Prompt-based Multi-interest Learning Method for Sequential Recommendation

TL;DR

PoMRec tackles the challenge of modeling multi‑faceted user interests in sequential recommendation by injecting learnable prompts into user interaction inputs to tailor the extractor and aggregator objectives. It introduces a centrality‑dispersion based multi‑interest extractor that combines mean‑centered semantics with dispersion information, and an attention‑based aggregator that yields per‑interest weights for final fusion. Empirical results on three public datasets show PoMRec consistently outperforms existing multi‑interest methods with fewer parameters, and ablations confirm the mutually beneficial roles of prompts and dispersion modeling. The approach offers a practical, adaptable framework for more accurate next‑item prediction and suggests promising directions for multimodal extensions.

Abstract

Multi-interest learning method for sequential recommendation aims to predict the next item according to user multi-faceted interests given the user historical interactions. Existing methods mainly consist of a multi-interest extractor that embeds the multiple user interests based on the user interactions, and a multi-interest aggregator that aggregates the learned multi-interest embeddings to derive the final user embedding, used for predicting the user rating to an item. Despite their effectiveness, existing methods have two key limitations: 1) they directly feed the user interactions into the multi-interest extractor and aggregator, while ignoring their different learning objectives, and 2) they merely consider the centrality of the user interactions to embed multiple interests of the user, while overlooking their dispersion. To tackle these limitations, we propose a prompt-based multi-interest learning method (PoMRec), where specific prompts are inserted into user interactions, making them adaptive to the extractor and aggregator. Moreover, we utilize both the mean and variance embeddings of user interactions to embed the user multiple interests for the comprehensively user interest learning. We conduct extensive experiments on three public datasets, and the results verify that our proposed PoMRec outperforms the state-of-the-art multi-interest learning methods.

Figures (4)

• Figure 1: Overview of the proposed prompt-based multi-interest learning method (PoMRec) for sequential recommendation (a). We insert certain prompt embeddings at the beginning of the user interaction sequence to make the model know whether it should focus on the contents of user interactions to derived the multi-interest embeddings or the preference over multiple interest to predict the aggregation weights. Besides, we propose a centrality-dispersion based multi-interest extractor (b) that derives the multi-interests embeddings based on both the centrality and dispersion of user interactions. Here we provide an example of the user that has three interests.
• Figure 2: Illustration of the user interactions with different dispersion. The blue points represent the user interacted items, and the grey triangles are their corresponding centrality representation. Intuitively, along with the dispersion of user interactions from low to high, the reliability of the centrality for representing the user interactions decreases.
• Figure 3: Performance of the proposed PoMRec with respect to the different hyper-parameters in ML-1M and Movie & TV datasets. The left vertical axis refers to the Recall@5, while the right vertical axis refers to the NDCG@5.
• Figure 4: Visualization of the learned user multi-interests in ML-1M dataset with the tool of t-SNE. The grey points represent the embeddings of all items in the dataset. The interactions of three users are highlighted with different colors. We visualize the learned user multi-interests with triangles (representing the centrality) and ellipses (representing the dispersion). To derive a deep understanding of the user multi-interests, we annotate the word cloud of user each interest generated with the item genres.