pith. sign in

arxiv: 1906.08511 · v2 · pith:ZLHXRK66new · submitted 2019-06-20 · 💻 cs.CV · cs.IR

From Zero-Shot Learning to Cold-Start Recommendation

Jingjing Li , Mengmeng Jing , Ke Lu , Lei Zhu , Yang Yang , Zi Huang This is my paper

Pith reviewed 2026-05-25 20:03 UTC · model grok-4.3

classification 💻 cs.CV cs.IR
keywords zero-shot learningcold-start recommendationrecommender systemsauto-encoderdomain shiftuser attributes
0
0 comments X

The pith

Cold-start recommendation can be reformulated as zero-shot learning to improve performance on both tasks.

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

Zero-shot learning and cold-start recommendation both require predicting classes or items never seen in training data, using one space for direct features and another for supplementary descriptions. The paper treats these as the same underlying problem and shows that a zero-shot approach can be directly applied to cold-start recommendation. A Low-rank Linear Auto-Encoder is introduced with an encoder that maps user behavior to attributes and a symmetric decoder that reconstructs behavior from attributes. This handles domain shift, spurious correlations, and efficiency while delivering better results on recommendation benchmarks than prior methods and also lifting zero-shot learning accuracy.

Core claim

This work formulates CSR as a ZSL problem for the first time and proposes the Low-rank Linear Auto-Encoder, which maps user behavior into user attributes via a low-rank encoder and reconstructs user behavior from attributes via a symmetric decoder.

What carries the argument

Low-rank Linear Auto-Encoder (LLAE) consisting of a low-rank encoder and symmetric decoder that maps between user behavior and attributes.

If this is right

  • CSR can be handled by ZSL models with significant performance improvement compared with conventional state-of-the-art methods.
  • The consideration of CSR can benefit ZSL performance as well.
  • The formulation addresses domain shift, spurious correlations, and computing efficiency in the joint setting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Other tasks that involve predicting unseen elements from paired feature and description spaces may admit similar cross-application of methods.
  • A single model trained under this equivalence could be tested on mixed recommendation and classification benchmarks to measure transfer gains.
  • Extending the low-rank constraint to nonlinear mappings might reveal whether the linearity assumption limits further gains.

Load-bearing premise

That zero-shot learning and cold-start recommendation are two extensions of the same intension, both attempting to predict unseen classes and involving two spaces for direct feature representation and supplementary description.

What would settle it

Running the proposed LLAE on standard cold-start recommendation benchmarks and finding no accuracy gain over conventional state-of-the-art CSR methods would falsify the performance benefit.

Figures

Figures reproduced from arXiv: 1906.08511 by Jingjing Li, Ke Lu, Lei Zhu, Mengmeng Jing, Yang Yang, Zi Huang.

Figure 1
Figure 1. Figure 1: Illustration of ZSL and CSR. It can be seen that [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of the proposed LLAE. Firstly, we [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: The effect of the low-rank and reconstruction con [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
read the original abstract

Zero-shot learning (ZSL) and cold-start recommendation (CSR) are two challenging problems in computer vision and recommender system, respectively. In general, they are independently investigated in different communities. This paper, however, reveals that ZSL and CSR are two extensions of the same intension. Both of them, for instance, attempt to predict unseen classes and involve two spaces, one for direct feature representation and the other for supplementary description. Yet there is no existing approach which addresses CSR from the ZSL perspective. This work, for the first time, formulates CSR as a ZSL problem, and a tailor-made ZSL method is proposed to handle CSR. Specifically, we propose a Low-rank Linear Auto-Encoder (LLAE), which challenges three cruxes, i.e., domain shift, spurious correlations and computing efficiency, in this paper. LLAE consists of two parts, a low-rank encoder maps user behavior into user attributes and a symmetric decoder reconstructs user behavior from user attributes. Extensive experiments on both ZSL and CSR tasks verify that the proposed method is a win-win formulation, i.e., not only can CSR be handled by ZSL models with a significant performance improvement compared with several conventional state-of-the-art methods, but the consideration of CSR can benefit ZSL as well.

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

3 major / 2 minor

Summary. The manuscript claims that zero-shot learning (ZSL) and cold-start recommendation (CSR) are structurally equivalent extensions of the same underlying problem—both require predicting unseen classes from a direct feature space plus a supplementary description space. It formulates CSR as a ZSL task for the first time and introduces the Low-rank Linear Auto-Encoder (LLAE): a low-rank encoder that maps user behavior vectors to user attributes and a symmetric decoder that reconstructs behavior from attributes. The method is said to address domain shift, spurious correlations, and efficiency; experiments on both ZSL and CSR benchmarks reportedly show that ZSL models yield significant gains over conventional CSR baselines while the CSR perspective also improves ZSL performance.

Significance. If the claimed structural equivalence is rigorously established and the LLAE construction transfers without domain-specific adjustments, the work would usefully bridge the computer-vision and recommender-systems literatures and supply an efficient linear baseline for cold-start settings. The explicit win-win experimental outcome would be a concrete contribution worth citing.

major comments (3)
  1. [Abstract / §1] Abstract and §1: the central claim that CSR is literally a ZSL problem rests on the assertion that both settings share 'unseen classes' and a two-space structure, yet no formal mapping or invariance argument is supplied showing that user-behavior vectors correspond to visual features and user attributes to semantic attributes while preserving the same domain-shift and spurious-correlation geometry. Without this, the LLAE solution derived for ZSL does not automatically solve the CSR version.
  2. [§3] §3 (LLAE derivation): the low-rank encoder-decoder is motivated by ZSL-specific issues (domain shift, spurious correlations). The manuscript does not demonstrate that the same low-rank constraint eliminates the analogous issues when the 'visual' space is replaced by user-behavior vectors, nor does it provide an ablation isolating the contribution of the low-rank term versus a plain linear auto-encoder on CSR data.
  3. [§4] §4 (experiments): the reported CSR performance gains are presented without the explicit protocol that converts a recommendation dataset into a ZSL-compatible pair of spaces (e.g., how 'unseen classes' are defined, how the attribute matrix is constructed). This prevents verification that the improvement is due to the ZSL formulation rather than dataset-specific tuning.
minor comments (2)
  1. [§2] Notation for the two spaces is introduced informally; a single table or diagram explicitly aligning ZSL symbols (visual features, semantic attributes) with CSR symbols (behavior vectors, user attributes) would improve readability.
  2. [Abstract] The abstract states 'extensive experiments on both ZSL and CSR tasks' but does not list the concrete datasets or metrics in the opening paragraph; moving this information forward would help readers assess scope immediately.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and constructive report. The comments highlight important points about formalizing the ZSL-CSR equivalence, validating the low-rank constraint in the CSR setting, and documenting the experimental protocol. We address each major comment below with clarifications and proposed revisions where appropriate.

read point-by-point responses

  1. Referee: [Abstract / §1] Abstract and §1: the central claim that CSR is literally a ZSL problem rests on the assertion that both settings share 'unseen classes' and a two-space structure, yet no formal mapping or invariance argument is supplied showing that user-behavior vectors correspond to visual features and user attributes to semantic attributes while preserving the same domain-shift and spurious-correlation geometry. Without this, the LLAE solution derived for ZSL does not automatically solve the CSR version.

    Authors: The equivalence is structural rather than a strict isomorphism: both problems require predicting labels for unseen classes using only a description space after training on paired feature-description data for seen classes. We map user behavior vectors to visual features (both high-dimensional, noisy observations) and user attributes to semantic attributes (both provide auxiliary information for generalization). Domain shift and spurious correlations arise analogously because the feature space distribution for seen vs. unseen items differs, and high-dimensional behavior vectors contain spurious correlations that low-rank regularization mitigates. We will add a formal problem formulation subsection (new §2.1) that defines the shared tuple (X_seen, A_seen, A_unseen) and shows the prediction task is identical under this correspondence. This makes the transfer of LLAE explicit without claiming identical geometry in every dataset. revision: yes

  2. Referee: [§3] §3 (LLAE derivation): the low-rank encoder-decoder is motivated by ZSL-specific issues (domain shift, spurious correlations). The manuscript does not demonstrate that the same low-rank constraint eliminates the analogous issues when the 'visual' space is replaced by user-behavior vectors, nor does it provide an ablation isolating the contribution of the low-rank term versus a plain linear auto-encoder on CSR data.

    Authors: The low-rank constraint addresses the same issues in CSR because user behavior vectors are high-dimensional and sparse, exactly as visual features are; the encoder projects them onto a low-dimensional attribute subspace to reduce overfitting to spurious correlations and to regularize against domain shift between training and cold-start users. While the original derivation is ZSL-motivated, the mathematics (nuclear-norm or rank constraint on the encoder) is domain-agnostic. We performed the requested ablation (LLAE vs. plain linear AE) on CSR datasets and will add an explicit table (new Table 4) and paragraph in §4.3 reporting the gains attributable to the low-rank term alone on MovieLens and LastFM, confirming the same benefit observed in ZSL. revision: partial

  3. Referee: [§4] §4 (experiments): the reported CSR performance gains are presented without the explicit protocol that converts a recommendation dataset into a ZSL-compatible pair of spaces (e.g., how 'unseen classes' are defined, how the attribute matrix is constructed). This prevents verification that the improvement is due to the ZSL formulation rather than dataset-specific tuning.

    Authors: We agree the protocol should be fully explicit. In the revised §4.2 we will insert a dedicated 'Dataset Conversion Protocol' subsection with numbered steps: (1) treat items as classes, (2) split items into seen/unseen with no overlapping users for unseen (cold-start definition), (3) construct user behavior matrix X from implicit/explicit feedback, (4) build attribute matrix A from provided side information (genres, tags, etc.), (5) ensure A_unseen is available at test time only. We will also release the exact preprocessing scripts and data splits used. This makes clear that gains come from the ZSL-style training (learning from seen items only) rather than tuning on cold-start data. revision: yes

Circularity Check

0 steps flagged

No circularity: central claim is a structural analogy mapping, not a self-referential derivation

full rationale

The paper's load-bearing step is the observation that ZSL and CSR share the properties of predicting unseen classes and operating over two spaces (direct features + supplementary descriptions), which is used to justify reformulating CSR as a ZSL problem and applying LLAE. This is an explicit problem-mapping argument presented in the abstract and introduction; it does not reduce any derived quantity (e.g., the low-rank encoder-decoder) to a fitted parameter or self-citation by construction. No equations are shown that equate a prediction to its own input, no uniqueness theorems are imported from the authors' prior work, and no ansatz is smuggled via citation. The derivation chain therefore remains self-contained against external benchmarks once the analogy is granted; any weakness lies in the empirical validity of the mapping rather than in definitional circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that ZSL and CSR share the same underlying structure; no free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption ZSL and CSR are two extensions of the same intension, both attempting to predict unseen classes and involving two spaces for direct feature representation and supplementary description.
    This premise, stated in the abstract, is what allows the reformulation of CSR as a ZSL problem.

pith-pipeline@v0.9.0 · 5770 in / 1108 out tokens · 28882 ms · 2026-05-25T20:03:41.212462+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel echoes
    ?
    echoes

    ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

    a low-rank encoder maps user behavior into user attributes and a symmetric decoder reconstructs user behavior from user attributes... low-rank constraint... helps reveal the dominant factors and filter out trivial connections, or in other words, spurious correlations

  • IndisputableMonolith/Foundation/BranchSelection.lean branch_selection echoes
    ?
    echoes

    ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

    the reconstruction part is effective in mitigating the domain shift problem... the demand for more truthful reconstruction from the attributes to behavior is generalizable across warm and cold domains

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages · 1 internal anchor

  1. [1]

    Bobadilla, J.; Ortega, F.; Hernando, A.; and Guti \'e rrez, A. 2013. Recommender systems survey. Knowledge-based systems 46:109--132

    work page 2013

  2. [2]

    Boureau, Y.-L.; Chopra, S.; LeCun, Y.; et al. 2007. A unified energy-based framework for unsupervised learning. In Artificial Intelligence and Statistics , 371--379

    work page 2007

  3. [3]

    L.; and Kuflik, T

    Cantador, I.; Brusilovsky, P. L.; and Kuflik, T. 2011. Second workshop on information heterogeneity and fusion in recommender systems (hetrec2011)

    work page 2011

  4. [4]

    Ding, Z.; Shao, M.; and Fu, Y. 2017. Low-rank embedded ensemble semantic dictionary for zero-shot learning. In CVPR . IEEE

    work page 2017

  5. [5]

    Ding, Z.; Shao, M.; and Fu, Y. 2018. Incomplete multisource transfer learning. IEEE TNNLS 29(2):310--323

    work page 2018

  6. [6]

    D.; Riedl, J

    Ekstrand, M. D.; Riedl, J. T.; Konstan, J. A.; et al. 2011. Collaborative filtering recommender systems. Foundations and Trends in Human--Computer Interaction 4(2):81--173

    work page 2011

  7. [7]

    Farhadi, A.; Endres, I.; Hoiem, D.; and Forsyth, D. 2009. Describing objects by their attributes. In CVPR , 1778--1785. IEEE

    work page 2009

  8. [8]

    Fern \'a ndez-Tob \' as, I.; Cantador, I.; Kaminskas, M.; and Ricci, F. 2012. Cross-domain recommender systems: A survey of the state of the art. In Spanish Conference on Information Retrieval , 24

    work page 2012

  9. [9]

    Fern \'a ndez-Tob \' as, I.; Braunhofer, M.; Elahi, M.; Ricci, F.; and Cantador, I. 2016. Alleviating the new user problem in collaborative filtering by exploiting personality information. User Modeling and User-Adapted Interaction 26(2-3):221--255

    work page 2016

  10. [10]

    Gantner, Z.; Drumond, L.; Freudenthaler, C.; Rendle, S.; and Schmidt-Thieme, L. 2010. Learning attribute-to-feature mappings for cold-start recommendations. In ICDM , 176--185. IEEE

    work page 2010

  11. [11]

    Kodirov, E.; Xiang, T.; Fu, Z.; and Gong, S. 2015. Unsupervised domain adaptation for zero-shot learning. In ICCV , 2452--2460

    work page 2015

  12. [12]

    Kodirov, E.; Xiang, T.; and Gong, S. 2017. Semantic autoencoder for zero-shot learning. arXiv preprint arXiv:1704.08345

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  13. [13]

    Krohn-Grimberghe, A.; Drumond, L.; Freudenthaler, C.; and Schmidt-Thieme, L. 2012. Multi-relational matrix factorization using bayesian personalized ranking for social network data. In ACM WSDM , 173--182. ACM

    work page 2012

  14. [14]

    H.; Nickisch, H.; and Harmeling, S

    Lampert, C. H.; Nickisch, H.; and Harmeling, S. 2014. Attribute-based classification for zero-shot visual object categorization. IEEE TPAMI 36(3):453--465

    work page 2014

  15. [15]

    Li, J.; Wu, Y.; Zhao, J.; and Lu, K. 2016. Low-rank discriminant embedding for multiview learning. IEEE TCYB

    work page 2016

  16. [16]

    Li, J.; Lu, K.; Huang, Z.; and Shen, H. T. 2017. Two birds one stone: On both cold-start and long-tail recommendation. In ACM MM , 898--906. ACM

    work page 2017

  17. [17]

    Li, J.; Lu, K.; Huang, Z.; Zhu, L.; and Shen, H. T. 2018a. Heterogeneous domain adaptation through progressive alignment. IEEE TNNLS

    work page

  18. [18]

    Li, J.; Lu, K.; Huang, Z.; Zhu, L.; and Shen, H. T. 2018b. Transfer independently together: A generalized framework for domain adaptation. IEEE TCYB

    work page

  19. [19]

    Li, J.; Zhu, L.; Huang, Z.; Lu, K.; and Zhao, J. 2018c. I read, i saw, i tell: Texts assisted fine-grained visual classification. In ACM MM , 663--671. ACM

    work page

  20. [20]

    Lin, J.; Sugiyama, K.; Kan, M.-Y.; and Chua, T.-S. 2013. Addressing cold-start in app recommendation: latent user models constructed from twitter followers. In ACM SIGIR , 283--292. ACM

    work page 2013

  21. [21]

    E.; and Hinton, G

    Mohamed, A.-r.; Dahl, G. E.; and Hinton, G. 2012. Acoustic modeling using deep belief networks. IEEE TASLP 20(1):14--22

    work page 2012

  22. [22]

    V.; Abbasnejad, E.; and Della Penna, N

    Noel, J.; Sanner, S.; Tran, K.-N.; Christen, P.; Xie, L.; Bonilla, E. V.; Abbasnejad, E.; and Della Penna, N. 2012. New objective functions for social collaborative filtering. In WWW , 859--868. ACM

    work page 2012

  23. [23]

    Patterson, G., and Hays, J. 2012. Sun attribute database: Discovering, annotating, and recognizing scene attributes. In CVPR , 2751--2758. IEEE

    work page 2012

  24. [24]

    A.; Kasirun, Z

    Rohani, V. A.; Kasirun, Z. M.; Kumar, S.; and Shamshirband, S. 2014. An effective recommender algorithm for cold-start problem in academic social networks. Mathematical Problems in Engineering 2014

    work page 2014

  25. [25]

    Romera-Paredes, B., and Torr, P. 2015. An embarrassingly simple approach to zero-shot learning. In ICML , 2152--2161

    work page 2015

  26. [26]

    Sedhain, S.; Sanner, S.; Braziunas, D.; Xie, L.; and Christensen, J. 2014. Social collaborative filtering for cold-start recommendations. In ACM RecSys , 345--348. ACM

    work page 2014

  27. [27]

    K.; Sanner, S.; Xie, L.; and Braziunas, D

    Sedhain, S.; Menon, A. K.; Sanner, S.; Xie, L.; and Braziunas, D. 2017. Low-rank linear cold-start recommendation from social data. In AAAI , 1502--1508

    work page 2017

  28. [28]

    Smith, B., and Linden, G. 2017. Two decades of recommender systems at amazon.com. IEEE Internet Computing 21(3):12--18

    work page 2017

  29. [29]

    Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; and Rabinovich, A. 2015. Going deeper with convolutions. In CVPR , 1--9

    work page 2015

  30. [30]

    Tang, L.; Wang, X.; and Liu, H. 2012. Scalable learning of collective behavior. TKDE 24(6):1080--1091

    work page 2012

  31. [31]

    Vincent, P.; Larochelle, H.; Bengio, Y.; and Manzagol, P. A. 2008. Extracting and composing robust features with denoising autoencoders. In ICML , 1096--1103

    work page 2008

  32. [32]

    Wah, C.; Branson, S.; Welinder, P.; Perona, P.; and Belongie, S. 2011. The caltech-ucsd birds-200-2011 dataset

    work page 2011

  33. [33]

    Yang, Y.; Luo, Y.; Chen, W.; Shen, F.; Shao, J.; and Shen, H. T. 2016. Zero-shot hashing via transferring supervised knowledge. In ACM MM , 1286--1295. ACM

    work page 2016

  34. [34]

    Zhang, Z., and Saligrama, V. 2015. Zero-shot learning via semantic similarity embedding. In ICCV , 4166--4174

    work page 2015

  35. [35]

    Zhang, Z., and Saligrama, V. 2016. Zero-shot learning via joint latent similarity embedding. In CVPR , 6034--6042

    work page 2016

  36. [36]

    Zhang, M.-L., and Zhou, Z.-H. 2014. A review on multi-label learning algorithms. IEEE TKDE 26(8):1819--1837

    work page 2014

  37. [37]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

    work page