----------------------- REVIEW 1 --------------------- PAPER: 226 TITLE: VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback AUTHORS: Ruining He and Julian McAuley Significance: 2 (modest or incremental contribution) Soundness: 3 (correct) Scholarship: 2 (relevant literature cited but could expand) Clarity: 3 (crystal clear) Breadth of Interest: 2 (interest limited to specialty area) SUMMARY RATING: 2 (++) CONFIDENCE: 3 (certain) ----------- SUMMARIZE THE MAIN CONTRIBUTION OF THE PAPER ----------- This paper proposes a scalable factorization model to incorporate visual signals into predictors of people's opinions. Their contributions are in three folds: 1. They claim that their approach is the first attempt that integrates the visual appearances of items into personalized ranking recommender system problem. 2. In order to uncover the visual factors, they used BPR [UAI'09] as the main baseline for which they gave derivation and analysis. 3. They validated their model by showing the improvement in the performance of their model compared to the previous works, such as BPR and MMMF. Besides, they showed that their embedding technique made success in representing the latent visual space. ----------- COMMENTS FOR THE AUTHORS ----------- The strengths of this work are... 1. The very first attempt to have incorporated visual signals into recommender system. 2. The authors clearly define the limitation of the previous works and suggest a novel method to overcome the limitations 3. Thoroughly written paper with no critical flaws in the writing. Easy to understand and to follow the flow of the paper. The weaknesses of this work are... 1. Although there exists a work (GBPR [IJCAI'13]) that improves the performance over BPR, they neither mentioned it in the related work nor did they compare the performance with their model. 2. Although the authors introduced the concept of "visual bias", they could not fully persuade the readers the rationale behind it. That is, I could not be convinced the intuition behind the "visual bias" term. As an example of this kind, in the paper(SVD++ [KDD'08]) that introduces user bias and item bias for the first time, the author tries to convince readers why the terms are necessary. 3. In the experiments section, they only computed the value of AUC in order to validate the performance of their model. However, there are other various ranking metrics such as NDCG, MRR and MAP, which could have been used in this paper. Overall opinion - Although this work seems to be a work that has only incremental improvements on the previous works(BPR), in my opinion, it is well worth enough to have incorporated visual signal into recommender system for the first time. However, due to the lack of technical depth (just adding a visual term into BPR and just using the pretrained feature of CNN), I could not give more score than ++. ----------------------- REVIEW 2 --------------------- PAPER: 226 TITLE: VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback AUTHORS: Ruining He and Julian McAuley Significance: 2 (modest or incremental contribution) Soundness: 2 (minor inconsistencies or small fixable errors) Scholarship: 2 (relevant literature cited but could expand) Clarity: 2 (more or less readable) Breadth of Interest: 3 (some interest beyond specialty area) SUMMARY RATING: 1 (+ (weak accept)) CONFIDENCE: 2 (reasonably confident) ----------- SUMMARIZE THE MAIN CONTRIBUTION OF THE PAPER ----------- The paper extends an existing work BPR by including visual features. The claim is that some product categories are visually driven, such as clothes and cell phones. Therefore, factoring in visual features gleaned from product images may help. ----------- COMMENTS FOR THE AUTHORS ----------- The extension over BPR essentially involves extending the user vectors and item vectors by another K dimensionalities. For the items, these additional dimensionalities are pre-filled with features extracted from images. For the users, these additional dimensionalities are learned, to reflect the users’ preference for visual features. The paper is reasonably clear and well-written. There are two weaknesses. First, the technical extension is very minor. In fact, a similar “trick” has been used in other occasions before. For example, the following paper extends matrix factorization, by including location features in exactly the same way. The only difference is the feature type. GeoMF: Joint Geographical Modeling and Matrix Factorization for Point-of-Interest Recommendation, KDD2014. Second, the baselines are very weak (MM-MF and BPR-MF). They incorporate only basic matrix factorizations without visual features. The proper baseline should be at least some form of combination of both MF and visual features. For example, we can use normal MF to learn the user’s preference for visual features, and then combine the results with the pairwise comparison (e.g., alpha x BPR-MF + (1-alpha) MF for visual features). ----------------------- REVIEW 3 --------------------- PAPER: 226 TITLE: VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback AUTHORS: Ruining He and Julian McAuley Significance: 2 (modest or incremental contribution) Soundness: 3 (correct) Scholarship: 3 (excellent coverage of related work) Clarity: 3 (crystal clear) Breadth of Interest: 3 (some interest beyond specialty area) SUMMARY RATING: 1 (+ (weak accept)) CONFIDENCE: 3 (certain) ----------- SUMMARIZE THE MAIN CONTRIBUTION OF THE PAPER ----------- This paper proposed to incorporate visual features from a pretrained convolutional neural network into the Bayesian Personal Ranking (BPR) formulation for enhancing the accuracy of product recommendation for items with associated with images. The main contributions are: (1) introducing visual features and modifying the BPR-MF formulation accordingly (2) thorough comparison of the proposed methods with the existing ones ----------- COMMENTS FOR THE AUTHORS ----------- - Using the visual features has been shown to be useful for the recommendation, which is not a surprise for the problem setting being considered. It is not clear how effective it will be for settings like not all items are associated with images (missing data) and the images are not as carefully created as what being considered in the paper. - A visual bias term is defined in Eq. 4. Is there any reason why this is not defined in the visual rating space but the visual feature space? ------------------------- METAREVIEW ------------------------ PAPER: 226 TITLE: VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback All three reviewers voted to accept this paper. The strengths were that it introduced visual features and that it was nicely written. The weakness was that the technical extension was not very challenging, and the baselines for comparison were weak.