Research map / Learning theory and algorithms / Combinatorial and search optimization

Goal-Conditioned Supervised Learning for Multi-Objective Recommendation

NeurIPS 2026 · Learning theory and algorithms › Combinatorial and search optimization · approach: Multi-objective optimization · 1 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.

Extends goal-conditioned supervised learning (GCSL) to multi-objective next-action recommendation by using vector-valued goals (cumulative click and purchase rewards), trained with cross-entropy, plus a CVAE-based goal-selection method for inference.

Problem

Multi-objective recommendation requires balancing conflicting objectives (click, purchase) with complex architectures or loss-weighting, which is costly and treats noisy interaction data uniformly.

Approach

Redefine GCSL goals as multi-dimensional vectors of cumulative per-objective rewards, so a single transformer trained with plain cross-entropy learves to reach multiple objectives; noisy low-goal sessions are implicitly discounted, and a CVAE models achievable goals to select high inference goals.

Key results

• On RetailRocket, MOGCSL reaches 65.43% HR@10 and 52.92% NG@10 for purchase, versus 63.14% and 51.02% for the best prior baseline PMORS.
• On Challenge15, MOGCSL reaches 56.82% HR@10 for purchase versus 54.98% for PMORS, while MMOE-PE is slightly better on click.
• MOGCSL uses a 9.1M-parameter model with about 3.0Ks training time on RetailRocket, versus 14.0M and 9.6Ks for Share-Fix and 17.5M and 100.2Ks for RMTL.

Surprising findings

• Fixed-weight optimization outperforms the more advanced DWA on several metrics.
• The CVAE-based goal selection does not consistently beat the simple statistical goal strategy; it is worse on Challenge15 because high-goal training data is sparse.

Limitations

• CVAE-based goal selection underperforms when high-goal trajectories are sparse.
• Denoising claim is scoped to noise correlated with low long-term goals, not arbitrary noise.
• No live-user A/B testing; evaluation is offline.

Experimental setup

Datasets: Challenge15 RetailRocket Tenrec

Benchmarks: Challenge15 RetailRocket Tenrec

Metrics: HR@10 NG@10 HR@20 NG@20

Baselines: Shared-Bottom MMOE PE DWA Nash-MTL FAMO RMTL PMORS MOPRL

Builds on: GCSL MMOE FAMO PRL

Closest papers by meaning

• Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for RL-based LLM Recommenders (NeurIPS 2026 · similarity 0.85)
• LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.84)
• Would You Like Your Data to Be Trained? A User Controllable Recommendation Framework (AAAI 2024 · similarity 0.84)
• Many-Objective Multi-Solution Transport (ICLR 2025 · similarity 0.83)
• A Goal Interaction Graph Planning Framework for Conversational Recommendation (AAAI 2024 · similarity 0.83)
• Think before Recommendation: Autonomous Reasoning-enhanced Recommender (NeurIPS 2025 · similarity 0.83)