Knowledge Graph-based Session Recommendation with Adaptive Propagation

TL;DR

This work builds a knowledge graph by connecting items with multi-typed edges to characterize various user-item interactions and adaptively aggregate items' neighbor information considering user intention within the learned session to address session recommendation backbones limitations.

Abstract

Session-based recommender systems (SBRSs) predict users' next interacted items based on their historical activities. While most SBRSs capture purchasing intentions locally within each session, capturing items' global information across different sessions is crucial in characterizing their general properties. Previous works capture this cross-session information by constructing graphs and incorporating neighbor information. However, this incorporation cannot vary adaptively according to the unique intention of each session, and the constructed graphs consist of only one type of user-item interaction. To address these limitations, we propose knowledge graph-based session recommendation with session-adaptive propagation. Specifically, we build a knowledge graph by connecting items with multi-typed edges to characterize various user-item interactions. Then, we adaptively aggregate items' neighbor information considering user intention within the learned session. Experimental results demonstrate that equipping our constructed knowledge graph and session-adaptive propagation enhances session recommendation backbones by 10%-20%. Moreover, we provide an industrial case study showing our proposed framework achieves 2% performance boost over an existing well-deployed model at The Home Depot e-platform.

Figures (4)

• Figure 1: Since the user in Session A (b) intends to decorate the garden, while the user in Session B (c) intends to decorate the kitchen, the corresponding neighbors from the knowledge graph (a) are different for the same flower. By our proposed session-adaptive propagation, the flower aggregates more information from the lopper/watering can in (b) while more information from the sink/table in (c).
• Figure 2: In (a)-(c), we first extract three types of edges from historical sessions to construct item knowledge graph. Then in (d), we forward the given session through the $1^{\text{st}}$ transformer layer to obtain items' contextual embeddings, which are used for query-relevant neighbors for GNNs to perform graph propagation. The propagated item embeddings are fed into the $2^{\text{nd}}$ transformer with a pooling layer afterward to obtain session embedding for the recommendation.
• Figure 3: (a)-(b): Comparing the performance improvement(%) over the baseline Top-N on predicting next-item meta-labels. Top-N: the simple heuristic recommending the most frequent meta-labels among previous items within one session; (2) X$_{m*}$ uses an untrained prediction head to predict the task-specific label of the next item; (3) P-X$_{m*}$ first predicts top-N next items and use their task-specific labels as recommendations; (4) M-X$_{m*}$ uses a trained prediction head that is trained by multi-task learning to predict the task-specific label of the next item. (c) N@10 on sessions of different sparsity. M-KGHT$_{qm*}$ is better than M-SASRec$_{m*}$ on sessions of the sparsity at the middle level.
• Figure 4: Case study of our proposed M-KGHT. (a) Comparing the Top5 recommendation by M2TRec and M-KGHT. By leveraging neighborhood information of Spray Mop1, the correct item Mop Refill appears in the recommendation list. (b)-(c) visualizes the learned attention of one attention head over co-view neighbors. Since users in both of these two sessions intend to clean the gardens, the session-adaptive graph propagation successfully learns the higher intention for garden-related items.