Research map / Label-efficient and robust learning / Recognition and long-tail

Ada-Retrieval: An Adaptive Multi-Round Retrieval Paradigm for Sequential Recommendations

AAAI 2024 · Label-efficient and robust learning › Recognition and long-tail · approach: Recommendation representation learning · 4 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Ada-Retrieval, a model-agnostic multi-round retrieval paradigm that iteratively refines user representations using item and user context adapters, improving five sequential recommenders on three datasets.

Problem

Single-round retrieval in sequential recommenders produces a fixed user representation, so top-k candidates are confined to a static region of the item space and may miss diverse or shifting user preferences.

Approach

Retrieve top-k items in n rounds, where each round's user representation is adapted using context from prior rounds: an item representation adapter (learnable filter plus context-aware attention over retrieved items) and a user representation adapter (GRU over prior user representations plus MLP fusion), trained with a decayed per-round cross-entropy loss.

Key results

• On Beauty, Ada-Retrieval improves SASRec NDCG@50 by 8.55% (6.506 to 7.062).
• Ada-Retrieval (GRU4Rec) averages about 8.37% NDCG@50 improvement over GRU4Rec across the three datasets.
• Ada-Retrieval (SASRec) outperforms multi-interest baselines ComiRec, MIND, SINE and DNN on all three datasets (e.g., Beauty HR@50 17.741 vs 14.394 for ComiRec).

Measured effects reported in the paper

Surprising findings

• Omitting the MLP fusion layer in the user adapter hurts performance more than removing the entire user representation adapter, suggesting naive context fusion adds noise.

Limitations

• Not compatible with matrix factorization models relying on user-item interaction matrices, per the authors.
• Performance can become unpredictable when the number of rounds T exceeds the optimal value.
• Later-round gains diminish, with a slight drop on Beauty in the final rounds.

Experimental setup

Datasets: Beauty Sports Yelp

Benchmarks: Beauty Sports Yelp

Metrics: HR@50 NDCG@50

Baselines: GRU4Rec SASRec NextItNet SRGNN FMLPRec DNN MIND ComiRec SINE

Builds on: SASRec GRU4Rec FMLPRec filter-enhanced MLP

Closest papers by meaning

• LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.87)
• SRA-CL: Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation (NeurIPS 2025 · similarity 0.86)
• LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.86)
• Augmenting Sequential Recommendation with Balanced Relevance and Diversity (AAAI 2025 · similarity 0.85)
• UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · similarity 0.85)
• Improving LLMs for Recommendation with Out-Of-Vocabulary Tokens (ICML 2025 · similarity 0.85)