Pattern-wise Transparent Sequential Recommendation

TL;DR

PTSR addresses the challenge of achieving both high accuracy and interpretability in sequential recommendation by decomposing user histories into multi-level patterns extracted with sliding windows and represented with probabilistic Beta/Gamma embeddings. It uses a probabilistic conjunction to fuse item patterns and KL-Divergence to measure their distance to candidate items, complemented by distance-based weights and a sequence-aware bias to implicitly emphasize key patterns without ground-truth annotations. Across five public datasets, PTSR delivers strong performance and provides interpretable explanations at both point- and union-level, as validated by comprehensive ablations and case studies. The work advances transparent recommender design and identifies avenues for future enhancements, including dynamic pattern selection and the incorporation of negation operators to model negative feedback.

Abstract

A transparent decision-making process is essential for developing reliable and trustworthy recommender systems. For sequential recommendation, it means that the model can identify key items that account for its recommendation results. However, achieving both interpretability and recommendation performance simultaneously is challenging, especially for models that take the entire sequence of items as input without screening. In this paper, we propose an interpretable framework (named PTSR) that enables a pattern-wise transparent decision-making process without extra features. It breaks the sequence of items into multi-level patterns that serve as atomic units throughout the recommendation process. The contribution of each pattern to the outcome is quantified in the probability space. With a carefully designed score correction mechanism, the pattern contribution can be implicitly learned in the absence of ground-truth key patterns. The final recommended items are those that most key patterns strongly endorse. Extensive experiments on five public datasets demonstrate remarkable recommendation performance, while statistical analysis and case studies validate the model interpretability.

Figures (7)

• Figure 1: Illustration of interpretability at the point level and union level. The former is commonly used in existing studies to identify key individual items, while the latter provides additional insights into the common interests exhibited by a user.
• Figure 2: Architecture of PTSR: The framework employs multi-scale sliding windows for pattern extraction, utilizes probabilistic embeddings and logical operators for pattern representation, and computes target scores through KL-Divergence with integrated weights and biases.
• Figure 3: Illustration of probabilistic conjunction operator. $S_1 \bigwedge S_2$ denotes the conjunction of distribution $S_1$ and $S_2$, achieved by the weighted product of the probability density function (PDF).
• Figure 4: NDCG@10 performance over various $\lambda$ on four datasets.
• Figure 5: A case study of a real user. The bottom panel presents the interaction history and the target item, while the bar chart above displays the correlation distribution of patterns at three levels (the higher the value, the stronger the correlation between the pattern and the target item, and vice versa). '$\land$' indicates a conjunction relationship.