DIFF: Dual Side-Information Filtering and Fusion for Sequential Recommendation

TL;DR

DIFF addresses noisy signals in sequential recommendations and underutilization of item attributes in SISR, enabling more accurate next-item predictions. It introduces frequency-based noise filtering to separate low-frequency, stable interests from high-frequency fluctuations, and dual multi-sequence fusion to jointly model intra-attribute and inter-attribute correlations; an additional representation alignment loss harmonizes ID and attribute spaces. Empirical results on Yelp and four Amazon-related datasets show up to $14.1\%$ recall gains at $R@20$ and $12.5\%$ gains at $NDCG@20$, outperforming state-of-the-art SR and SISR baselines. The method demonstrates robustness to noisy histories and improved performance in cold-start and tail-item scenarios, highlighting practical benefits for real-world recommender systems.

Abstract

Side-information Integrated Sequential Recommendation (SISR) benefits from auxiliary item information to infer hidden user preferences, which is particularly effective for sparse interactions and cold-start scenarios. However, existing studies face two main challenges. (i) They fail to remove noisy signals in item sequence and (ii) they underutilize the potential of side-information integration. To tackle these issues, we propose a novel SISR model, Dual Side-Information Filtering and Fusion (DIFF), which employs frequency-based noise filtering and dual multi-sequence fusion. Specifically, we convert the item sequence to the frequency domain to filter out noisy short-term fluctuations in user interests. We then combine early and intermediate fusion to capture diverse relationships across item IDs and attributes. Thanks to our innovative filtering and fusion strategy, DIFF is more robust in learning subtle and complex item correlations in the sequence. DIFF outperforms state-of-the-art SISR models, achieving improvements of up to 14.1% and 12.5% in Recall@20 and NDCG@20 across four benchmark datasets.

Figures (9)

• Figure 1: (i) Frequency signals and (ii) fusion types in side information integrated sequential recommendation. Frequency-based noise filtering removes the fourth item with inconsistent signals. Intermediate fusion (blue) highlights items aligned with key signals, while early fusion (green) captures broader combinations.
• Figure 2: Comparison of side information fusion methods. Existing methods are broadly categorized into (a) early, (b) late, and (c) intermediate fusion. We introduce (d) dual fusion, which benefits from early and intermediate fusion.
• Figure 3: An overview of DIFF. DIFF processes both independent sequences and early fused sequences via $L$ layers of two components: (i) Frequency-based Noise Filtering and (ii) Dual Multi-sequence Fusion. DIFF yields filtered user representations that fully integrates item attributes. Multi-task learning with representation alignment ensures smooth ID-attribute fusion.
• Figure 4: Performance comparison on different target item popularity groups. The target items of Head group are the top 10% most popular items, while the Tail group includes sequences with less popular target items.
• Figure 5: Performance comparison on different sequence length groups. The Short group consists of sequences with a length of five (43% of Yelp and 51% of Beauty dataset), while the Long group includes sequences longer than five.