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

Action-On-Item Preference Flow: A Shared Event Schema for Predictive and Generative Personalization

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Recommendation representation learning · 0 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 a shared action-on-item event schema and the MTSH/PerTIDE multi-timescale SSM history encoder that learns reusable user-history updates for both predictive ranking and personalized generation, with invariance/stability/pooling theory and controlled transfer experiments.

Problem

Personalization histories differ in native actions and objectives (watch/skip, click/ignore, request/response), so separate history encoders are typically built per action space; it is unclear whether one reusable update mechanism can learn from and transfer across these histories.

Approach

Rewrite every native interaction as a mapped action role plus an item content embedding; apply a learned action gate, three lag-biased SSM traces (long, short, episodic), adaptive fusion with a residual path, and command-conditioned readout, so shared update parameters maintain separate user states usable by predictive or generative consumers.

Key results

• On MIND, PerTIDE reaches 46.65 MRR vs 42.53 for a same-input three-branch SSM control (Tri-SSM-B), a 4.12-point gain.
• In a PENS-to-MovieLens frozen-core experiment, a PENS-trained core with trained head reaches 16.344 MRR vs 1.115 for a random frozen core with the same trained head.
• On PENS headline generation, PerTIDE reaches 0.366 PSE-JSD vs 0.263 for DeepSeek-R1-Distill-Qwen-32B (2-shot).

Surprising findings

• The episodic trace alone reaches only 24.18 MRR on MIND, yet adding it to L+S raises MRR from 41.26 to 46.65.
• An untrained MovieLens head on the PENS-trained frozen core reaches only 1.371 MRR, showing target alignment is required.

Limitations

• Reuse depends on supplied action mappings, informative content embeddings, and compatible downstream semantics.
• OpenAI-Reddit uses constructed chronology; MPChat shows deficits in response prediction and speaker identification.
• Stability and trace structure do not certify readout contraction or semantic disentanglement; objective weights are globally fixed.

Experimental setup

Datasets: MovieLens-1M MIND PENS OpenAI-Reddit MPChat PerAugy augmentation

Benchmarks: MovieLens-1M MIND PENS OpenAI-Reddit MPChat

Metrics: MRR nDCG@5 nDCG@10 HR@10 PSE-JSD PSE-SU4 PSE-METEOR BLEU-4 BERTScore

Baselines: UniSRec RecGURU SSD4Rec MINER LSTUR Tri-SSM-B GRU-B DeepSeek-R1-Distill-Qwen-32B

Builds on: UIG PerAugy UniSRec RecGURU LaMP

Closest papers by meaning

• Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents (NeurIPS 2026 · similarity 0.87)
• HYDRA: Model Factorization Framework for Black-Box LLM Personalization (NeurIPS 2024 · similarity 0.86)
• Eliciting Human Preferences with Language Models (ICLR 2025 · similarity 0.86)
• UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · similarity 0.85)
• BehaviorBench: Modeling Real-World User Decisions from Behavioral Traces (NeurIPS 2026 · similarity 0.85)
• PersonalLLM: Tailoring LLMs to Individual Preferences (ICLR 2025 · similarity 0.85)