{"id":"61f9eab9-df06-45d2-bd52-a37d6fec1599","arxiv_id":"2501.12175","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"IBMRec applies feature-level and graph-level information bottleneck losses to strip irrelevant image/text features from multimedia recommenders, improving Top-20 ranking on three Amazon datasets.","lead":"A new recommender module, IBMRec, uses an information bottleneck objective to keep only the image and text features that predict user interactions, and it cleans the item similarity graph the same way. The authors report accuracy gains over existing multimedia recommenders on three Amazon datasets and show the module can be added to several prior models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GIB rests on unproven equality Eq. 26 and a mismatched HSIC surrogate Eq. 28; without an ablation isolating the HSIC term, the graph-level Information Bottleneck claim is not established.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing point: Eq. 26 has no proof or bound relating graph-level mutual information to node-embedding mutual information, and Eq. 28 replaces it with an HSIC term that is not a valid surrogate. My reading of the manuscript confirms this. The GIB module is the part of the method most directly tied to the claim of 'graph-level Information Bottleneck denoising,' and the derivation is not merely missing a constant or a regularity condition; the asserted equality appears conceptually wrong because the two embeddings share the user-item graph and input features, so their dependence does not measure graph-structure information. The empirical ablation in Table 4 shows a drop when GIB is removed, but that drop could be due to the preference-guided edge masking alone, since the paper does not isolate the HSIC regularization. Therefore the central claim is not established as stated. Because the paper is an engineering contribution with consistent empirical gains, I do not think the correct verdict is outright rejection; the right response is to require the proposed ablation and a corrected derivation, which is what the reader's CONDITIONAL verdict already does. My stress test therefore leaves the verdict unchanged rather than moving it.","tokens_in":20887,"tokens_out":5836,"duration_ms":64874,"concrete_test":"Recompute Table 4 on Clothing, Sports, and Baby using the VLATTICE backbone, retaining the preference-guided edge-masking MLP (Eqs. 23-24) but replacing the GIB HSIC term with (a) zero and (b) a simple L2 consistency regularizer ||Y'-Y||^2. If the NDCG@20 differences between IBMRec and variant (a) are within about 1% relative, the graph-level bottleneck is not responsible for the gains and the IB interpretation of GIB fails. If the differences are larger, the HSIC term is doing real work and the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that IBMRec removes task-irrelevant features via two IB modules. The feature-level module uses a standard HSIC surrogate for I(Z;M) in Eq. 22, but the graph-level module depends on two unjustified moves. Eq. 26 asserts Min I(S';S) = sum_i I(Y';Y), where Y' and Y are GNN embeddings computed from S' and S. This is not an identity or a bound: mutual information between graph structures is not the sum of mutual information between node embeddings, and because both Y' and Y depend on the shared user-item graph A and the same multimedia features M, I(Y';Y) can be large even when S' and S are independent, or small when S' carries structure that the GNN does not preserve. Eq. 28 then replaces the graph-level objective with HSIC(Y';Y); even setting aside the apparent typo (HSIC(S';S') in Eq. 28 versus HSIC(S';S) in Algorithm 1), HSIC on paired node embeddings is not an upper bound or a controlled approximation of I(S';S). Consequently the GIB loss may be doing no information-bottleneck work at all; the reported gains in Tables 3 and 4 could come entirely from the preference-guided edge-masking MLP in Eqs. 23-24. This is load-bearing because the paper's stated novelty is specifically the IB interpretation of both denoising levels, not just the architectural components.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IBMRec, a multimedia recommendation model built around the Information Bottleneck (IB) principle. Two regularizers are added to a LATTICE-style backbone: a feature-level IB module (FIB) that maximizes mutual information between the learned multimedia representation and the rating signal while using HSIC to reduce dependence between the representation and the original pre-trained multimedia features, and a graph-level IB module (GIB) that first re-weights the item-item graph with a preference-guided MLP and then uses an HSIC term to reduce dependence between embeddings obtained from the original and refined graphs. Experiments on Amazon Clothing, Sports, and Baby report consistent accuracy gains over existing baselines, e.g., 14.25%, 8.53%, and 8.96% relative NDCG@20 improvements over MICOR, and improved performance when IBMRec is attached to several backbones.","tokens_in":21201,"tokens_out":7548,"duration_ms":78615,"significance":"The empirical portion is the strongest part of the manuscript: the reported gains are consistent across three datasets, several baselines, and multiple backbones, and the paper includes ablation studies and hyper-parameter sensitivity analyses. If the IB interpretation were fully justified, the two-level denoising framework would be a useful and reasonably novel contribution to multimedia recommendation. However, the central theoretical derivation is not currently established. The graph-level IB objective relies on an unjustified equality between graph-level mutual information and a sum of node-level mutual informations, and the HSIC surrogate is not a controlled approximation of that objective. In addition, the ablations do not isolate the contribution of the HSIC regularizer from the contribution of the preference-guided edge-masking MLP. The paper therefore reports a promising empirical recipe, but the load-bearing claim that it implements Information Bottleneck denoising at both feature and graph levels needs substantial revision.","major_comments":[{"comment":"The derivation of the InfoNCE-style objective contains an unstated expectation condition. In Eq. (18), the step from the sum over negative samples to the factor (N-1) requires E_j[p(r_ai | z_j, a, i) / p(r_ai | a, i)] = 1, but this is not implied by the preceding definitions and is not generally true; p(r_ai | a, i) is defined through f(a,i), not as the marginal of p(r_ai | z_j, a, i) integrated over the batch distribution of z_j. Relatedly, Eq. (19) replaces p(r_ai | a, i) with an average over batch items z_j, which is a heuristic rather than a consequence of the probabilistic model. The final objective Eq. (20) is a familiar InfoNCE lower bound, but the derivation as written does not prove it. Please either state the missing condition, provide a correct derivation, or explicitly present Eq. (20) as a standard contrastive lower bound rather than as a quantity derived from Eqs. (17)-(19).","section":"§4.2.1, Eq. (18)"},{"comment":"The equality Min I(S'; S) = Σ_i I(Y'; Y) is asserted without proof and is not an identity or a bound. Mutual information between two graphs does not decompose into a sum of mutual informations between node embeddings produced by a GNN. As a concrete failure mode, if S' = S then the left-hand side is H(S) (for discrete graphs), while the right-hand side is I(Y; Y) = H(Y); these are generally different. Conversely, if S' and S are independent but Y' and Y are both computed from the same user-item graph A and the same multimedia features M, the two embedding sets can be strongly dependent even when the graph structures are independent. Because Eq. (26) is the step that converts the graph-level IB objective into a computable node-level objective, the graph-level IB claim is not supported without a derivation or a bound relating these quantities.","section":"§4.3.2, Eq. (26)"},{"comment":"The replacement of the graph-level objective by HSIC(Y'; Y) is not justified as an approximation of I(S'; S). HSIC on paired node embeddings measures dependence between embedding sets, not between the graph structures S' and S, and no upper bound or controlled-error statement connects the two. The notation is also internally inconsistent: Eq. (28) writes LGIB = \\hat{HSIC}(S'; S') ≈ \\hat{HSIC}(Y'; Y), while Algorithm 1 line 8 says \\hat{HSIC}(S', S) and Eq. (30) writes \\hat{HSIC}(\\tilde Y'; \\tilde Y). The authors should either prove a relation between the HSIC term and the graph-level IB objective, or explicitly reframe GIB as a heuristic regularizer that is only motivated by, rather than derived from, the Information Bottleneck principle.","section":"§4.3.2, Eq. (28) and Algorithm 1"},{"comment":"The ablation study does not isolate the contribution of the HSIC term in GIB from the contribution of the preference-guided edge-masking MLP. The GIB module in Eqs. (23)-(24) and Eq. (28) contains two distinct components: the construction of S' through the MLP-gated mask, and the HSIC regularization on Y' and Y. The 'IBMRec-w/o GIB' variant removes both components together, so the improvement of the full model over this variant could be entirely due to the edge-masking MLP rather than to the information-bottleneck regularizer. Please add an ablation that keeps the preference-guided masking but removes only the HSIC term, and ideally also a variant that applies the HSIC term to the original graph S instead of the masked graph S'. Without such an ablation, the empirical evidence does not establish that graph-level Information Bottleneck denoising, rather than the architectural masking component, is responsible for the reported gains.","section":"§5.3.2, Table 4"}],"minor_comments":[{"comment":"There are several notation and naming inconsistencies: Eq. (28) uses HSIC(S'; S') while Algorithm 1 uses HSIC(S', S) and Eq. (30) uses a garbled \\hat{HSIC}(\\tilde Y'; \\tilde Y); the baseline name appears as both MICRO and MICOR; SGFP in the related work is SGFD in the experiments. Please unify these.","section":"Throughout"},{"comment":"The symbol Y is used both for the item embedding matrix after user-item graph convolution and for the final fused item representation in Eq. (12); this makes Eq. (27) and Eq. (30) harder to follow. Please use distinct symbols.","section":"§4.1.2, Eq. (12)"},{"comment":"The text says 'As shown in Fig. 6' when discussing the FIB parameters (alpha, sigma^2), but the corresponding panel is Fig. 5; the GIB parameter discussion also refers to Fig. 6, which is likely the intended panel. Please correct the cross-references.","section":"§5.4.2, Figs. 5 and 6"},{"comment":"The evaluation reports means over 10 repeated runs but no standard deviations or significance tests. Given that the reported relative improvements are substantial, adding variance or significance information would strengthen the comparison.","section":"§5.1.2"},{"comment":"The contribution statement claims to be the first to introduce the IB principle to multimedia recommendation. Given existing IB-based recommendation and graph IB works cited in the paper itself (e.g., Refs. [23], [46]), please soften or qualify the novelty claim.","section":"§1, contributions"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a useful empirical finding but currently overclaims the theoretical grounding of the graph-level IB module. I would encourage the editor to require the authors to either provide a correct derivation for Eq. (26) and a justification for the HSIC surrogate, or explicitly downgrade the IB interpretation to a heuristic motivation, plus add the missing ablation that separates the HSIC term from the edge-masking MLP. If the authors cannot supply a bound or a more honest framing, the graph-level IB claim should be removed from the central contributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the empirical core is believable and useful, but the Information Bottleneck branding is doing more work than the math supports. If you read one thing, read Table 3: the method consistently improves four backbones by 4 to 28 percent, and the ablations show both modules contribute. That is a genuinely useful plug-in denoiser for multimedia recommendation.\n\nWhat is new is the combination: feature-level HSIC bottleneck plus graph-level HSIC bottleneck with preference-guided edge masking, applied to the pre-trained feature noise problem in multimedia recommendation. The FIB part is a reasonable repackaging of standard pieces (BPR plus InfoNCE for the max side, HSIC for the min side), and the derivation, while informal, at least points in the right direction. The GIB part is the risk. Eq. 26 asserts that I(S';S) equals the sum of node-level I(Y';Y), with no proof or bound; the next step replaces that with HSIC(Y';Y), which is not a controlled approximation of either graph-level quantity. Eq. 28 even appears to contain a typo. So the graph-level “Information Bottleneck” might be doing nothing beyond the MLP edge mask, and the ablation only removes the whole GIB module, not the HSIC term, so we cannot tell.\n\nThe other soft spots are smaller. The paper claims to be first to introduce IB to multimedia recommendation, but reference [23] already applies IB to recommendation; the qualifier “multimedia” may rescue the claim, but it deserves a sharper comparison. No code is released, and the “10 repeated experiments” are reported without variance. The parameter sensitivity figures are helpful but only for one backbone.\n\nThe citation pattern is not a problem: the core components are honestly traced to prior work, and the self-citations are to legitimate graph-recommendation papers.\n\nWho this is for: people working on denoising multimedia recommendation will find the recipe useful even if they ignore the IB story. It deserves a serious referee because the empirical contribution is solid enough to survive a rewrite. My recommendation: send it out, require a major revision that (1) fixes or removes Eqs. 26–28, (2) adds an ablation isolating the graph-level HSIC term, (3) reports error bars, and (4) releases code.","headline":"Real, plausible gains from a two-level HSIC denoiser, but the Information Bottleneck framing overreaches—especially the graph-level module, whose key equality is asserted, not proven.","tokens_in":21742,"tokens_out":2658,"would_cite":false,"duration_ms":26263,"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":"Applying information bottleneck denoising at both feature and graph levels removes irrelevant multimedia signals and lifts recommendation accuracy.","keywords":["multimedia recommendation","information bottleneck","feature denoising","graph structure learning","Hilbert-Schmidt independence criterion","multimodal recommendation","collaborative filtering"],"falsifier":"On a small dataset, compute a tight variational or exact upper bound of $I(S'; S)$ alongside the HSIC($Y'$, $Y$) surrogate the paper minimizes across training; if the surrogate falls while the true bound stays flat or rises, the graph-level module is not bottlenecking graph information. A simpler experiment: hold the preference-guided edge-masking network fixed and delete only the HSIC regularization; unchanged accuracy would mean the gains come from the masking MLP rather than from Information Bottleneck compression.","tokens_in":1808,"feed_emoji":"✂️","tokens_out":4712,"duration_ms":101437,"temperature":0.7,"pith_summary":"The paper argues that multimedia recommenders are held back by task-irrelevant content that generic pre-trained extractors inject into item representations, such as image backgrounds and redundant text, and that the fix is to compress both features and item-item structure toward what actually predicts user preference. It proposes IBMRec, built on the Information Bottleneck principle, with a feature-level module (FIB) that maximizes the mutual information between multimedia representations and ratings while minimizing dependence on the raw pre-trained features, and a graph-level module (GIB) that rebuilds the item-item graph from preference affinity and then minimizes dependence between the original and refined graphs. On the Amazon Clothing, Sports, and Baby datasets, IBMRec is reported to outperform all baselines, with relative NDCG@20 gains of 14.25%, 8.53%, and 8.96% over the strongest baseline, MICOR, and to improve several existing recommenders when attached to them. The intended upshot is that deliberately discarding information from rich multimedia inputs makes recommenders more accurate, and the Information Bottleneck supplies a task-agnostic way to decide what to discard.","feed_headline":"Trimming irrelevant features lifts multimedia recommendation","feed_subtitle":"Applying information bottleneck at feature and graph levels beats strongest baseline by up to 14.25% NDCG@20.","key_machinery":"The load-bearing machinery is the Information Bottleneck objective restated for recommender data: maximize $I(R; Z)$ while minimizing $I(M; Z)$ for features, and maximize task-relevant structure while minimizing $I(S'; S)$ for the item-item graph. Direct computation is handled in three moves: first, $I(R; U, V, Z)$ is decomposed as $I(R; U, V) + I(R; Z | U, V)$ and each term is replaced by a tractable lower bound (BPR ranking loss and a conditional noise-contrastive loss); second, the intractable minimization of $I(M; Z)$ is approximated by HSIC computed in a reproducing kernel Hilbert space with RBF kernels; third, the graph-level objective relaxes $I(S'; S)$ to node-level $I(Y'; Y)$ between the embeddings produced by the original and refined graphs, and again uses HSIC. The preference-guided edge-masking MLP, which keeps each item-item edge with probability $w_{ij}$ learned from denoised multimedia and collaborative representations, is what injects collaborative information into the graph refinement before the HSIC penalty is applied.","core_discovery":"The central claim is that the Information Bottleneck principle ported to multimedia recommendation removes irrelevant multimedia features and thereby improves ranking accuracy. The paper treats pre-trained multimedia features $M$ as the input, the rating matrix $R$ as the task signal, and the learned multimedia representation $Z$ plus the refined item-item graph $S'$ as the bottleneck variables, optimizing $\\max I(R; Z) - \\alpha I(M; Z)$ at the feature level and $\\max I(R; U, V, S') - \\beta I(S'; S)$ at the graph level. Because direct mutual information estimation is intractable, the paper decomposes $I(R; U, V, Z)$ into lower bounds realized by a BPR ranking loss and a conditional noise-contrastive loss, and replaces the minimization terms with HSIC, a kernel-based statistical independence criterion. The experiments report consistent gains over all baselines on three datasets, and ablations show that removing either IB module degrades performance, while adding more modalities helps IBMRec even though it hurts its VLATTICE backbone. The discovery, on the paper's own terms, is that less information, chosen by an Information Bottleneck objective, can be more signal for the recommendation task.","pith_inferences":["An implication the paper leaves implicit is that the mechanism can be stress-tested with controlled noise: adding synthetic irrelevant dimensions to image or text features should make IBMRec's advantage over non-denoised baselines grow monotonically, and if it does not, the reported gains are not caused by removing irrelevant features.","The paper does not isolate the contribution of the preference-guided masking MLP from the HSIC penalty; an ablation that keeps the MLP but removes only the HSIC terms would tell whether Information Bottleneck compression or the added collaborative edge mask drives the results.","Because the GIB relaxation measures embedding-level dependence rather than graph-level mutual information, a stricter estimator or an exact small-scale computation could reveal whether the method truly minimizes graph information or merely regularizes embeddings.","The reported gains are concentrated on sparse, small datasets; testing on denser or larger-scale multimedia benchmarks would show whether the denoising benefit survives when collaborative signal alone is already strong."],"forward_implications":["Multimedia recommenders that rely on generic pre-trained extractors leave accuracy on the table: compressing the extracted features toward rating-predictive content should improve ranking without retraining the extractor.","Structure-level denoising matters even after feature-level compression, because an item-item graph built from noisy features amplifies the noise through graph convolution; GIB targets that amplification directly.","IBMRec behaves as a plugin: attaching FIB and GIB to backbones such as VBPR, VLightGCN, LATTICE, and VLATTICE is reported to improve each one, with the largest relative gains on the noisiest or weakest settings.","More modalities become useful once noise is removed: the paper reports that combining visual and textual features hurts its VLATTICE backbone but helps IBMRec, suggesting the bottleneck turns modality fusion from a liability into an asset."],"supporting_citations":[{"why":"Supplies the original Information Bottleneck formulation that the paper says it is porting to recommendation.","marker":"[35]"},{"why":"Defines the minimal-sufficient representation objective that FIB and GIB instantiate for recommendation data.","marker":"[36]"},{"why":"Provides the HSIC bottleneck approximation the paper uses to replace intractable mutual information minimization in FIB.","marker":"[26]"},{"why":"Supports using HSIC as an information bottleneck objective for robustness, the basis for the graph-level surrogate.","marker":"[39]"},{"why":"Supplies the collaborative-plus-multimedia preference decomposition and serves as a backbone that IBMRec improves with FIB.","marker":"[16]"},{"why":"Provides the semantic item-item graph construction from modality features that GIB refines and also serves as a backbone.","marker":"[51]"},{"why":"MICOR is the strongest baseline in the comparisons, and the reported NDCG@20 gains are measured against it.","marker":"[52]"},{"why":"Provides the BPR ranking loss used as the practical lower bound for maximizing $I(R; U, V)$ and as a baseline model.","marker":"[31]"}],"fun_headline_variants":["Information Bottleneck clears noise in multimedia recommendations","Less irrelevant media data, better recommendations via IB","Denoising multimedia features with Information Bottleneck boosts ranking","Cutting task-irrelevant media features via Information Bottleneck","Information Bottleneck filters noise for better media recommendations"],"cache_read_input_tokens":23808,"weakest_assumption_plain":"The load-bearing premise is that measuring the statistical dependence between the original and denoised item graphs can be swapped for measuring the dependence between the embeddings those graphs produce; if that swap does not upper-bound the actual graph-level information, the graph-level module is not enforcing an information bottleneck.","fun_headline_variants_meta":{"raw":{"variants":["Information Bottleneck clears noise in multimedia recommendations","Less irrelevant media data, better recommendations via IB","Denoising multimedia features with Information Bottleneck boosts ranking","Cutting task-irrelevant media features via Information Bottleneck","Information Bottleneck filters noise for better media recommendations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001037,"raw_usage":{"total_tokens":4422,"prompt_tokens":1059,"completion_tokens":3363,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":675,"completion_tokens_details":{"reasoning_tokens":3289}},"tokens_in":675,"tokens_out":3363,"duration_ms":24137,"temperature":1.0,"reasoning_tokens":3289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:26:38.581739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small dataset, compute a tight variational or exact upper bound of $I(S'; S)$ alongside the HSIC($Y'$, $Y$) surrogate the paper minimizes across training; if the surrogate falls while the true bound stays flat or rises, the graph-level module is not bottlenecking graph information. A simpler experiment: hold the preference-guided edge-masking network fixed and delete only the HSIC regularization; unchanged accuracy would mean the gains come from the masking MLP rather than from Information Bottleneck compression.","supporting_citations":[{"cited_title":"Tishby and N","cited_arxiv_id":null,"evidence_quote":"Defines the minimal-sufficient representation objective that FIB and GIB instantiate for recommendation data."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the HSIC bottleneck approximation the paper uses to replace intractable mutual information minimization in FIB."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports using HSIC as an information bottleneck objective for robustness, the basis for the graph-level surrogate."},{"cited_title":"He and J","cited_arxiv_id":null,"evidence_quote":"Supplies the collaborative-plus-multimedia preference decomposition and serves as a backbone that IBMRec improves with FIB."},{"cited_title":"Zhang, Y","cited_arxiv_id":null,"evidence_quote":"Provides the semantic item-item graph construction from modality features that GIB refines and also serves as a backbone."},{"cited_title":"Zhang, Y","cited_arxiv_id":null,"evidence_quote":"MICOR is the strongest baseline in the comparisons, and the reported NDCG@20 gains are measured against it."},{"cited_title":"Rendle, C","cited_arxiv_id":null,"evidence_quote":"Provides the BPR ranking loss used as the practical lower bound for maximizing $I(R; U, V)$ and as a baseline model."}],"review_version":1}