A preprint describes D3ER, a multimodal recommendation method whose authors report the strongest offline results across three Amazon review datasets, compared with 15 listed recommendation models. The reported leaders differed by dataset: PGL+D3ER was reported best on Sports and Outdoors and Clothing and Shoes, while DiffMM+D3ER had the highest Recall and NDCG on Baby.
The authors also report a 7.58% improvement in the offline ranking measure Recall@50 over the MGCN model used as a base. The evaluation reported Recall and NDCG at cutoffs of 20 and 50.
How the system works
D3ER separates information shared by its two data types from information specific to each one. In this evaluation, the data types were visual and textual item features, and the feature-component disentanglement (FCD) step used inter-modal alignment and intra-modal separation to make that split.
It then uses separate recommenders for shared information, visual-specific information and textual-specific information. KDBoost, the knowledge-distillation-enhanced gradient-boosting stage, adds boosting and space-efficient distillation and applies global-correction regularization.
The evaluation used three real-world Amazon review datasets: Baby, Sports and Outdoors, and Clothing and Shoes. Visual features came from CLIP-ViT for Baby and Sports or VGG-16 for Clothing, while Sentence-BERT supplied textual features for all three categories.
What the tests showed
For DiffMM+D3ER, the reported Recall@20 values were 0.1018, 0.1062 and 0.0849 across the three datasets; for PGL+D3ER, they were 0.0961, 0.1124 and 0.1058. The reported improvement rows were positive for every displayed metric in both backbone blocks.
The analysis marked an improvement as statistically significant when a paired t-test gave p<0.05. Exact p-values, standard deviations and confidence intervals were not reported for the comparison.
With DiffMM as the backbone, variants that omitted FCD or either alignment loss showed lower performance, while the full model retaining both alignment losses had the best result among those variants. The version without boosting and distillation showed drops across all datasets, and the version without global correction had the poorest performance among the variants.
Performance initially rose and then fell as the two alignment-loss weights increased, with the best weight contributions varying by dataset. Very small distance thresholds were reported to miss discriminative information, while very large thresholds added noise.
A two-dimensional t-SNE visualization showed clear separation among the three feature types: shared, visual-specific and textual-specific. The authors treat that pattern as support for the design, but t-SNE is exploratory and does not quantify disentanglement quality.
The authors report inference speed and memory use comparable to the backbone. Disentanglement added negligible training overhead, while distillation was the main additional training cost; exact speed, memory and storage measurements were not reported in the extracted results.
What remains untested
The study's evidence is limited to offline ranking metrics on the three Amazon datasets. It therefore does not show whether D3ER would change user satisfaction, retention or welfare in an online service, and it does not establish performance beyond the tested Amazon categories, visual and textual modalities, or backbones.
Dataset sizes, user, item and interaction counts, train/validation/test splits, independent-run counts, standard deviations and confidence intervals were not reported in the supplied analysis. Open questions include whether the gains survive different splits, random initializations and hyperparameter choices, and whether distillation's extra training cost is justified by its storage savings.
The document is marked arXiv v1 and dated 26 Aug 2026; its front matter also includes an ACM MM '26 proceedings citation and DOI.
Paper data and sources
Original title: D3ER: Supporting Multi-Modal Recommendation via Disentangle and Distillation-based Dynamic Ensemble
Authors: Bingnan Wang, Yi Li, Xiongxin Tang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text