FEDIN: Frequency-Enhanced Deep Interest Network for CTR Prediction
Researchers have introduced FEDIN, a Frequency-Enhanced Deep Interest Network aimed at predicting click-through rates. Traditional sequential recommendation systems face challenges in recognizing latent periodic patterns in user interests due to noise in the time domain. While frequency-domain analysis provides a comprehensive view, current methods often analyze user sequences separately, neglecting the context of target items. The research team found that user attention scores exhibit unique spectral entropy distributions when linked to positive versus negative target items: genuine interests reveal concentrated spectral patterns with lower entropy, whereas irrelevant actions manifest as high-entropy noise. FEDIN incorporates a frequency-domain component with a target-aware spectrum filtering mechanism to extract periodic patterns. The study is accessible on arXiv with the identifier 2605.01726.
Key facts
- FEDIN stands for Frequency-Enhanced Deep Interest Network.
- It addresses click-through rate prediction.
- Sequential recommendation models fail to capture latent periodic patterns due to time-domain noise.
- Frequency-domain analysis provides a global perspective.
- Existing approaches treat user sequences in isolation.
- User attention scores have distinct spectral entropy distributions for positive vs. negative items.
- True interests show concentrated spectral patterns with lower entropy.
- FEDIN uses a target-aware spectrum filtering mechanism.
Entities
Institutions
- arXiv