pith. sign in

arxiv: 1907.01645 · v1 · pith:DGPFWN4Gnew · submitted 2019-06-29 · 💻 cs.IR · cs.LG· cs.SI· stat.ML

Adaptive Deep Learning of Cross-Domain Loss in Collaborative Filtering

Dimitrios Rafailidis , Gerhard Weiss This is my paper

Pith reviewed 2026-05-25 12:30 UTC · model grok-4.3

classification 💻 cs.IR cs.LGcs.SIstat.ML
keywords cross-domain recommendationcollaborative filteringadaptive deep learningloss balancinggradient magnitudesneural architecturebackpropagationuser behavior transfer
0
0 comments X

The pith

ADC balances cross-domain recommendation losses by tuning gradient magnitudes to each domain's complexity during backpropagation.

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

The paper presents ADC, a neural model for cross-domain collaborative filtering that learns non-linear user preferences across domains such as movies and music while transferring knowledge from multiple user accounts. It adds an algorithm that directly adjusts gradient magnitudes and learning rates on the fly based on the relative scales and complexities of the source domains. This mechanism lets the optimizer automatically control how much each domain influences the shared model parameters. Experiments across six public cross-domain tasks show ADC outperforming prior methods by handling domains that would otherwise dominate or under-contribute. A reader would care because real recommendation systems routinely face uneven domain sizes and user behavior shifts that standard multi-task losses fail to manage.

Core claim

The ADC model formulates a cross-domain loss function inside a neural architecture and introduces an efficient balancing algorithm that tunes gradient magnitudes and adapts learning rates according to domain complexities when training via backpropagation; this allows the model to control and adjust each domain's contribution to the parameter updates, resulting in effective knowledge transfer and superior performance on cross-domain recommendation tasks compared with state-of-the-art methods.

What carries the argument

The adaptive cross-domain loss balancing algorithm that tunes gradient magnitudes and adapts learning rates to domain complexities/scales during backpropagation.

Load-bearing premise

That directly tuning gradient magnitudes according to domain complexities will produce better optimization without introducing new biases or requiring extensive extra hyperparameter search.

What would settle it

Training ADC on the six public tasks and observing no measurable gain in recommendation accuracy or failure to equalize the effective contribution of domains with differing scales would falsify the central claim.

Figures

Figures reproduced from arXiv: 1907.01645 by Dimitrios Rafailidis, Gerhard Weiss.

Figure 1
Figure 1. Figure 1: An overview of the proposed ADC model. The MLP network with [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Effect on recall when varying the top-N recommendations. of our ADC model, demonstrating the importance of deep learning when generating cross-domain recommendations. In this set of experiments, we fix the number of negative samples |I− u (k) |=5 for each observed rating/sample of each domain, and the number of latent dimensions d=100. Notice that we keep the number of negative samples and latent dimension… view at source ↗
Figure 3
Figure 3. Figure 3: Effect on NDCG when varying the number of shared hidden layers [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect on NDCG when varying the asymmetry parameter γ in the loss function Lgrad. factors of ADC are (i) to capture the non-linear associations of user preferences across domains by formulating a joint cross-domain loss function in our deep learning strategy and (ii) to adjust and weigh the influence of each domain when optimizing the model parameters based on the domains’ complexities by applying an adapt… view at source ↗
read the original abstract

Nowadays, users open multiple accounts on social media platforms and e-commerce sites, expressing their personal preferences on different domains. However, users' behaviors change across domains, depending on the content that users interact with, such as movies, music, clothing and retail products. In this paper, we propose an adaptive deep learning strategy for cross-domain recommendation, referred to as ADC. We design a neural architecture and formulate a cross-domain loss function, to compute the non-linearity in user preferences across domains and transfer the knowledge of users' multiple behaviors, accordingly. In addition, we introduce an efficient algorithm for cross-domain loss balancing which directly tunes gradient magnitudes and adapts the learning rates based on the domains' complexities/scales when training the model via backpropagation. In doing so, ADC controls and adjusts the contribution of each domain when optimizing the model parameters. Our experiments on six publicly available cross-domain recommendation tasks demonstrate the effectiveness of the proposed ADC model over other state-of-the-art methods. Furthermore, we study the effect of the proposed adaptive deep learning strategy and show that ADC can well balance the impact of the domains with different complexities.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper proposes ADC, an adaptive deep learning model for cross-domain collaborative filtering. It presents a neural architecture and cross-domain loss function to capture non-linear user preferences and transfer knowledge across domains (e.g., movies, music). An algorithm is introduced to balance the loss by directly tuning gradient magnitudes and adapting learning rates according to domain complexities/scales during backpropagation, thereby controlling each domain's contribution. Experiments on six public cross-domain recommendation tasks report superiority over state-of-the-art methods, with additional analysis claiming that ADC effectively balances domains of differing complexities.

Significance. If the adaptive balancing mechanism is demonstrated to be the causal driver of the gains (rather than architecture or tuning alone), the approach could offer a practical, automated solution for multi-domain recommendation systems where domains vary in scale and complexity, reducing reliance on manual hyperparameter search while improving preference transfer.

major comments (2)
  1. [Abstract] Abstract: the central claim that the algorithm 'directly tunes gradient magnitudes and adapts the learning rates based on the domains' complexities/scales' and thereby 'controls and adjusts the contribution of each domain' rests on an undefined procedure for quantifying or measuring those complexities/scales; without an explicit definition, formula, or measurement method, it is impossible to verify that the adaptation differs from standard per-domain learning-rate search or to reproduce the balancing effect.
  2. [Abstract] Abstract: the reported superiority on six tasks and the claim of effective domain balancing require ablation experiments that isolate the adaptive gradient-tuning component against a non-adaptive baseline (shared architecture with fixed or grid-searched learning rates); absent such controls, the causal link between the proposed balancing algorithm and the performance gains cannot be established.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's insightful comments on the abstract and the need for clearer definitions and stronger experimental controls. We will revise the manuscript to address these points.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the algorithm 'directly tunes gradient magnitudes and adapts the learning rates based on the domains' complexities/scales' and thereby 'controls and adjusts the contribution of each domain' rests on an undefined procedure for quantifying or measuring those complexities/scales; without an explicit definition, formula, or measurement method, it is impossible to verify that the adaptation differs from standard per-domain learning-rate search or to reproduce the balancing effect.

    Authors: We agree that the abstract does not include an explicit definition or formula for quantifying domain complexities/scales. The body of the paper describes the adaptive algorithm, but to enhance clarity and reproducibility as noted, we will revise the abstract to provide a brief definition of the measurement method and how it differs from standard approaches. revision: yes

  2. Referee: [Abstract] Abstract: the reported superiority on six tasks and the claim of effective domain balancing require ablation experiments that isolate the adaptive gradient-tuning component against a non-adaptive baseline (shared architecture with fixed or grid-searched learning rates); absent such controls, the causal link between the proposed balancing algorithm and the performance gains cannot be established.

    Authors: We acknowledge that the current experiments, while showing superiority and including analysis of the adaptive strategy, do not include the specific ablation against a non-adaptive baseline with grid-searched learning rates. We will add these ablation experiments in the revised manuscript to better establish the causal contribution of the adaptive balancing algorithm. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper introduces a neural architecture and cross-domain loss, then describes an adaptive balancing algorithm that tunes gradient magnitudes and learning rates according to domain complexities/scales. No quoted equations or steps reduce the balancing mechanism to a fitted parameter or self-citation by construction; the algorithm is presented as a novel procedure whose effect is validated through experiments on six tasks. The central claim of improved balancing and superiority over SOTA rests on empirical results rather than definitional equivalence or load-bearing self-citation. The derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no specific free parameters, axioms, or invented entities can be identified from the provided information.

pith-pipeline@v0.9.0 · 5729 in / 993 out tokens · 52835 ms · 2026-05-25T12:30:19.596029+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · 1 internal anchor

  1. [1]

    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, Aberdeen, UK, April 8-13, pages 291–303, 2017

    work page 2017

  2. [2]

    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 Learn- ing and Knowledge Discovery in Databases - European Conference, ECML PKDD, Riva del Garda, Italy, September 19-23 Proceedings, Part II, pages 426–441, 2016

    work page 2016

  3. [3]

    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 Interna- tional Conference on Information and Knowledge Management, CIKM, Singapore, November 06 - 10 , pages 2263–2266, 2017

    work page 2017

  4. [4]

    Performance of recommender algorithms on top-n recommendation tasks

    Paolo Cremonesi, Yehuda Koren, and Roberto Turrin. Performance of recommender algorithms on top-n recommendation tasks. In Proceed- ings of the ACM Conference on Recommender Systems RecSys , pages 39–46, 2010

    work page 2010

  5. [5]

    Herlocker, Joseph A

    Jonathan L. Herlocker, Joseph A. Konstan, Loren G. Terveen, and John Riedl. Evaluating collaborative filtering recommender systems. ACM Transaction on Information Systems , 22(1):5–53, 2004

    work page 2004

  6. [6]

    Bell, and Chris V olinsky

    Yehuda Koren, Robert M. Bell, and Chris V olinsky. Matrix factorization techniques for recommender systems. IEEE Computer , 42(8):30–37, 2009

    work page 2009

  7. [7]

    Nonlinear dimensionality reduction for efficient and effective audio similarity searching

    Dimitrios Rafailidis, Alexandros Nanopoulos, and Yannis Manolopou- los. Nonlinear dimensionality reduction for efficient and effective audio similarity searching. Multimedia Tools Appl., 51(3):881–895, 2011

    work page 2011

  8. [8]

    Cluster-based joint matrix fac- torization hashing for cross-modal retrieval

    Dimitrios Rafailidis and Fabio Crestani. Cluster-based joint matrix fac- torization hashing for cross-modal retrieval. In Proceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval, SIGIR, Pisa, Italy, July 17-21 , pages 781–784, 2016

    work page 2016

  9. [9]

    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 Proceedings of the International Joint Conference on Artificial Intelligence, IJCAI, pages 2052–2057, 2009

    work page 2052

  10. [10]

    Cross-domain recommender systems

    Paolo Cremonesi, Antonio Tripodi, and Roberto Turrin. Cross-domain recommender systems. In Proceedings of the IEEE International Con- ferencecon Data Mining Workshops (ICDMW) , pages 496–503, 2011

    work page 2011

  11. [11]

    Nima Mirbakhsh and Charles X. Ling. Improving top-n recommendation for cold-start users via cross-domain information. ACM Transactions on Knowledge Discovery from Data , 9(4):33:1–33:19, 2015

    work page 2015

  12. [12]

    Cross domain recommender systems: A systematic literature review

    Muhammad Murad Khan, Roliana Ibrahim, and Imran Ghani. Cross domain recommender systems: A systematic literature review. ACM Comput. Surv., 50(3):36:1–36:34, 2017

    work page 2017

  13. [13]

    Distributed collaborative filtering with domain specialization

    Shlomo Berkovsky, Tsvi Kuflik, and Francesco Ricci. Distributed collaborative filtering with domain specialization. In Proceedings of the ACM Conference on Recommender Systems RecSys , pages 33–40, 2007

    work page 2007

  14. [14]

    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 Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases, ECML PKDD, Part II, pages 161–176, 2013

    work page 2013

  15. [15]

    Transfer learning in collaborative filtering for sparsity reduction

    Weike Pan, Evan Wei Xiang, Nathan Nan Liu, and Qiang Yang. Transfer learning in collaborative filtering for sparsity reduction. In Proceedings of the AAAI Conference on Artificial Intelligence, AAAI , 2010

    work page 2010

  16. [16]

    Personalized recommendation via cross-domain triadic factor- ization

    Liang Hu, Jian Cao, Guandong Xu, Longbing Cao, Zhiping Gu, and Can Zhu. Personalized recommendation via cross-domain triadic factor- ization. In Proceedings of the ACM International Conference on World Wide Web, WWW, pages 595–606, 2013

    work page 2013

  17. [17]

    Cross-domain collaborative filtering with factorization machines

    Babak Loni, Yue Shi, Martha Larson, and Alan Hanjalic. Cross-domain collaborative filtering with factorization machines. In Proceedings of the European Conference on Information Retrieval, ECIR , pages 656–661, 2014

    work page 2014

  18. [18]

    A multi-view deep learning approach for cross domain user modeling in recommen- dation systems

    Ali Mamdouh Elkahky, Yang Song, and Xiaodong He. A multi-view deep learning approach for cross domain user modeling in recommen- dation systems. In Proceedings of the ACM International Conference on World Wide Web, WWW , pages 278–288, 2015

    work page 2015

  19. [19]

    Conet: Collaborative cross networks for cross-domain recommendation

    Guangneng Hu, Yu Zhang, and Qiang Yang. Conet: Collaborative cross networks for cross-domain recommendation. In Proceedings of the ACM International Conference on Information and Knowledge Management, CIKM, pages 667–676, 2018

    work page 2018

  20. [20]

    Multi-Task Feature Learning for Knowledge Graph Enhanced Recommendation

    Hongwei Wang, Fuzheng Zhang, Miao Zhao, Wenjie Li, Xing Xie, and Minyi Guo. Multi-task feature learning for knowledge graph enhanced recommendation. CoRR, abs/1901.08907, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1901

  21. [21]

    A joint many-task model: Growing a neural network for multiple NLP tasks

    Kazuma Hashimoto, Caiming Xiong, Yoshimasa Tsuruoka, and Richard Socher. A joint many-task model: Growing a neural network for multiple NLP tasks. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, EMNLP , pages 1923–1933, 2017

    work page 2017

  22. [22]

    Deep neural networks employing multi-task learning and stacked bottleneck features for speech synthesis

    Zhizheng Wu, Cassia Valentini-Botinhao, Oliver Watts, and Simon King. Deep neural networks employing multi-task learning and stacked bottleneck features for speech synthesis. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP, pages 4460–4464, 2015

    work page 2015

  23. [23]

    Girshick

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross B. Girshick. Mask R-CNN. In Proceedings of the IEEE International Conference on Computer Vision, ICCV , pages 2980–2988, 2017

    work page 2017

  24. [24]

    YOLO9000: better, faster, stronger

    Joseph Redmon and Ali Farhadi. YOLO9000: better, faster, stronger. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 6517–6525, 2017

    work page 2017

  25. [25]

    Multi-task learning using uncertainty to weigh losses for scene geometry and semantics

    Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceeding ot the IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 7482–7491, 2018

    work page 2018

  26. [26]

    Cross-stitch networks for multi-task learning

    Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, CVPR , pages 3994–4003, 2016

    work page 2016

  27. [27]

    Factorization machines with libfm

    Steffen Rendle. Factorization machines with libfm. ACM Transactions on Intelligent Systems and Technology , 3(3):57:1–57:22, 2012

    work page 2012

  28. [28]

    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 Proceedings of the Conference on Uncertainty in Artificial Intelligence, UAI, pages 452–461, 2009

    work page 2009

  29. [29]

    Adamic, and Bernardo A

    Jure Leskovec, Lada A. Adamic, and Bernardo A. Huberman. The dynamics of viral marketing. ACM Transactions on the Web , 1(1):5, 2007

    work page 2007

  30. [30]

    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 Proceedings of the ACM Conference on Recommender Systems, RecSys , pages 5–13, 2017

    work page 2017

  31. [31]

    Strintzis, and Petros Daras

    Dimitrios Rafailidis, Theodoros Semertzidis, Michalis Lazaridis, Michael G. Strintzis, and Petros Daras. A data-driven approach for social event detection. In Proceedings of the MediaEval 2013 Multimedia Benchmark Workshop, Barcelona, Spain, October 18-19 , 2013

    work page 2013

  32. [32]

    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

    work page 2014

  33. [33]

    Crossing the bound- aries of communities via limited link injection for information diffusion in social networks

    Dimitrios Rafailidis and Alexandros Nanopoulos. Crossing the bound- aries of communities via limited link injection for information diffusion in social networks. In Proceedings of the 24th International Conference on World Wide Web Companion, WWW 2015, Florence, Italy, May 18- 22, - Companion Volume , pages 97–98, 2015

    work page 2015

  34. [34]

    Modeling trust and distrust information in rec- ommender systems via joint matrix factorization with signed graphs

    Dimitrios Rafailidis. Modeling trust and distrust information in rec- ommender systems via joint matrix factorization with signed graphs. In Proceedings of the 31st Annual ACM Symposium on Applied Computing, Pisa, Italy, April 4-8 , pages 1060–1065, 2016

    work page 2016

  35. [35]

    Collaborative ranking with so- cial relationships for top-n recommendations

    Dimitrios Rafailidis and Fabio Crestani. Collaborative ranking with so- cial relationships for top-n recommendations. In Proceedings of the 39th International ACM SIGIR conference on Research and Development in Information Retrieval, SIGIR, Pisa, Italy, July 17-21 , pages 785–788, 2016

    work page 2016

  36. [36]

    Prefer- ence dynamics with multimodal user-item interactions in social media recommendation

    Dimitrios Rafailidis, Pavlos Kefalas, and Yannis Manolopoulos. Prefer- ence dynamics with multimodal user-item interactions in social media recommendation. Expert Syst. Appl. , 74:11–18, 2017

    work page 2017

  37. [37]

    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 - Companion Volume, pages 99–100, 2015

    work page 2015

  38. [38]

    Translation-based factorization machines for sequential recommendation

    Rajiv Pasricha and Julian McAuley. Translation-based factorization machines for sequential recommendation. In Proceedings of the ACM Conference on Recommender Systems, RecSys , pages 63–71, 2018

    work page 2018

  39. [39]

    Personalizing session-based recommendations with hierar- chical recurrent neural networks

    Massimo Quadrana, Alexandros Karatzoglou, Bal ´azs Hidasi, and Paolo Cremonesi. Personalizing session-based recommendations with hierar- chical recurrent neural networks. In Proceedings of the ACM Conference on Recommender Systems, RecSys , pages 130–137, 2017

    work page 2017

  40. [40]

    Jose, and Xiangnan He

    Fajie Yuan, Alexandros Karatzoglou, Ioannis Arapakis, Joemon M. Jose, and Xiangnan He. A simple convolutional generative network for next item recommendation. In Proceedings of the ACM International Conference on Web Search and Data Mining, WSDM , pages 582–590, 2019

    work page 2019