DDMSR: Dual-Level Denoising Boosts Multi-Modal Recommendation by 10%
New framework purifies both features and user sequences using graph smoothing and FFT filters.
Multi-modal Sequential Recommendation (SR) systems struggle with a "Dual-Noise Dilemma": feature-level redundancy from generic pre-trained representations and sequence-level stochasticity from accidental clicks. The new paper from Jie Luo et al., accepted at ACM MM 2026, introduces DDMSR (Dual-level Denoising Multi-modal Sequential Recommendation) to systematically purify signals from both perspectives.
DDMSR uses a graph-based module with Laplacian smoothing as a structural low-pass filter to suppress high-frequency semantic noise while preserving salient features. For sequences, it applies Fast Fourier Transform and a learnable frequency filter to attenuate anomalous interaction signals adaptively. A multi-modal contrastive alignment objective enforces cross-modal consistency. On four public datasets, DDMSR beats current SOTA methods, demonstrating high robustness and efficiency. Code is available on GitHub.
- DDMSR addresses both feature-level redundancy and sequence-level stochasticity using dual denoising modules.
- Graph-based Laplacian smoothing filters high-frequency semantic noise; FFT with learnable filters cleans sequence anomalies.
- Outperforms state-of-the-art baselines on 4 benchmark datasets; accepted at ACM MM 2026.
Why It Matters
Cleaner signals in multi-modal recommendation mean less noise from accidental clicks and irrelevant features, improving personalization accuracy.