Reviewer 4 (chair) Expertise Expert The Meta-Review This short paper explores the problem of recommending user-generated content, where users are not only consumers but also producers of items. It proposes a matrix factorization method extended by producer latent factors and an additional term to model the consumer-producer preference (in additional to the consumer-item preference). The paper includes a short discussion of related work, in particular the Vista method [2]. A preliminary experimental evaluation on two datasets demonstrates that the proposed method outperforms several baselines in terms of AUC. The problem is important. The proposed approach is simple but principled. The paper is well-written and clear. However, it is unclear how the proposed method compares to the many existing methods for the recommendation of blogs, tweets, images, which are also types of user-generated content, since these methods are neither discussed nor experimentally compared against. Given that this is a short paper, the paper is borderline. It may be acceptable, if there remains enough space. Final Recommendation after Meta-Review Probably accept: I would argue for accepting this paper. Reviewer 1 (PC) Review Rating Probably accept: I would argue for accepting this paper. Expertise Passing Knowledge Contribution Novel idea of modeling user-producer affinity to improve MF results in UGC platforms. Experimental results prove its effectiveness. Relevance to RecSys Platforms like Reddit and Pinterest are growing a lot and are more popular than ever. This paper fits well several core topics as social recommenders and novel algorithms. Your Review This short paper describes a novel matrix factorization algorithm modeling user-producer affinity for improving recommendation accuracy in UGC platforms such as Reddit, Pinterest, etc. Offline evaluations on 2 medium-size datasets prove the effectiveness of the algorithm. The authors adopt the biased Matrix Factorization algorithm and add a consumer-producer appreciation based on two different embeddings, one for the consumer and one for the producer role. Those embeddings are obtained from a single core user embedding by applying two separate transformation matrices. The objective function is based on the BPR framework which is a learning to rank method for Matrix Factorization. It would also be interesting to see how the CPRec algorithm would perform when changing the BPR objective function. The weighted ALS algorithm proposed in "Collaborative Filtering for Implicit Feedback Datasets" has comparable performances to BPR. I overall like the idea and the experiments prove its effectiveness. I have some minor reservations about the structure and content of the paper. The authors could probably motivate better the intuition of modeling user-producer affinity in the Introduction Section. Also the derivation of the W^c and W^p matrices could use some more space. I would personally remove part of the Related Work Section on Recommender Systems which can be obvious for the RecSys conference and add more details to the core contributions, especially to 3.2 and 3.3 Sections. Reviewer 2 (PC) Review Rating Probably accept: I would argue for accepting this paper. Expertise Knowledgeable Contribution The paper presents CPRec, which is a method for recommending contents on UGC platforms. Since users in UGC platforms can serve as producers and consumers at the same time, the authors propose to learn a base embedding for each user and two transformation matrices to project the base embedding to two role embeddings. They also propose a variation of the classical matrix factorization approach to model consumer-item preference and consumer-producer appreciation simultaneously. Relevance to RecSys The authors present a novel approach for content recommendation in UGC platforms, which is highly relevant to the RecSys community. Your Review ----------- Strong Points ----------- 1) A reasonable approach to address an important problem 2) Very well written and easy to follow 3) Comprehensive experiments and significant improvements over existing methods ----------- Weak Points ----------- 1) The proposed approach is a bit incremental ----------- Detailed Review ----------- This paper addresses an important problem of making recommendations in UGC platforms. Noticing that each user assumes both the role of a consumer and a producer, the authors propose to learn two role embeddings for each user. Concretely, this is achieved by learning a core embedding vector per user and two global transformation matrices to transform the core embeddings into the corresponding consumer embeddings and producer embeddings. Overall, the idea of CPRec is straightforward but well motivated. In terms of novelty, the proposed method is a little bit incremental when compared to Vista / FMs, but the differences are clearly discussed in Section 3.4. In terms of experiments, the performance gain over the baseline methods is significant. Reviewer 3 (PC) Review Rating Borderline: Overall I would not argue for accepting this paper. Expertise Knowledgeable Contribution The key contribution of the paper is that the authors proposed to provide content recommendation in UGC-based platforms. Specifically, the proposed model learn a core embedding for each user and two transformation matrices to project the user’s core embedding into two ‘role’ embeddings (i.e., a producer and consumer role). Relevance to RecSys The paper is relevant to recsys. Your Review In this paper, the authors proposed to conduct content recommendation in UGC platforms, where the users are content producers and at the same time content consumers. To do so, the authors proposed a recommendation method CPRec, which learns two role embeddings (for consumer and producer roles) derived from the same core user embedding. My key concerns with the paper is that, the authors did not clearly point out the difference between the proposed research and traditional social content recommendation research such as twit recommendation or blog post recommendation. In twit recommendation for example, the users also serve as the consumer and producers roles at the same time. In this sense, content recommendation in UGC platform is not a new task. This is reflected on the datasets used for evaluation. The have been a vast amount of work providing recommendations based on Pinterest and Reddit dataset, and they are working on very similar or even the same problem. The authors should clearly point out the difference with these research, and compare with the other models on the same dataset. In terms of experiments, the authors only used AUC as the evaluation measure, however, AUC is not a very frequently used measure and it not easy to intuitively understand the performance of the proposed method based on AUC. The authors are suggested to use more intuitive measures such as precision, recall, f-measure, NDCG, etc.