Learning User and Item Representations from Heterogeneous Feedback for Accurate and Explainable Recommendation

Abstract

Representation learning is central to recommender systems because the information encoded in user and item representations directly affects how effectively user preferences and item characteristics can be modelled. Recommender systems increasingly learn from heterogeneous feedback, including behavioural interactions, numerical ratings, and textual reviews. These feedback sources provide complementary information, but they also raise an important representation learning problem: the same user or item may need to be represented differently depending on the type of feedback being modelled or the particular interaction for which a prediction is made. This thesis investigated this problem from the perspective of accurate and aspect-level explainable recommendation.

The thesis examined two related research questions. The first concerned how implicit and explicit feedback could be integrated while preserving their different meanings and allowing information to be exchanged between their respective user and item representations. The second concerned how user and item representations could depend on their current interaction counterpart in aspect-level recommendation.

To investigate these questions, the thesis adopted a polymorphic representation learning perspective, under which the same entity can have different representations across feedback types or interaction counterparts. For the first research question, we developed Interdependent Embedding Learning (IEL), a GCN-based framework that maintained separate user and item representations for implicit and explicit feedback. IEL used intra-graph and selected inter-graph message passing to allow these feedback-specific representations to exchange collaborative information and was evaluated for both item ranking and rating prediction. For the second research question, we developed Dynamic Representation Learning (DRL), which combined general collaborative representations learned from ratings with aspect-sentiment information extracted from textual reviews for the current interaction. DRL constructed an item-dependent user representation and a user-dependent item representation for rating and positive-aspect prediction. A prompt-based large language model (LLM) procedure was used to transform reviews into structured aspect-sentiment information for this study.

Offline experiments on public Amazon product-review datasets provided empirical support for both approaches. IEL improved recommendation performance relative to the compared unified recommendation baselines, with the strongest and most consistent improvements observed for item ranking. DRL achieved lower rating-prediction MSE than the compared baselines across all three evaluated datasets, with an average reduction of 5.1\% relative to the strongest compared baseline. It also improved Hit@1 for positive-aspect prediction by an average of 4.8\%. The latter evaluation measured the model's ability to predict aspects subsequently appearing in held-out reviews and therefore served as an operational measure of aspect-level explanation.

The main contribution of this thesis is a unified representation-learning perspective that connects feedback-dependent and interaction-dependent representations. The IEL and DRL studies demonstrate how heterogeneous feedback can be preserved and integrated without requiring a single invariant representation of each user or item. In addition, the thesis constructed three derived Amazon review datasets containing LLM-extracted aspect-sentiment information for aspect-level recommendation research. Overall, the findings suggest that allowing user and item representations to vary according to the feedback type and the interaction being modelled provides a useful direction for learning from heterogeneous feedback in accurate and explainable recommender systems. Show more