{"id":"ce3d2b2d-dd41-4a50-a63f-781a4af7f552","arxiv_id":"2411.09286","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"CDTM uses dual embeddings plus a learned transfer matrix and attention to improve click-through prediction in a target domain by transferring from multiple heterogeneous source domains, with reported offline AUC and online CTR/eCPM gains on NetEase data.","lead":"This paper presents CDTM, a model that transfers knowledge from several advertising domains at once to improve click prediction in a sparse target domain. It combines domain-specific and shared embeddings, a transfer matrix, and an attention mechanism to handle mismatched feature sets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The transfer-matrix objective in Eq. (7) is in direct tension with the combination mechanism in Eq. (1): if it works, T⊙Gc collapses toward Ec and the GSE contribution vanishes; if it does not, the claimed latent-space alignment is untested.","rationale":"The reader's weakest assumption already identifies the transfer-matrix alignment as the most fragile point: a single learned map T minimizing Euclidean distance between Ec and T⊗Gc can collapse the shared embedding into the domain-specific embedding. My reading agrees and sharpens the problem: the paper's own Eq. (1) and Eq. (7) are in direct tension. The auxiliary loss is designed to make the mapped GSE equal to Ec, while the combination layer is supposed to keep them as complementary information sources. The text even states both requirements explicitly, so this is not an external objection but an internal inconsistency. Because the paper provides no ablation of λ and no measurement of the alignment distance, the central mechanism is unverified. This concern does not necessarily falsify the empirical claim that the CDTM architecture improves CTR: the improvements in Table IV could still come from multi-source data volume or from the dual-embedding/attention components. However, the paper's stated contribution is specifically the transfer-matrix solution to latent-space heterogeneity, and that contribution is exactly what remains untested. A conditional verdict remains appropriate: the authors should add an ablation varying λ, report whether the auxiliary loss actually brings T⊗Gc close to Ec, and show that the attention weights are non-degenerate. Since the reader already reached CONDITIONAL, I recommend no change to the verdict.","tokens_in":9016,"tokens_out":6935,"duration_ms":68530,"concrete_test":"Retrain CDTM on F1–F4 with λ = 0 and with λ = 10^-4 and λ = 10^-2, keeping all other settings fixed, and report Table IV plus the converged mean Frobenius norm of Ec − T⊗Gc and the distribution of attention weights A for the target domain. If λ = 0 reproduces the multi-source gains, then Eq. (7) is not load-bearing; if large λ drives the distance to zero, then Eq. (1) contains no effective GSE contribution and the stated mechanism collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C1 defines the transfer matrix by minimizing ||Ec − T⊗Gc||^2 (Eq. 2) and adds this as the auxiliary loss λΣ||Ec−T⊗Gc||^2 in Eq. (7). But Section III-C uses the same T⊗Gc in the combination E = Ec⊗A + T⊗Gc⊗(1−A) (Eq. 1) and explicitly states fact (i): DSE and GSE 'should be different because they keep different information.' If Eq. (7) is effective, T⊗Gc converges toward Ec, so the combined embedding collapses to approximately Ec regardless of A, and the claimed adaptive combination of global knowledge is vacuous. If Eq. (7) is weak (and the experimental section reports a regularization parameter of only 0.0001, if that is λ), the transfer matrix is not actually trained to align the latent spaces, and the paper's remedy for latent-space heterogeneity has no empirical support. The paper never reports the distance ||Ec−T⊗Gc||, never ablates λ, and never shows that the attention weights A are non-degenerate. This is not just a missing baseline: it is an internal consistency problem in the central mechanism. The multi-source gains in Table IV could then come from training with more source data or from the dual embedding, not from the proposed transfer-matrix alignment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CDTM, a centralized-distributed transfer model for multi-source cross-domain CTR prediction. The architecture gives each domain a domain-specific embedding (DSE) and all domains a jointly trained global shared embedding (GSE). Transferable features are represented by both embeddings, and the combination layer maps GSE through a transfer matrix T, then adaptively combines it with DSE using an attention vector A (Eq. (1)). The transfer matrix is intended to align the heterogeneous latent spaces of DSE and GSE by minimizing the Euclidean distance in Eq. (2), and an auxiliary loss version of this objective is added to the total training loss in Eqs. (7)-(8). Experiments on a proprietary NetEase Cloud Music advertising dataset compare CDTM with DCN, CoNet, SCoNet, DDTCDR, DTCDR, and GA-DTCDR under single-source and multi-source settings, include a four-domain extension, and report an online A/B test with 5.1% CTR and 6.6% eCPM improvements over DCN.","tokens_in":9255,"tokens_out":6031,"duration_ms":63453,"significance":"If the central claims were fully established, the paper would make a useful empirical contribution to multi-source heterogeneous cross-domain recommendation. The strongest evidence is Table IV, where CDTM is compared against CDTM-H and CDTM-J under the same architecture: this directly supports the multi-source-over-single-source claim. The paper also usefully identifies feature-dimensional and latent-space heterogeneities that many earlier CDR methods ignore, and the online deployment is an encouraging practical signal. However, the central transfer-matrix mechanism is not empirically isolated. There is no same-architecture source-free baseline, no ablation that removes the transfer matrix or the auxiliary loss, no report of the achieved alignment distance, no diagnostic of the learned attention weights, and no error bars or significance tests on the offline AUC comparisons. Because the reported gaps are often very small (for example, 0.0021 between CDTM and GA-DTCDR-H on F2 in Table II), the evidence for the headline claims is currently incomplete.","major_comments":[{"comment":"The auxiliary loss in Eq. (7) directly minimizes the same Euclidean distance between Ec and T⊗Gc that Eq. (2) defines, while Eq. (1) combines the two quantities with attention A. If the auxiliary loss is effective, T⊗Gc converges toward Ec and the combined embedding in Eq. (1) becomes approximately Ec regardless of A, making the GSE pathway and the attention mechanism vacuous. If the auxiliary loss is weak—and the reported regularization parameter in Section IV-A3 is only 0.0001, presumably λ—then the paper's stated remedy for latent-space heterogeneity has no empirical support. The manuscript reports neither the achieved value of ||Ec−T⊗Gc||, nor a sensitivity analysis over λ, nor the learned attention values. This is load-bearing for the central claim that the transfer matrix addresses latent-space heterogeneity, and the current experiments do not reveal which of the two regimes holds.","section":"Section III-C1, Eqs. (1), (2), (7)"},{"comment":"There is no source-free CDTM baseline. CDTM-H and CDTM differ from the DCN Base model not only in the use of source-domain data but also in architecture: they add the dual embedding structure, the combination attention, and the transfer matrix. Therefore the gains over Base cannot be attributed specifically to knowledge transfer from source domains rather than to the added capacity. To support the transfer claim, the authors should compare CDTM trained with target-domain data only against CDTM-H and CDTM under identical architecture and training settings.","section":"Section IV-B, Tables II-IV"},{"comment":"All offline results are reported as point estimates without standard errors, confidence intervals, or significance tests. This matters because several reported improvements are small; for example, on F2 in Table II the CDTM-H gain over DTCDR-H is 0.0028 AUC, and on F4 in Table III the CDTM-J gain over GA-DTCDR-J is 0.0005 AUC. The statement in Section IV-B3 that CDTM 'performs significantly' than CDTM-DA is not supported by any statistical test. The authors should provide error bars over multiple runs (or an equivalent uncertainty quantification) for the main comparisons.","section":"Section IV-B, Tables II-V and Fig. 2"},{"comment":"The ablation study and the extensibility study are incomplete relative to the paper's stated contributions. The ablation removes only the combination attention (CDTM-DA), leaving the transfer matrix and the auxiliary loss unablated; contribution 3 in the introduction is therefore not tested. In Task 4, CDTM4 is compared only against the Base DCN model, not against the multi-domain CTR methods STAR and DADNN cited in Section II-B, so the claim that CDTM can be extended to multi-source and multi-target CDR is not benchmarked against existing multi-domain approaches.","section":"Section IV-B3, Fig. 2 and Table V"}],"minor_comments":[{"comment":"The first sentence says 'The transferable features have only one type of embedding (i.e., DSE), while the nontransferable features have two (DSE and GSE).' The following sentences state the opposite for both cases. The sentence appears to have transferable and nontransferable swapped and should be corrected, as the architecture description is otherwise inconsistent.","section":"Section III-B1"},{"comment":"Eq. (2) writes L = argmin_T Σ ||Ec − T⊗Gc||²; this should be the minimum value of the objective or the minimizing argument, not the loss L equated to the argmin. The notation also conflicts with the use of L as a loss symbol elsewhere.","section":"Eq. (2)"},{"comment":"The table captions are inconsistent with the task numbering: Table III is labeled 'TASK 2' but reports Task 1 using J as the source, and Table IV is labeled 'TASK 3' but reports the Task 2 multi-source comparison. Please align the captions with the task definitions in Section IV-A2.","section":"Table captions"},{"comment":"The notation for T⊗Gc and Ec⊗A is not fully specified: A is called an attention matrix but described as a vector, and the dimensions of T relative to the embedding matrices are not stated. The paper should clarify whether ⊗ is element-wise multiplication and give the shapes of T and A.","section":"Section III-C, Eq. (1)"},{"comment":"The dataset description reports the number of transferable feature fields as 386 for H and 555 for J and F1-F4, but the text does not explain how the common transferable feature fields between each source-target pair are selected, nor how the differing feature fields are handled in the shared GSE table. A few sentences on this would improve reproducibility.","section":"Section IV-A1"},{"comment":"There are several typographical errors, including 'clcik' in the Index Terms, 'domians' in Section II-B, and 'CTR estimation' in the abstract used where 'cross-domain recommendation' seems intended. A careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"This is an industrial paper based on proprietary data, and the authors explicitly note the lack of suitable public datasets. The multi-source versus single-source comparison in Table IV is the most credible part of the empirical case. My main concern is that the transfer-matrix mechanism may be decorative: the auxiliary loss in Eq. (7) could collapse T⊗Gc toward Ec, or it may be too weak to matter, and the paper provides no diagnostic to distinguish these cases. This is fixable with additional experiments (source-free same-architecture baseline, λ sensitivity, alignment distance, attention-weight statistics, and uncertainty quantification), so I recommend major revision rather than rejection. The citation list is adequate, and I do not see a novelty-disclosure problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper for two reasons: it reports a deployed industrial system with a 5.1% CTR and 6.6% eCPM lift over DCN, and it has a clean same-architecture comparison (Table IV) showing that using two source domains beats either single source across four targets. That multi-source result is the paper's real contribution, and it is credible.\n\nWhat is genuinely new: the specific packaging of dual DSE/GSE embeddings, a per-domain transfer matrix, and attention gating into a single multi-source heterogeneous CDR model. The individual pieces are known, but this configuration appears in no prior work, and the industrial setting gives it practical value. The baselines show frequent negative transfer, which supports the authors' claim that heterogeneity matters.\n\nNow the soft spots. The stress-test note is on target. Eq. (7) adds an auxiliary loss that minimizes ||Ec − T⊗Gc||^2, while Eq. (1) combines Ec and T⊗Gc with attention A. If the auxiliary loss works, T⊗Gc collapses toward Ec and the combination becomes approximately Ec regardless of A, making the GSE path and attention vacuous. If the auxiliary loss is weak—and λ = 0.0001 suggests it is—then the transfer matrix is never shown to actually align the latent spaces. The paper reports neither the alignment distance, nor an ablation over λ, nor the learned attention values. This is an internal inconsistency in the central mechanism, not just a missing baseline. The multi-source gains in Table IV may come from added capacity or the dual embedding, not from the proposed transfer-matrix alignment.\n\nOther issues are standard but real: no source-free CDTM baseline, no error bars or significance tests, proprietary data with no code release, and a couple of minor text inconsistencies (Table III is labeled Task 2 but is part of Task 1, for example). None of these, on its own, would sink the paper; the missing mechanism validation is the load-bearing one.\n\nBottom line: treat the multi-source empirical result as plausible and the mechanism explanation as unverified. This paper deserves a serious referee, but the authors should be pushed to add a no-transfer CDTM variant, report variance, and at minimum show the learned T and A behave as claimed.\n\nRecommendation: accept for peer review, expecting major revision.","headline":"A credible industrial multi-source CDR paper whose multi-source claim is supported by same-architecture comparisons, but whose central transfer-matrix mechanism has an internal consistency problem the authors never test.","tokens_in":9818,"tokens_out":1819,"would_cite":false,"duration_ms":47261,"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":"CDTM claims that a centralized-distributed dual-embedding architecture lets a target domain borrow from many source domains at once, fixing feature and latent-space heterogeneity and lifting ad CTR by 5.1% online.","keywords":["cross-domain recommendation","click-through rate prediction","multi-source transfer learning","heterogeneous transfer learning","dual embedding","combination attention","negative transfer","advertising system"],"falsifier":"Run CDTM with the transfer matrix replaced by a per-field nonlinear mapping, or with the auxiliary alignment loss removed, and compare AUC on the same F1-F4 targets; if the gains over single-source transfer persist, the claimed latent-space-alignment mechanism is not what drives the results. Alternatively, inspect the learned embeddings directly: if $T \\otimes G_c$ nearly equals $E_c$ for most features, the auxiliary loss has collapsed the two spaces and the attention weights carry the entire effect.","tokens_in":8795,"feed_emoji":"🎯","tokens_out":6186,"duration_ms":56674,"temperature":0.7,"pith_summary":"The paper proposes CDTM, a centralized-distributed transfer model that lets a target advertising domain borrow knowledge from several source domains at once, rather than from a single source. It argues that the two main causes of negative transfer in cross-domain recommendation, different feature fields across domains and mismatched learned feature spaces, can be addressed by a dual embedding structure plus a learned transfer matrix and attention weights. The model is trained jointly across domains: each domain keeps its own domain-specific embedding, while all domains share a global shared embedding, and the target model adaptively mixes the two. Offline experiments on commercial advertising data report AUC gains over the single-domain baseline and over single-source transfer baselines on every tested target domain, and a two-week online A/B test reports 5.1% higher CTR and 6.6% higher eCPM than a DCN baseline.","feed_headline":"Multi-source transfer lifts ad CTR by 5.1% online","feed_subtitle":"Dual embeddings plus a transfer matrix beat single-source transfer on every target and extend to multi-domain setups.","key_machinery":"The load-bearing mechanism is the dual embedding structure combined with a linear transfer matrix and an attention gate. Each domain has a domain-specific embedding (DSE) table for its unique features, and all domains share a global shared embedding (GSE) table for transferable features; nontransferable features use only DSE, while transferable features get both. The combination layer merges the two embeddings for each transferable feature with $E = E_c \\otimes A + T \\otimes G_c \\otimes (1-A)$, where the transfer matrix $T$ maps GSE into DSE's latent space to fix latent-space heterogeneity and the attention vector $A$, computed from $[E_c, E_c \\otimes G_c, E_c \\oplus G_c, G_c]$, weights which embedding matters more per feature. The auxiliary loss drives $T \\otimes G_c$ toward $E_c$, and the total loss is a weighted sum over all domains of prediction loss plus this alignment term.","core_discovery":"On the paper's own terms, the discovery is that a symmetric centralized-distributed architecture can make multi-source, heterogeneous cross-domain transfer work in CTR prediction. For each transferable feature, CDTM combines the domain-specific embedding $E_c$ with the global shared embedding $G_c$ through $E = E_c \\otimes A + T \\otimes G_c \\otimes (1-A)$, where $T$ is a transfer matrix that maps $G_c$ into the latent space of $E_c$ and $A$ is a learned combination-attention vector. An auxiliary loss $\\lambda \\sum \\|E_c - T \\otimes G_c\\|^2$ is added so the transfer matrix is optimized to bring the two embedding spaces into alignment. The paper reports that this structure beats single-source variants on all four target domains, avoids the negative transfer seen in several baselines, and extends to a four-domain jointly trained setup that improves all four domains at once.","pith_inferences":["A natural next step the paper does not take is to measure whether each source domain's contribution is positive online; the reported A/B test compares only against DCN, not against single-source variants, so per-source negative transfer in production remains untested.","Because the transfer matrix is shared element-wise across all transferable features, a per-field or per-feature transfer map would be a direct extension and could capture cases where the relation between global and domain spaces differs across features.","The gains reported on commercial data leave open how the method behaves on public multi-domain benchmarks; a testable extension is to apply CDTM to a public dataset with more than two source domains and report confidence intervals across seeds."],"forward_implications":["Multi-source transfer consistently beats single-source transfer: CDTM outperforms both CDTM-H and CDTM-J on every target domain in Table IV.","Handling heterogeneity prevents negative transfer: several baselines lose AUC on at least one target domain, while CDTM improves on all of them in every configuration.","The symmetric design extends to multi-target scenarios: a four-domain jointly trained CDTM 4 improves all four domains simultaneously over the single-domain base model.","The method survives deployment: a two-week online A/B test reports 5.1% CTR lift and 6.6% eCPM lift over the DCN baseline.","Both components matter: removing combination attention still beats the base model, and adding it back improves results further in the ablation study."],"supporting_citations":[{"why":"First embedding-and-mapping framework for cross-domain recommendation that CDTM extends to the multi-source case.","marker":"[16]"},{"why":"CoNet baseline that must be beaten on single-source transfer tasks.","marker":"[22]"},{"why":"DTCDR baseline for dual-target cross-domain transfer.","marker":"[17]"},{"why":"DDTCDR baseline for deep dual transfer.","marker":"[23]"},{"why":"GA-DTCDR unified baseline for cross-domain and cross-system recommendation.","marker":"[20]"},{"why":"DCN defines the single-domain Base model used for all AUC comparisons and the online A/B baseline.","marker":"[9]"},{"why":"STAR topology adaptive recommender, a multi-domain CTR method whose network-sharing limitation motivates CDTM's separable embeddings.","marker":"[28]"},{"why":"DADNN multi-scene CTR method representing the multi-domain approach CDTM distinguishes itself from.","marker":"[27]"}],"fun_headline_variants":["Multi-source transfer model beats single-source in all domains","CDTM: no negative transfer with multi-source heterogeneous learning","Dual embeddings plus transfer matrix lift CTR across four domains","Centralized-distributed transfer avoids negative transfer in CDR","CDTM multi-source transfer: +5.1% ad CTR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes a single learned linear element-wise transfer matrix can align the global shared embedding with each domain's own embedding, and that the alignment loss does not just collapse the shared embedding into the domain-specific one.","fun_headline_variants_meta":{"raw":{"variants":["Multi-source transfer model beats single-source in all domains","CDTM: no negative transfer with multi-source heterogeneous learning","Dual embeddings plus transfer matrix lift CTR across four domains","Centralized-distributed transfer avoids negative transfer in CDR","CDTM multi-source transfer: +5.1% ad CTR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001016,"raw_usage":{"total_tokens":4284,"prompt_tokens":932,"completion_tokens":3352,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":3270}},"tokens_in":548,"tokens_out":3352,"duration_ms":23905,"temperature":1.0,"reasoning_tokens":3270,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:48:59.532942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CDTM with the transfer matrix replaced by a per-field nonlinear mapping, or with the auxiliary alignment loss removed, and compare AUC on the same F1-F4 targets; if the gains over single-source transfer persist, the claimed latent-space-alignment mechanism is not what drives the results. Alternatively, inspect the learned embeddings directly: if $T \\otimes G_c$ nearly equals $E_c$ for most features, the auxiliary loss has collapsed the two spaces and the attention weights carry the entire effect.","supporting_citations":[{"cited_title":"Conet: Collaborative cross networks for cross-domain recommendation,","cited_arxiv_id":null,"evidence_quote":"CoNet baseline that must be beaten on single-source transfer tasks."},{"cited_title":"Dtcdr: A framework for dual-target cross-domain recommendation,","cited_arxiv_id":null,"evidence_quote":"DTCDR baseline for dual-target cross-domain transfer."},{"cited_title":"Ddtcdr: Deep dual transfer cross domain recommendation,","cited_arxiv_id":null,"evidence_quote":"DDTCDR baseline for deep dual transfer."},{"cited_title":"A unified framework for cross-domain and cross-system recommendations,","cited_arxiv_id":null,"evidence_quote":"GA-DTCDR unified baseline for cross-domain and cross-system recommendation."},{"cited_title":"Deep & cross network for ad click predictions,","cited_arxiv_id":null,"evidence_quote":"DCN defines the single-domain Base model used for all AUC comparisons and the online A/B baseline."},{"cited_title":"One model to serve all: Star topology adaptive recommender for multi-domain ctr prediction,","cited_arxiv_id":null,"evidence_quote":"STAR topology adaptive recommender, a multi-domain CTR method whose network-sharing limitation motivates CDTM's separable embeddings."}],"review_version":1}