Robust sequential recommendation against unreliable data

<p dir="ltr">In recent years, Sequential Recommender Systems (SRSs) have gained significant attention due to their ability to capture the dynamic and evolving preferences of users. By leveraging the historical sequence of user interactions, SRSs aim to predict the next item a user is likely to interact with. Therefore, SRSs are typically trained to predict the next item as the <i>target </i>given its preceding (and succeeding) items as the <i>input</i>. Such a paradigm assumes that every input-target pair is matched and thus is reliable for training. However, in real-world scenarios, users can be induced by external distractions (e.g. friends’ recommendations) to click on items that are inconsistent with their true preferences, resulting in unreliable training instances, i.e., mismatched input-target pairs. These unreliable instances can mislead the training process of SRSs and thus undermine the recommendation accuracy. To address this issue, this thesis presents a series of Robust Sequential Recommender Systems (RSRSs) that can resist unreliable data during training. </p><p dir="ltr">First, this thesis conducts data analysis to verify the existence and severity of unreliable instances across various real-world datasets. Subsequently, this data analysis guides the design of a BERD framework, which enhances SRSs <b><u>B</u></b>y <b><u>E</u></b>liminating un<b><u>R</u></b>eliable <b><u>D</u></b>ata during training. Specifically, BERD identifies unreliable instances by modeling both the loss and uncertainty of each instance via a Gaussian distribution. It is motivated by two insightful observations: 1) unreliable instances generally exhibit high training loss, and 2) high-loss instances are not always unreliable but can also be uncertain ones with blurry sequential patterns. By capturing these characteristics, BERD effectively identifies and eliminates unreliable instances, leading to improved recommendation accuracy. </p><p dir="ltr">Second, to detect unreliable instances more accurately, the thesis introduces the BERD+ framework. It extends BERD by leveraging the rich information provided by item attributes to rectify instance loss and uncertainty at a finer granularity. However, the integration of item attributes poses an additional challenge, i.e., handling the disturbance caused by attributes. To overcome this issue, BERD+ employs a heterogeneous uncertainty-aware graph convolutional network (HU-GCN) to learn accurate embeddings for different entities, thereby mitigating the impact of uncertain attribute values. Furthermore, an explicit preference extractor (EPE) is proposed to capture users’ explicit preferences and mitigate the disturbance of less-focused attribute types. These designs enable a more precise estimation of instance loss and uncertainty, thus facilitating the identification of unreliable instances with higher accuracy. </p><p dir="ltr">Third, to investigate the theoretical guarantees of RSRSs and avoid the data sparsity issue caused by directly eliminating unreliable instances, the thesis proposes a theoretically guaranteed <b><u>Bi</u></b>di<b><u>r</u></b>ectional <b><u>D</u></b>ata <b><u>Rec</u></b>tification (BirDRec) framework. Bir- DRec can effectively rectify unreliable instances with theoretically guaranteed strategies to handle both partially and completely mismatched instances. Specifically, it introduces a bidirectional architecture that leverages two SRSs in opposite directions to delete unreliable input and replace unreliable targets, respectively. In addition, to enhance the stability of the rectification process and reduce the space complexity, Bir- DRec incorporates a self-ensemble mechanism to aggregate model predictions from different training epochs. Moreover, to reduce the time complexity of BirDRec, a novel rectification sampling strategy is devised, which avoids ranking the full item set when searching for the candidates for replacement. </p><p dir="ltr">Extensive experiments on multiple real-world datasets and representative SRS backbones demonstrate the superiority of the proposed approaches in improving the robustness and effectiveness of SRSs against unreliable data. This thesis provides novel insights and solutions for building RSRSs, contributing to the advancement of sequential recommendation research. </p>