REVIEW 3 major objections 4 minor 18 references
CosRec: 2D Convolutional Neural Networks for Sequential Recommendation
T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read CosRec claims that encoding item histories as pairwise tensors lets a standard 2D CNN outperform prior sequential recommenders on MovieLens-1M and Gowalla.
desk verdict A clean, small architectural idea with a credible ablation, but the 'state-of-the-art' claim outruns the experiments because SASRec and BERT4Rec are cited but never run and there are no repeated trials. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the pairwise encoding module that converts an ordered sequence into an image-like three-way tensor. For the previous $L$ items, the tensor $T^L_{(u,t)}\in\mathbb{R}^{L\times L\times 2d}$ has entry $(i,j)$ equal to the concatenated embedding $[\mathbf{e}_i;\mathbf{e}_j]$, so every ordered pair of recent items becomes a position in a feature map. Standard 2D convolutional kernels then slide over this map; because pairs like (camera, lens) are adjacent in the tensor even when an unrelated item (bike) sits between them in the original sequence, small filters such as $3\times3$ can capture skip correlations that a chain-structured model would have to propagate through the intervening item. The same encoding turns the problem into one that can reuse ordinary CNN architectures, including stacking and padding.
What would settle it
Run the released code on the same 80/20 split and the same T=3 prediction protocol with the same hyperparameter search for every baseline, and check whether CosRec still exceeds Caser, GRU4Rec, and FPMC on both ML-1M and Gowalla; if the margins disappear under a matched protocol, the central claim would be refuted.
Extended reading notes
Core claim
On the paper's own terms, the discovery is that pairwise encoding of a user's recent history is a stronger substrate for sequential recommendation than the ordered embedding matrix used by Caser and NextItNet. CosRec constructs a tensor $T^L_{(u,t)}\in\mathbb{R}^{L\times L\times 2d}$ whose $(i,j)$-th vector is the concatenation $[\mathbf{e}_i;\mathbf{e}_j]$ of the embeddings of the $i$-th and $j$-th previous items, applies two blocks of $1\times1$ and $3\times3$ 2D convolutions, concatenates the resulting sequential vector with the user embedding, and predicts the next $T=3$ items under binary cross-entropy with negative sampling. The paper reports that this outperforms all tested baselines on MovieLens-1M and Gowalla, including a 25.0% MAP gain over Caser on ML-1M, and that even the MLP-based CosRec-base beats prior methods on ML-1M.
Load-bearing premise
The load-bearing premise is that the evaluation protocol is exactly comparable to Caser's—first 80% of each user's actions for training, last 20% for testing, with all baselines tuned under the same conditions; if any baseline was under-tuned or the metric computation differs, the reported margins would be overstated.
Editorial extensions
If this is right
- Sequential recommenders can stop relying on special vertical/horizontal or dilated 1D filter shapes and instead reuse standard 2D CNN architectures, including deeper stacks.
- Irrelevant or noisy 'skip' items in a user's history no longer need to interrupt the signal between two related items, because the pairwise tensor makes those items adjacent to each other.
- Because the architecture supports stacked convolutions with padding, it can be extended to longer histories and larger data streams than a single-layer vertical-filter design.
- Even a plain MLP over the pairwise encoding can outperform strong sequential baselines on dense data, suggesting that how the sequence is represented matters as much as the convolution operator.
Reading between the lines
- The pairwise tensor has $L^2$ entries, so CosRec's memory and compute grow quadratically with the Markov order $L$; a natural extension the paper leaves implicit is replacing the dense tensor with sparse or windowed pairs, which would keep the skip-capture property at much longer horizons.
- The reported gap is much larger on ML-1M than Gowalla, suggesting the benefit of pairwise encoding is tied to data density and reliable item co-occurrence; this predicts that on sparser or noisier datasets the margin over Caser would shrink, and regularizing the pairwise weights might restore it.
- Since CosRec-base already beats prior methods on ML-1M, a focused ablation varying the number of 2D convolution blocks and filter sizes could separate how much of the gain is the pairwise encoding and how much is the convolution itself; the paper does not report such an ablation.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes CosRec, a sequential recommendation model that encodes the last L items into an L x L x 2d tensor of pairwise concatenated item embeddings, applies standard 2D convolutions to this tensor, and conditions the resulting sequential representation on a user embedding for next-item prediction. The model is trained with binary cross-entropy loss and negative sampling. Experiments on ML-1M and Gowalla compare CosRec against PopRec, BPR, FMC, FPMC, GRU4Rec, Caser, and an MLP-based ablation CosRec-base, reporting MAP, Precision@N, and Recall@N. The authors claim state-of-the-art performance and include a filter visualization intended to show that pairwise encoding captures non-adjacent, "skip" dependencies.
Significance. If the empirical results are robust, the paper makes a useful empirical contribution: pairwise encoding followed by standard 2D convolution is a simple and plausible alternative to Caser-style vertical/horizontal filters and NextItNet-style 1D dilated convolutions. The authors should be credited for releasing code, following the Caser evaluation protocol, including an ablation of the 2D CNN module, and reporting metrics on two public benchmarks. However, the contribution is primarily empirical, and the central claim of state-of-the-art performance is currently supported only against the baselines that were actually run; the strongest cited recent sequence models are not evaluated, and no uncertainty quantification is provided.
major comments (3)
- [Section 4.2, Table 3] The abstract and Section 5 claim that CosRec 'outperforms both conventional methods and recent sequence-based approaches' and achieves 'state-of-the-art performance,' but SASRec and BERT4Rec, both cited in Section 2 as recent sequence models, are absent from Table 3. Because those models use different evaluation conventions in their original papers, published numbers cannot be substituted for a direct comparison under the protocol of Section 4.1.1. The claim should be either supported by direct same-protocol comparisons with these models or explicitly restricted to the evaluated baselines.
- [Section 4.2, Table 3] All metrics in Table 3 appear to be single-run point estimates; no error bars, confidence intervals, or significance tests are reported. Given that Section 4.2 describes the improvement over Caser as 'significant' (e.g., MAP 0.1883 vs. 0.1507 on ML-1M), the absence of variance information makes it impossible to determine whether the margins are beyond run-to-run noise. The authors should provide results over multiple random seeds and, where appropriate, paired significance tests.
- [Section 4.1.1 and 4.1.3] The experimental protocol is not fully specified in ways that directly affect comparability. The manuscript states that the first 80% of each user's actions are used for training and validation, but it does not state the validation split or early-stopping rule, nor does it report how the baseline hyperparameters were selected or tuned. In addition, the paper says the model predicts the next T=3 items at once, but it does not explain how multiple target timestamps per user are incorporated into the reported MAP, Precision@N, and Recall@N. Without these details, the reader must assume that the protocol is exactly the Caser protocol; if any part differs, the reported margins over the baselines could be overstated.
minor comments (4)
- [Eq. (1)] The formula contains 'loд' instead of '\log', and the notation 'j < Su' for sampled negative items is introduced without a formal definition; please clarify the negative sampling notation.
- [Section 3.2] The definition of the pairwise tensor says the (i,j)-th vector is the concatenation of embeddings of 'the item pair (i,j), i,j in (S^u_{t-L}, ..., S^u_{t-1})', which is imprecise because i and j are indices, not items; please index the tensor positions explicitly.
- [Table 1] The kernel sizes are listed as '1 x 1' and '3 x 3' but the first layer is described as 'enriching feature representations'; it would help to state explicitly that these are 2D kernels applied over the L x L spatial dimensions with 2d input channels.
- [Figure 3] The caption says 'Darker colors indicate higher values' while the colorbar shows values increasing from 0.000 to 0.027, but the actual colormap direction is not labeled on the figure; please clarify or label the colorbar.
Circularity Check
No significant circularity: CosRec's claims are empirical, measured on held-out test data against external baselines.
full rationale
This is an empirical architecture paper rather than a formal derivation, so the circularity patterns do not apply. The central claim is that CosRec outperforms existing baselines, and this is supported by Table 3, which reports MAP, Precision@N, and Recall@N on held-out test data (the last 20% of each user's actions) from two public datasets under a protocol explicitly inherited from the external Caser evaluation. No fitted parameter is repackaged as a prediction: hyperparameters such as latent dimension d, Markov order L, and learning rate are chosen by the authors, but the reported metrics are measured on a test split not used for training, and the baselines are external methods (Caser, GRU4Rec, FPMC, FMC, BPR, PopRec). The CosRec-base ablation is an architectural variant, not a circular validation. The self-citations in the reference list (e.g., SASRec and translation-based recommendation) concern related work and are not load-bearing justifications for the experimental outcome; the comparisons that support the central claim are against externally developed baselines. The absence of SASRec and BERT4Rec from Table 3, despite being mentioned in related work, is a legitimate completeness and comparability risk, but it is a correctness risk about the strength of the state-of-the-art claim, not a circularity: omitting a strong baseline does not make the reported numbers equal to the inputs by construction. No equation in the paper defines CosRec's output in terms of its evaluation metrics, and no uniqueness theorem or prior author result is invoked to force the architecture. Therefore the paper is self-contained in its empirical evaluation and the correct circularity score is 0.
Assumptions & free parameters
free parameters (7)
- item/user latent dimension d =
50 (ML-1M), 100 (Gowalla)
- Markov order L =
5
- negative sampling rate =
3
- dropout rate =
0.5
- learning rate =
0.001
- batch size =
512
- number of 2D convolution blocks =
2
assumptions (5)
- domain assumption A user's next action can be predicted from the previous L items alone.
- domain assumption Presence of a review or rating is treated as implicit positive feedback, and timestamps define sequence order.
- domain assumption The first 80% of each user's sequence is a valid training proxy for the last 20% (test) behavior.
- ad hoc to paper Concatenating the embeddings of item pairs into an LxLx2d tensor preserves the information needed for 2D convolutions to learn useful sequential patterns.
- standard math Adam with a fixed learning rate converges to a good local optimum for this architecture.
Cite this review
Pith. "Pith review of CosRec: 2D Convolutional Neural Networks for Sequential Recommendation." pith.science (2026). https://pith.science/paper/NNAGY6U5
@misc{pith2026190809972,
author = {Pith},
title = {Pith review of: CosRec: 2D Convolutional Neural Networks for Sequential Recommendation},
year = {2026},
howpublished = {\url{https://pith.science/paper/NNAGY6U5}},
note = {Machine review of arXiv:1908.09972}
}
read the original abstract
Sequential patterns play an important role in building modern recommender systems. To this end, several recommender systems have been built on top of Markov Chains and Recurrent Models (among others). Although these sequential models have proven successful at a range of tasks, they still struggle to uncover complex relationships nested in user purchase histories. In this paper, we argue that modeling pairwise relationships directly leads to an efficient representation of sequential features and captures complex item correlations. Specifically, we propose a 2D convolutional network for sequential recommendation (CosRec). It encodes a sequence of items into a three-way tensor; learns local features using 2D convolutional filters; and aggregates high-order interactions in a feedforward manner. Quantitative results on two public datasets show that our method outperforms both conventional methods and recent sequence-based approaches, achieving state-of-the-art performance on various evaluation metrics.
Figures
Reference graph
Works this paper leans on
-
[1]
Eunjoon Cho, Seth A Myers, and Jure Leskovec. 2011. Friendship and mobility: user movement in location-based social networks. In KDD
work page 2011
-
[2]
F Maxwell Harper and Joseph A Konstan. 2016. The movielens datasets: History and context. Acm transactions on interactive intelligent systems (TIIS) (2016)
work page 2016
- [3]
-
[4]
Xiangnan He, Xiaoyu Du, Xiang Wang, Feng Tian, Jinhui Tang, and Tat-Seng Chua. 2018. Outer Product-based Neural Collaborative Filtering. In IJCAI
2018
-
[5]
Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk
-
[6]
Wang-Cheng Kang, Chen Fang, Zhaowen Wang, and Julian McAuley. 2017. Visually-Aware Fashion Recommendation and Design with Generative Image Models. In ICDM
work page 2017
-
[7]
Wang-Cheng Kang and Julian McAuley. 2018. Self-Attentive Sequential Recom- mendation. In ICDM
work page 2018
-
[8]
Dong Hyun Kim, Chanyoung Park, Jinoh Oh, Sungyoung Lee, and Hwanjo Yu. 2016. Convolutional Matrix Factorization for Document Context-Aware Recommendation. In RecSys
work page 2016
Show all 18 references
-
[9]
Diederik Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimiza- tion. arXiv preprint arXiv:1412.6980 (2014)
2014 arXiv
-
[10]
Koren, R
Y. Koren, R. Bell, and C. Volinsky. 2009. Matrix factorization techniques for recommender systems. IEEE Computer (2009)
2009
-
[11]
Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme
-
[12]
Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme. 2010. Factor- izing personalized markov chains for next-basket recommendation. In WWW
2010
-
[13]
Fei Sun, Jun Liu, Jian Wu, Changhua Pei, Xiao Lin, Wenwu Ou, and Peng Jiang
-
[14]
Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommendation via convolutional sequence embedding. In WSDM
2018
-
[15]
Jose, and Xiangnan He
Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis, Joemon M. Jose, and Xiangnan He. 2019. A Simple Convolutional Generative Network for Next Item Recommendation. In WSDM
2019
-
[2009]
BPR: Bayesian personalized ranking from implicit feedback. In UAI
-
[2015]
Session-based recommendations with recurrent neural networks. In ICLR
-
[2019]
arXiv preprint arXiv:1904.06690 (2019)
BERT4Rec: Sequential Recommendation with Bidirectional Encoder Repre- sentations from Transformer. arXiv preprint arXiv:1904.06690 (2019)
2019 arXiv
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.