----------------------- REVIEW 1 --------------------- PAPER: 562 TITLE: Ups and Downs: Modeling the Visual Evolution of Fashion Trends with One-Class Collaborative Filtering AUTHORS: Ruining He and Julian McAuley OVERALL EVALUATION: 2 (accept) ----------- REVIEW ----------- The paper focuses on addressing the one-class collaborative filtering problem in settings where temporal dynamics are present and visual awareness is desirable. In particular, the paper focuses on modeling fashion trends in amazon transaction records. The presented approach is based on matrix factorization and a Deep Convolutional Neural Network model capturing visual dimensions of shopping products from photos. A pretrained CNN model is used. Evaluation is based on hold-out experiments. Different MF methods are instantiated and compared. The quantitative evaluation results presented are interesting. Qualitatively, it seems like the visual patterns pick up not only on fashion trends (as intended) but also on different model poses or photo setups (full body shots vs. upper body shots in Fig3, photos with a model vs. photos without a model in Fig 5) which are not related to fashion trends. It seems like in a scenario with standardized photo shooting setups this might not be a problem, but in the amazon dataset this seems to represent an unintended side effect of the model. The authors should address this issue in the final version of the paper. Overall, a benefit of the presented model lies in the fact that because of the visual layer, recommendations can be explained via fashion trends that change over time. So in addition to recommendation, the model could easily be adapted to study the rise and fall of fashion (or temporal) trends in given data. The paper tackles an important problem, the evaluation setup is sound, the approach is technically interesting and the paper overall is well written. ----------------------- REVIEW 2 --------------------- PAPER: 562 TITLE: Ups and Downs: Modeling the Visual Evolution of Fashion Trends with One-Class Collaborative Filtering AUTHORS: Ruining He and Julian McAuley OVERALL EVALUATION: 2 (accept) ----------- REVIEW ----------- This paper presents a method to model fashion-aware preferences of users. I really enjoyed the paper and found it very interesting. I appreciate the effort the authors have made in the presentation, since, despite the many equations needed, it is easy to follow the process and know at each point what is being referenced. I have some comments I will detail next, but my main concerns are: a) why is the review rating of the user not being used? I understand the goal is One Class CF, but since the dataset is being created by considering the users' review histories, I wonder how different the results could be if items with low ratings are removed from the user profiles. It is true that this would impact to every algorithm being tested by the authors, but I think it could be interesting to incorporate this into the data processing stage. b) how different is the work done in [26] from the one presented here? It is mentioned incidentally in Section 4.1, and I find more details should be included at this point (even though in Section 5 it is mentioned that a different function is learnt). c) Section 3.3.2 should be explained in more detail, right now it is not clear how the dynamic programming procedure is applied nor why/how the sequence segmentation problem is important. Other comments: - It is not true that BPR-TMF always get an improvement over BPR-MF: in the Men dataset the performance is worse. - I am not 100% confident I understood the evaluation methodology presented in Section 4.3: is it true that each user appears only once in each validation and test set? If this is the case, maybe an explanatory sentence could be added, something regarding the size of those sets (equal to the number of users). If that is not the case, then the sentence "an item ... for validation ... and another for testing" should be rewritten. - Minor comments: * I think there is a typo in the last line before Section 4.6.2: instead of "last epoch" should be "last row". * In the first paragraphs of Sections 3 and 4, the use of 'before' seems incorrect. * There are also several cases where contractions (it's, doesn't, what's) are used, change them to their expanded form. ----------------------- REVIEW 3 --------------------- PAPER: 562 TITLE: Ups and Downs: Modeling the Visual Evolution of Fashion Trends with One-Class Collaborative Filtering AUTHORS: Ruining He and Julian McAuley OVERALL EVALUATION: 1 (weak accept) ----------- REVIEW ----------- This paper presents an approach to include visual appeal of fashion items into the predictive modelling component of a recommender systems. The paper has a particular focus on fashion items, and area the authors claim to be less infommed by user ratings and more informed by appeal and personal preference. I like this paper. It is well written, excellently communicated and a pleasure to read. I like, and appreciate the area that the authors are working in as a real world challenge, outside the typical recommender system space. The authors focus on improving the existing MF approach to incorporate global and local factors pertaining to images of clothing items. The authors should explain the abbreviations used for algorithms in the trial using the terms in the abbreviations for easy of reading. My hesitation with his work lies in the examination of an entire image to determine preference for one item in the image. In teh cases shown in Figure 3 there are other commonalities in teh images beyond the t-shirts - like the full length shots of the model, the position of the models arms etc. These do impacts a person's desire for an items but it's not actually related to the item but the way that it is presented. I'd be keen to see the performance of the approach on a more limited set of images that were more closely aligned. I'd also like to see how the photography and style of images has changed in the 10 years. This could plan a very large part in the epochs being identified and could significantly skew the results obtained. The authors have examples that illustrate when their approach models the real world. I'd be keen to know more from the authors on plans to validate this in other ways. I encourage the authors to continue this work, I believe the area is a key on for the recommender systems area and a good starting point, but there is much more to do. ------------------------- METAREVIEW ------------------------ PAPER: 562 TITLE: Ups and Downs: Modeling the Visual Evolution of Fashion Trends with One-Class Collaborative Filtering This paper is novel, well-presented, and has an interesting evaluation comparing to many state-of-the-art approaches. While there are some minor concerns, the paper is well suited and ready for acceptance at WWW. Please address some of the comments from the reviewers in the camera-ready version.