Click-Through Rate Prediction Based on Filtering-Enhanced with Multi-head Attention

Click-Through Rate (CTR) prediction is a crucial task in recommend systems (RSs), particularly in large-scale industrial applications. CTR models based on graph neural networks (GNNs) is currently the mainstream technology, however, it also encounters challenges in terms of feature interactions and user interests. In the context of RSs, the individual attribute information associated with each entity holds significant importance beyond the inherent interactions between users and items. Rational utilization and capturing of dependencies among these attributes can substantially enhance the predictive accuracy of the model. Therefore, we propose two attention mechanisms of interactions, including internal interaction and external interaction, to handle the information exchange within the attributes of either users or items individually, as well as the interactions between the attributes of users and items. Simultaneously, we introduce a multi-head attention (MHA) mechanism for the interaction selection, to capture features from different dimensions. Furthermore, the behavioral data reflecting user interests unavoidably contains noise, we attenuates noise by utilizing fast Fourier transform (FFT) filtering algorithms to transfer information from the spatial domain to the frequency domain. Experimental results on three benchmarks show that our model outperforms existing SOTA models.