Firzen: Firing Strict Cold-Start Items with Frozen Heterogeneous and Homogeneous Graphs for Recommendation

TL;DR

A unified framework incorporating multi-modal content of items and KGs to effectively solve both strict cold-start and warm-start recommendation termed Firzen is proposed, which extracts the user-item collaborative information over frozen heterogeneous graph (collaborative knowledge graph), and exploits the item-item semantic structures and user-user behavioral association over frozen homogeneous graphs.

Abstract

Recommendation models utilizing unique identities (IDs) to represent distinct users and items have dominated the recommender systems literature for over a decade. Since multi-modal content of items (e.g., texts and images) and knowledge graphs (KGs) may reflect the interaction-related users' preferences and items' characteristics, they have been utilized as useful side information to further improve the recommendation quality. However, the success of such methods often limits to either warm-start or strict cold-start item recommendation in which some items neither appear in the training data nor have any interactions in the test stage: (1) Some fail to learn the embedding of a strict cold-start item since side information is only utilized to enhance the warm-start ID representations; (2) The others deteriorate the performance of warm-start recommendation since unrelated multi-modal content or entities in KGs may blur the final representations. In this paper, we propose a unified framework incorporating multi-modal content of items and KGs to effectively solve both strict cold-start and warm-start recommendation termed Firzen, which extracts the user-item collaborative information over frozen heterogeneous graph (collaborative knowledge graph), and exploits the item-item semantic structures and user-user behavioral association over frozen homogeneous graphs (item-item relation graph and user-user co-occurrence graph). Furthermore, we build four unified strict cold-start evaluation benchmarks based on publicly available Amazon datasets and a real-world industrial dataset from Weixin Channels via rearranging the interaction data and constructing KGs. Extensive empirical results demonstrate that our model yields significant improvements for strict cold-start recommendation and outperforms or matches the state-of-the-art performance in the warm-start scenario.

Figures (8)

• Figure 1: Performance comparison (MRR@20) of Firzen and existing methods on Amazon Beauty dataset for both strict cold-start and warm-start scenarios.
• Figure 2: Illustrated example of users' preferences for items reflected on the visual content, textual content and knowledge graphs.
• Figure 3: Warm-start, normal item cold-start and strict item cold-start scenarios. The filled and diagonal blocks indicate that there is an interaction record between the user and item at the training and inference phase, respectively.
• Figure 4: The model flow of the proposed Firzen architecture. The KGs, user-item interaction graph and multi-modal content are utilized to construct frozen graphs. Then the graphs together with the initial user/item/entity ID embeddings and multi-modal features are fed into SAHGL and MSHGL in succession to obtain final representations. $\lambda_{k}, \lambda_{m}$ represent the importance of knowledge-aware and modality-aware representations, respectively.
• Figure 5: Illustration of the constructed knowledge graphs for Amazon datasets.