{"id":"be432ff4-c7db-4317-984a-95c8a495200e","arxiv_id":"2506.05044","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MACL augments item images and text to create contrastive views for items and sessions, weights them by a learned MLP, and outperforms existing session-based recommenders on three datasets.","lead":"Session-based recommenders predict what an anonymous shopper will click next from a short browsing session. This paper adds multi-modal augmentation, using product images and text to create training views, plus an adaptive contrastive loss that weights informative signals, and reports gains over prior methods on three datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim of consistent superiority is not fully established because the baseline set omits a strong multi-modal SBR method from the authors' own prior work (Ref. [39]); a direct comparison is needed.","rationale":"The reader's conditional verdict is reasonable, and I agree that the paper needs stronger empirical support. However, I identify a different load-bearing concern than the one listed as the reader's weakest assumption. The semantic-preservation assumption is important, but even if some augmentations change item semantics, MACL could still enjoy empirical gains from multi-modal features, as many contrastive methods do with imperfect augmentations. The more direct threat to the central claim of 'consistent superiority over all baselines' is the omission of a strong multi-modal baseline from the authors' own prior work (Ref. [39]). Since MACL's contribution is fundamentally about leveraging image/text features, a comparison against a method that already uses those features is necessary to attribute the gains to the proposed augmentation and adaptive loss rather than to the extra input signals. The reader's rationale does mention this omission as a supporting concern, so my disagreement is partial rather than complete. The proposed test, running Ref. [39] on the same datasets and metrics, would settle whether the omitted baseline changes the headline conclusion. If the test shows that MACL still wins clearly, the concern is resolved and the conditional verdict can be upgraded; if not, the claim must be qualified. Therefore, I recommend keeping the verdict unchanged at CONDITIONAL until this comparison is performed.","tokens_in":16444,"tokens_out":9334,"duration_ms":108723,"concrete_test":"Run the official implementation of Ref. [39] (or reproduce it from the paper) on the same Cellphones, Grocery, and Instacart splits and with the same preprocessing and metrics as Table 2. Report Prec@10/20 and MRR@10/20. If MACL's margins over this baseline match the margins shown in Table 2, the concern is resolved; if the baseline matches or beats MACL, the 'consistent superiority' claim fails and the paper's comparison must be expanded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, stated in Section 6.1, is that MACL achieves consistent superiority over all baselines on all metrics and datasets. The baseline set, however, omits 'Beyond Co-occurrence: Multi-modal Session-based Recommendation' (Reference [39]), which is the authors' own prior work and a directly relevant competitor: it also uses item images and text for session-based recommendation. This omission is load-bearing because MACL's main design elements are the incorporation of image/text features and the use of those features for augmentation. The observed gains over ID-only baselines (SASRec, CL4SRec, etc.) could therefore be attributed to the extra modalities themselves rather than to the proposed multi-modal augmentation or adaptive contrastive loss. The ablation in Section 6.2 compares MACL with MACL_com (dropout/crop/mask/reorder on IDs), but this does not isolate the effect of using image/text features, because MACL_com has neither the multi-modal fusion in Equation (5) nor the multi-modal augmentation. Without a comparison to a strong multi-modal SBR baseline, the superiority claim is not fully supported; if Reference [39] outperforms MACL on these datasets, the headline claim would be false. This is a standard baseline-completeness concern, not a question of internal consistency.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MACL, a multi-modal adaptive contrastive learning framework for session-based recommendation. MACL augments items and sessions using image and text features rather than item IDs, and introduces an adaptive contrastive loss in which per-signal weights α_i and β_i are produced by small MLPs. The method is evaluated on Cellphones, Grocery, and Instacart, with reported improvements over ID-only supervised and self-supervised baselines. The paper claims that MACL achieves consistent superiority over all baselines on all metrics and datasets, and that it specifically improves handling of item-level long-tail and session-level short-session sparsity.","tokens_in":16698,"tokens_out":7074,"duration_ms":82987,"significance":"If the reported results hold, the paper makes a useful contribution by demonstrating that multi-modal features can serve as a source of semantically consistent augmentations for contrastive session-based recommendation, and by proposing a signal-adaptive reweighting of contrastive terms. The release of source code, the ablations isolating item- and session-level augmentation, and the explicit study of long-tail and short-session scenarios are strengths. However, the absence of strong multi-modal baselines and the underspecified and potentially ill-conditioned adaptive loss leave the central claims not fully established.","major_comments":[{"comment":"The baseline set omits 'Beyond Co-occurrence: Multi-modal Session-based Recommendation' (Ref. [39]), a directly relevant multi-modal SBR method from the same group, as well as the related multi-modal method in Ref. [40]. Because MACL's main design elements are the incorporation of image/text features and their use for augmentation, the observed gains over ID-only baselines (e.g., SASRec, CL4SRec) could be attributable to the extra modalities alone. The ablation MACL_com in Section 6.2 does not isolate this effect, since it removes both the multi-modal fusion of Eq. (5) and the multi-modal augmentation. Please add these baselines, or explicitly justify their exclusion and provide a variant that uses ID-based augmentation with the same fusion network to separate the modality effect from the augmentation effect.","section":"Section 5.4, Table 2"},{"comment":"The contrastive loss is not fully specified. It is written as -Σ α_i <e_i,e_i^+>/Σ_k <e_i,e_{i,k}^->, with no logarithm, no temperature, and no explicit normalization; Section 4.3 refers to 'cosine distance' as the similarity function, which is ambiguous between distance and similarity. More importantly, α_i is the raw output of an MLP with no constraints or regularization, so the objective is not bounded below: scaling α_i can arbitrarily reduce the loss, and the MLP could minimize the loss without learning meaningful signal utility. The claim that the network 'evaluates the contribution of each signal' therefore needs a constraint discussion (e.g., non-negativity, normalization, or a bounded range) and a clear statement of how gradients flow through α_i.","section":"Eq. (14) (and Eq. (1), Eq. (16))"},{"comment":"The paper states that MACL's improvements over the best baseline are significant by a t-test at p<0.01, but no standard deviations, confidence intervals, or number of independent runs are provided. Without these, the significance claim and the 'consistent superiority' assertion in Section 6.1 cannot be verified. Please report mean±std over multiple seeds and describe the exact t-test procedure used.","section":"Table 2"},{"comment":"For session-level augmentation, the positive session embedding s^+ is computed from single-modality embeddings (imgEmb(·) or txtEmb(·)) of the augmented items, whereas the anchor session embedding s uses the fused multi-modal item embeddings from Eq. (5). The same asymmetry appears at item level between Eq. (12) and Eq. (5). The paper does not explain why the positive view is uni-modal while the anchor is multi-modal; this makes the claimed 'semantic consistency' of the augmented views with the anchor less direct and may cause the contrastive loss to align different representation spaces. Please clarify the design or use the fused representation for the augmented views as well.","section":"Section 4.2.2, Eq. (13)"},{"comment":"The paper claims that the selected augmentations 'preserve an item's original semantics' and 'guarantee semantic consistency', but this is asserted without validation. Operations such as Gaussian Noise, Max Pooling, Deletion, Substitution, and Insertion can plausibly change an item's meaning (for example, deleting a color or brand word from the text). Please provide a quantitative or qualitative validation of semantic preservation (e.g., whether the original item can be retrieved from the augmented view) or temper the guarantee.","section":"Section 4.2.1"}],"minor_comments":[{"comment":"The caption says 'over four datasets' but only three datasets are reported in the table; please correct this.","section":"Table 2 caption"},{"comment":"The probability y_i is defined as softmax(e_i s), but the loss L_rec is written as binary cross-entropy over items. Please clarify whether a sigmoid and negative sampling are intended, or use the standard cross-entropy with softmax.","section":"Eqs. (10)-(11)"},{"comment":"The sub-figures in Figure 5, Figure 6, and Figure 7 are very small and the text and numbers are nearly illegible in the provided PDF; please enlarge them and increase the font size.","section":"Figures 5-7"},{"comment":"There are several typos and grammatical issues, e.g., 'duo to' should be 'due to' in Section 2.1, and 'scratched' should be 'scraped' in Section 5.2. A careful proofread is needed.","section":"Throughout"},{"comment":"The runtime comparison reports only MACL and MACL-adp; to put the efficiency claim in context, please also report the training/inference time of the most competitive baselines.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The most pressing issue is the omission of the authors' own prior multi-modal SBR work (Ref. [39], and possibly Ref. [40]) from the baseline table. Since the novelty claim centers on multi-modal augmentation, a reader could plausibly attribute the gains to simply using image and text features. I would ask the editor to require a direct comparison in the revision. The adaptive loss in Eq. (14) also needs to be specified with constraints and normalization before the method can be considered sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a solid within-subfield contribution. It identifies three genuine weaknesses in contrastive learning for session-based recommendation: item-level sparsity, semantic consistency of augmentation, and uniform weighting of contrastive signals. The proposed MACL uses item images and text to build item- and session-level augmentations, and an MLP-weighted adaptive contrastive loss. The ablations support both components, and the long-tail and short-session analyses are thoughtful. If the results hold, the MRR gains are substantial, especially on Cellphones.\n\nThe main problem is the baseline set. The paper cites the authors' own prior work, 'Beyond Co-occurrence: Multi-modal Session-based Recommendation' (Ref [39]), which also uses image and text features for SBR, but it is not included in Table 2. The baselines are almost all ID-only, so the consistent superiority claim could be explained by adding modalities rather than by the proposed augmentation or adaptive loss. The ablation against MACL_com does not isolate this, because MACL_com removes both the multi-modal fusion (Eq. 5) and the multi-modal augmentation. The authors need to add a strong multi-modal baseline, at minimum their own, before the headline claim is supportable.\n\nSecondary issues: no standard deviations or confidence intervals are reported despite a t-test significance claim; the contrastive loss notation in Eq. (1) and (14) is underspecified; and the semantic-preservation of augmentations like horizontal flipping product images or word substitution in descriptions is asserted but not validated. These are fixable and less severe than the missing baseline, but they should be addressed.\n\nOverall, this is a legitimate method paper with a clear contribution and sensible experiments, but the empirical case as reported is incomplete. It deserves a serious referee, and I would read a revised version with interest. If you work on side information or contrastive learning for SBR, it is worth citing, but I would wait before treating its quantitative claims as established.","headline":"Useful method paper with a clear contribution, but the missing multi-modal baseline from the authors' own prior work leaves the superiority claim unproven.","tokens_in":17266,"tokens_out":2436,"would_cite":true,"duration_ms":27339,"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":"Session-based recommendation improves when contrastive learning augments item images and text rather than scrambling item IDs, and when each contrastive signal is weighted by its learned utility, as shown by the MACL framework on three…","keywords":["session-based recommendation","contrastive learning","multi-modal augmentation","adaptive contrastive loss","data sparsity","long-tail items","short sessions","self-supervised learning"],"falsifier":"Hold out 1,000 items and have human annotators flag each augmentation output that changes the product's semantic category; then train a version of MACL that discards those flagged positive pairs. If performance does not drop, the semantic-preservation assumption is not doing the work.","tokens_in":16210,"feed_emoji":"🛒","tokens_out":10187,"duration_ms":105154,"temperature":0.7,"pith_summary":"Session-based recommendation is the task of predicting the next item an anonymous user will interact with given only a short history of clicks. This paper argues that contrastive learning, a self-supervised method that brings augmented views of the same data closer and separates unrelated data, has been applied to this task in a limited way: it augments sessions by scrambling item IDs, ignores the sparsity of long-tail items, and treats every positive–negative comparison as equally useful. To fix this, the paper proposes MACL, which augments item images and texts with computer-vision and NLP operations to create semantically consistent views at both the item and session level, and uses an adaptive contrastive loss that learns a weight for each signal set. On Cellphones, Grocery, and Instacart, MACL reports consistent gains over all baselines, with the largest improvements on mean reciprocal rank (up to 25.8% over the best baseline). The paper's insight is that self-supervised learning for recommendations should be built on item semantics, not on opaque identifiers.","feed_headline":"Multi-modal views beat ID-based contrastive session models","feed_subtitle":"MACL augments images and text and adaptively weighs contrastive pairs, lifting MRR by up to 25.8%.","key_machinery":"The key mechanism is a multi-modal augmentation pool A, a set of nine image and text operations that the paper asserts preserve item semantics; applying one of these operations to an item's image or description yields a positive view while applying the same operation to other items yields negatives, and applying a single operation uniformly across a session yields a positive session. The second mechanism is the adaptive contrastive loss: instead of summing unweighted log-ratios as in standard contrastive losses, MACL multiplies each term by a scalar alpha_i (for items) or beta_i (for sessions) produced by an MLP on the concatenated anchor, positive, and mean negative embeddings. The learned weights are intended to emphasize informative signal sets and suppress uninformative ones during self-supervised training.","core_discovery":"The paper's central claim is that the utility of a contrastive view in session-based recommendation is determined by whether the augmentation preserves the item's meaning, and that both item-level and session-level sparsity can be mitigated by augmenting an item's image and text rather than its ID. MACL maintains a pool A of nine operations (Hflip, Cropping, Gaussian Noise, Gaussian Blur, Max Pooling for images; Swap, Deletion, Substitution, Insertion for text). For each anchor item, one operation is randomly chosen and applied to its image or text to create the positive view; the same operation is applied to other items to form negatives. For a session, the chosen operation is applied to every item in the session to create a positive session, and other sessions in the batch are transformed similarly as negatives. The item-level and session-level contrastive losses are then each reweighted by scalar weights produced by a two-layer MLP from the anchor, positive, and mean negative embeddings. The paper demonstrates that this design outperforms ten baselines on three datasets and that item-level contrastive learning alone yields larger gains than session-level contrastive learning, suggesting that long-tail item sparsity is the more binding constraint.","pith_inferences":["The paper asserts but does not verify semantic preservation for each of the nine augmentations; measuring the semantic drift of positive pairs and correlating it with the learned adaptive weights would directly test the stated mechanism.","The adaptive weight is computed from anchor, positive, and mean negative embeddings; this could be a form of hardness weighting, and a simpler heuristic (e.g., weighting by cosine similarity between anchor and positive minus anchor and mean negative) might achieve comparable performance without training an MLP.","Since the encoders are frozen, the reported gains may partly reflect the quality of pre-trained image and text features rather than the augmentation logic; a test that unfreezes the encoders or uses randomly initialized encoders would separate these effects.","In settings without rich multi-modal features, one could approximate the method by treating embeddings from a co-occurrence-based model as a pseudo-modality and applying dropout-style augmentations to them, but the semantic-consistency advantage would likely weaken."],"forward_implications":["If MACL is correct, contrastive learning for session-based recommendation should move away from item-ID perturbations and toward augmenting item features that carry meaning, such as images and text.","Because item-level contrastive learning produced larger gains than session-level in the paper's ablations, future methods should explicitly target long-tail item sparsity rather than focusing only on short-session augmentation.","The adaptive contrastive loss is a drop-in modification that could be applied to other self-supervised learning settings where the usefulness of positive–negative pairs varies, such as graph or multi-modal representation learning.","The reported MRR improvements (up to 25.8%) imply that the target item appears much higher in the recommendation list, which is the metric that matters most for user-facing ranking.","The dependency on pre-trained image and text encoders means that MACL's approach is most actionable in domains where such features already exist or can be cheaply obtained."],"supporting_citations":[{"why":"Supplies the contrastive learning formulation and the augmentation assumption that MACL revises with semantically consistent multi-modal views.","marker":"[2]"},{"why":"Provides the SASRec sequence encoder used as the backbone for session embedding and augmentation.","marker":"[10]"},{"why":"Provides the GoogLeNet image encoder that embeds item images and image-based augmented views.","marker":"[23]"},{"why":"Provides the BERT text encoder that embeds item descriptions and text-based augmented views.","marker":"[5]"},{"why":"A representative ID-based contrastive baseline whose Crop, Mask, and Reorder augmentations MACL argues are semantically inconsistent and against which MACL is compared.","marker":"[30]"},{"why":"Example of substitute/insert ID-based augmentation that relies on co-occurrence; cited as a source of noisy or semantically inconsistent views.","marker":"[17]"},{"why":"Provides the Dropout and Retrieval augmentations used by a strong baseline (DuoRec) and represents the family of methods that MACL claims fail to distinguish signal utility.","marker":"[19]"}],"fun_headline_variants":["Adaptive multi-modal contrast beats ID tricks in session rec","Multi-modal views fix contrastive session gaps","MACL: meaning-preserving views for session recommendations","Item-level sparsity yields to multi-modal contrastive views","Adaptive loss weighs contrastive pairs in session rec"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the nine augmentation operations preserve the semantic meaning of an item, so an augmented image or text still describes the same product; if a horizontal flip, crop, blur, or word substitution alters what the item actually is, the contrastive objective aligns views that no longer share the same intent.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive multi-modal contrast beats ID tricks in session rec","Multi-modal views fix contrastive session gaps","MACL: meaning-preserving views for session recommendations","Item-level sparsity yields to multi-modal contrastive views","Adaptive loss weighs contrastive pairs in session rec"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1239,"prompt_tokens":968,"completion_tokens":271,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":195}},"tokens_in":584,"tokens_out":271,"duration_ms":3487,"temperature":1.0,"reasoning_tokens":195,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:26:34.301721+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out 1,000 items and have human annotators flag each augmentation output that changes the product's semantic category; then train a version of MACL that discards those flagged positive pairs. If performance does not drop, the semantic-preservation assumption is not doing the work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the contrastive learning formulation and the augmentation assumption that MACL revises with semantically consistent multi-modal views."},{"cited_title":"Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabi- novich","cited_arxiv_id":null,"evidence_quote":"Provides the GoogLeNet image encoder that embeds item images and image-based augmented views."},{"cited_title":"InThe Thirty-Third AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"A representative ID-based contrastive baseline whose Crop, Mask, and Reorder augmentations MACL argues are semantically inconsistent and against which MACL is compared."},{"cited_title":"Yu, Julian J","cited_arxiv_id":null,"evidence_quote":"Example of substitute/insert ID-based augmentation that relies on co-occurrence; cited as a source of noisy or semantically inconsistent views."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Dropout and Retrieval augmentations used by a strong baseline (DuoRec) and represents the family of methods that MACL claims fail to distinguish signal utility."}],"review_version":1}