{"id":"6bea0f56-1c45-4ebc-86bb-b83ce73a2b9d","arxiv_id":"2502.05523","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Domain-conditioned generation of per-user sequence item and candidate query representations improves multi-domain target-aware attention in recommenders.","lead":"A ByteDance team proposes Adaptive Domain Scaling (ADS), a plug-in module that reshapes user behavior sequences and candidate items according to the current business domain before attention scoring. In offline tests on Taobao and two Douyin datasets and in online A/B tests, ADS improved ranking accuracy and revenue metrics by roughly 0.3% to 1.0%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ADS's claimed gain is driven by E_D, but E_D's 'implicit-domain-indicator' statistical features are never defined; if they are label-adjacent user statistics, the baseline comparison is confounded, and if they are coarse, the per-user personalization claim collapses.","rationale":"The reader's weakest assumption correctly identified E_D as the linchpin. I agree that the paper's failure to specify the 'implicit-domain-indicator' statistical features is the most load-bearing concern: every adaptive component of ADS is conditioned on E_D, so the source of the reported gains is unverifiable without knowing what E_D contains. The concern is not merely about reproducibility; it can change the interpretation of Table 2. If E_D includes user-level statistical features that are absent from the baselines, then the comparison is confounded because the baselines lack the same input signal. If E_D is only the explicit domain indicator, then PSRG is a per-domain affine transformation applied uniformly within a domain, and the paper's 'personalized' sequence representation claim is overstated. The paper earns credit for consistent offline gains across three attention backbones and three datasets, for reporting online A/B results, and for a concrete deployment narrative; these are real evidence that something in the deployed system improved. However, they do not disentangle the architecture from the undisclosed features. The significance reporting issue (no error bars, heuristic significance thresholds) is secondary: even if the point estimates are accepted, the E_D ambiguity remains. A faithful reproduction on a public dataset with a fully specified E_D, plus the baseline-with-same-features arm, would settle the concern. The verdict remains CONDITIONAL as the reader stated, so no adjustment is needed.","tokens_in":15379,"tokens_out":6328,"duration_ms":67386,"concrete_test":"Run a controlled ablation on the public Taobao data: (i) implement ADS with E_D containing only the explicit domain indicator (item category embedding); (ii) implement ADS with the full disclosed E_D; and (iii) concatenate the same statistical features into the input of each backbone (DIN/MHA/CAN) without PSRG/PCRG. If (ii) materially beats (i), the implicit features are doing the work; if (iii) closes the gap against (ii), the comparison is confounded. If neither, publish the E_D feature list so the test can be reproduced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Both ADS modules (PSRG and PCRG) take E_D as their only source of adaptivity (Eqs. 2, 3, 4, 6). The explicit part of E_D is just a scenario/category indicator, while the 'implicit-domain-indicator features' are described only as 'engineer-constructed statistical features' (Sec. 2.2.1). This leaves two unresolved threats to the main claim that the share-and-private architecture, rather than input features, causes the reported lifts. First, if these statistics are user-level activity or label-adjacent counts, they may encode supervised signal that the DIN/MHA/CAN/FRNet/PEPNet baselines never receive, making the Table 2 comparison unfair regardless of network structure. Second, if E_D is effectively coarse (e.g., scenario id), then Eq. 5 applies the same affine map to every sequence position for every user in a domain, so the claimed per-user personalization of item representations reduces to a handful of domain-specific linear layers; the paper's motivating claim that 'the same item occurred in different user's sequence has different representations' is then not realized. The paper provides no analysis of E_D's content, dimensionality, or effect size, and no ablation that isolates E_D from the rest of ADS.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Adaptive Domain Scaling (ADS), a plug-and-play module for multi-domain target-aware sequence modeling in recommender systems. ADS contains two components: Personalized Sequence Representation Generation (PSRG), which applies a learnable affine transformation to user behavior item embeddings using domain-related features, and Personalized Candidate Representation Generation (PCRG), which generates one query per sequence position by combining the target item embedding with domain features. The adapted keys, values, and queries are fed into standard attention backbones (DIN, MHA, and CAN). The authors evaluate on the public Taobao dataset and two billion-scale industrial datasets from ByteDance, report consistent AUC gains over strong baselines across all backbone groups, provide ablations removing each module, and present online A/B experiments in Douyin Ads and Douyin Ecom with business-metric lifts of 0.52%-1.00%.","tokens_in":15602,"tokens_out":4193,"duration_ms":74089,"significance":"If the reported results hold, ADS is a simple and architecture-agnostic contribution that extends multi-domain personalization to sequence modeling, an area the paper correctly identifies as under-served relative to feature-interaction multi-domain methods. The paper's strengths are its broad offline evaluation across three backbones and three datasets, the inclusion of a public dataset, and the deployment evidence with clearly named business metrics and user counts. Ablations confirm that both modules contribute, and the chunked variant of PCRG addresses efficiency concerns. The main caveats are (i) the undefined 'implicit-domain-indicator' features that form the backbone of the adaptivity mechanism, and (ii) unsupported statistical significance claims without error bars. These are fixable with additional description and reporting, but they currently prevent a full assessment of whether the gains come from the architecture or from input-feature asymmetries.","major_comments":[{"comment":"E_D is the sole input that makes PSRG and PCRG adaptive, yet the 'implicit-domain-indicator features' are described only as 'engineer-constructed statistical features' and are never enumerated. If these features include user-activity or label-adjacent statistics that the DIN/MHA/CAN/FRNet/PEPNet baselines do not receive, the Table 2 comparison is confounded because the input features, not the network structure, may explain the gains; if E_D is essentially a scenario ID, then Eq. (5) applies the same affine map to all sequence positions and the claimed per-user personalization reduces to a small set of domain-level linear transforms. Please specify the exact feature set, its dimensionality, and include an ablation that removes or randomizes E_D to isolate the architecture.","section":"Section 2.2.1, Eqs. (2), (4), (6)"},{"comment":"The caption states that gray-shaded results are significant at p < 0.05, but no standard errors, confidence intervals, number of repeated runs, or statistical test are reported anywhere, and the same issue affects the p < 0.01 claims in Tables 4 and 5. With a single evaluation split per dataset, significance cannot be established; please report the variance across repeated training or, if only one run is feasible for billion-scale data, state this explicitly and replace the p-value claim with a threshold-based statement.","section":"Section 3.2 and Table 2"},{"comment":"The ablation only removes modules from the full ADS and never varies the content of E_D, so it cannot distinguish the effect of the domain-conditioned transformation from the effect of the extra parameters introduced by the two MLP-based generation networks. A control that applies the same generated transformation to a random or fixed domain indicator would strengthen the claim that multi-domain conditioning, rather than added capacity, drives the gains.","section":"Section 3.3, Table 3"}],"minor_comments":[{"comment":"There are several typographical errors, including 'accomodate', 'distrbutions', and 'senarios' (in Section 3.1); the paper should be carefully proofread.","section":"Throughout"},{"comment":"The definition of AUC Imp. normalizes by AUC = 0.5 as the random baseline, but the surrounding text's repeated statement that 'a random strategy yields AUC value at 0.5' is redundant; consider clarifying that this is the standard normalized improvement metric.","section":"Section 3.1, Eq. (14)"},{"comment":"The figure is dense and the data-flow arrows connecting scenario-related features to PSRG and PCRG are not labeled with the corresponding tensor shapes or equation numbers; please annotate the figure to make it self-contained.","section":"Figure 2"},{"comment":"The claim that 'a 0.1% Overall Imp. in Douyin Ads and 0.2% Overall Imp. in Douyin Ecom is considered to be significant' is informal; please define the offline significance threshold more precisely and indicate whether it is calibrated against historical online A/B outcomes.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a strong industrial systems paper, but the missing definition of E_D is the central gap; the editors should require the authors to disclose the implicit features or adjust the claims accordingly. The significance claims also need methodological support. If ByteDance considers these features proprietary, the paper should at least provide a public-data ablation with a simple proxy (e.g., user activity buckets) to demonstrate that the mechanism, not the extra input features, is responsible for the reported gains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is real: instead of the usual bitwise or linear domain scaling, ADS generates a full weight-and-bias transform for sequence-item embeddings (PSRG) and a set of per-position or per-chunk queries for target attention (PCRG). That combination is not in the cited prior art, and it is a reasonable way to inject domain information into target-aware sequence modeling. The paper also gives the architecture clearly, with equations for every step.\n\nWhat the paper does well is the evaluation package. Offline results are consistent: ADS beats the second-best method in all three backbone groups (DIN, MHA, CAN) on Taobao, Douyin Ads, and Douyin Ecom, with ablations showing both modules contribute. The chunk sensitivity analysis is sensible, and the online A/B lifts (0.52% CPM, 1.00% ADVV in Ads; 0.79% GMV/U in Ecom) are plausible for a deployed system. For an industrial paper, that is a solid benchmark.\n\nThe soft spot is the domain feature vector E_D, which is load-bearing for both modules (Eqs. 2, 3, 4, 6). The paper specifies the explicit part (scenario/category IDs) but not the 'implicit-domain-indicator features' beyond calling them engineer-constructed statistical features. That matters for two reasons. First, if those statistics are user-level activity or label-adjacent counts and are not given to the baselines, then the comparison is unfair regardless of network structure. Second, if E_D is effectively coarse, the per-user personalization claim reduces to a handful of domain-specific linear layers. The paper should define E_D, state its dimensionality, and show whether it is already part of the shared input features.\n\nAlso missing are error bars or repeated-run details. The p < 0.05 claim in Table 2 and the '0.1% / 0.2% Imp is significant' heuristic in Section 3.1 are not substitutes for actual variance reporting. Online A/B details are also thin, though the p<0.01 statement is at least explicit. No code or data is released, which is common for industrial papers but still limits reproducibility.\n\nOverall, the central architecture is sound and the empirical story is consistent. The E_D underspecification is the one issue that could change the interpretation of the gains, so it needs to be fixed before publication. This deserves serious peer review; I would send it out and ask for a major revision that specifies E_D and reports variance.","headline":"A competent industrial multi-domain recommender paper with a genuinely new domain-conditioned 'scale and query' mechanism, held back by an underspecified domain feature vector and missing variance reporting.","tokens_in":16190,"tokens_out":2195,"would_cite":true,"duration_ms":22653,"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":"The paper claims that domain-conditioned personalization of both user behavior and candidate representations in target-aware attention consistently improves multi-domain ranking, and reports significant online revenue lifts of 1.00% ADVV…","keywords":["Multi-domain recommendation","Sequential modeling","Target-aware attention","Personalized representation","Click-through rate prediction","Parameter generation","Industrial recommender system"],"falsifier":"Run ADS with $E_D$ reduced to the explicit scenario ID alone (dropping the engineered statistical features) on Douyin Ads; if the AUC gain over the backbone does not largely disappear, the claim that implicit domain features carry the personalization would be weakened. Alternatively, compute the variance of $\\mathbf{W}_{generated}$ across samples within one scenario: near-zero variance would contradict the per-user personalization story.","tokens_in":15144,"feed_emoji":"📈","tokens_out":6207,"duration_ms":56379,"temperature":0.7,"pith_summary":"The paper proposes Adaptive Domain Scaling (ADS), a plug-in for target-aware attention in multi-domain recommender ranking. Its two modules, PSRG and PCRG, generate domain-conditioned versions of the user behavior sequence embeddings and of the candidate item queries before attention is applied. The central claim is that this recovers personalization lost when a shared embedding table and a single shared query are used across all domains. Offline, ADS beats the second-best approach in every backbone group on Taobao, Douyin Ads, and Douyin Ecom; online, it reports a 1.00% ADVV lift in Douyin Ads and a 0.79% GMV/U lift in Douyin Ecom. If true, it is a low-cost, backbone-agnostic way to inject multi-domain personalization into sequential modeling.","feed_headline":"Per-domain attention lifts Douyin ads 1% and e-commerce GMV 0.8%","feed_subtitle":"A plug-in that rewrites user behavior and candidate embeddings per domain beats strong baselines and is live at ByteDance.","key_machinery":"Two generation networks built on a share-and-private design. Sequence-Weight Gen-Net computes $\\mathbf{W}_{private} = \\text{Sigmoid}(\\text{ReLU}(E_D \\mathbf{W}_1^T + \\mathbf{b}_1) \\mathbf{W}_2^T + \\mathbf{b}_2)$, then $\\mathbf{W}_{generated} = \\eta (\\mathbf{W}_{shared} \\odot \\mathbf{W}_{private})$ with a learnable global $\\mathbf{W}_{shared}$; Sequence-Bias Gen-Net produces $\\mathbf{b}_{generated}$ similarly. PSRG applies the resulting affine map to the sequence embedding $E_S$. Multi-Query Gen-Net computes $E_{Q\\_private} = \\text{ReLU}([E_D \\oplus E_Q] \\mathbf{W}_{q1}^T + \\mathbf{b}_{q1}) \\mathbf{W}_{q2}^T + \\mathbf{b}_{q2}$, and PCRG forms per-position queries $E_{Q\\_private} + \\text{tile}(E_Q)$. A chunked variant groups adjacent actions to lower cost. These personalized queries, keys, and values feed standard target attention (DIN, MHA, CAN) and then the ranking MLP.","core_discovery":"The paper's central claim is that two shortcomings of multi-domain target-aware attention — the shared sequence embedding that ignores the domain a user is acting in, and the single candidate query that ignores which history item it is attending to — can be repaired by generating representations conditioned on domain features $E_D$. PSRG learns a global weight matrix $\\mathbf{W}_{shared}$ that is element-wise scaled by a sigmoid-gated private weight $\\mathbf{W}_{private}$ (with a tuned factor $\\eta$) and combined with a generated bias; the resulting affine transform maps each sequence item embedding to a personalized one, so the same item differs across user-domain contexts. PCRG builds a private query per sequence position from $[E_D \\oplus E_Q]$ through a two-layer MLP and adds it residually to the tiled original query, letting the candidate ask different questions of each history item. The paper reports consistent wins over the second-best method in every backbone group on three datasets, and online A/B tests with statistical significance at the level reported.","pith_inferences":["If $E_D$ is the bottleneck, replacing the hand-built statistical domain features with learned user-level or cluster-level domain embeddings could amplify the gains, since PSRG and PCRG are agnostic to what $E_D$ encodes.","The per-position query mechanism resembles a form of multi-aspect target attention; one could test whether the generated queries specialize into interpretable aspects such as price-sensitive versus creator-sensitive behavior by probing or clustering the queries.","The method should transfer to retrieval or cold-start settings because query generation depends only on the target embedding and domain features, not on the candidate corpus, though the chunked variant would be needed to control cost.","The paper's offline significance convention treats 0.1%–0.2% overall Imp as meaningful without error bars; re-running with confidence intervals on repeated seeds would sharpen the comparison."],"forward_implications":["Any target-aware attention backbone can be upgraded by prepending PSRG and PCRG; gains appear with DIN, MHA, and CAN alike.","Sequence items become context-dependent: identical items are embedded differently across user domains, allowing the attention to reflect scenario-specific attraction.","The candidate becomes multi-faceted: different history positions are queried with different representations of the same target item.","Chunk size tunes the personalization-compute tradeoff; smaller chunks (down to per-item) monotonically improve AUC at higher FLOPs.","Industrial deployment is feasible at low overhead: +3.6% training time and no latency change, with online revenue lifts of 1.00% ADVV and 0.79% GMV/U."],"supporting_citations":[{"why":"Supplies the DIN backbone that ADS personalizes and the main target-attention baseline.","marker":"[33]"},{"why":"Provides the multi-head attention mechanism used as one of the three backbones and as the running example in the method.","marker":"[21]"},{"why":"Supplies the CAN backbone, a third target-attention baseline that ADS must beat.","marker":"[2]"},{"why":"The strongest multi-domain personalization baseline (PEPNet) that ADS compares against in all backbone groups.","marker":"[6]"},{"why":"FRNet baseline: context-aware feature refinement that ADS outperforms in the experiments.","marker":"[22]"},{"why":"Releases the Taobao dataset used as the public evaluation benchmark with category-as-domain splits.","marker":"[34]"}],"fun_headline_variants":["Domain-scaling embeddings lift Douyin ads 1%, e-commerce 0.8%","Adaptive domain scaling lifts Douyin ads 1% and e-commerce 0.8%","Personalized domain scaling lifts Douyin ads 1%, e-commerce 0.8%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire adaptation is driven by the domain feature vector $E_D$; if those features are too coarse to separate users or scenarios at the instance level, the generated transforms become nearly constant and the claimed per-user personalization collapses to a fixed per-scenario scaling.","fun_headline_variants_meta":{"raw":{"variants":["Domain-scaling embeddings lift Douyin ads 1%, e-commerce 0.8%","Adaptive domain scaling lifts Douyin ads 1% and e-commerce 0.8%","Personalized domain scaling lifts Douyin ads 1%, e-commerce 0.8%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001511,"raw_usage":{"total_tokens":6072,"prompt_tokens":975,"completion_tokens":5097,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":5022}},"tokens_in":591,"tokens_out":5097,"duration_ms":36334,"temperature":1.0,"reasoning_tokens":5022,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:58:32.280772+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ADS with $E_D$ reduced to the explicit scenario ID alone (dropping the engineered statistical features) on Douyin Ads; if the AUC gain over the backbone does not largely disappear, the claim that implicit domain features carry the personalization would be weakened. Alternatively, compute the variance of $\\mathbf{W}_{generated}$ across samples within one scenario: near-zero variance would contradict the per-user personalization story.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the DIN backbone that ADS personalizes and the main target-attention baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FRNet baseline: context-aware feature refinement that ADS outperforms in the experiments."}],"review_version":1}