Cross-Domain Collaborative Filtering via Translation-based Learning

Dimitrios Rafailidis

REVIEW 4 major objections 4 minor 19 references

Cross-Domain Collaborative Filtering via Translation-based Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper proposes that modeling cross-domain user preferences as translation vectors in a shared embedding space improves recommendation accuracy, and reports gains on all six Amazon tasks.

desk verdict A plausible new combination of translation-based learning and co-clustering for cross-domain recommendation, but the evaluation has a potentially load-bearing test-leakage ambiguity. read the letter →

arxiv 1908.06169 v1 pith:UCSJSKJ5 submitted 2019-08-11 cs.IR cs.LGstat.ML

Dimitrios Rafailidis This is my paper · ORCID

classification cs.IRcs.LGstat.ML

keywords cross-domainrecommendationtranslation-basedlearningcollaborativefilteringfactorizationmachinesdeepBayesianpersonalizedrankinguserco-clusteringAmazondataset

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that cross-domain recommendation works better when knowledge transfer is modeled as translation in an embedding space rather than as shared latent factors or raw feature concatenation. The proposed CDT model—named for cross-domain translation—learns an embedding vector and a translation vector for each feature, so that an item in a source domain lies near the target-domain item shifted by the learned offset; this makes similarity propagate transitively to pairs with no observed interaction. A deep variant, DeepCDT, stacks hidden layers on top of the translation-aware factorization machine to capture nonlinear user behavior. On six Amazon cross-domain tasks with partial user overlap, the paper reports that DeepCDT attains the highest recall in every task and beats all compared baselines at p<0.05. If the claim is right, recommenders can exploit sparse ratings in auxiliary domains to rank items in a target domain without relying on item overlap or dense shared features.

What carries the argument

The load-bearing mechanism is the translation-based factorization machine built on cluster-weighted source features. Each feature dimension receives an embedding vector $v$ and a translation vector $v'$, and the relation $v_i+v'_i \approx v_h$ states that item $h$ in a source domain is a neighbor of item $i$ in the target domain shifted by the learned offset. Replacing inner products with $d^2(v_i+v'_i, v_h)$ makes interaction strength behave like a metric, so transitive similarity propagates to feature pairs with no observed interaction. User co-clustering supplies the cross-domain weights: the matrices $C^{(p)}$ and $C^{(t)}$ come from graph Laplacian clustering, $Y^{(pt)}$ is learned under an $L_{2,1}$ penalty with orthogonality, and $Q^{(pt)}$ averages cluster similarities over users who interacted with each source item. DeepCDT feeds these translated features through five hidden layers trained by mini-batch Adam.

What would settle it

Re-run the six Amazon tasks while computing the cluster matrices $C^{(t)}$ and the similarity matrix $Q^{(pt)}$ strictly from the 50% training split, then compare recall with the reported values; if DeepCDT no longer beats ScoNet and CDCF, the advantage depends on leakage rather than on translation.

Watch

Extended reading notes

Core claim

The central discovery claimed is that replacing the inner product of a factorization machine with a squared Euclidean distance between translated embeddings creates a transitive similarity structure across domains: for observed cross-domain feature pairs with high interaction weight, unobserved neighboring pairs become close as well. The paper embodies this in the model equation $\hat{y}(x^{(t)}) = w_0 + \sum_i w_i x^{(t)}_i + \sum_{h>i} d^2(v_i+v'_i, v_h) x^{(t)}_i x^{(t)}_h$, trained with a Bayesian Personalized Ranking objective and negative sampling. Cross-domain knowledge enters through user co-clustering: graph Laplacian clustering produces user-cluster matrices in each domain, an $L_{2,1}$-sparse co-clustering step yields cluster similarities, and the matrix $Q^{(pt)}$ reweights source-domain ratings before they are appended to the target feature vector. The authors report that on six Amazon tasks, DeepCDT—the five-hidden-layer version—achieves the best recall on all six domains and improves significantly over every competitor at $p<0.05$.

Load-bearing premise

The reported gains assume that the user cluster assignments and cross-domain similarity weights are computed from training ratings only, with no information from the target-domain test set entering the feature construction.

Editorial extensions

If this is right

Source-domain ratings become usable ranking evidence in the target domain even with partial user overlap, because each source interaction is weighted by the cluster-level similarity between users.
Unobserved cross-domain item pairs inherit closeness from observed pairs through the translation relation, so the model can rank items that have no direct interaction evidence.
The comparison between DeepCDT and the deep single-domain baseline indicates that translating the feature input before feeding a deep network contributes to the reported gain.
Because the objective is a Bayesian Personalized Ranking loss with negative sampling, the method extends naturally to implicit feedback and top-n ranking, not only explicit ratings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct test the paper does not run: randomizing the entries of $Q^{(pt)}$ while keeping the same embeddings should push recall back toward the single-domain baselines; if it does not, the cross-domain weighting is not the active ingredient.
The learned translation vectors could transfer to sequential cross-domain recommendation, which the paper lists as future work, by treating the offset from a user's last item to the next item as the same kind of translation.
The evaluation does not isolate cold-start users, so testing on users with very few target-domain ratings would show whether the transfer helps exactly where the paper motivates it.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Let me give you my take on this one. The paper proposes CDT and its deep variant DeepCDT, which apply translation-based metric learning to cross-domain collaborative filtering. The key idea is to compute user clusters in source and target domains, build a cluster-level similarity matrix, and use that to add source-domain features to target interactions. Then a translation vector per target item lets the model learn embeddings such that v_i + v'_i ≈ v_h for related target and source items. That is a new combination of existing ideas, and it's not a trivial one: the cluster-weighted feature augmentation is a reasonable way to transfer knowledge under partial user overlap. The reported gains over FM, CML, FNN, CBT, CLFM, CDCF, and ScoNet on the six Amazon tasks are sizable, especially for DeepCDT.

The problem is the empirical section. The biggest question is whether the user clusters in the target domain are computed on the training split only. Section II-A and the evaluation setup in Section III-A2 never say. If the graph-Laplacian clusters use the full target rating matrix, then the features in Eq. (3) encode test ratings, and the BPR ranking is trained on leaked information. That would make the reported improvements artifacts. The ambiguity is easy to miss, and it is load-bearing because the whole case for the method is empirical.

There are other soft spots. The co-clustering objective in Eq. (1) has no solver described; you don't know how Y is optimized under the orthogonality constraint. The translation vector is defined only for target items; what happens for user or source features is left vague. No error bars are reported, and the paired t-test over five runs is not enough to back the strong significance claim. The dataset citation points to a paper about viral marketing, not the Amazon review dataset, which is a red flag. No code is released.

Despite these, the modeling idea is coherent and the literature is cited fairly. The paper deserves a serious referee, but only if the authors can clarify the clustering split and otherwise tighten the evaluation. As written, I would not recommend acceptance in its current form. It might be a useful comparison point for cross-domain metric learning work, but I wouldn't cite it until the leakage question is resolved.

Referee Report

4 major / 4 minor

Summary. The paper proposes two cross-domain collaborative filtering models, CDT and DeepCDT. CDT computes user cluster assignments per domain via a graph-Laplacian method, builds a cluster-level similarity matrix Y(pt) by optimizing a co-clustering objective, and then augments target-domain feature vectors with source-domain interactions weighted by the resulting matrix Q(pt). The model is trained with a Bayesian Personalized Ranking objective using a translation-based factorization-machine scoring function, where a translation vector is learned so that item embeddings in the target and source domains satisfy an approximate additive relation. DeepCDT extends CDT by feeding the translation-based features into a deep neural network. The experiments on six Amazon domains report recall and NDCG improvements over single-domain and cross-domain baselines, with a claim that DeepCDT is statistically superior to all competitors at p<0.05.

Significance. If the claims are correct, the paper contributes a plausible extension of translation-based metric learning to cross-domain recommendation and demonstrates its usefulness on a standard public dataset against a reasonable set of baselines. The proposed formulation is clearly stated at a high level, and the use of a ranking objective is appropriate for top-n recommendation. However, the empirical central claim currently rests on a protocol ambiguity that could inflate the reported gains, and several model components are under-specified, so the results are not yet reproducible as reported. The contribution is potentially useful but requires clarification and experimental verification before it can be accepted.

major comments (4)

[II-A and III-A2] The paper never states whether the user cluster matrices C(t) and C(p), and consequently the cross-domain similarity matrix Q(pt) in Eq. (2), are computed on the training split only or on the full target rating matrix. Section III-A2 defines a 50/10/40 train/validation/test split, but the text only says that models are trained on 50% of the target domain and all source ratings. If C(t) is computed from the full R(t), then the augmented feature vector x(t) in Eq. (3) contains information about target test items through Q(pt)(h,u), and the recall/NDCG comparisons in Table II would be inflated. This is a load-bearing issue because the central empirical claim is the superiority of CDT and DeepCDT. Please state explicitly that all clustering and similarity computations are restricted to the training and validation splits, and if this is not currently the case, rerun the experiments with the corrected protocol.
[II-A, Eq. (1)] The co-clustering objective in Eq. (1) is stated as a constrained optimization problem with an L2,1 penalty, an orthogonality constraint, and a nonnegativity constraint, but no solver, initialization, or stopping criterion is described. Since Y(pt) feeds directly into Eq. (2) and therefore into the model features, the method cannot be reproduced without knowing how Eq. (1) is minimized. Please provide the optimization algorithm, the update rules or the library used, and the value of lambda.
[II-B4 and III-B] DeepCDT is only sketched: the paper states that Eq. (4) is the bottom layer, that there are e=5 hidden layers, and that mini-batch Adam with batch size 512 and learning rate 1e-4 is used, but it does not specify activation functions, layer dimensions, the mapping from the factorization-machine output to the hidden layers, or how the final prediction is computed from the deep network. Given that Table II reports DeepCDT as the best method on all six tasks, the model is under-specified to a degree that prevents independent verification. Please provide the full architecture and training details.
[III-B, Table II] Table II reports only point estimates, and the text contains one general statement that DeepCDT is superior for p<0.05 without reporting variances, the number of paired observations, or whether the test was applied to all pairwise comparisons or only to DeepCDT versus each baseline. Because some differences, such as CDT versus ScoNet, are described as comparable, the reader cannot assess which reported improvements are meaningful. Please report standard deviations or confidence intervals across the five repeated runs and specify the exact paired t-test procedure.

minor comments (4)

[II-A, Eq. (2)] The notation C(p)(k) and C(t)(u) is ambiguous: it is not clear whether these denote cluster indices, row vectors, or one-hot assignments, and the expression Y(pt)(C(p)(k), C(t)(u)) should be defined explicitly.
[II-B2] There is a typo in 'according to a metric-based metric' and another in 'in the the latent space' in the same section; please correct these.
[III-A2] The description of the training set says 'we trained the examined models on the 50% of the target domain and all the ratings of the source domains as training set'; it would be clearer to state explicitly whether the source-domain ratings are all used in their entirety and how the validation split is used for early stopping or model selection.
[II-B4] The sentence 'we used five negative samples for each positive/observed sample in the target domain t' reports a fixed choice without showing the sensitivity analysis that would justify it; since negative sampling affects the BPR objective, a small ablation or reference would strengthen the presentation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: CDT/DeepCDT are empirical ranking models whose components come from external prior work, and no fitted input is renamed as a prediction.

full rationale

The paper's central claim is empirical: CDT and DeepCDT outperform baselines on held-out test ratings (Section III-B, Table II). The derivation chain is the feature vector x(t) in Eq. (3), built from the co-clustering similarity matrix Q(pt) of Eq. (2), which in turn is fitted to the user-overlap matrix O(pt) in Eq. (1). The ranking objective in Eq. (6) then learns embedding and translation vectors on the training interactions. There is no equation in which a predicted quantity coincides by construction with an input, and no parameter fitted to a subset is reported as a prediction of that subset. The translation-based idea is attributed to external work [9]-[11], the graph-Laplacian clustering to [12], and the deep variant to [15]; the author's own prior work [5], [6], [17] appears only as background and as an evaluation detail, so it is not load-bearing. The one legitimate concern is not circularity but potential leakage: Section III-A2 defines a 50/10/40 train/validation/test split, but Section II-A does not explicitly say the graph-Laplacian user clusters C(t) are computed from the training split only. If the full target matrix were clustered, Q(pt) would encode test information. That would be an experimental validity problem, not a self-referential derivation, and no text in the paper asserts that full-matrix clustering was used.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The model introduces no new physical or conceptual entities; all components are standard latent vectors, translation vectors, and cluster matrices. The main free parameters are hyperparameters tuned on validation data, so they are not independent evidence. The clustering and distance assumptions are borrowed from prior work.

free parameters (6)

Number of user clusters per domain (cp, ct) = not reported
Tuned via cross-validation; affects co-clustering matrix Y(pt) and feature weights in Eq. (2).
Latent dimension q = 10 to 100 by step 10, selected by grid search
Embedding and translation vector dimension in Eq. (4), chosen via cross-validation.
Regularization weight lambda for L2,1 norm = not reported
In Eq. (1) forces sparsity in Y(pt).
L2 regularization Omega(Theta) = not reported
In Eq. (6) to penalize model parameters.
Number of negative samples = 5
Set to five per positive sample based on pilot experiments (Section II-B3).
DeepCDT hidden layers, batch size, learning rate = e=5, batch=512, lr=1e-4
Fixed in Section II-B4 without sensitivity analysis.

assumptions (4)

domain assumption The graph Laplacian method produces meaningful user clusters in each domain.
Used to build C(p) and C(t) in Section II-A; no validation of cluster quality.
domain assumption Euclidean distance with translation vectors captures interaction strength between features.
Core modeling choice in Eq. (4), borrowed from CML [9] and translation-based FMs [10, 11].
ad hoc to paper The co-clustering objective in Eq. (1) has a computable solution with the given constraints.
No algorithm or solver is provided; the paper only says it follows a co-clustering strategy.
domain assumption Partial user overlap across domains is sufficient for knowledge transfer.
The paper assumes partial overlaps as in [7], but does not quantify the overlap in the six tasks.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cross-Domain Collaborative Filtering via Translation-based Learning." pith.science (2026). https://pith.science/paper/UCSJSKJ5

@misc{pith2026190806169,
  author       = {Pith},
  title        = {Pith review of: Cross-Domain Collaborative Filtering via Translation-based Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UCSJSKJ5}},
  note         = {Machine review of arXiv:1908.06169}
}

read the original abstract

With the proliferation of social media platforms and e-commerce sites, several cross-domain collaborative filtering strategies have been recently introduced to transfer the knowledge of user preferences across domains. The main challenge of cross-domain recommendation is to weigh and learn users' different behaviors in multiple domains. In this paper, we propose a Cross-Domain collaborative filtering model following a Translation-based strategy, namely CDT. In our model, we learn the embedding space with translation vectors and capture high-order feature interactions in users' multiple preferences across domains. In doing so, we efficiently compute the transitivity between feature latent embeddings, that is if feature pairs have high interaction weights in the latent space, then feature embeddings with no observed interactions across the domains will be closely related as well. We formulate our objective function as a ranking problem in factorization machines and learn the model's parameters via gradient descent. In addition, to better capture the non-linearity in user preferences across domains we extend the proposed CDT model by using a deep learning strategy, namely DeepCDT. Our experiments on six publicly available cross-domain tasks demonstrate the effectiveness of the proposed models, outperforming other state-of-the-art cross-domain strategies.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 18 canonical work pages

[1]

Factorization machines

Steffen Rendle. Factorization machines. In ICDM, pages 995–1000, 2010

work page 2010
[2]

Per- sonalized keyword boosting for venue suggestion based on multiple lbsns

Mohammad Aliannejadi, Dimitrios Rafailidis, and Fabio Crestani. Per- sonalized keyword boosting for venue suggestion based on multiple lbsns. In Advances in Information Retrieval - 39th European Conference on IR Research, ECIR 2017, Aberdeen, UK, April 8-13, 2017 , pages 291–303, 2017

work page 2017
[3]

Can movies and books collaborate? cross-domain collaborative ﬁltering for sparsity reduction

Bin Li, Qiang Yang, and Xiangyang Xue. Can movies and books collaborate? cross-domain collaborative ﬁltering for sparsity reduction. In IJCAI, pages 2052–2057, 2009

work page 2009
[4]

Cross-domain collaborative ﬁltering with factorization machines

Babak Loni, Yue Shi, Martha Larson, and Alan Hanjalic. Cross-domain collaborative ﬁltering with factorization machines. In ECIR, pages 656– 661, 2014

work page 2014
[5]

Top-n recommendation via joint cross-domain user clustering and similarity learning

Dimitrios Rafailidis and Fabio Crestani. Top-n recommendation via joint cross-domain user clustering and similarity learning. In Machine Learning and Knowledge Discovery in Databases - European Confer- ence, ECML PKDD 2016, Riva del Garda, Italy, September 19-23, 2016, Proceedings, Part II , pages 426–441, 2016

work page 2016
[6]

A collaborative ranking model for cross-domain recommendations

Dimitrios Rafailidis and Fabio Crestani. A collaborative ranking model for cross-domain recommendations. In Proceedings of the ACM Con- ference on Information and Knowledge Management, CIKM, Singapore, November 06 - 10, 2017 , pages 2263–2266, 2017

work page 2017
[7]

Conet: Collaborative cross networks for cross-domain recommendation

Guangneng Hu, Yu Zhang, and Qiang Yang. Conet: Collaborative cross networks for cross-domain recommendation. In CIKM, pages 667–676, 2018

work page 2018
[8]

Cross-domain recommendation via cluster-level latent factor model

Sheng Gao, Hao Luo, Da Chen, Shantao Li, Patrick Gallinari, and Jun Guo. Cross-domain recommendation via cluster-level latent factor model. In ECML PKDD , pages 161–176, 2013

work page 2013

Show all 19 references

[9]

Belongie, and Deborah Estrin

Cheng-Kang Hsieh, Longqi Yang, Yin Cui, Tsung-Yi Lin, Serge J. Belongie, and Deborah Estrin. Collaborative metric learning. In WWW, pages 193–201, 2017

2017
[10]

Translation-based factorization machines for sequential recommendation

Rajiv Pasricha and Julian McAuley. Translation-based factorization machines for sequential recommendation. In RecSys, pages 63–71, 2018

2018
[11]

Translation-based recommendation

Ruining He, Wang-Cheng Kang, and Julian McAuley. Translation-based recommendation. In RecSys, pages 161–169, 2017

2017
[12]

Laplacian sparse coding, hypergraph laplacian sparse coding, and applications

Shenghua Gao, Ivor Wai-Hung Tsang, and Liang-Tien Chia. Laplacian sparse coding, hypergraph laplacian sparse coding, and applications. IEEE Trans. Pattern Anal. Mach. Intell. , 35(1):92–104, 2013

2013
[13]

Cross-domain recommender systems

Paolo Cremonesi, Antonio Tripodi, and Roberto Turrin. Cross-domain recommender systems. In ICDMW, pages 496–503, 2011

2011
[14]

Bpr: Bayesian personalized ranking from implicit feedback

Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt-Thieme. Bpr: Bayesian personalized ranking from implicit feedback. In UAI, pages 452–461, 2009

2009
[15]

Deep learning over multi- ﬁeld categorical data - - A case study on user response prediction

Weinan Zhang, Tianming Du, and Jun Wang. Deep learning over multi- ﬁeld categorical data - - A case study on user response prediction. In ECIR, pages 45–57, 2016

2016
[16]

Adamic, and Bernardo A

Jure Leskovec, Lada A. Adamic, and Bernardo A. Huberman. The dynamics of viral marketing. TWEB, 1(1):5, 2007

2007
[17]

Learning to rank with trust and distrust in recommender systems

Dimitrios Rafailidis and Fabio Crestani. Learning to rank with trust and distrust in recommender systems. In RecSys, pages 5–13, 2017

2017
[18]

Link injection for boosting information spread in social networks

Stefanos Antaris, Dimitrios Rafailidis, and Alexandros Nanopoulos. Link injection for boosting information spread in social networks. Social Netw. Analys. Mining , 4(1):236, 2014

2014
[19]

Repeat consumption recommendation based on users preference dynamics and side informa- tion

Dimitrios Rafailidis and Alexandros Nanopoulos. Repeat consumption recommendation based on users preference dynamics and side informa- tion. In Proceedings of the 24th International Conference on World Wide Web Companion, WWW 2015, Florence, Italy, May 18-22, 2015 - Companion ...

2015

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Factorization machines

Steffen Rendle. Factorization machines. In ICDM, pages 995–1000, 2010

work page 2010

[2] [2]

Per- sonalized keyword boosting for venue suggestion based on multiple lbsns

Mohammad Aliannejadi, Dimitrios Rafailidis, and Fabio Crestani. Per- sonalized keyword boosting for venue suggestion based on multiple lbsns. In Advances in Information Retrieval - 39th European Conference on IR Research, ECIR 2017, Aberdeen, UK, April 8-13, 2017 , pages 291–303, 2017

work page 2017

[3] [3]

Can movies and books collaborate? cross-domain collaborative ﬁltering for sparsity reduction

Bin Li, Qiang Yang, and Xiangyang Xue. Can movies and books collaborate? cross-domain collaborative ﬁltering for sparsity reduction. In IJCAI, pages 2052–2057, 2009

work page 2009

[4] [4]

Cross-domain collaborative ﬁltering with factorization machines

Babak Loni, Yue Shi, Martha Larson, and Alan Hanjalic. Cross-domain collaborative ﬁltering with factorization machines. In ECIR, pages 656– 661, 2014

work page 2014

[5] [5]

Top-n recommendation via joint cross-domain user clustering and similarity learning

Dimitrios Rafailidis and Fabio Crestani. Top-n recommendation via joint cross-domain user clustering and similarity learning. In Machine Learning and Knowledge Discovery in Databases - European Confer- ence, ECML PKDD 2016, Riva del Garda, Italy, September 19-23, 2016, Proceedings, Part II , pages 426–441, 2016

work page 2016

[6] [6]

A collaborative ranking model for cross-domain recommendations

Dimitrios Rafailidis and Fabio Crestani. A collaborative ranking model for cross-domain recommendations. In Proceedings of the ACM Con- ference on Information and Knowledge Management, CIKM, Singapore, November 06 - 10, 2017 , pages 2263–2266, 2017

work page 2017

[7] [7]

Conet: Collaborative cross networks for cross-domain recommendation

Guangneng Hu, Yu Zhang, and Qiang Yang. Conet: Collaborative cross networks for cross-domain recommendation. In CIKM, pages 667–676, 2018

work page 2018

[8] [8]

Cross-domain recommendation via cluster-level latent factor model

Sheng Gao, Hao Luo, Da Chen, Shantao Li, Patrick Gallinari, and Jun Guo. Cross-domain recommendation via cluster-level latent factor model. In ECML PKDD , pages 161–176, 2013

work page 2013