Making Recommender Systems More Knowledgeable: A Framework to Incorporate Side Information

TL;DR

This work addresses the underutilization of side information in session-based recommender systems by introducing a general, architecture-agnostic framework that injects item-side information into existing SBRS encoders. The approach fuses side information through dedicated encoders and concatenation with item embeddings, and adds an attention-focused regularization term to guide the model's focus. Empirical results across RepeatNet and SR-GNN on datasets such as Diginetica, Ta Feng, MovieLens, and Last.FM show consistent performance gains and faster convergence, including state-of-the-art results on Diginetica. The framework demonstrates practical impact by enabling knowledge-enhanced recommendations with minimal architectural changes and reasonable training efficiency.

Abstract

Session-based recommender systems typically focus on using only the triplet (user_id, timestamp, item_id) to make predictions of users' next actions. In this paper, we aim to utilize side information to help recommender systems catch patterns and signals otherwise undetectable. Specifically, we propose a general framework for incorporating item-specific side information into the recommender system to enhance its performance without much modification on the original model architecture. Experimental results on several models and datasets prove that with side information, our recommender system outperforms state-of-the-art models by a considerable margin and converges much faster. Additionally, we propose a new type of loss to regularize the attention mechanism used by recommender systems and evaluate its influence on model performance. Furthermore, through analysis, we put forward a few insights on potential further improvements.

Figures (6)

• Figure 1: RepeatNet model. This picture is taken from the RepeatNet paper. $\forall j \in \{1, \cdots, t\}, i_j$ represents the item at timestamp $j$, and $h_j$ represents the embedding vector of $i_j$ extracted from a GRU that handles items
• Figure 2: SR-GNN model. This picture is a modification of the original SR-GNN paper. $v$'s represent items in a session and $h$'s represents the embedding vectors of $v$'s extracted from a Graph Neural Network that handles items
• Figure 3: RepeatNet model with side information. $\forall j \in \{1, \cdots, t\}, s_j$ represents the side information at timestamp $j$, and $h'_j$ represents the embedding vector of $s_j$ extracted from a separate GRU that handles side information. The model would concatenate $h_j$ with $h'_j$, and the concatenated vector would be used by the remaining parts of the model
• Figure 4: SR-GNN model with side information. $s$'s represent the side information in a session and $h'$'s represents the embedding vectors of $s$'s extracted from a Graph Neural Network that handles side information. The model would concatenate $h_j$ with $h'_j$, and the concatenated vector would be used by the remaining parts of the model
• Figure 5: RepeatNet Performance Maximum highlighted