Modeling User Interests Based on Dynamic Graph Convolution and Temporal-Aware Attention

This paper introduces GLSTI, a time-aware user interest modeling framework leveraging graph convolutional networks for personalized search. To address information decay and dynamic interest capture, GLSTI incorporates a multi-head attention mechanism with time-decay in LSTM, employing an exponential function to dynamically adjust the weights of historical behaviors. A dynamic gated residual connection within LightGCN is utilized to mitigate over-smoothing issues, while a query-aware fusion mechanism integrates interest features effectively. Experimental results demonstrate significant improvements of 4.25% (MRR@20) and 3.33% (NDCG@20) over baseline models, with ablation studies validating the effectiveness of each module.