DTAMLP: time-aware MLP boosts session recommendations by denoising clicks

A new arXiv paper by Jiamu Zheng and Xiaojun Shan introduces DTAMLP (Denoise Time-aware MLP), a unified framework for session-based recommendation (SBR) that tackles two overlooked problems in existing models. First, the authors identify "sporadic noise": very short dwell times often reflect accidental clicks that carry little preference signal, yet popular models like TiSASRec and SR-GNN treat every click-time interval as equally informative. To fix this, they designed a lightweight, plug-and-play weight fusion module that blends a model's attention weight with a threshold-capped time-interval weight. This module can be inserted into existing time-aware or GNN-based architectures with almost no structural changes, yielding consistent accuracy gains across models.

The paper's second contribution is an interpretive explanation for a known but poorly understood result from FMLP-Rec: learnable frequency-domain filters on item embeddings improve accuracy. The authors conjecture that time-domain user behavior mixes multiple entangled psychological preferences, and applying FFT-based filtering in the frequency domain lets the model naturally separate and down-weight preference noise. Building on both insights, DTAMLP combines weight fusion with FFT filtering in an entirely MLP-based architecture. The authors validate DTAMLP on the Diginetica and RetailRocket datasets, noting the system-level design reflects the state of the field circa 2023 rather than a state-of-the-art claim. Ablation studies confirm the two mechanisms provide complementary, non-redundant improvements, making the paper most valuable for its directly verifiable weight fusion module.