Semantic-enhanced Co-attention Prompt Learning for Non-overlapping Cross-Domain Recommendation

TL;DR

This work tackles many-to-one non-overlapping cross-domain sequential recommendation by abandoning explicit item ID alignment in favor of semantic text representations of items. It introduces TCPLP, a text-enhanced prompt learning framework with domain-shared and domain-specific prompts encoded by a co-attention network, trained in a two-stage pre-training and prompt-tuning workflow. Empirical results on Amazon subsets (Office-Arts, Cell-Toys-Automotive) and cross-platform data (OR-Pantry-Instruments) show TCPLP achieving state-of-the-art performance, highlighting the value of text semantics and prompt-based transfer for cross-domain recommendation. The method demonstrates robustness across language models (Longformer/BERT) and offers a flexible paradigm for cross-domain transfer without explicit domain alignment, with potential applicability to related prediction tasks.

Abstract

Non-overlapping Cross-domain Sequential Recommendation (NCSR) is the task that focuses on domain knowledge transfer without overlapping entities. Compared with traditional Cross-domain Sequential Recommendation (CSR), NCSR poses several challenges: 1) NCSR methods often rely on explicit item IDs, overlooking semantic information among entities. 2) Existing CSR mainly relies on domain alignment for knowledge transfer, risking semantic loss during alignment. 3) Most previous studies do not consider the many-to-one characteristic, which is challenging because of the utilization of multiple source domains. Given the above challenges, we introduce the prompt learning technique for Many-to-one Non-overlapping Cross-domain Sequential Recommendation (MNCSR) and propose a Text-enhanced Co-attention Prompt Learning Paradigm (TCPLP). Specifically, we capture semantic meanings by representing items through text rather than IDs, leveraging natural language universality to facilitate cross-domain knowledge transfer. Unlike prior works that need to conduct domain alignment, we directly learn transferable domain information, where two types of prompts, i.e., domain-shared and domain-specific prompts, are devised, with a co-attention-based network for prompt encoding. Then, we develop a two-stage learning strategy, i.e., pre-train & prompt-tuning paradigm, for domain knowledge pre-learning and transferring, respectively. We conduct extensive experiments on three datasets and the experimental results demonstrate the superiority of our TCPLP. Our source codes have been publicly released.

Figures (5)

• Figure 1: The semantic space differences among different forms of items in various domains. (a) illustrates the different semantic spaces of ID-based items in different domains. (b) illustrates that the text-based item can ensure that items in different domains are in the same semantic space by exploiting the generality of text in different domains. (c) compares the average cosine distances of ID-based and Text-based item representations from MCRPL and from TCPLP intra-domain and inter-domain in the Office-Arts dataset.
• Figure 2: The system architecture of TCPLP in the pre-training stage, which mainly comprises three components: the text sequence encoder, prompt encoder, and the learning framework. At the same time, we utilize multiple domain text sequences as input for pre-training and prior knowledge learning to train the text sequence encoder, prompt encoder, and MLP network.
• Figure 3: The workflow of TCPLP in the prompt-tuning stage. During this stage, we freeze the text sequence encoder, domain-shared prompt parameters, and the co-attention network based on domain-shared prompt. We only update the domain-specific prompt, the co-attention network based on domain-specific prompt, and the MLP network.
• Figure 4: Impact of the hyper-parameter $d_W$ on the Office-Arts dataset.
• Figure 5: Impact of the hyper-parameter $\tau$ on the Office-Arts dataset.