A recommender model designed for sparse, multimodal data reported its clearest advantage when the system had to recommend items with their training interactions hidden. In tests on three benchmark datasets, MOTIF posted higher reported cold-item scores than the listed comparison methods, with gains of about 25% to 27% across Recall@20 and NDCG@20.
The overall gains were smaller but consistent in the reported results. At 20 recommendations, MOTIF reached Recall@20/NDCG@20 scores of 0.1083/0.0472 on Amz-Baby, 0.1146/0.0508 on Amz-Sports, and 0.0938/0.0429 on MicroLens. The reported improvements over the comparison results ranged from 4.92% to 6.07%.
A graph built from inferred interests
MOTIF combines semantic motivation reasoning, knowledge-enhanced graph reconstruction, weighted graph contrastive learning, and semantic-structural alignment. Language-model-generated semantics guide supervision, while final prediction uses graph embeddings.
The pipeline infers user and item motivations from sparse multimodal context, then turns those signals into an item-to-item graph. That reconstructed topology is intended to support propagation to cold items. MOTIF also uses weighted graph contrastive learning to make representations more robust and coordinates semantic information with graph structure.
The language-model reasoning, semantic encoding, and graph reconstruction are performed offline. At recommendation time, the system makes no language-model call, according to the reported setup.
Tests in highly sparse settings
The evaluation used three preprocessed datasets. Amazon-Baby contained 19,445 users, 7,050 items, and 160,792 interactions, with 99.88% sparsity. Amazon-Sports contained 35,598 users, 18,357 items, and 296,337 interactions, with 99.95% sparsity. MicroLens-50K contained 50,000 users, 19,220 items, and 359,708 interactions, with 99.96% sparsity.
The study examined overall recommendation accuracy, cold-start performance, the effect of individual model components, graph-quality measures, sensitivity to parameter choices, and efficiency. It classified users with one or two visible training interactions as extreme-cold, those with three to five as cold, and those with more than five as warm. Cold items were created by masking training interactions while keeping their multimodal content available; the masked interactions were reserved for validation and testing.
Performance was measured with Recall and NDCG, two ranking measures, at recommendation lists of 10 and 20 items. Each experiment was run with five random seeds and averaged, and paired t-tests were used for significance testing.
The gap widens when items are cold
MOTIF was reported to achieve the best performance across all the listed cold-start settings, including extreme-cold users, cold users, and cold items. The authors associate motivation reasoning with the sparse-user cases and reconstructed topology with the cold-item cases, but the benchmark comparison itself does not establish that either component causes the gains.
For cold items, MOTIF's reported Recall@20/NDCG@20 scores were 0.0613/0.0264 on Amz-Baby, 0.0671/0.0297 on Amz-Sports, and 0.0512/0.0226 on MicroLens. The reported improvements were 25.61%/26.32%, 24.95%/26.92%, and 25.49%/26.97%, respectively.
That pattern matters because the model's central design choice is to reconstruct links between items before making recommendations. The results therefore point to a specific strength in the tested setting: recovering useful structure when direct interaction data for an item are deliberately limited. They do not establish that the inferred motivations or graph links are functionally correct in human terms.
A benchmark result, not a user trial
The evidence remains an offline comparison on three preprocessed datasets, under the stated splits, masking rules, baselines, ablations, and five-seed averages. It supports comparative results in those settings, not a conclusion about how real users would respond or whether MOTIF would generalize to other domains.
The reported cold-start summary does not give exact subgroup sizes or exact significance results. The overall tables omit standard deviations for readability, while the cold-item table reports standard deviations only as being below 0.0016. The study also adds offline semantic-reasoning and graph-reconstruction work, so the reported recommendation gains do not by themselves show that MOTIF is cheaper overall.
The next test is whether the gains survive on additional datasets and in online recommendation settings. The study also leaves open whether motivation vectors and reconstructed relations are semantically accurate, and whether its graph diagnostics predict downstream recommendation value.
Paper data and sources
Original title: MOTIF: Motivation-guided Topology Inference for Cold-start Multimodal Recommendation
Authors: Yurui Shi, Yuchen Miao, Ximing Hu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text