{"id":"97be632d-40a3-4393-ba59-eccdec1adeee","arxiv_id":"2504.16352","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DGMRec disentangles modality features into shared and unique parts and generates the missing ones, outperforming prior multimodal recommenders on four datasets.","lead":"This paper introduces DGMRec, a recommender system that generates missing image, text, or audio features instead of dropping items or injecting crude averages. It reports consistent performance gains over existing multimodal recommender systems in missing-modality and new-item scenarios, with code released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Specific-feature generation via Eq. 13 depends on P_{i,m}, which is empty for new items; Table 2's largest gains occur in exactly that new-item setting.","rationale":"The strongest claim is broad: DGMRec consistently outperforms existing MRSs and MMA-RSs in both the missing-modality setting and the combined missing-modality plus new-item setting. The paper supports this with tables, ablations, and parameter sensitivity analyses, and the generation losses are plausible as a design. Still, the mechanism for specific-feature generation in Eq. 13 is the linchpin, and it is least secure exactly where the paper claims its biggest advantage. The reader's weakest_assumption already identifies the dependence on user preference aggregates; my read sharpens it: in the new-item protocol, the aggregate does not exist. This is not a stylistic objection; it creates an internal gap between the stated method and the headline result. The proposed test is a direct way to decide whether the authors handle N_i = empty by a documented default or by leaking test interactions. If the former, the significant improvement in Table 2's lower half is unexplained because the generated specific feature is item-independent; if the latter, the comparison is not a fair new-item evaluation. Since the central claim explicitly includes new items, this concern is load-bearing. I nevertheless recommend keeping the reader's CONDITIONAL verdict: the missing-modality result for existing items may survive, and the authors can resolve the concern by clarifying the protocol and adding the requested ablation. My agreement with the reader's weakest_assumption is partial: it is the same Eq. 13 mechanism, but the empty-N_i case is a sharper and more decisive failure mode.","tokens_in":20307,"tokens_out":6178,"duration_ms":65165,"concrete_test":"In the released code, trace the data split for the new-items protocol and log |N_i| for each new item when Eq. 13 is evaluated. Then re-run Table 2's lower half under two conditions: (a) P_{i,m} for new items set to a learned global vector, and (b) P_{i,m} computed only from training interactions. If Recall@20 or NDCG@20 in the new-item setting changes materially relative to the reported numbers, the protocol is either leaky or Eq. 13 contributes nothing; report the comparison and the resulting drop.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3.2 defines specific-feature generation as Ê^s_m = G^s_m(P_{i,m}), where P_{i,m} is the mean of preference embeddings of users in N_i (Eq. 2). This is the only route for generating modality-specific content when all modalities are missing, and it is also the route used in the Missing Modality + New Items setting. But in that setting, new items are defined as appearing only in the test set, so N_i is empty by construction: the average in Eq. 2 is undefined, and no default is specified. If P_{i,m} is replaced by a constant or zero vector for such items, Ê^s_m is item-independent, leaving only the general-feature path (itself constant when all modalities are missing) plus the untrained ID embedding. Yet Table 2 shows DGMRec's largest relative gains in this setting (e.g., Baby Recall@20 +14.06% vs +7.43% in the plain missing-modality setting). The paper therefore either uses test-time interactions to compute P_{i,m}, which would be a leak, or the reported new-item gains cannot be explained by the generation mechanism as written. Even for non-new items, the premise that an average of user preference embeddings carries item-specific visual or textual attributes is asserted but not verified; the new-item case makes the missing link concrete.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DGMRec, a multimodal recommender system designed for missing-modality scenarios. The model splits each modality feature into a general component shared across modalities and a specific component unique to the modality, using CLUB and InfoNCE losses to enforce disentanglement and cross-modal alignment. Missing modality features are generated by a reconstruction-based module: general features are produced from other available modalities, while specific features are generated from an item-level aggregation of user modality preference embeddings (Eq. 13). The generated features are used to refine an item-item graph, and two alignment losses connect modality representations with collaborative filtering. Experiments on Baby, Sports, Clothing, and TikTok compare DGMRec with CF, MRS, and MMA-RS baselines in two settings (missing modality and missing modality plus new items), together with ablations, missing-ratio analysis, sensitivity analysis, time complexity, and cross-modal retrieval.","tokens_in":20553,"tokens_out":5955,"duration_ms":62679,"significance":"If the empirical claims are reliable, the paper makes a useful contribution: it offers a generation-based alternative to injection-based and robustness-based handling of missing modalities, reports directionally consistent gains over a wide set of baselines, and enables a cross-modal retrieval task that injection methods cannot support. The release of code, the broad baseline coverage, and the inclusion of ablations and sensitivity analyses are strengths. The main caveats are that the reported numbers carry no variance information and that the mechanism underlying the new-item results is under-specified; as written, the claimed new-item gains cannot be traced to the generation module without additional clarification.","major_comments":[{"comment":"For the \"Missing Modality + New Items\" setting, p_{i,m} is undefined for new items. New items are defined in Section 4.1 as appearing only in the test set, so their neighbor set N_i is empty when the training split is used, and the manuscript does not specify any default for P_{i,m} in this case. If P_{i,m} is computed from test interactions, that would be label leakage; if it is replaced by a constant or omitted, the specific-feature generator G^s_m(P_{i,m}) produces an item-independent vector, and the generation mechanism cannot explain the large new-item gains in Table 2 (e.g., Baby R@20 +14.06%). Please clarify how P_{i,m} is obtained for new items and report results on the new-item subset separately, because the current text makes the central new-item claim untestable.","section":"§3.1, Eq. (2); §3.3.2, Eq. (13); §4.1"},{"comment":"The claim that specific features generated via the preference-based approach remain effective when no modality is available is not supported for items with no interactions. When all modalities are missing, the general-feature generator in Eq. (12) receives only mean-imputed inputs and outputs a constant; when the item is also new, the specific-feature generator in Eq. (13) is constant by the argument above. The only item-dependent signal left is a randomly initialized ID embedding that was not trained for new items. The paper should either show that P_{i,m} is well-defined and trained for such items or provide an analysis restricted to items that are simultaneously new and have zero available modalities; otherwise the reported gains in this cell cannot be attributed to the proposed generation module.","section":"§3.3.2; §4.2.2"},{"comment":"No standard deviations, confidence intervals, or significance tests are reported. The central claim is an empirical superiority claim (Section 4.2.1), and several reported improvements are small (e.g., TikTok R@20 +4.69% in the missing-modality setting). Without any uncertainty quantification, the reader cannot distinguish a systematic advantage from run-to-run variation. Please add multiple-seed results with standard deviations and, where appropriate, paired significance tests.","section":"Tables 2-4; Figures 3-4"}],"minor_comments":[{"comment":"The text says \"Across all missing ratios, DGMRec consistently outperformed all baselines, except for the 0% baseline,\" which is internally contradictory; please rephrase and discuss the 0% case explicitly.","section":"§4.2.3"},{"comment":"The cross-modal retrieval setup is under-specified: Hit@10 and Hit@20 are not defined for the retrieval task, and the NN baseline is absent for the two-missing-modalities case, so the comparison is incomplete.","section":"Table 3"},{"comment":"There are several typographical issues, including \"Disentangling and Generating Modality Recommender\" without spacing, \"TikTook\" instead of \"TikTok\", and an unusual denominator symbol in Eq. (8) that should be a summation sign.","section":"Abstract and §1"},{"comment":"The TikTok dataset's raw features are not publicly available; this is disclosed, but the resulting reproducibility limitation should be addressed explicitly in the experimental section, for example by noting which downloaded feature files are required to run the released code.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The skeptical concern about Eq. (13) and the new-item setting is well grounded in the manuscript text alone. If the released code computes P_{i,m} from the full interaction matrix including test interactions, the new-item results would be invalid and would require a much more serious intervention; if not, the manuscript still needs a clear specification and new-item-only metrics. Please ask the authors to state the exact computation path for P_{i,m} in the new-item setting and to verify it against the code before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper does something real: it separates shared and modality-specific information using mutual information bounds (CLUB and InfoNCE), then generates missing modalities with decoders. That combination in one recommend system is new to me, and the ablations say each piece contributes. The gains over strong baselines on the Amazon and TikTok benchmarks are roughly 5-11% in recall/NDCG, which is in the range practitioners care about. Second, there is a load-bearing hole in the new-item setting. For items that appear only in the test set, the interacted-user set N_i is empty, so Eq. 2 and Eq. 13 are undefined. The paper does not say what the code substitutes. If it uses a zero or constant vector, the generated specific features are item-independent, so the mechanism cannot explain the larger relative gains in the right half of Table 2. If it uses test-time interactions to compute P_{i,m}, that is a leak. Either way, the authors need to specify what is implemented. This is not a cosmetic issue: the new-item scenario is a core part of the practical claim. Other soft spots are more routine: no standard deviations or significance tests, so the word \"consistently\" is stronger than the evidence; at 0% missing ratio DGMRec actually trails GUME slightly in Figure 4(a); and the TikTok raw features are not released, so that part cannot be independently verified. The code is available but without a commit hash, so replication is partial. None of this makes me think the method is fake. The plain missing-modality setting, where items have training interactions, is well-defined and the gains there are credible. The disentanglement visualization and parameter sensitivity analysis are also encouraging. The cross-modal retrieval result is a nice byproduct, though it would be stronger with a second content-based baseline. Bottom line: this paper deserves a serious referee. I would send it to review, asking the authors to clarify the new-item protocol, add variance estimates, and temper the \"consistently\" wording. If the new-item mechanism turns out to be leaky, the contribution shrinks considerably; if it is just under-specified, it is a fixable reporting gap.","headline":"DGMRec is a genuinely integrated method for missing-modality recommendation, but the new-item experiments are unexplained by the mechanism as written and need clarification before the headline claim holds.","tokens_in":21111,"tokens_out":3701,"would_cite":true,"duration_ms":35918,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DGMRec claims that missing modality features in recommender systems should be generated, not injected, and that disentangling each modality into shared and unique parts makes this generation accurate enough to beat existing multimodal…","keywords":["Multi-modal recommendation","Missing modalities","Feature disentanglement","Collaborative filtering","Modality generation","Cross-modal retrieval","Mutual information","Graph refinement"],"falsifier":"Take items whose modalities are fully available, artificially mask their image and text features entirely, and compare DGMRec's generated features for those items against the true held-out features using cosine similarity or MSE, against a baseline that uses the global mean or nearest-neighbor features; if the generated features are no closer to the truth than the baseline, the mechanism behind the reported gains does not work as described. A second check is to replace the user-modality preference embedding $P_{i,m}$ in Eq. (13) with random vectors of the same dimension; if recommendation performance does not drop substantially, then the specific-feature generator is not actually using item content information.","tokens_in":20067,"feed_emoji":"🧩","tokens_out":5114,"duration_ms":45094,"temperature":0.7,"pith_summary":"The paper argues that the right way to handle missing modalities in recommender systems is to generate the missing features rather than to inject crude substitutes or ignore them. It proposes DGMRec, which splits each modality's feature vector into a general part shared across modalities and a specific part unique to that modality, then reconstructs missing modality features by combining aligned general features from available modalities with user modality preferences. The claim is that this generation-based approach yields consistently better ranking quality than existing multimodal and missing-modality-aware recommenders, including when all modalities of an item are missing and when new items appear at test time. It also enables cross-modal retrieval, a task that existing multimodal recommenders cannot perform at all.","feed_headline":"Recommender regenerates missing images and text to outrank rivals","feed_subtitle":"DGMRec splits each modality into shared and unique parts, then generates what is missing—even when every modality is gone.","key_machinery":"The central object is the disentangled modality representation: each modality $m$ is passed through a shared general encoder $f_g$ and a modality-specific encoder $f^s_m$ to yield general features $E^g_m$ and specific features $E^s_m$. Two information-based losses drive the separation: the Contrastive Log-ratio Upper Bound (CLUB) minimizes mutual information between general and specific features within the same modality, while InfoNCE maximizes mutual information between general features across modalities. For missing modalities, a general generator $G^g_m$ maps concatenated aligned general features from other modalities to the missing general feature, and a specific generator $G^s_m$ maps the item's user-modality preference embedding $P_{i,m}$ (the average of preference embeddings of users who interacted with the item) to the missing specific feature; a per-modality decoder then reconstructs the raw feature, and the generated features refine the item-item graph with an adaptive blend of old and new adjacency matrices.","core_discovery":"On its own terms, the paper establishes that a recommender that explicitly separates modality features into general and specific components, and generates missing components rather than injecting average or nearest-neighbor features, outperforms state-of-the-art multimodal recommenders and missing-modality-aware recommenders in missing-modality settings, new-item settings, varying missing ratios, and extreme cases where no modality remains. The key result is that specific features can be generated from user modality preference embeddings alone, so even an item with every modality missing can still receive a useful modality representation. The paper reports consistent improvements over the best baseline across four datasets in both evaluation settings, with relative Recall@20 gains ranging from roughly 4.7% to 14.1%, and it shows the model can retrieve similar items through generated features when all modalities are absent.","pith_inferences":["If the preference-embedding generation works as claimed, the same user-modality preference embeddings could serve as a portable representation for cold-start users, transferring modality preferences across domains without retraining.","The directed-edge graph refinement strategy, which only adds edges from available modalities to missing ones, suggests a general recipe for any graph-based recommender to ingest generated features without contaminating original signals.","A direct testable extension is to hide an available modality at test time and check whether DGMRec's generated feature for that modality approximates the true feature better than nearest-neighbor injection; if it does not, the ranking gains must come from the alignment or disentanglement losses rather than from faithful generation.","The disentanglement losses could transfer to other multimodal tasks, such as retrieval with mixed text-and-image queries, wherever modality-specific detail matters."],"forward_implications":["Recommenders that generate missing modalities retain more ranking quality as the missing ratio grows; the paper reports a relative performance drop of 10.2% at 80% missing modality, compared with 15.1% for the best baseline.","New items with partial or no modality data can still be represented through generated features, directly addressing the cold-start item recommendation problem.","Cross-modal retrieval becomes possible: given a query item with a missing modality, generated features allow retrieving similar items even when every modality is absent.","Disentangling modality features into shared and unique parts improves the quality of item representations for ranking, beyond the benefit of generation alone.","Modality knowledge and collaborative knowledge reinforce each other through alignment losses, so the generated features contribute to the recommendation score rather than remaining isolated content representations."],"supporting_citations":[{"why":"Supplies the nearest-neighbor injection baseline and the motivating observation that injection still causes performance drops under missing modalities.","marker":"[13]"},{"why":"CI2MG is the main prior generation-based missing-modality recommender using optimal transport; DGMRec must outperform it and contrasts its approach.","marker":"[12]"},{"why":"MILK provides the invariant-learning baseline and the new-item evaluation setup used in the missing-modality plus new-items setting.","marker":"[1]"},{"why":"SIBRAR is the single-branch embedding network baseline for missing-modality and cold-start scenarios.","marker":"[5]"},{"why":"LGMRec is the recent multimodal recommender used in the Figure 1 analysis of injection failure and as a state-of-the-art baseline.","marker":"[6]"},{"why":"GUME is a state-of-the-art hybrid multimodal recommender that degrades sharply under missing modalities, serving as a key comparison point.","marker":"[11]"},{"why":"DAMRS is a denoising-and-aligning multimodal recommender baseline that DGMRec must beat in all experiments.","marker":"[29]"},{"why":"CLUB provides the contrastive log-ratio upper bound used to minimize mutual information between general and specific features within a modality.","marker":"[2]"},{"why":"InfoNCE provides the lower bound on mutual information used to align general features across different modalities.","marker":"[15]"},{"why":"LightGCN is the graph convolution backbone used for modality feature propagation and also a strong collaborative filtering baseline.","marker":"[9]"}],"fun_headline_variants":["Missing modalities? Recommender regenerates them to rank better","Even with zero modalities, recommender still recommends","Disentangle and generate: MRS beats rivals with missing data","Shared vs specific features: generate missing ones for better recs","Generating lost modalities outranks state-of-the-art recommenders"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the average of the modality-preference embeddings of the users who interacted with an item contains enough item-specific information to regenerate that item's unique modality attributes, so that even an item with no available modalities can receive a faithful modality representation.","fun_headline_variants_meta":{"raw":{"variants":["Missing modalities? Recommender regenerates them to rank better","Even with zero modalities, recommender still recommends","Disentangle and generate: MRS beats rivals with missing data","Shared vs specific features: generate missing ones for better recs","Generating lost modalities outranks state-of-the-art recommenders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000895,"raw_usage":{"total_tokens":3850,"prompt_tokens":929,"completion_tokens":2921,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":2837}},"tokens_in":545,"tokens_out":2921,"duration_ms":19876,"temperature":1.0,"reasoning_tokens":2837,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:05:05.915458+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take items whose modalities are fully available, artificially mask their image and text features entirely, and compare DGMRec's generated features for those items against the true held-out features using cosine similarity or MSE, against a baseline that uses the global mean or nearest-neighbor features; if the generated features are no closer to the truth than the baseline, the mechanism behind the reported gains does not work as described. A second check is to replace the user-modality preference embedding $P_{i,m}$ in Eq. (13) with random vectors of the same dimension; if recommendation performance does not drop substantially, then the specific-feature generator is not actually using item content information.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the nearest-neighbor injection baseline and the motivating observation that injection still causes performance drops under missing modalities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CI2MG is the main prior generation-based missing-modality recommender using optimal transport; DGMRec must outperform it and contrasts its approach."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MILK provides the invariant-learning baseline and the new-item evaluation setup used in the missing-modality plus new-items setting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SIBRAR is the single-branch embedding network baseline for missing-modality and cold-start scenarios."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GUME is a state-of-the-art hybrid multimodal recommender that degrades sharply under missing modalities, serving as a key comparison point."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DAMRS is a denoising-and-aligning multimodal recommender baseline that DGMRec must beat in all experiments."}],"review_version":1}