Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
MEMOIR is a novel recommendation framework that segments user interaction histories into temporal windows and utilizes a large language model to create semantic behavioral memories for each period, aggregating user state, evolution direction, and predicted future into a unified representation. Evaluated on Amazon Reviews, MEMOIR achieved competitive performance against the leading baseline, UniSRec, by excelling in ranking-quality metrics among users at the extremes of the preference-drift spectrum. Notably, the framework's enhancements are attributed to its ability to capture nuanced user behavior over time, revealing significant performance gains in specific user segments rather than across the board.
MEMOIR reveals that tailored user representations can significantly enhance recommendation quality, especially for users with extreme preference drift.
We propose MEMOIR, a framework that segments user interaction histories into temporal windows, generates semantic behavioral memory for each period using an LLM, and aggregates current state, evolution direction, and predicted future into a single user representation. On the Electronics and Clothing_Shoes_and_Jewelry categories of Amazon Reviews 2023, MEMOIR is statistically tied with UniSRec, the strongest baseline, on aggregate NDCG@10 (0.0643 vs. 0.0641), splitting the four reported metrics 2-2: MEMOIR leads NDCG@10 and MRR, UniSRec leads HR@10 and HR@20. An ablation study finds that no single architectural component - the evolution-preserving contrastive loss, its directional-consistency term, or temporal window segmentation itself - individually explains much of MEMOIR's approximately 18% relative gain over ID-based SASRec; all four ablations land within 2% of the full model on aggregate NDCG@10. Stratifying test performance by a composite preference-drift score instead reveals where the gain concentrates: MEMOIR leads on ranking-quality metrics (NDCG@10, MRR) specifically among users at the high- and low-drift extremes of the distribution, while UniSRec leads the volume-oriented HR@10/HR@20 metrics across all drift strata and edges out MEMOIR on ranking quality in the middle band. We report this drift-stratified pattern, rather than the near-tied aggregate numbers or any single ablated component, as MEMOIR's most substantive and reproducible finding, and surface why it holds as an open question for future work.