Source-linked AI summary
MOTIF: Motivation-guided Topology Inference for Cold-start Multimodal Recommendation
Yurui Shi, Yuchen Miao, Ximing Hu, Zijun Wang, Chang Han
TL;DR
Cold-start multimodal recommendation must address sparse user intent evidence, isolated cold items, and semantic drift from similarity-based graphs. MOTIF uses offline LLM reasoning to infer motivations, reconstruct item topology, and learn robust graph embeddings without injecting generated text into prediction. Experiments on three multimodal benchmarks show consistent gains over several baseline families, with up to 6.07% relative improvement over the strongest recent baseline.
Problem
Cold-start multimodal recommendation faces sparse interactions that obscure user intent, isolated cold items, and semantic drift from similarity-based item graphs.
Method
MOTIF uses offline LLM reasoning to infer motivation semantics, reconstruct transferable item-item topology, and learn robust graph embeddings through weighted graph contrastive learning and semantic-structural alignment.
Results
Experiments on three multimodal benchmarks show consistent gains over graph-based, multimodal, cold-start, and LLM-enhanced baselines, with up to 6.07% relative improvement over the strongest recent baseline.
Takeaways & Limitations
Translating LLM-derived knowledge into graph structure and representation supervision outperforms direct semantic fusion while prediction uses only graph embeddings.
Takeaways & Limitations
Motivation text and semantic vectors are obtained using only training interactions to avoid information leakage.
Abstract
from arXiv · showhide
Cold-start multimodal recommendation faces three coupled challenges: (i) sparse interactions obscure user intent, (ii) cold items remain topologically isolated, and (iii) similarity-based item graphs may cause semantic drift. To address these issues, we propose MOTIF, a Motivation-guided Topology Inference framework for cold-start multimodal recommendation. MOTIF integrates Semantic Motivation Reasoning, Knowledge-enhanced Graph Reconstruction, Weighted Graph Contrastive Learning, and Semantic-Structural Alignment. It uses offline LLM reasoning to infer motivation semantics, reconstructs transferable item-item topology, and learns robust graph embeddings without injecting generated text into prediction. Experiments on three multimodal benchmarks show consistent gains over graph-based, multimodal, cold-start, and LLM-enhanced baselines, with up to 6.07% relative improvement over the strongest recent baseline.
1 Introduction
Cold-start multimodal recommendation suffers from sparse intent evidence, isolated cold items, and semantic drift in similarity-based graphs. MOTIF jointly infers motivations, reconstructs transferable topology, and learns robust representations.
- Motivation: Sparse interactions obscure user motivations, while cold items lack sufficient interaction edges for graph propagation.
- Motivation: Existing approaches often separate semantic modeling, graph reconstruction, and representation robustness, limiting their joint treatment of cold-start challenges.
- Motivation: Similarity-based item graphs may connect visually similar but functionally different items, causing semantic drift.
- MOTIF: MOTIF infers user and item motivations from sparse multimodal contexts and transforms them into knowledge-enhanced item-item topology.
- MOTIF: MOTIF learns robust representations through weighted graph contrastive learning and semantic-structural coordination.
- Contributions: The work formulates cold-start multimodal recommendation as joint semantic and topological reconstruction over motivations and graph connectivity.
2 Related Work
Related work spans multimodal graph learning, graph collaborative filtering, cold-start adaptation, and LLM-enhanced recommendation. MOTIF targets their underexplored interaction through motivation-aware functional relations and cold-start graph reconstruction.
- Multimodal Recommendation: Multimodal methods build user-item or item graphs from modality-aware features, but appearance similarity can hide different functions and cause semantic drift.
- Graph Recommendation: Graph collaborative filtering and contrastive methods improve propagation robustness, but usually assume reliable neighborhoods unavailable for sparse users and cold items.
- Cold-start Recommendation: Cold-start methods use side information, generated representations, or rapid adaptation, yet rarely reconstruct graph structures for cold-item message passing.
- LLM-enhanced Recommendation: LLM-enhanced recommenders support semantic understanding, profile or graph augmentation, alignment, and motivation extraction.
- Research Gap: The interaction among motivation reasoning, cold-start graph reconstruction, and graph contrastive learning remains underexplored.
3 Methodology
MOTIF formulates cold-start multimodal recommendation as semantic-topological co-reconstruction, using motivation semantics to infer transferable item connectivity and robust graph representations while keeping prediction graph-based.
- 3.1 Problem Formulation: The task ranks candidate items for each user despite sparse behavioral evidence, limited cold-item connectivity, and multimodal information spanning text, visual, and attributes.MOTIF treats multimodal content as motivation-level evidence rather than only auxiliary item information.
- 3.2 Framework Overview: MOTIF combines Semantic Motivation Reasoning, Knowledge-enhanced Graph Reconstruction, Weighted Graph Contrastive Learning, and Semantic-Structural Alignment.LLM-generated semantics guide graph reconstruction and alignment supervision.
- 3.3 Semantic Motivation Reasoning: Semantic Motivation Reasoning converts sparse multimodal evidence into compact motivation vectors using an offline LLM, whose outputs are encoded before training.For items, the LLM summarizes functional role, usage scenario, target audience, and complementary or substitutable relations.
- 3.4 Knowledge-enhanced Graph Reconstruction: The reconstructed item graph combines raw multimodal similarity, motivation semantics, and relation priors to improve cold-item connectivity.Cosine similarity is balanced between motivation semantics and original multimodal representations; relation scores are computed within retrieved candidate sets and each item retains top-Kg neighbors.
- 3.5 Weighted Graph Contrastive Learning: Weighted Graph Contrastive Learning improves robustness against uncertain reconstructed edges by perturbing final-layer representations and recalibrating contrastive views.The contrastive loss operates on node representations from stochastic views, with mini-batch negatives and temperature τ.
- 3.6 Semantic-Structural Alignment: Semantic-Structural Alignment projects semantic vectors and graph embeddings into a shared space, while final recommendation scores use only graph embeddings.LLM-derived semantics therefore supervise reconstruction and alignment without being directly injected into prediction.
4 Experiments
MOTIF is evaluated across three multimodal datasets using standard accuracy, cold-start, ablation, graph-quality, sensitivity, and efficiency analyses. It consistently outperforms baseline groups, improves reconstructed cold-item topology, and separates offline semantic costs from online prediction.
- Experimental Setup: Experiments use Amazon-Baby, Amazon-Sports, and MicroLens-50K, with user- and item-side cold-start settings constructed from the training graph.User-side and item-side reasoning are restricted to training-visible evidence or item content to prevent validation and test leakage.
- Experimental Setup: MOTIF is compared with graph collaborative filtering, graph contrastive learning, multimodal, cold-start, and LLM-enhanced recommendation baselines.Recall@K and NDCG@K are evaluated at K ∈{10, 20}, with results averaged over five random seeds.
- Overall Performance: MOTIF consistently outperforms all baseline groups in overall performance, supporting reconstructed cold-start topology and semantic alignment with collaborative graph embeddings.Overall comparisons are reported in Tables 2 and 3.
- Cold-start Performance: MOTIF achieves the best performance across extreme-cold users, cold users, and cold items.The results associate motivation reasoning with sparse-user performance and reconstructed topology with cold-item performance.
- Ablation Study: Removing LLM reasoning or graph reconstruction degrades performance, while Direct-Fusion performs worse than using LLM semantics as auxiliary supervision.The Similarity-Graph comparison supports motivation-aware reconstruction, and the ablations cover weighted contrastive learning, recalibration, and alignment components.
- Reconstructed Graph Analysis: The reconstructed graph reduces isolated cold items, increases their average degree, and achieves higher NU@10 than similarity-based graphs.The analysis compares the original interaction graph, a multimodal similarity graph, and the proposed reconstructed graph.
- Hyperparameter Sensitivity: Moderate hyperparameter values generally perform best, balancing regularization, semantic alignment, connectivity, and graph noise.Very small λ1 or λ2 weaken their respective objectives, whereas overly large values may dominate recommendation supervision.
- Efficiency: MOTIF adds offline motivation-reasoning and topology-reconstruction cost, but cached semantic vectors and graphs enable online prediction without LLM calls and with comparable inference latency.Offline and online costs are reported separately; online prediction uses only graph embeddings.
5 Conclusion
The paper presents MOTIF as an offline LLM-enhanced framework that translates motivation semantics into graph structure and representation supervision. Across three datasets, it reports consistent gains while avoiding online LLM calls, and identifies adaptive graph priors and more efficient reasoning as future directions.
- Conclusion: MOTIF infers motivation semantics offline, reconstructs transferable item-item topology, and learns graph embeddings through weighted contrastive learning and semantic-structural alignment.Prediction uses only graph embeddings rather than generated text.
- Conclusion: Experiments on three datasets show consistent gains over graph-based, multimodal, cold-start, and LLM-enhanced baselines.Ablations validate motivation-aware reconstruction and auxiliary alignment.
- Conclusion: Translating LLM-derived knowledge into graph structure and representation supervision outperforms direct semantic fusion while avoiding online LLM calls.Future work will explore adaptive graph priors and more efficient semantic reasoning.