{"id":"94648a03-4ecb-4963-a4db-fcedc03a9677","arxiv_id":"2504.16077","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"InDiRec generates intent-aligned augmented views with a conditional diffusion model and uses them in contrastive learning, outperforming prior sequential recommenders on five public datasets.","lead":"This paper proposes InDiRec, a sequential recommendation model that uses K-means clustering to infer a user's purchasing intent and a diffusion model to generate intent-matched training views for contrastive learning. It reports large gains over strong baselines on five public datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The intent clusters in Eq. (9) are never validated as recovering latent purchasing intents; if they instead encode sequence length, popularity, or the target sequence itself, the guidance signal is not intent and the central mechanism is unsupported.","rationale":"The reader's conditional verdict is appropriate, but the most consequential uncertainty should be made explicit. The empirical numbers in Table 2 could be reproducible and still not demonstrate the paper's mechanism. The genuine novelty claim is not that diffusion helps contrastive sequential recommendation—that already exists in CaDiRec—but that intent guidance is what makes the generated views better. If the K-means clusters do not capture intent, then the guidance signal is arbitrary, and the observed gains may come from the stochastic denoising process alone or from self-conditioning. The proposed permutation test isolates the information content of the cluster assignment; it is feasible because the code is released. I therefore keep the verdict conditional rather than accepting at face value, but I do not reject: the method may still be an effective diffusion-based augmentation, and the empirical comparison could remain valid. The added condition should be that the authors demonstrate cluster validity, for example by the permutation test or by measuring cluster purity against item category metadata. This extends, rather than replaces, the reader's conditions about error bars and hyperparameter tuning.","tokens_in":17340,"tokens_out":6815,"duration_ms":69956,"concrete_test":"Create an InDiRec variant in which the intent cluster assignments used for guidance are permuted at random: for each target sequence, draw the guidance sequence s_e from a randomly chosen cluster, or from a random sequence not sharing the target's cluster, while leaving all other components including the diffusion and contrastive losses identical. Evaluate HR@20 and ND@20 on all five datasets. If the permuted-guidance variant matches InDiRec within run-to-run variance, the reported gains do not depend on genuine intent recovery; if it degrades substantially, the intent assumption is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the K-means clusters in Eq. (9) correspond to latent purchasing intents and that any sequence retrieved from the same cluster shares the target's intent (Eq. (10), Section 3.2.2). This premise is never validated. Sequence representations from the dynamic prefix segmentation D(·) may cluster by length, position, or popularity rather than by intent; in that case s_e carries no intent information and the generated views are not semantically more consistent than random augmentations. Two details make the failure mode concrete. First, the intent prototypes are recomputed from the current encoder every epoch and are not checked for stability or semantic purity. Second, Algorithm 1 line 6 samples RandomSample(c) from the target's own cluster without excluding the target sequence itself, so the guidance signal can be the target's own representation; the diffusion model then degenerates to a denoising autoencoder conditioned on the input, and the contrastive pair is an almost-trivial reconstruction rather than an intent-based augmentation. The t-SNE visualization in Section 4.6 cannot settle this because contrastive training itself pulls same-cluster representations together regardless of whether clusters are semantically meaningful.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InDiRec, a sequential recommendation model that replaces random contrastive augmentation with intent-guided diffusion. It segments training sequences with dynamic incremental prefix segmentation (Eq. 7), encodes the subsequences with a Transformer, and runs K-means on the resulting representations to obtain intent prototypes (Eq. 9). For a target sequence, the nearest prototype is queried (Eq. 10), a same-cluster sequence is sampled to form the guidance signal s_e, and a conditional diffusion model generates an augmented embedding view (Eqs. 12-19). The view is used as a positive pair in a contrastive loss (Eq. 20) that is jointly optimized with the next-item cross-entropy loss (Eqs. 21-23). Experiments on Beauty, Sports, Toys, Video, and ML-1M report consistent gains over general, contrastive, and diffusion baselines, plus ablations, robustness tests, and t-SNE visualizations.","tokens_in":17588,"tokens_out":7871,"duration_ms":78492,"significance":"The paper addresses a real weakness of stochastic augmentation in contrastive sequential recommendation and, if the intent mechanism is sound, offers a reusable design. The submission is strong on empirical breadth: five public datasets, released code, component ablations, robustness to sparsity and noise, and paired t-tests. The gains over baselines are substantial and consistent. However, the central mechanism is not directly validated: the K-means clusters in Eq. (9) are called intent prototypes without evidence that they capture purchasing intents, and the same-cluster sampling can collapse to self-conditioning. The contribution is therefore conditional on additional cluster-level validation and on fixing the sampling degeneracy.","major_comments":[{"comment":"The intent-guided signal can be the target sequence itself. Because D(·) in Eq. (7) includes the full sequence among the subsequences for sequences of length at most n, the target training sequence is a member of the pool that K-means clusters, and RandomSample(c_e) is not constrained to exclude the target. If the target is selected, s_e is approximately the target's own representation h_e, so the conditional diffusion in Eq. (12) is conditioned on the input, the diffusion loss in Eq. (17) approaches a self-reconstruction objective, and the contrastive pair in Eq. (20) is a near-trivial reconstruction rather than an intent-based augmentation. Please exclude the target sequence (and its duplicate subsequences) from RandomSample(c_e), report how often this exclusion changes the sampled guidance, and rerun the main comparisons under this exclusion.","section":"Section 3.2.2, Algorithm 1 line 6"},{"comment":"The K-means clusters are called intent prototypes but are never validated as recovering latent purchasing intents. They may instead encode sequence length, item popularity, position, or artifacts of the current encoder, in which case the guidance signal s_e carries no genuine intent information and the central mechanism is unsupported. The t-SNE visualization in Section 4.6 cannot settle this question because the contrastive loss in Eq. (20) actively pulls same-cluster representations together, so clustered t-SNE is partly a consequence of the loss. Please add cluster-level diagnostics, such as category purity of cluster members, intra-cluster item overlap, cluster stability across epochs, and a control experiment that uses randomly assigned or length-stratified clusters for guidance while keeping everything else fixed.","section":"Section 3.1.4, Eq. (9)"},{"comment":"The definition of Ldiff is ambiguous. Eq. (17) defines a per-step loss at time t, but Section 3.5 states that Ldiff is the cumulative loss across all T sampling steps, and Algorithm 1 line 9 says to calculate Ldiff via Eq. (17) after one T-step noising pass. Please state explicitly whether Ldiff is a sum over t = 1..T, an expectation over randomly sampled t, or a single-step loss, and confirm that the released code implements the stated objective; this determines the gradient scale and the role of the weight λ.","section":"Section 3.5 and Eq. (17)"}],"minor_comments":[{"comment":"The abstract says the method generates item sequences, but the diffusion process in Eqs. (11)-(19) operates on the sequence embedding e0, not on discrete item sequences; please rephrase to say embedding-level augmented views.","section":"Abstract and Section 1"},{"comment":"The variance term is written as (1 - \\bar\\alpha)I; it should be (1 - \\bar\\alpha_t)I.","section":"Eq. (3)"},{"comment":"There is an unbalanced parenthesis in \\|e0 - f_theta(e_t, s_e, t))\\|^2; the extra closing parenthesis should be removed.","section":"Eq. (17)"},{"comment":"The caption reports paired t-tests with p < 0.05, but no actual p-values or standard deviations are given; please report mean and standard deviation over multiple seeds, or at least the exact p-values for the headline comparisons.","section":"Table 2"},{"comment":"The per-dataset hyperparameter choices reported for K, T, ω, λ, γ, and dropout should be summarized in a single table, and the text should state whether the same validation procedure was applied to all baselines before the test results in Table 2 were obtained.","section":"Section 4.5"},{"comment":"Please clarify the computational cost of the training loop: lines 7-10 appear to run a T-step forward noising pass, a T-step denoising pass, and a separate T-step sampling pass in every batch; the implementation details should state the actual number of diffusion passes per batch and whether the reported runtime reflects this.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a camera-ready SIGIR 2025 paper. If it is under journal review, the revisions requested above should be addressed in the submitted version. I found no evidence of fabrication; the main risks are omitted validation of the intent clusters and a sampling degeneracy that can trivialize the diffusion guidance. The released code will make the requested checks straightforward."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: I agree with the reader's conditional verdict, with one important nuance. The stress-test concern is real but not disqualifying. The paper's empirical core is strong, but the intent mechanism is under-validated.\n\nWhat's new and good: the combination of ICLRec-style intent clustering with conditional diffusion to generate contrastive views is genuinely new, and the design is sensible. They train the diffusion during training and keep inference fast, which is a practical plus. The experiments are broad: five public datasets, full-rank evaluation, ablations, robustness to noise and cold-start, and the code is released. The consistent improvements over strong baselines, including CaDiRec and DuoRec, are hard to explain away by randomness alone.\n\nThe soft spots. First, the intent prototypes are never validated. The t-SNE in Figure 6 is post-hoc and, as the stress-test note says, contrastive training itself pulls same-cluster representations together, so it cannot prove semantic purity. But the ablation w/o IGS shows a clear drop, which suggests the guidance signal carries something useful—so the mechanism is plausible, just not proven. Second, Algorithm 1 line 6 samples randomly from the entire cluster without excluding the target sequence itself. If it draws the target, the diffusion model degenerates into a denoising autoencoder conditioned on the input, and the contrastive pair is trivial. That is a real flaw in the method description, even if the average behavior over the cluster is not degenerate. Third, the hyperparameter p is set to 0.1 in implementation but never explained; that is a minor but obvious omission. Fourth, Table 2 has no error bars or confidence intervals; the paired t-test is mentioned but not reported numerically. Fifth, the abstract says the method 'generate[s] item sequences' when it actually generates embeddings; that overclaim should be fixed.\n\nNone of these are fatal. The gains are large and consistent, the ablations support each component, and the identified issues are addressable in revision. The weakest part is the claim that the clusters correspond to latent purchasing intents; a serious referee should push for cluster validation and the self-sampling exclusion.\n\nWho this is for: researchers working on contrastive learning or diffusion augmentation for sequential recommendation will find it useful, both as a baseline and as a recipe. I would send it to review without hesitation. Recommended action: accept with minor-to-major revisions, focusing on the intent validation and the sampling issue.","headline":"InDiRec is a solid empirical paper with a plausible but under-validated intent mechanism; the gains look real, but the 'intent' story needs stronger evidence before the mechanism is taken at face value.","tokens_in":18117,"tokens_out":1575,"would_cite":true,"duration_ms":17220,"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":"A diffusion model that generates intent-preserving augmented views lifts next-item recommendation across five datasets.","keywords":["Sequential Recommendation","Contrastive Learning","Diffusion Model","Intent Representation Learning","Data Augmentation","Next-Item Prediction","K-means Clustering","Robustness"],"falsifier":"Take a dataset with explicit per-interaction intent labels, run InDiRec's clustering, and check whether sequences assigned to the same intent prototype share the labelled intent significantly more often than chance; a negative result, or a result no better than random sequence assignment, would show that the guidance signal $s_e$ is not capturing genuine intent. Alternatively, substitute a same-cluster sequence with a random other sequence as $s_e$ while holding everything else fixed; if HR@20 does not drop, the intent guidance is not what drives the reported gains.","tokens_in":17121,"feed_emoji":"🎯","tokens_out":10769,"duration_ms":89991,"temperature":0.7,"pith_summary":"Sequential recommenders predict the next item from a user's history, and contrastive learning is used to make them robust to sparse data; the catch is that the augmented views are usually made by random deletion, masking, or replacement, which can destroy the latent intent behind a purchase. This paper proposes InDiRec, which first clusters sequence representations into K-means intent prototypes, then uses a sequence drawn from the target's prototype as a conditioning signal for a diffusion model to generate a positive view intended to share the same purchasing intent. That view is paired with the original sequence in a contrastive loss, so the encoder is trained to treat intent-aligned views as consistent. The paper reports that on Beauty, Sports, Toys, Video, and ML-1M, InDiRec beats every compared baseline on HR and NDCG at 5 and 20, with average relative gains of 13.17% (HR) and 20.68% (NDCG), and keeps the advantage under noise injection and short interaction histories. It matters because replacing stochastic augmentation with semantically guided generation directly addresses a known failure mode in contrastive sequential recommendation.","feed_headline":"Intent-aware diffusion beats random augmentation in recommendations","feed_subtitle":"InDiRec clusters sequences by intent, then generates matching views; gains hold even with 20% noise and short histories.","key_machinery":"The load-bearing object is the intent-aware conditional diffusion process over sequence embeddings. After splitting training sequences into prefix-like subsequences and encoding them with a Transformer sequence encoder, K-means clustering produces $K$ intent prototypes; a query function assigns each target representation to its nearest prototype, and the encoder output of another sequence in that cluster becomes the guidance signal $s_e$. An MLP predicts the clean representation under classifier-free guidance, and the denoised sample is used as the positive view in the contrastive loss. This machinery replaces random data augmentation: instead of deleting or cropping items blindly, the model samples from a learned conditional distribution anchored to a prototype that contains sequences of the same intent.","core_discovery":"The central claim is that contrastive sequential recommendation improves when the positive augmented views are generated from the target sequence's own intent distribution rather than by random perturbation. InDiRec operationalizes this by training a conditional DDPM on sequence embeddings: the forward process adds noise to the target embedding $e_0$, and the reverse process denoises it under a guidance signal $s_e = Encoder(sequence)$ taken from a randomly selected training sequence whose K-means prototype matches the target. The generated view $\\hat{e}_0$ is encoded and pulled toward the original sequence by a contrastive loss, while the diffusion loss and a cross-entropy next-item loss are trained jointly. Evaluated by ranking all items without negative sampling, the model reports the best HR@5, HR@20, NDCG@5, and NDCG@20 on all five datasets, and ablation studies show that removing either the intent-guided signal, the diffusion loss, or the contrastive loss lowers performance.","pith_inferences":["The method's dependence on K-means granularity is hidden in K: with per-dataset tuning from 32 to 1024 prototypes, part of the gain may come from cluster-count selection rather than from the diffusion mechanism itself.","Replacing the randomly sampled same-cluster sequence $s_e$ with the cluster centroid or a learned prototype embedding would separate the effect of the prototype signal from the effect of the particular sampled example.","Because the guidance signal is just another sequence's encoding, the framework can sit on top of any sequence encoder, so it is a drop-in upgrade for existing contrastive SR models rather than a new architecture.","A direct intent-preservation test on labeled data, such as gift purchases versus self-use purchases, would quantify whether the generated views actually keep intent and would turn the paper's central assumption into a measurable statistic."],"forward_implications":["Intent-guided view generation should apply to any contrastive sequential recommender that currently relies on stochastic augmentation, since InDiRec reports consistent gains over the CL4SRec, DuoRec, and MCLRec baselines.","On the shortest interaction histories (five items), InDiRec still outperforms the strongest baselines, so the benefit is largest exactly where data sparsity is worst.","With 20% random noise inserted at test time, InDiRec degrades less than SASRec, DuoRec, MCLRec, DiffuRec, and CaDiRec, indicating that the intent-conditioned views stay semantically consistent under input corruption.","Because diffusion sampling happens only during training, the model's prediction-time cost matches the underlying SASRec encoder with no extra sampling at inference."],"supporting_citations":[{"why":"SASRec supplies the Transformer sequence encoder and backbone that InDiRec builds on for sequence representations and prediction.","marker":"[13]"},{"why":"DDPM provides the forward-noising and reverse-denoising machinery and the simplified MSE loss that intent-aware diffusion adapts.","marker":"[10]"},{"why":"Classifier-free guidance is the scheme used to strengthen or weaken the intent signal during augmented view generation.","marker":"[11]"},{"why":"ICLRec is the intent-clustering contrastive baseline whose negative-sampling strategy, excluding examples from the same intent cluster, InDiRec borrows.","marker":"[3]"},{"why":"CaDiRec is the nearest conditional-diffusion contrastive baseline that InDiRec compares against and extends.","marker":"[4]"},{"why":"Full-item ranking without negative sampling is the evaluation protocol that produces the reported HR and NDCG numbers.","marker":"[16]"},{"why":"DiffuRec is the diffusion-based sequential recommendation baseline that motivates learning user intent distributions.","marker":"[19]"},{"why":"DuoRec is the contrastive baseline addressing representation degeneration that motivates intent-level contrastive learning.","marker":"[25]"},{"why":"Dynamic incremental prefix segmentation supplies the subsequence construction used to create the sequences that are clustered into intents.","marker":"[30]"},{"why":"GPU-accelerated K-means clustering is used to build the intent prototypes that anchor the guidance signal.","marker":"[12]"}],"fun_headline_variants":["Intent-guided diffusion creates better views for recommendation","Diffusion with intent clustering boosts sequential recs","InDiRec: intent-aware views beat random augmentation","Generating intent-aligned sequences improves contrastive rec","Noise-free augmentation via intent diffusion in recsys"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that K-means clusters of learned sequence representations correspond to real purchasing intents, so any other sequence pulled from the same cluster genuinely shares the target sequence's intent; if the clusters are arbitrary, the diffusion guidance signal carries no intent information and the generated views are no better than random augmentations.","fun_headline_variants_meta":{"raw":{"variants":["Intent-guided diffusion creates better views for recommendation","Diffusion with intent clustering boosts sequential recs","InDiRec: intent-aware views beat random augmentation","Generating intent-aligned sequences improves contrastive rec","Noise-free augmentation via intent diffusion in recsys"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1276,"prompt_tokens":996,"completion_tokens":280,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":207}},"tokens_in":612,"tokens_out":280,"duration_ms":3025,"temperature":1.0,"reasoning_tokens":207,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:10:47.453865+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset with explicit per-interaction intent labels, run InDiRec's clustering, and check whether sequences assigned to the same intent prototype share the labelled intent significantly more often than chance; a negative result, or a result no better than random sequence assignment, would show that the guidance signal $s_e$ is not capturing genuine intent. Alternatively, substitute a same-cluster sequence with a random other sequence as $s_e$ while holding everything else fixed; if HR@20 does not drop, the intent guidance is not what drives the reported gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SASRec supplies the Transformer sequence encoder and backbone that InDiRec builds on for sequence representations and prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DDPM provides the forward-noising and reverse-denoising machinery and the simplified MSE loss that intent-aware diffusion adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ICLRec is the intent-clustering contrastive baseline whose negative-sampling strategy, excluding examples from the same intent cluster, InDiRec borrows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Full-item ranking without negative sampling is the evaluation protocol that produces the reported HR and NDCG numbers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DiffuRec is the diffusion-based sequential recommendation baseline that motivates learning user intent distributions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DuoRec is the contrastive baseline addressing representation degeneration that motivates intent-level contrastive learning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Dynamic incremental prefix segmentation supplies the subsequence construction used to create the sequences that are clustered into intents."}],"review_version":1}