GPT4Rec: Graph Prompt Tuning for Streaming Recommendation

Abstract

In the realm of personalized recommender systems, the challenge of adapting\nto evolving user preferences and the continuous influx of new users and items\nis paramount. Conventional models, typically reliant on a static training-test\napproach, struggle to keep pace with these dynamic demands. Streaming\nrecommendation, particularly through continual graph learning, has emerged as a\nnovel solution. However, existing methods in this area either rely on\nhistorical data replay, which is increasingly impractical due to stringent data\nprivacy regulations; or are inability to effectively address the over-stability\nissue; or depend on model-isolation and expansion strategies. To tackle these\ndifficulties, we present GPT4Rec, a Graph Prompt Tuning method for streaming\nRecommendation. Given the evolving user-item interaction graph, GPT4Rec first\ndisentangles the graph patterns into multiple views. After isolating specific\ninteraction patterns and relationships in different views, GPT4Rec utilizes\nlightweight graph prompts to efficiently guide the model across varying\ninteraction patterns within the user-item graph. Firstly, node-level prompts\nare employed to instruct the model to adapt to changes in the attributes or\nproperties of individual nodes within the graph. Secondly, structure-level\nprompts guide the model in adapting to broader patterns of connectivity and\nrelationships within the graph. Finally, view-level prompts are innovatively\ndesigned to facilitate the aggregation of information from multiple\ndisentangled views. These prompt designs allow GPT4Rec to synthesize a\ncomprehensive understanding of the graph, ensuring that all vital aspects of\nthe user-item interactions are considered and effectively integrated.\nExperiments on four diverse real-world datasets demonstrate the effectiveness\nand efficiency of our proposal.\n