Beyond Recency Bias: Combining Sequential and Global Collaborative Signals in LLM-Based Generative Recommendation for Sparse Data

Behrad Ghiasi, CI Ezeife
Proceedings of the The 39th Canadian Conference on Artificial Intelligence, PMLR 318:284-295, 2026.

Abstract

Recommender systems often use multi-stage retrieval and ranking pipelines, where errors made early cannot be fixed later, which hurts overall recommendation quality. LLM-based generative recommendation avoids this by directly generating item identi- fiers, but many methods represent each item as a token sequence, which creates two concrete problems: generation is slow because tokens must be produced step by step, and it can fail due to beam-search local optima, where the correct item is dropped early because its first token has low probability. SETRec addresses these issues by represent- ing each item as an order-agnostic set of tokens and using query-guided simultaneous token generation, so the item’s CF and semantic tokens are generated in parallel without intra-item token dependency. However, SETRec still uses only one collaborative filtering (CF) token, and when that CF signal comes from a sequence-aware model, it is vulnera- ble to recency bias, especially in sparse and cold-start settings where recent interactions dominate. CFs-SETRec reduces this recency bias by adding a second CF token chosen to capture long-term preferences, and combining it with the sequential CF signal, which preserves both short-term behavior and long-term affinity and leads to more balanced recommendations under sparse data.

Cite this Paper

BibTeX
@InProceedings{pmlr-v318-ghiasi26a, title = {Beyond Recency Bias: Combining Sequential and Global Collaborative Signals in LLM-Based Generative Recommendation for Sparse Data}, author = {Ghiasi, Behrad and Ezeife, CI}, booktitle = {Proceedings of the The 39th Canadian Conference on Artificial Intelligence}, pages = {284--295}, year = {2026}, editor = {Bouzar-Benlabiod, Lydia and Leung, Carson}, volume = {318}, series = {Proceedings of Machine Learning Research}, month = {25--29 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v318/main/assets/ghiasi26a/ghiasi26a.pdf}, url = {https://proceedings.mlr.press/v318/ghiasi26a.html}, abstract = {Recommender systems often use multi-stage retrieval and ranking pipelines, where errors made early cannot be fixed later, which hurts overall recommendation quality. LLM-based generative recommendation avoids this by directly generating item identi- fiers, but many methods represent each item as a token sequence, which creates two concrete problems: generation is slow because tokens must be produced step by step, and it can fail due to beam-search local optima, where the correct item is dropped early because its first token has low probability. SETRec addresses these issues by represent- ing each item as an order-agnostic set of tokens and using query-guided simultaneous token generation, so the item’s CF and semantic tokens are generated in parallel without intra-item token dependency. However, SETRec still uses only one collaborative filtering (CF) token, and when that CF signal comes from a sequence-aware model, it is vulnera- ble to recency bias, especially in sparse and cold-start settings where recent interactions dominate. CFs-SETRec reduces this recency bias by adding a second CF token chosen to capture long-term preferences, and combining it with the sequential CF signal, which preserves both short-term behavior and long-term affinity and leads to more balanced recommendations under sparse data.} }
Endnote
%0 Conference Paper %T Beyond Recency Bias: Combining Sequential and Global Collaborative Signals in LLM-Based Generative Recommendation for Sparse Data %A Behrad Ghiasi %A CI Ezeife %B Proceedings of the The 39th Canadian Conference on Artificial Intelligence %C Proceedings of Machine Learning Research %D 2026 %E Lydia Bouzar-Benlabiod %E Carson Leung %F pmlr-v318-ghiasi26a %I PMLR %P 284--295 %U https://proceedings.mlr.press/v318/ghiasi26a.html %V 318 %X Recommender systems often use multi-stage retrieval and ranking pipelines, where errors made early cannot be fixed later, which hurts overall recommendation quality. LLM-based generative recommendation avoids this by directly generating item identi- fiers, but many methods represent each item as a token sequence, which creates two concrete problems: generation is slow because tokens must be produced step by step, and it can fail due to beam-search local optima, where the correct item is dropped early because its first token has low probability. SETRec addresses these issues by represent- ing each item as an order-agnostic set of tokens and using query-guided simultaneous token generation, so the item’s CF and semantic tokens are generated in parallel without intra-item token dependency. However, SETRec still uses only one collaborative filtering (CF) token, and when that CF signal comes from a sequence-aware model, it is vulnera- ble to recency bias, especially in sparse and cold-start settings where recent interactions dominate. CFs-SETRec reduces this recency bias by adding a second CF token chosen to capture long-term preferences, and combining it with the sequential CF signal, which preserves both short-term behavior and long-term affinity and leads to more balanced recommendations under sparse data.
APA
Ghiasi, B. & Ezeife, C.. (2026). Beyond Recency Bias: Combining Sequential and Global Collaborative Signals in LLM-Based Generative Recommendation for Sparse Data. Proceedings of the The 39th Canadian Conference on Artificial Intelligence, in Proceedings of Machine Learning Research 318:284-295 Available from https://proceedings.mlr.press/v318/ghiasi26a.html.

Related Material