{"id":"fbea5e6f-69de-449b-be8d-eef3b1566fd9","arxiv_id":"2505.12283","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TDM trains diffusion recommenders on histories with deliberately removed items, using dual-side Thompson sampling guided by continuity and stability, to improve robustness to real missing data.","lead":"This paper proposes TDM, a diffusion-based recommender that copes with missing interaction data by deliberately removing some items from user histories and training the model to stay consistent despite the gaps. The approach could make generative recommenders more robust to incomplete user logs and is reported to beat several baselines on five datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The extrapolation proof's alignment assumption (Sec. 4.3) is unverified and likely false for non-preference-preserving missingness; TDM's robustness may not transfer to real missing data.","rationale":"The paper's contribution is primarily empirical: TDM outperforms several baselines on six datasets, and ablations show DTS helps even for traditional recommenders. This is genuine evidence. However, the headline theoretical justification in Sec. 4.3 is the extrapolation inequality (19). I re-derived the argument and found that the only bridge from simulated missingness to real missingness is the assertion that the two missing-data processes 'align well,' so that ḡ−ĝ ≈ ĝ−g̃. No structure in DTS guarantees this alignment; in fact DTS is engineered to preserve preferences, whereas real missingness mechanisms (privacy, exposure, popularity) are not. The paper's own Conclusion lists exposure bias and popularity bias as directions for 'targeted simulation,' which concedes that the current simulation is not aligned with those mechanisms. Under a mismatched mechanism, the ratio of inner products can be unbounded, so the proof provides no guarantee. This concern does not require dismissing the empirical results: the improvements may come from data augmentation and regularization rather than from a validated extrapolation guarantee. A conditional acceptance that asks for the proof to be softened or for controlled missingness experiments is appropriate. My concrete test would settle whether the alignment assumption matters in practice.","tokens_in":17828,"tokens_out":6820,"duration_ms":67154,"concrete_test":"Use the fully-observed KuaiRec dataset to create corrupted versions with four missing mechanisms: (a) DTS-style preference-preserving deletion, (b) uniform random deletion, (c) popularity-biased deletion (low-popularity items removed), and (d) low-continuity/preference-shift deletion (items with small con_n removed). Train TDM and DreamRec/PDRec on each corrupted version and compare HR@20/NDCG@20. If TDM's advantage over baselines shrinks or disappears under (b)-(d), the alignment assumption fails and the generalization claim is not supported. Additionally, on validation sequences with ground-truth complete histories, compute the coefficient ratio |(ḡ−ĝ)ᵀ∇f| / |(g̃−ĝ)ᵀ∇f| for the trained model; a ratio substantially larger than a small constant under any real mechanism would directly falsify inequality (19).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TDM extrapolates from simulated missingness to real missingness rests on the coefficient bound in Eq. (19) of Sec. 4.3. The bound requires the ratio ((ḡ−ĝ)ᵀ∇f)/((g̃−ĝ)ᵀ∇f) to be bounded, and the only justification is the sentence: 'if the missing data process from e1:N−1 to e1:N−1⊖δ′ can align well with that from e1:N−1⊕δ to e1:N−1 ... the coefficient will be close to 1.' This is an assumption, not a proof. DTS (Eqs. 6-11) is deliberately designed to remove high-continuity items from high-stability sequences so that preference evolution is preserved. But real missing data are caused by privacy, technical limitations, exposure bias, and popularity bias (as the paper's Conclusion acknowledges). These mechanisms are not preference-preserving: exposure bias can preferentially remove low-popularity items, and privacy can remove sensitive items, both of which may be precisely the items that mark preference shifts. When the real missingness removes low-continuity or sequence-destabilizing items, the two pairs of guidance differences need not be approximately equal; the ratio can be large or the inner product in the denominator can be near zero, breaking inequality (19). In that case training on preference-preserving simulated deletions does not imply robustness to real missingness. The proof is therefore circular: it assumes the simulated missingness mechanism matches the real one, which is essentially the conclusion to be established. The consistency-regularization bound in Eq. (20) is also only an upper bound, not direct optimization, but the alignment assumption is the more serious gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes TDM, a diffusion-based generative sequential recommender designed to be robust to missing entries in user interaction sequences. Rather than imputing missing interactions, TDM deliberately introduces additional artificial missingness: a dual-side Thompson sampling procedure (Section 4.1) scores items by local continuity and sequences by global stability, and edits sequences by removing high-continuity items from high-stability sequences. The edited sequence is encoded as guidance for a DDIM-based conditional diffusion model trained with classifier-free guidance (Section 4.2). The authors claim that training on such edited sequences provides consistency regularization and that, by extrapolation, the model becomes robust to real missing data. Section 4.3 gives a Taylor-expansion argument intended to bound the effect of real missingness by the effect of simulated missingness. Experiments on YooChoose, KuaiRec, Zhihu, Steam, Beauty, and Toys compare TDM with sequential and generative baselines, including recovering-based methods, and report consistent improvements; ablations and a DTS extension to non-diffusion recommenders are also provided.","tokens_in":18171,"tokens_out":7959,"duration_ms":77778,"significance":"If the central transfer claim were rigorously supported, TDM would be a valuable alternative to recovery-based approaches: it is simple, computationally efficient, and shows consistent improvements across several datasets and baseline families. The paper has concrete strengths: the experimental comparison is broad (Tables 2 and 3), ablations isolate the continuity and stability metrics (Table 4), the DDIM/DDPM comparison (Table 5) is useful, and the code and data are released. The main weakness is that the theoretical justification in Section 4.3 does not actually establish extrapolation from simulated to real missingness; as written, it assumes the key property it sets out to prove. The empirical evidence is encouraging, but the transfer mechanism is not validated under missingness mechanisms that differ from DTS's preference-preserving deletions.","major_comments":[{"comment":"The extrapolation argument is circular. The coefficient ratio is bounded only when 'the missing data process from e_{1:N-1} to e_{1:N-1}⊖δ′ can align well with that from e_{1:N-1}⊕δ to e_{1:N-1};' this is exactly the claim that simulated deletions mimic real missing data, and it is not derived from the model or from any stated assumption. The final paragraph of Section 6 acknowledges that real missing data can arise from exposure or popularity bias, which are not preference-preserving processes of the type DTS is designed to emulate. Because the alignment premise is the load-bearing assumption, the section should either be reframed explicitly as intuition or supplied with a formal, testable condition under which the two guidance differences are approximately equal.","section":"Section 4.3, Eq. (19)"},{"comment":"The consistency regularization claim is not supported by the inequality presented. Equation (20) is a trivial consequence of ||a-b||_2^2 ≤ 2(||a-y||_2^2 + ||b-y||_2^2); any two functions that both approximate y satisfy it. The training objective L in Eq. (14) is a sum of per-sample reconstruction losses and contains no term that directly penalizes disagreement between f(e,ĝ) and f(e,g̃). Hence the text does not show that L 'achieves consistency regularization' in the sense of enforcing invariance to simulated missingness. If consistency is a stated contribution, the authors should either add an explicit consistency penalty and ablate it, or weaken the wording to describe a property of the objective.","section":"Section 4.3, Eq. (20)"},{"comment":"The synthetic missingness experiment does not establish transfer to real missingness as described. The text reports creating datasets with 10-30% missing data but does not specify the deletion mechanism. If the deletion is uniform random, it is a different process from DTS and, more importantly, it does not test the claimed extrapolation for non-preference-preserving mechanisms (e.g., exposure or popularity bias). Please specify the mechanism and include at least one condition in which missingness is correlated with item popularity or recency, or provide an evaluation with naturally missing observations from a dataset with a known complete ground truth.","section":"Section 5.5"}],"minor_comments":[{"comment":"The 'Improv.' rows are computed relative to TDM's own score rather than the baseline score. For example, on YooChoose HR the table reports 9.85% for TDM over PDRec, which is (6.90−6.22)/6.90; the standard relative improvement is (6.90−6.22)/6.22 ≈ 10.93%. Please recompute all improvement percentages.","section":"Tables 2 and 3"},{"comment":"The paper says 'five datasets,' but Table 1 lists six datasets (YooChoose, KuaiRec, Zhihu, Steam, Beauty, and Toys); the counts should be made consistent.","section":"Abstract, Section 5, and Table 1"},{"comment":"The procedure in Eqs. (9)-(11) uses fixed thresholds λ1 and λ2 after drawing probabilities, with no posterior update or exploration-exploitation tradeoff as described for Thompson sampling in Section 3.2. Either define the sampling distribution F(v,p) concretely or rename the procedure to avoid the mismatch.","section":"Section 4.1.2 and Section 3.2"},{"comment":"No statistical significance tests are reported for the main comparisons. Given the reported standard deviations, pairwise significance tests (e.g., paired bootstrap) would make the improvement claims more convincing.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"I believe the paper has a useful empirical contribution, but the theoretical framing overstates what is proven. The authors should be asked to either strengthen the theory or carefully demote it to a heuristic motivation. I would not reject on the empirical results alone; the main tables are structurally reproducible and the gains are consistent. The improvement-percentage error is easy to fix and should be corrected. With those revisions, the paper could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick read of Mao et al. on diffusion-based recommendation with missing data. The idea is genuinely new: instead of recovering or imputing missing items, they simulate extra missingness with dual-side Thompson sampling—removing high-continuity items from high-stability sequences—and train the diffusion guidance to be insensitive to those edits. Empirically it works: consistent HR/NDCG gains over strong baselines across six datasets (the Table 1 header says five, a typo), ablations support the two sampling criteria, and DTS also helps non-diffusion recommenders. That last generalization test is a nice touch and makes the contribution broader than a single model.\n\nWhere it gets soft is Section 4.3. The 'extrapolation proof' is a Taylor expansion with an unproven coefficient bound. The sentence 'if the missing data process... can align well...' is doing all the load-bearing work; that alignment is exactly what the paper has not established. Real missingness from exposure bias or privacy may remove low-continuity or preference-shifting items, not the safe high-continuity ones DTS targets. In that regime the coefficient in (19) has no reason to be bounded. This is not a minor gap; it is a circular argument. The consistency-regularization bound in (20) is also just the parallelogram inequality, an upper bound, not evidence that the reconstruction loss minimizes the desired distance. The empirical robustness results (RQ4) use synthetic missingness that likely resembles the simulation, so they don't close the gap either.\n\nThat said, the paper's empirical contribution stands on its own. The DTS strategy is well-ablated, the improvements are large on several datasets, and the authors are honest in the conclusion about preference dynamics and missing-data mechanisms exceeding their models. The citation pattern looks fine; the method is built directly on DreamRec, and they acknowledge that lineage. Code is promised in the abstract, which is good.\n\nWho is this for: people working on generative recommenders or robustness to missing interaction data. They will get a useful, seemingly reproducible method and a clear reminder that theoretical 'validation' can be a dressed-up assumption. I would not block the paper on the theory if the claims were softened to intuition or a conjecture; as written, the proof claim overstates what is actually shown.\n\nRecommendation: worth a serious referee. Send it out; ask for a revised theoretical section that either removes the proof claim, proves a meaningful bound under stated conditions, or tests against missingness mechanisms that violate the alignment assumption.","headline":"Novel and well-tested idea for diffusion recommenders, but the theoretical 'proof' is a circular assumption; the empirical results carry the paper.","tokens_in":18697,"tokens_out":2738,"would_cite":false,"duration_ms":27247,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Deliberately deleting extra items from a user's history makes diffusion recommenders tolerate the gaps that are already there, and a Thompson-sampling edit rule keeps the preference trajectory intact.","keywords":["diffusion models","sequential recommendation","missing data","Thompson sampling","consistency regularization","denoising diffusion implicit models","generative recommendation"],"falsifier":"A controlled experiment can test the extrapolation claim directly: generate test-set missingness by a mechanism opposite to DTS, such as deleting low-continuity items or hiding items that were never exposed, and compare TDM with a no-simulation diffusion recommender such as DreamRec. If TDM's advantage shrinks or reverses under such missingness, the claim that consistency on simulated gaps transfers to real gaps is falsified; a positive result would require TDM to keep its gain across random, preference-preserving, and exposure-based missingness.","tokens_in":17629,"feed_emoji":"🎯","tokens_out":7260,"duration_ms":70519,"temperature":0.7,"pith_summary":"Diffusion-based sequential recommenders generate the next item by denoising under guidance from a user's observed history, so missing entries in that history weaken the generated item. This paper argues that recovering the missing entries is impractical because missingness is uncertain in both occurrence and content, and instead proposes TDM, which simulates additional missing data by deliberately removing items from the guidance sequence. The removals are chosen by dual-side Thompson sampling using local continuity and global stability scores, so the user-preference trajectory stays approximately unchanged while the sequence acquires gaps like the ones the model should tolerate. Training the diffusion model to reconstruct the next item from both observed and edited sequences acts as consistency regularization, and a Taylor-expansion argument says that robustness to simulated gaps transfers to real gaps by extrapolation. The paper reports consistent gains over prior sequential and generative recommenders on six datasets.","feed_headline":"Simulating missing data beats recovering it for diffusion recommenders","feed_subtitle":"A dual-side Thompson sampling edit preserves preference drift while making diffusion recommenders stable to real missing items.","key_machinery":"The load-bearing mechanism is dual-side Thompson sampling built from two probability models. Locally, each item $n$ in a sequence gets a continuity score $\\mathrm{con}_n=\\exp(\\mathrm{sim}(\\mathbf e_n,\\mathbf e_{n+1}))/\\sum_{n'}\\exp(\\mathrm{sim}(\\mathbf e_{n'},\\mathbf e_{n'+1}))$; globally, each sequence $k$ gets a stability score $\\mathrm{sta}_k$ obtained by softmax-normalizing the entropy of its continuity scores. These scores parameterize two Thompson-sampling distributions, and thresholds $\\lambda_1,\\lambda_2$ decide whether a sequence is edited and whether an item is replaced by a dummy token. The edited sequence is encoded into guidance $\\mathbf g$, the denoising network predicts the next-item embedding directly, and the reconstruction loss doubles as consistency regularization because observed and edited versions are perturbed pairs. The extrapolation proof applies a first-order Taylor expansion to show $\\|f_\\theta(\\cdot,\\bar{\\mathbf g},\\cdot)-f_\\theta(\\cdot,\\hat{\\mathbf g},\\cdot)\\| \\le C\\|f_\\theta(\\cdot,\\hat{\\mathbf g},\\cdot)-f_\\theta(\\cdot,\\tilde{\\mathbf g},\\cdot)\\|$ whenever the simulated missingness mechanism aligns with the real one.","core_discovery":"The paper's central claim is that the right response to missing interaction history is not to repair it but to make the recommender indifferent to gaps, by injecting simulated missingness during training. TDM's dual-side Thompson sampling computes a continuity score for every adjacent pair (softmax-normalized cosine similarity) and an entropy-based stability score for each whole sequence, then drops high-continuity items from high-stability sequences with probabilities governed by those scores; replacing an item with a dummy token and re-encoding the edited sequence by a Transformer yields the guidance for the diffusion denoiser. Because the reconstruction loss also bounds the difference between predictions made from observed and edited guidance, the model becomes invariant to preference-preserving perturbations, and the paper's extrapolation argument extends that invariance to real missing data. Empirically, TDM improves over the best baselines in hit ratio and NDCG on YooChoose, KuaiRec, Zhihu, Steam, Beauty, and Toys, and applying DTS to non-diffusion recommenders yields gains in most settings.","pith_inferences":["If the alignment assumption holds, a general recipe emerges: for any conditional generative model with partially observed conditioning, train by self-inflicted, distribution-preserving deletion of conditioning tokens rather than by imputation.","Because the entropy-based stability score identifies sequences where preference is volatile, the same signal could be used at serving time to lower the removal threshold or to demand more evidence before recommending, a use the paper does not explore.","The method's success should depend on the quality of the embedding space in which continuity is measured; a testable extension is to pair DTS with content-enriched item embeddings and check whether the gains grow as cosine similarity becomes more semantically meaningful.","The paper's own limitation suggests a natural successor: replace the hand-defined continuity and stability metrics with a learned model of preference evolution while keeping the Thompson-sampling framework, and test whether robustness improves further."],"forward_implications":["DTS is a model-agnostic editing layer: it can be attached to any recommender that encodes a history sequence, and the paper shows gains on GRU4Rec, SASRec, CL4SRec, Caser, AdaRanker, and DiffuASR as well as on diffusion recommenders.","Recovering-based pipelines such as imputation or insertion are not necessary for missing-data robustness; simulating gaps plus consistency regularization is cheaper and avoids the risk of injecting guessed items.","The consistency-regularization view implies that the reconstruction loss itself can be reused as the invariance objective, so TDM adds no auxiliary head or adversarial loss.","Using DDIM acceleration, the inference cost drops dramatically relative to DDPM-based DreamRec while keeping performance, which makes generative sequential recommendation practical at scale."],"supporting_citations":[{"why":"DreamRec: establishes the guided-diffusion sequential recommendation setup—noising the next-item embedding and denoising under history guidance—that TDM inherits and exceeds.","marker":"[49]"},{"why":"DDIM: supplies the non-Markovian denoising diffusion implicit model used to accelerate the reverse process.","marker":"[35]"},{"why":"Thompson sampling tutorial: provides the probability-model formulation of Thompson sampling on which dual-side Thompson sampling is built.","marker":"[33]"},{"why":"Classifier-free guidance: provides the conditional and unconditional training scheme TDM uses with the edited guidance.","marker":"[11]"},{"why":"DDPM: defines the denoising diffusion objective whose reconstruction loss underlies the consistency-regularization bound.","marker":"[10]"},{"why":"PDRec: a recovering-based approach that generates supplement items with diffusion models, serving as the main alternative that TDM outperforms.","marker":"[23]"},{"why":"Consistency regularization: supplies the idea that a model should give similar outputs under perturbed inputs, motivating the observed-versus-edited prediction difference minimization.","marker":"[51]"},{"why":"DiffuRec: an earlier diffusion-based sequential recommender that shapes the guidance-from-history paradigm and motivates the missing-data problem.","marker":"[17]"}],"fun_headline_variants":["Simulate missing data, don't recover it, to harden diffusion recommenders","Training diffusion recommenders on simulated gaps beats patching real ones","Don't recover missing data; simulate it to strengthen diffusion recommenders","Make diffusion recommenders missing-proof by simulating gaps during training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole robustness argument rests on the assumption that real missing data in user histories behaves like the preference-preserving gaps DTS creates, so that the guidance differences from real and simulated missingness have comparable direction and size; if real gaps come from exposure bias, popularity, or other non-preference mechanisms, the extrapolation guarantee does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Simulate missing data, don't recover it, to harden diffusion recommenders","Training diffusion recommenders on simulated gaps beats patching real ones","Don't recover missing data; simulate it to strengthen diffusion recommenders","Make diffusion recommenders missing-proof by simulating gaps during training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00054,"raw_usage":{"total_tokens":2590,"prompt_tokens":949,"completion_tokens":1641,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1575}},"tokens_in":565,"tokens_out":1641,"duration_ms":12422,"temperature":1.0,"reasoning_tokens":1575,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:36:28.765590+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment can test the extrapolation claim directly: generate test-set missingness by a mechanism opposite to DTS, such as deleting low-continuity items or hiding items that were never exposed, and compare TDM with a no-simulation diffusion recommender such as DreamRec. If TDM's advantage shrinks or reverses under such missingness, the claim that consistency on simulated gaps transfers to real gaps is falsified; a positive result would require TDM to keep its gain across random, preference-preserving, and exposure-based missingness.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DreamRec: establishes the guided-diffusion sequential recommendation setup—noising the next-item embedding and denoising under history guidance—that TDM inherits and exceeds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DDIM: supplies the non-Markovian denoising diffusion implicit model used to accelerate the reverse process."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Classifier-free guidance: provides the conditional and unconditional training scheme TDM uses with the edited guidance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PDRec: a recovering-based approach that generates supplement items with diffusion models, serving as the main alternative that TDM outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Consistency regularization: supplies the idea that a model should give similar outputs under perturbed inputs, motivating the observed-versus-edited prediction difference minimization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DiffuRec: an earlier diffusion-based sequential recommender that shapes the guidance-from-history paradigm and motivates the missing-data problem."}],"review_version":1}