Hybrid Personalized Sequence Recommendation based on LSTM and Filter Enhancement

Sequence recommendation captures dynamic user intent based on historical interaction sequences and has increasingly become a prominent area of research. Models such as CNN and Transformer have become mainstream approaches for sequence recommendation. However, single-model methods face clear limitations in data utilization and noise filtering. To effectively leverage valuable information and filter out noise, we propose a novel Hybrid Personalized Sequence Recommendation based on LSTM and Filter Enhancement, termed as LFPRec. LFPRec comprises two core components: the FT block and the LSTM block. The Filter Enhancement module applies Fast Fourier Transform (FFT) and Inverse Fast Fourier Transform (IFFT) to perform time-domain and frequency-domain transformations, isolating low-frequency trends reflective of stable user preferences while filtering out high-frequency noise, such as transient or erroneous interactions. Utilizing its gating mechanism, the LSTM module models long-term dependencies in sequences to capture dynamic user intent. The processed data from these modules is subsequently integrated using a learnable, adaptive hybrid approach. Theoretical analysis suggests that frequency-domain filtering enhances the signal-to-noise ratio, enabling more accurate modeling of user behavior patterns critical to recommendation tasks. To evaluate the effectiveness of LFPRec, we conducted experiments on three public datasets. Experimental results demonstrate that LFPRec consistently outperforms six state-of-the-art recommendation models in terms of recommendation accuracy and robustness, highlighting its enhanced capabilities in data utilization and noise reduction.