{"id":"1d470697-dfcc-4eb2-bbbd-cafdd05867e0","arxiv_id":"2412.01378","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey that categorizes 2020-2024 deep learning approaches to collaborative filtering by architecture and summarizes their datasets, metrics, and open challenges.","lead":"This survey organizes roughly 80 recent papers on deep neural network methods for collaborative filtering into seven architecture families and reviews the datasets and metrics those papers use. It is a starting point for newcomers who want a map of the DNN-based recommendation literature and its common evaluation practices.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 misclassifies GCN papers [79]–[83] as CNN, contradicting §3.2.1/§3.2.3 and undermining the claimed comprehensive categorization.","rationale":"The paper's intended contribution is a structured map of DNN-based CF. For that map to have value, the category assignments must be correct and consistent. I found a concrete violation: Table 1 lists [79]–[83] under CNN, but these are all graph convolutional network methods; they are not discussed in the CNN section, and they are absent from the GNN section. This is more decisive than the general corpus-selection concern because it shows the taxonomy itself is unreliable, independently of how the corpus was chosen. The duplicated [37]/[45] citation and the N3S/HN3S name mismatch are additional symptoms of the same underlying reliability problem. The survey is not beyond repair — these are correctable errors — so conditional acceptance with a required revision is the appropriate verdict. I partially agree with the reader: both of us identify reliability threats, but I pinpoint a specific, checkable categorization error rather than relying only on the unrepresentative-corpus concern.","tokens_in":17405,"tokens_out":3898,"duration_ms":30355,"concrete_test":"Construct the reference-to-category mapping implied by Table 1 and verify each entry against the paper's own section discussions. For refs [79]–[83], confirm from their titles/abstracts that they are graph convolutional networks, then check whether they appear in §3.2.1 or §3.2.3. If they are GCNs and are not discussed in either section, the taxonomy is internally inconsistent; the fix would be to move [79]–[83] to the GNN row and add a paragraph in §3.2.3 covering graph convolutional approaches.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is a comprehensive and reliable categorization of DNN families in CF. That claim fails internally: Table 1 assigns references [79]–[83] to the CNN row, but all five are graph convolutional network papers (linear residual GCN, incremental GCN, LightGCN, simplified graph-based CF, and hyperbolic GCN). §3.2.1 (CNN) discusses only [8]–[16], never mentioning [79]–[83]; §3.2.3 (GNN) covers [27]–[46], also omitting them. Thus the taxonomy contradicts the body: a reader following Table 1 would misclassify a major GNN subfamily as CNN, while a reader following the GNN section would miss these methods entirely. This is not a typo; it means the advertised 'comprehensive analysis and categorization' is unreliable for exactly the purpose stated. The duplicated [37]/[45] entry and the N3S/HN3S name mismatch are additional symptoms of the same reliability problem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper surveys the use of deep neural network architectures in collaborative filtering recommender systems. It covers seven architecture families (MLP, CNN, RNN, GNN, autoencoder, GAN, RBM), describes representative methods under each, and provides tables of datasets, evaluation metrics, and a high-level classification of application challenges. The stated contribution is a comprehensive analysis and categorization of DNN applications in CF, filling a perceived gap left by surveys that focus on a single model family.","tokens_in":17567,"tokens_out":5888,"duration_ms":43476,"significance":"If the internal inconsistencies were corrected, the survey could serve as a useful entry point for researchers, particularly because it spans multiple architecture families and compiles dataset statistics and metric usage from a substantial set of recent papers. The paper's strengths include the breadth of covered methods and the explicit tables linking references to architecture and metrics. However, the current version contains several factual and editorial errors that undermine its reliability as a reference, so the significance depends on a revision that resolves these issues.","major_comments":[{"comment":"Table 1 lists references [79]–[83] in the CNN row, but all five are graph convolutional network papers (linear residual GCN, incremental GCN, LightGCN, simplified graph-based CF, and hyperbolic GCN). Section 3.2.1 discusses only [8]–[16] for CNN, and Section 3.2.3 discusses [27]–[46] for GNN, omitting [79]–[83]. This contradicts the advertised comprehensive categorization: a reader relying on Table 1 would misclassify a major GNN subfamily as CNN, while a reader following the GNN section would miss these methods entirely.","section":"Table 1"},{"comment":"The text near Equation (2) says 'Figure 2 shows the architecture diagram of NCF,' but Figure 2 was already introduced in Section 1 as the percentage distribution by year. The NCF architecture is not shown in Figure 2; this cross-reference is incorrect and should be fixed or replaced with an actual architecture diagram.","section":"§2.1 (NCF paragraph)"},{"comment":"The GNN section states 'The rule is: (4)' and then gives a verbal description of node embedding updates, but Equation (4) is not actually displayed. Since the paper aims to introduce GNNs to readers, the missing equation is a substantive technical gap and should be added.","section":"§3.2.3 (GNN update rule)"},{"comment":"References [37] and [45] are the same work (Xia et al., 2023, 'Graph-less collaborative filtering'), but the text cites them separately for SimRec with different descriptions: [37] 'transferring knowledge from a teacher GNN to a lightweight student network' and [45] 'combines knowledge distillation and contrastive learning.' This duplication and inconsistency should be resolved.","section":"§3.2.3 (SimRec)"},{"comment":"The text attributes the 'N3S' model to Zhang et al. [47], but reference [47] is titled 'HN3S: A Federated AutoEncoder framework for Collaborative Filtering via Hybrid Negative Sampling and Secret Sharing.' The model name in the text does not match the cited source; this is a factual error that should be corrected.","section":"§3.2.4 (Privacy Protection)"},{"comment":"The methodology states that the authors used Google Scholar to collect 'approximately 80 relevant papers' from four publishers, but it reports no search strings, inclusion or exclusion criteria, or screening process. Without this information, the claim that the survey provides a comprehensive and reliable categorization cannot be verified or reproduced; the corpus may be unrepresentative.","section":"§1 (Approach to papers collection)"}],"minor_comments":[{"comment":"The abstract contains a grammatical error: 'a examination' should be 'an examination.'","section":"Abstract"},{"comment":"The caption reads 'Figure 5 shows a sampe of Multilayer Perceptron architecture'; 'sampe' should be 'sample.'","section":"Figure 5 caption"},{"comment":"The bullet 'Data Sparsity: Techniques used in models like FEDNCF [4] and SRSCCNN [13]' appears to cite the wrong reference: FedNCF is reference [1], not [4], and the model name should be SRSCNN as used in §3.2.1, not SRSCCNN.","section":"§4.3 (Data Sparsity bullet)"},{"comment":"Table 3 lists references [79], [82], and [83] under metric rows, but these GCN papers are not discussed in the architecture sections and are misclassified in Table 1; the metric table should be cross-checked with the corrected taxonomy.","section":"Table 3"},{"comment":"The bullet 'Adaptive Propagation: Techniques in models like CARA [17]' cites reference [17], which is the CRCF model described in §3.2.2, not a model named CARA; the model name should be verified.","section":"§4.3 (Adaptive Propagation bullet)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful scope but needs a careful revision to address the internal inconsistencies, especially the Table 1 misclassification and the missing equation. The issues are fixable in a revision, but they currently affect the survey's credibility as a reference. I would also suggest a rigorous proofreading pass, as the number of small errors suggests the manuscript was not carefully checked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this survey. Its value is as a categorized reading list of roughly eighty 2020-2024 papers on deep neural networks for collaborative filtering — not as a source of new methods or findings. And its central claim, a comprehensive and reliable categorization, is not fully earned: Table 1 assigns references [79]–[83] to the CNN row, but all five are graph convolutional network papers (linear residual GCN, incremental GCN, LightGCN, simplified graph-based CF, hyperbolic GCN). The CNN section never mentions them; the GNN section omits them. A reader trusting the table will misclassify a major GNN subfamily.\n\nWhat the paper does well: the per-architecture summaries are largely accurate and cover a reasonable slice of recent applied work. The dataset table with sparsity and links is genuinely useful, as is the metrics table mapping papers to evaluation protocols. The \"why DNNs\" list of application areas (privacy, cold start, data sparsity) gives a newcomer a workable orientation.\n\nWhere it gets soft. The paper collection method is non-reportable: \"Google Scholar, approximately 80 papers, four publishers\" with no search strings, screening criteria, or inclusion/exclusion rules. That makes the claimed comprehensiveness unverifiable. Then there are concrete internal errors: equation (4) for the GNN update rule is missing; Figure 2 is labeled as the NCF architecture but actually shows the year distribution; reference [37] and [45] are the same paper; and the N3S model in the text is HN3S in the reference list. On their own each is a minor typo, but together they corrode confidence in a reference work whose job is accurate mapping. The GNN/CNN table misclassification is the real problem because it directly breaks the advertised categorization.\n\nWho this is for: a student or practitioner who wants a quick names-and-titles map of DNN-CF work from the last five years. It should not be used as a definitive taxonomy without cross-checking the primary sources. It deserves a serious referee: the scope is useful, nothing here is irreparable, and a careful revision could turn it into a dependable entry point. I would recommend peer review with major revision, specifically asking the authors to fix the table taxonomy, restore the missing equation, correct figure and reference errors, and report a reproducible search protocol.","headline":"A useful but sloppy survey: good reading list, but the taxonomy table misclassifies five GNN papers as CNN and the collection method is non-reportable.","tokens_in":18123,"tokens_out":2226,"would_cite":false,"duration_ms":19207,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A new survey argues that deep neural networks, organized into seven architecture families, effectively address the scalability and nonlinearity limitations of traditional collaborative filtering.","keywords":["collaborative filtering","recommendation systems","deep neural networks","deep learning","neural collaborative filtering","survey","graph neural networks","autoencoders"],"falsifier":"A reader who re-runs the literature search with explicit queries and inclusion criteria and finds that major, frequently cited DNN-CF models are missing from the survey's seven categories would show that the coverage claim does not hold; alternatively, checking whether the cited papers' architectures match their placement in the taxonomy would reveal the accuracy of the categorization.","tokens_in":17205,"feed_emoji":"🤖","tokens_out":5178,"duration_ms":41069,"temperature":0.7,"pith_summary":"This paper is a survey that tries to establish that deep neural networks, applied within collaborative filtering, overcome the scalability and flexibility limits of traditional matrix-factorization methods, and that the resulting work can be usefully organized into seven architecture families. It reviews roughly eighty papers from 2020 to 2024 and categorizes them as MLP, CNN, RNN, GNN, autoencoder, GAN, or RBM based, while also compiling the datasets and evaluation metrics these works use. The reason this matters is that prior surveys cover only a single deep-learning family, so practitioners lack a single structured map of the whole DNN-CF landscape; this paper aims to provide that map. If the survey is accurate, it gives researchers a reliable entry point for choosing architectures and for locating open problems such as data sparsity, cold start, and interpretability.","feed_headline":"Survey maps deep-learning collaborative filtering across seven architectures","feed_subtitle":"One centralized review of MLP, CNN, RNN, GNN, autoencoder, GAN, and RBM methods for recommendation systems.","key_machinery":"The central mechanism is the classification framework itself: a taxonomy of DNN architectures applied to collaborative filtering, with the neural collaborative filtering formulation $\\hat{r}_{ui} = \\sigma(h^T f(P_u, Q_i))$ as the bridge from dot-product matrix factorization to deep interaction models. This taxonomy organizes the literature into MLP, CNN, RNN, GNN, autoencoder, GAN, and RBM families, and it is what lets the survey make comparative claims across models, datasets, and metrics.","core_discovery":"The survey establishes that deep neural networks can be systematically organized into seven architecture families, each of which addresses specific collaborative-filtering challenges such as sparsity, cold start, and implicit feedback. It argues that these models extend classical matrix factorization by learning nonlinear user-item interactions, and it provides a structured map of the field's models, datasets, and evaluation metrics. The paper positions its contribution as bridging the gap between collaborative filtering and deep learning surveys, which previously covered only one architecture family at a time.","pith_inferences":["The taxonomy could be extended to transformer-based and large-language-model recommendation methods, which have emerged since the paper's 2020-2024 corpus window and are not covered in the seven families.","The survey's emphasis on architecture families rather than training paradigms suggests an orthogonal classification by learning signal, such as implicit versus explicit feedback or centralized versus federated training, might also be informative.","If the collected corpus is representative, the clustering of papers around GNNs and autoencoders indicates where the field's momentum currently lies, a trend the paper only partially makes explicit."],"forward_implications":["Practitioners can use the seven-family taxonomy to choose an architecture based on their data type: sequential data points to RNNs, graph-structured interactions point to GNNs, and reconstruction-based learning points to autoencoders.","The consolidated dataset and metric tables give a standard benchmark set for evaluating new deep-neural collaborative-filtering models.","The paper's challenge list, including sparsity, cold start, interpretability, privacy, and adversarial robustness, defines a concrete research agenda for the field.","The survey's claim that DNNs capture nonlinear relationships implies that purely linear methods are expected to underperform on complex interaction data, guiding future method comparisons."],"supporting_citations":[{"why":"Supplies the prior deep-learning recommender survey that this paper says covers content-based systems rather than focusing on collaborative filtering.","marker":"[85]"},{"why":"A prior graph-neural-network recommender survey cited as one of the single-model surveys this work broadens.","marker":"[86]"},{"why":"Another prior graph-neural-network recommender survey, used to illustrate the single-architecture limitation in existing literature.","marker":"[87]"},{"why":"A prior CNN recommender survey, cited to show that existing surveys cover only one deep-learning architecture family.","marker":"[88]"},{"why":"The neural collaborative filtering model that defines the transition from matrix factorization to neural interaction modeling in this survey.","marker":"[90]"},{"why":"A comparison of neural collaborative filtering and matrix factorization that grounds the discussion of traditional CF limitations.","marker":"[89]"},{"why":"A foundational survey of collaborative, content-based, and hybrid recommenders that frames the paper's taxonomy of recommender types.","marker":"[78]"}],"fun_headline_variants":["Seven DNN families mapped for collaborative filtering","Deep learning survey bridges CF and seven neural architectures","Survey unifies seven deep learning styles for recommenders","From MLP to GAN: deep learning for recommendations","How seven DNN families tackle sparse recommendation data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's usefulness depends on the assumption that the roughly eighty papers it collected over 2020 to 2024 constitute a representative sample of deep-learning collaborative-filtering research, since no search strings, screening rules, or quality filters are reported.","fun_headline_variants_meta":{"raw":{"variants":["Seven DNN families mapped for collaborative filtering","Deep learning survey bridges CF and seven neural architectures","Survey unifies seven deep learning styles for recommenders","From MLP to GAN: deep learning for recommendations","How seven DNN families tackle sparse recommendation data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000333,"raw_usage":{"total_tokens":1787,"prompt_tokens":819,"completion_tokens":968,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":435,"completion_tokens_details":{"reasoning_tokens":909}},"tokens_in":435,"tokens_out":968,"duration_ms":7157,"temperature":1.0,"reasoning_tokens":909,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:24:07.515205+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader who re-runs the literature search with explicit queries and inclusion criteria and finds that major, frequently cited DNN-CF models are missing from the survey's seven categories would show that the coverage claim does not hold; alternatively, checking whether the cited papers' architectures match their placement in the taxonomy would reveal the accuracy of the categorization.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the prior deep-learning recommender survey that this paper says covers content-based systems rather than focusing on collaborative filtering."},{"cited_title":"& Li, Y.(2023).Asurvey of graph neural networks for recommender systems: Challenges, methods, and directions","cited_arxiv_id":null,"evidence_quote":"A prior graph-neural-network recommender survey cited as one of the single-model surveys this work broadens."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Another prior graph-neural-network recommender survey, used to illustrate the single-architecture limitation in existing literature."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A prior CNN recommender survey, cited to show that existing surveys cover only one deep-learning architecture family."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The neural collaborative filtering model that defines the transition from matrix factorization to neural interaction modeling in this survey."},{"cited_title":"(2020, September)","cited_arxiv_id":null,"evidence_quote":"A comparison of neural collaborative filtering and matrix factorization that grounds the discussion of traditional CF limitations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A foundational survey of collaborative, content-based, and hybrid recommenders that frames the paper's taxonomy of recommender types."}],"review_version":1}