Pith. sign in

REVIEW 3 major objections 6 minor 55 references

Improving Outfit Recommendation with Co-supervision of Fashion Generation

T0 review · 3 major / 6 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 FARM co-trains a fashion generator with an outfit recommender, and the added supervision improves ranking on two large datasets.

desk verdict Genuinely new co-supervision architecture with systematic ablations, but the missing item-overlap analysis leaves the headline gains vulnerable to a memorization-based alternative explanation. read the letter →

arxiv 1908.09104 v1 pith:GZS2MHT5 submitted 2019-08-24 cs.IR cs.CV

Yujie Lin , Pengjie Ren , Zhumin Chen , Zhaochun Ren , Jun Ma , Maarten de Rijke This is my paper · ORCID
classification cs.IRcs.CV
keywords outfitrecommendationfashiongenerationco-supervisionlearningvisualcompatibilitylayer-to-layermatchingvariationalautoencoderBayesianPersonalizedRanking
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

The paper argues that outfit recommendation is held back because visual features are learned only from match/no-match labels, which miss aesthetic cues like style and design. It proposes FARM, a framework that trains a fashion generator alongside a recommender: given a top and a text description, the generator produces a bottom image, and the recommender scores candidate bottoms partly by matching their features to the generated image at several neural layers. The authors claim this co-supervision makes the learned features carry more aesthetic information, and that the layer-to-layer matching lets the recommendation task use generation without being dominated by image quality. If the claim holds, generation is not just a side task but a source of compatibility knowledge, and the same setup improves both top and bottom recommendation on two public datasets.

What carries the argument

The key machinery is the co-supervision architecture plus the layer-to-layer matching mechanism. A variational transformer (an encoder CNN together with a deconvolutional generator) turns a top image and a bag-of-words description into a Gaussian latent code, from which a two-stage DCNN (including an SRResNet-style refinement stage) generates a bottom image; the generation loss forces the encoder to retain aesthetic information. The layer-to-layer matching mechanism then compares, at several selected layers, the feature maps of a candidate bottom with the feature maps of the generated bottom, summing these per-layer dot products into an additional matching score. This mechanism is what lets the recommendation part exploit the generated image without the model focusing only on generation quality.

What would settle it

Re-run FARM's experiments on a split in which no test-set positive bottom appears among the training generation targets; if the AUC and MRR gains over DVBPR$_{d}$ shrink to near zero, the reported improvement comes from memorizing seen items rather than from learning aesthetic compatibility.

Watch

Extended reading notes

Core claim

The central claim is that jointly training outfit recommendation with fashion generation improves recommendation accuracy. FARM consists of a variational-transformer-based generator that maps a top image and a bottom description into a generated bottom image, and a recommender that scores a candidate bottom by three matching signals: direct visual matching with the top, description matching with the text, and layer-to-layer matching between the candidate's CNN features and the generator's DCNN features at corresponding layers. The generator is trained to reconstruct the ground-truth positive bottom, while the recommender is trained with a Bayesian Personalized Ranking loss on positive and negative pairs; the two losses are optimized jointly. Experiments on FashionVC and ExpFashion show FARM outperforming strong baselines such as DVBPR$_{d}$, and ablations indicate that removing either the generation supervision or the layer-to-layer matching substantially reduces performance.

Load-bearing premise

The generation target during training is the actual positive bottom that will later be recommended, so the extra supervision only helps if the generator learns general compatibility rather than memorizing training images; the paper does not analyze whether test positive bottoms overlap with the generation targets seen during training.

Editorial extensions

If this is right

  • If FARM's central claim is correct, then generation can serve as an auxiliary supervision signal for visual understanding in recommendation, not merely as a standalone generative task.
  • The layer-to-layer matching design implies that the benefit of generation comes from intermediate visual abstractions, not from the final generated pixels alone.
  • The reported gains on both top and bottom recommendation, on both FashionVC and ExpFashion, suggest the method transfers across dataset sizes and description lengths.
  • Because the paper observes that poor generated images reduce recommendation quality, improving the generator should directly improve the recommender within this framework.
  • The framework's ability to condition on text descriptions means that cold-start or sparse-annotation settings could still benefit, as long as descriptions are available.

Reading between the lines

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

  • A direct test of the memorization concern would be to split the data so that no test-set positive bottom ever appears as a generation target during training; if the gain over DVBPR$_{d}$ vanishes, the reported improvement comes from reconstructing seen items rather than from learning general compatibility.
  • One can extend FARM by replacing the pixel-reconstruction loss with a perceptual or adversarial loss; if the aesthetic-supervision hypothesis is right, such a change should preserve or increase the recommendation gain while improving generated-image quality.
  • The layer-to-layer matching idea could transfer to other paired-image recommendation tasks where a generator can produce a plausible counterpart, such as shoes, bags, or whole outfits.
  • The paper's failure cases hint that description quality is a bottleneck; a testable extension is to measure how recommendation AUC changes as description vocabulary is artificially degraded or enriched.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes FARM (FAshion Recommendation Machine), an end-to-end framework for outfit recommendation. Given a top image and a free-text description of the desired bottom, a variational transformer encodes the pair into a latent distribution, and a DCNN generates a 'reference' bottom image. A recommender scores each candidate bottom by a sum of three terms: visual matching between top and bottom, description matching, and a layer-to-layer matching between the candidate and the generated reference at multiple CNN layers. The model is trained jointly with a generation loss (pixel-wise reconstruction of the ground-truth positive bottom plus a KL divergence) and a BPR recommendation loss. Experiments on FashionVC and ExpFashion report AUC and MRR improvements over several baselines, and ablations show that both the generation co-supervision and the layer-to-layer matching contribute.

Significance. Conditional on the results being free of memorization artifacts, FARM makes a useful contribution by demonstrating that a generation objective can act as a co-supervision signal for visual recommendation, and the layer-to-layer matching is a reasonable way to connect generation and recommendation. The authors ship code and data, and the evaluation is extensive, with paired t-tests and ablations (Tables 2-4). However, the central mechanism—that the generated reference is an independent aesthetic guide—requires that the generator generalizes rather than memorizes training positives, and this is not established. If the overlap threat is ruled out by additional experiments, the paper would provide a solid advance; without that evidence, the reported gains may be partly due to retrieval of seen images.

major comments (3)
  1. [Section 3.4, Eq. (20); Section 3.3.3, Eqs. (17)-(18); Section 4.1, Table 1] The reported improvements may be inflated by memorization of training positive bottoms. The generator is trained to minimize the pixel distance between the generated image and the ground-truth positive bottom Ip (Eq. 20), and at test time the reference is generated deterministically. The layer-to-layer matching (Eqs. 17-18) then scores candidates by their feature similarity to this reference. If a test positive bottom was used as a generation target during training—which is very likely given that ExpFashion contains 853,991 outfits but only 117,668 distinct bottoms, so a random outfit split will place the same bottom in both training and test—the generator can reproduce a memorized image and the matching term will favor it, raising AUC/MRR without adding compatibility knowledge. The paper provides no overlap statistics, no item-disjoint split, and no per-query analysis conditioned on whether the positive bottom was seen in training. The same issue affects the FARM-G comparison in Table 3, where generation-based retrieval would also benefit from memorized positives. I request: (i) report the fraction of test positives (and test queries) that appear in the training generation target set; (ii) run the main experiments under an item-disjoint split; and (iii) report results separately for seen versus unseen positives.
  2. [Section 5, Table 2] The significance claims are not fully supported by the reported statistics. The paper states that FARM significantly outperforms DVBPR_d using paired t-tests with p < 0.05, but it reports only point estimates with no standard deviations, confidence intervals, or number of runs. This makes it impossible to gauge the variability of the improvements or to verify that the t-test was conducted correctly. Please report mean ± standard deviation over multiple random seeds (or bootstrap confidence intervals) and state the number of runs used for the significance tests.
  3. [Abstract and Sections 6.3-6.4; Conclusion] The paper's claim that FARM 'generate[s] high quality images as references' and that generation quality affects recommendation is not quantitatively supported. Generation quality is only illustrated with a handful of qualitative examples (Figures 3-5), and no quantitative generation metrics (e.g., FID, IS, or human evaluation) or feature-level analyses are given. This is load-bearing because the proposed mechanism assumes the generated reference carries genuine aesthetic information; if the generator produces blurry or memorized images, the layer-to-layer matching gains could be an artifact of the reconstruction training. Please add quantitative generation-quality evaluation and, if possible, an analysis of how generation quality relates to recommendation performance.
minor comments (6)
  1. [Eq. (20)] The first term uses a squared L2 norm (with factor 1/2) while the second term uses a non-squared L2 norm; please clarify the intended loss (MSE versus L1) and whether this asymmetry is deliberate.
  2. [Section 4.2] The description of the layer-to-layer matching mapping is unclear: 'The 7th, the 6th and the 5th layers of the encoder CNN are adopted to compute the layer-to-layer matching with the input, the 1st and the 2nd layers of the generator DCNN.' Please specify exactly which encoder layer is paired with which DCNN layer.
  3. [Eq. (23)] The total loss sums Lgen, Lkl, and Lbpr without any weighting coefficients. Since the pixel-reconstruction term is likely to have a different scale than the BPR loss, please state whether these weights were tuned and what the final values are.
  4. [Section 4.1] Please clarify the train/validation/test split procedure (for example, by outfit or by query) and state the split ratio.
  5. [Eq. (30)] The AUC definition uses a strict indicator delta(s_p > s_n); ties are treated as failures. If ties occur, please define how they are handled.
  6. [Section 6.1, Table 3] The statement that FARM-G 'performs better than IBRd in most settings' is supported by Table 3, but the comparison is not like-for-like: FARM-G uses a different retrieval protocol (AlexNet features over generated images) than the trained end-to-end models. Please clarify the comparison protocol.

Circularity Check

1 steps flagged · score 6.0 of 10

The generation reference used in layer-to-layer matching is a reconstruction fitted to ground-truth positive images (Eq. 20), and the evaluation does not use an item-disjoint split, so the reported recommendation gains can partly reduce to matching against memorized training images.

  1. fitted input called prediction [Section 3.2 (Eq. 20), Section 3.3.3 (Eqs. 17-19), Section 4.1 (evaluation protocol)]
    "During training, we regard the image Ip of a positive bottom p, which not only matches the given top It but also meets the given description d, as the generation target. ... Lgen(t,d,p) = 1/2 ||I1_hat − Ip||2_2 + ||I2_hat − Ip||. ... The generated image can be seen as a reference for recommendation. ... s_hat = Σ_{l∈L} s_l_hat ... s = sv + sd + s_hat."

    The reference image scored in Eqs. 17-18 is not an independent aesthetic prior: it is produced by a generator trained with Eq. 20 to minimize pixel distance to the ground-truth positive bottom Ip of the same (t,d) query. The layer-to-layer term s_hat therefore measures candidate similarity to a reconstruction fitted on the label that the ranking is supposed to predict. When a test query's positive bottom also served as a generation target during training, ranking can be driven by retrieving that memorized image through s_hat rather than by generalizing compatibility. ExpFashion has 117,668 distinct bottoms but 853,991 outfits, so random outfit splits make such overlap very likely; the paper reports no item-disjoint split and no overlap analysis.

full rationale

FARM's architecture is not circular in the sense of a tautological derivation: the generator is a legitimate auxiliary network, the BPR loss is a standard external objective, and the baselines are implemented independently. The circularity is localized to the experimental support for the central claim. The generator's output is trained by Eq. 20 against the same positive bottom image that the recommender is asked to rank, and the layer-to-layer matching score of Eq. 19 then rewards candidates whose features resemble that fitted reconstruction. If train and test positives overlap, which the dataset statistics make very likely and which the paper neither prevents nor measures, the reported AUC/MRR improvements partly reduce to image retrieval of seen positives rather than to new compatibility knowledge. The paper's stated limitation, that recommendation performance is affected by generated image quality, does not address this train/test overlap issue. I do not find load-bearing self-citation: the self-created ExpFashion dataset is released and the method is implemented against external benchmarks, so the self-citation is not the mechanism forcing the result. The score of 6 reflects partial circularity in the evaluation chain: one prediction channel reduces to matching against a fitted reconstruction, while the overall framework retains independent content.

Assumptions & free parameters 6 free parameters · 5 assumptions · 1 invented entities

The central claim rests on several hand-chosen hyperparameters, the validity of outfit co-occurrence as a compatibility label, and the assumption that a generated image trained on positive targets can act as an independent matching reference. The invented generated-image entity is model-internal and carries no external evidence.

free parameters (6)
  • latent variable size k = 100
    Hand-chosen in Section 4.2; controls the capacity of the variational transformer and is not swept.
  • representation size e = 100
    Hand-chosen in Section 4.2 for visual, semantic, and layer representations.
  • layer-to-layer layer set L = Encoder layers 7, 6, 5 matched with generator input, 1st, and 2nd layers
    Hand-chosen in Section 4.2; the paper does not ablate alternative layer sets.
  • loss weights for Lgen, Lkl, Lbpr = 1.0 each
    Set implicitly in Eq. 23 with no tunable weights; no sensitivity analysis is reported.
  • candidate set size for evaluation = 100
    Randomly sampled negatives per query in the validation and test protocol, following Section 4.1.
  • description vocabulary size Dd = 547
    Result of frequency filtering plus manual curation in Section 4.2; the manual selection is a researcher choice that influences the description matching score.
assumptions (5)
  • standard math Variational inference ELBO and Gaussian posterior and prior assumptions are valid for the generative model.
    Eqs. 2 to 4 rely on standard variational autoencoder derivations following Kingma and Welling.
  • domain assumption Co-occurrence of items in Polyvore outfits is a sufficient proxy for fashion compatibility.
    Section 4.1 uses outfit co-occurrence to define positive pairs; no human compatibility ratings are used.
  • ad hoc to paper A generated bottom image can serve as a meaningful reference for ranking real candidate bottoms.
    Sections 3.3.3 and 6.2 assume the generator's output captures matching-relevant aesthetics; the paper provides only qualitative evidence for this assumption.
  • ad hoc to paper The pixel reconstruction loss on positive images transfers aesthetic knowledge to the recommendation encoder.
    Section 3.2 and Eq. 20 hypothesize that generation supervision improves visual understanding; this is the paper's core, unproven assumption.
  • domain assumption Randomly sampled negatives from B \ B_{t,d} provide a valid BPR training signal.
    Eq. 22 uses one randomly sampled negative per positive; this is standard but assumes most negatives are genuinely incompatible.
invented entities (1)
  • Generated reference image I_hat
    purpose: Serves as a pseudo-bottom that bridges generation and recommendation through layer-to-layer matching.
    I_hat is produced by the generator trained to reconstruct the ground-truth positive bottom with Lgen (Eq. 20). It is an internal construct with no falsifiable handle outside the model, so it cannot independently validate the aesthetic claims.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Improving Outfit Recommendation with Co-supervision of Fashion Generation." pith.science (2026). https://pith.science/paper/GZS2MHT5

@misc{pith2026190809104,
  author       = {Pith},
  title        = {Pith review of: Improving Outfit Recommendation with Co-supervision of Fashion Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GZS2MHT5}},
  note         = {Machine review of arXiv:1908.09104}
}
read the original abstract

The task of fashion recommendation includes two main challenges: visual understanding and visual matching. Visual understanding aims to extract effective visual features. Visual matching aims to model a human notion of compatibility to compute a match between fashion items. Most previous studies rely on recommendation loss alone to guide visual understanding and matching. Although the features captured by these methods describe basic characteristics (e.g., color, texture, shape) of the input items, they are not directly related to the visual signals of the output items (to be recommended). This is problematic because the aesthetic characteristics (e.g., style, design), based on which we can directly infer the output items, are lacking. Features are learned under the recommendation loss alone, where the supervision signal is simply whether the given two items are matched or not. To address this problem, we propose a neural co-supervision learning framework, called the FAshion Recommendation Machine (FARM). FARM improves visual understanding by incorporating the supervision of generation loss, which we hypothesize to be able to better encode aesthetic information. FARM enhances visual matching by introducing a novel layer-to-layer matching mechanism to fuse aesthetic information more effectively, and meanwhile avoiding paying too much attention to the generation quality and ignoring the recommendation performance. Extensive experiments on two publicly available datasets show that FARM outperforms state-of-the-art models on outfit recommendation, in terms of AUC and MRR. Detailed analyses of generated and recommended items demonstrate that FARM can encode better features and generate high quality images as references to improve recommendation performance.

Figures

Figures reproduced from arXiv: 1908.09104 by the authors.

Figure 1
Figure 1. Overview of FARM. The fashion generator (top) uses [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Details of the encoder and the generator in FARM, wh [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Comparison between real and generated images. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Case studies of recommendation. The items high [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Case studies of generation. Each case is in the form [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 53 canonical work pages

  1. [1]

    Timo Ahonen, Abdenour Hadid, and Matti Pietikainen. 200 6. Face Description with Local Binary Patterns: Application to Face Recognitio n. IEEE Trans on Pat- tern Analysis and Machine Intelligence (TPAMI) 28, 12 (2006)

    work page 2006

  2. [2]

    Lempitsky

    Artem Babenko, Anton Slesarev, Alexander Chigorin, and Victor S. Lempitsky

    work page

  3. [3]

    Jianmin Bao, Dong Chen, Fang Wen, Houqiang Li, and Gang Hu a. 2017. CV AE- GAN: Fine-Grained Image Generation through Asymmetric Tra ining. In Inter- national Conf. on Computer Vision (ICCV’17) . 2764–2773

    work page 2017

  4. [4]

    Blei, Alp Kucukelbir, and Jon D

    David M. Blei, Alp Kucukelbir, and Jon D. Mcauliffe. 2017. Variational Inference: A Review for Statisticians. Journal of the American Statistical Association (JASA) 112, 518 (2017)

    work page 2017

  5. [5]

    Bourdev, Subhransu Maji, and Jitendra Malik

    Lubomir D. Bourdev, Subhransu Maji, and Jitendra Malik. 2011. Describing peo- ple: A Poselet-based Approach to Attribute Classification. In International Conf. on Computer Vision (ICCV’11) . 1543–1550

    work page 2011

  6. [6]

    Lei Cai, Hongyang Gao, and Shuiwang Ji. 2017. Multi-Stag e Variational Auto- Encoders for Coarse-to-Fine Image Generation. CoRR abs/1705.07202 (2017). http://arxiv.org/abs/1705.07202

    work page Pith review arXiv 2017

  7. [7]

    Ken Chatfield, Karen Simonyan, Andrea Vedaldi, and Andre w Zisserman. 2014. Return of the Devil in the Details: Delving Deep into Convolu tional Nets. In British Machine Vision Conf. (BMVC’14)

    work page 2014

  8. [8]

    Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minji e Wang, Tianjun Xiao, Bing Xu, Chiyuan Zhang, and Zheng Zhang. 2015. MXNet: A Flexible and Efficient Machine Learning Library for Heterogeneous Distri buted Systems. In Annual Conf. on Neural Information Processing Systems (NIPS’ 15)

    work page 2015

Show all 55 references
  1. [9]

    Navneet Dalal and Bill Triggs. 2005. Histograms of Orien ted Gradients for Human Detection. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR’05). 886–893

    2005

  2. [10]

    Xavier Glorot and Yoshua Bengio. 2010. Understanding t he Difficulty of Train- ing Deep Feedforward Neural Networks. Journal of Machine Learning Research (JMLR) 9 (2010), 249–256

    2010

  3. [11]

    Xintong Han, Zuxuan Wu, Yu-Gang Jiang, and Larry S. Davi s. 2017. Learning Fashion Compatibility with Bidirectional LSTMs. In ACM International Conf. on Multimedia (MM’17). 1078–1086

    2017

  4. [12]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2 016. Deep Residual Learning for Image Recognition. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR’16)

  5. [13]

    Wei Lin Hsiao and Kristen Grauman. 2018. Creating Capsu le Wardrobes from Fashion Images. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR’18)

    2018

  6. [14]

    Yang Hu, Xi Yi, and Larry S. Davis. 2015. Collaborative F ashion Recommenda- tion: A Functional Tensor Factorization Approach. In ACM International Conf. on Multimedia (MM’15) . 129–138

    2015

  7. [15]

    Tomoharu Iwata, Shinji Watanabe, and Hiroshi Sawada. 2 011. Fashion Coor- dinates Recommender System Using Photographs from Fashion Magazines. In International Joint Conf. on Artificial Intelligence (IJCA I’11). 2262–2267

  8. [16]

    Vignesh Jagadeesh, Robinson Piramuthu, Anurag Bhardw aj, Wei Di, and Neel Sundaresan. 2014. Large Scale Visual Recommendations from Street Fashion Images. In ACM Knowledge Discovery and Data Mining (KDD’14) . 1925–1934

    2014

  9. [17]

    Gareth James, Daniela Witten, Trevor Hastie, and Rober t Tibshirani. 2013. An Introduction to Statistical Learning . Springer

    2013

  10. [18]

    Shatha Jaradat. 2017. Deep Cross-Domain Fashion Recom mendation. In ACM Conf. on Recommender Systems (RecSys’17) . 407–410

    2017

  11. [19]

    Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Kara yev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. 2014 . Caffe: Convolu- tional Architecture for Fast Feature Embedding. In ACM International Conf. on Multimedia (MM’14). 675–678

    2014

  12. [20]

    Wang-Cheng Kang, Chen Fang, Zhaowen Wang, and Julian Mc Auley. 2017. Visually-Aware Fashion Recommendation and Design with Gen erative Image Models. In International Conf. on Data Mining (ICDM’17) . 207–216

    2017

  13. [21]

    Hadi Kiapour, Xufeng Han, Svetlana Lazebnik, Alexan der C

    M. Hadi Kiapour, Xufeng Han, Svetlana Lazebnik, Alexan der C. Berg, and Tamara L. Berg. 2015. Where to Buy It: Matching Street Clothi ng Photos in Online Shops. In International Conf. on Computer Vision (ICCV’15) . 3343–3351

    2015

  14. [22]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2015. Adam: A Method for Stochas- tic Optimization. In International Conf. on Learning Representations (ICLR’15 ). http://arxiv.org/abs/1412.6980

    2015 arXiv

  15. [23]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. 2014. Auto-encodin g Variational Bayes. In International Conf. on Learning Representations (ICLR’14 )

    2014

  16. [24]

    Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caba llero, Andrew Cunning- ham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Joha nnes Totz, and Zehan Wang. 2017. Photo-Realistic Single Image Super-Reso lution Using a Gen- erative Adversarial Network. InIEEE Conf. on Co...

    2017

  17. [25]

    Jing Li, Pengjie Ren, Zhumin Chen, Zhaochun Ren, Tao Lia n, and Jun Ma. 2017. Neural Attentive Session-based Recommendation. In ACM International Conf. on Information and Knowledge Management (CIKM’17) . 1419–1428

    2017

  18. [26]

    Yuncheng Li, Liangliang Cao, Jiang Zhu, and Jiebo Luo. 2 017. Mining Fashion Outfit Composition Using an End-to-End Deep Learning Approa ch on Set Data. IEEE Transactions on Multimedia (TMM) 19, 8 (2017), 1946–1955

    2017

  19. [27]

    Kevin Lin, Huei Fang Yang, Jen Hao Hsiao, and Chu Song Che n. 2015. Deep Learning of Binary Hash Codes for Fast Image Retrieval. In IEEE Conf. on Com- puter Vision and Pattern Recognition (CVPR’15) . 27–35

    2015

  20. [28]

    Ren, Zhumin Chen, Zhaochun Ren, Jun Ma, and Maarten de Rijke

    Yujie Lin, Pengjie. Ren, Zhumin Chen, Zhaochun Ren, Jun Ma, and Maarten de Rijke. 2019. Explainable Outfit Recommendation with Joint Outfit Matching and Comment Generation. IEEE Transactions on Knowledge and Data Engineering (2019)

    2019

  21. [29]

    Si Liu, Jiashi Feng, Zheng Song, Tianzhu Zhang, Hanqing Lu, Changsheng Xu, and Shuicheng Yan. 2012. Hi, Magic Closet, Tell Me What to Wea r!. In ACM International Conf. on Multimedia (MM’12) . 619–628

    2012

  22. [30]

    Yihui Ma, Jia Jia, Suping Zhou, Jingtian Fu, Yejun Liu, a nd Zijian Tong. 2017. Towards Better Understanding the Clothing Fashion Styles: A Multimodal Deep Learning Approach. In AAAI Conf. on Artificial Intelligence (AAAI’17) . 38–44

    2017

  23. [31]

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton van den Hengel

  24. [32]

    Takuma Nakamura and Ryosuke Goto. 2018. Outfit Generati on and Style Extrac- tion via Bidirectional LSTM and Autoencoder. In ACM Knowledge Discovery and Data Mining (KDD’18)

    2018

  25. [33]

    Jose Oramas and Tinne Tuytelaars. 2016. Modeling Visua l Com- patibility through Hierarchical Mid-level Elements. CoRR (2016). http://arxiv.org/abs/1604.00036

    2016 arXiv

  26. [34]

    Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. 2013 . On the Difficulty of Training Recurrent Neural Networks. In International Conf. on Machine Learning (ICML’13). III–1310–III–1318

    2013

  27. [35]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner , and Lars Schmidt- Thieme. 2009. BPR: Bayesian Personalized Ranking from Impl icit Feedback. In International Conf. on Uncertaintyin Artificial Intellige nce (UAI’09). 452–461

    2009

  28. [36]

    Steffen Rendle and Lars Schmidt-Thieme. 2010. Pairwise Interaction Tensor Fac- torization for Personalized Tag Recommendation. In ACM International Conf. on Web Search and Data Mining (WSDM’10) . 81–90

    2010

  29. [37]

    Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wiers tra. 2014. Stochas- tic Backpropagation and Approximate Inference in Deep Gene rative Models. In International Conf. on Machine Learning (ICML’14) . 1278–1286

    2014

  30. [38]

    Edgar Simo-Serra, Sanja Fidler, Francesc Moreno-Nogu er, and Raquel Urtasun

  31. [39]

    Karen Simonyan and Andrew Zisserman. 2015. Very Deep Co nvolutional Net- works for Large-Scale Image Recognition. In International Conf. on Learning Rep- resentations (ICLR’15)

    2015

  32. [40]

    Xuemeng Song, Fuli Feng, Xianjing Han, Xin Yang, Wei Liu , and Liqiang Nie

  33. [41]

    In IEEE Conf

    Neuroaesthetics in fashion: Modeling the perception of fashionability. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR’1 5), Vol. 00. 869– 877

  34. [42]

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, I lya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: A Simple Way to Prevent Neura l Networks from Overfitting. Journal of Machine Learning Research (JMLR) 15, 1 (2014), 1929– 1958

    2014

  35. [43]

    Andreas Veit, Balazs Kovacs, Sean Bell, Julian Mcauley , Kavita Bala, and Serge Belongie. 2015. Learning Visual Clothing Style with Hetero geneous Dyadic Co- Occurrences. In International Conf. on Computer Vision (ICCV’15) . 4642–4650

    2015

  36. [44]

    Kota Yamaguchi. 2012. Parsing Clothing in Fashion Phot ographs. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR’12) . 3570–3577

    2012

  37. [45]

    Xuemeng Song, Fuli Feng, Jinhuan Liu, Zekun Li, Liqiang Nie, and Jun Ma. 2017. NeuroStylist: Neural Compatibility Modeling for Clothing Matching. In ACM International Conf. on Multimedia (MM’17) . 753–761

    2017

  38. [46]

    Wenhui Yu, Huidi Zhang, Xiangnan He, Xu Chen, Li Xiong, a nd Zheng Qin

  39. [47]

    Zeiler and Rob Fergus

    Matthew D. Zeiler and Rob Fergus. 2014. Visualizing and Understanding Convo- lutional Networks. In European Conf. on Computer Vision (ECCV’14) . 818–833

    2014

  40. [48]

    Zeiler, Graham W

    Matthew D. Zeiler, Graham W. Taylor, and Rob Fergus. 201 1. Adaptive Decon- volutional Networks for Mid and High Level Feature Learning . In International Conf. on Computer Vision (ICCV’11) . 2018–2025

    2018

  41. [49]

    Hadi Kiapour, Luis E

    Kota Yamaguchi, M. Hadi Kiapour, Luis E. Ortiz, and Tama ra L. Berg. 2015. Re- trieving Similar Styles to Parse Clothing. IEEE Trans on Pattern Analysis and Machine Intelligence (TPAMI) 37, 5 (2015), 1028–1040

    2015

  42. [50]

    Hanwang Zhang, Zheng-Jun Zha, Yang Yang, Shuicheng Yan , Yue Gao, and Tat- Seng Chua. 2013. Attribute-augmented Semantic Hierarchy: Towards Bridging Semantic Gap and Intention Gap in Image Retrieval. In ACM International Conf. on Multimedia (MM’13) . 33–42

    2013

  43. [51]

    In International World Wide Web Conferences (WWW’18)

    Aesthetic-based Clothing Recommendation. In International World Wide Web Conferences (WWW’18). 649–658

  44. [54]

    Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaoga ng Wang, Xiaolei Huang, and Dimitris Metaxas. 2017. StackGAN: Text to Photo- Realistic Image Synthesis with Stacked Generative Adversarial Networks. I n International Conf. on Computer Vision (ICCV’17) . 5908–5916

    2017

  45. [2014]

    In European Conf

    Neural Codes for Image Retrieval. In European Conf. on Computer Vision (ECCV’14)

  46. [2015]

    In International Conf

    Image-Based Recommendations on Styles and Substitut es. In International Conf. on Research on Development in Information Retrieval ( SIGIR’15). 43–52

  47. [2018]

    In International Conf

    Neural Compatibility Modeling with Attentive Knowle dge Distillation. In International Conf. on Research on Development in Informat ion Retrieval (SI- GIR’18)

Pith tools

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