Answers are grounded in this paper and cite their sources.
Multimodal recommenders usually assume that fusing collaborative, visual, and textual signals improves accuracy. We treat this as a diagnostic question rather than an accuracy contest. Under interaction-level leave-one-out evaluation on three sparse Amazon categories, three findings emerge. First, text is the only modality that transfers; removing it lowers AUC by 0.23–0.25, while removing the visual stream changes AUC by at most 0.01 and removing collaborative ID features leaves it unchanged. Second, the apparent value of collaborative features is protocol-dependent. Full models reach AUC 0.97–0.99 under random splitting but only 0.71–0.74 under leave-one-out. The gap vanishes under a user-disjoint split, which is consistent with identity-side shortcuts rather than transferable signal. Third, white-box routing agrees with modality ablations at the dataset level, but is not seed-stable enough to serve as a per-instance explanation. We introduce counterfactual modality masking, which yields signed per-item modality attributions and shows that a modality can help on average while acting as noise for a specific item; gradient-based attribution, which aggregates absolute magnitudes, cannot express this sign. Against recent 2025–2026 multimodal models re-implemented under the identical protocol, LCFC is significantly more accurate than OrthoRec and statistically indistinguishable from I3MRec. Its contribution is not a leaderboard claim but a leakage-aware evaluation protocol together with an auditable diagnostic instrument. All deltas carry five-seed dispersion and paired significance tests.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: