance the denoising ability of ADT.