Congratulations! We are pleased to inform you that your submission has been accepted to WSDM 2017: 238: Bartering Books to Beers: A Recommender System for Exchange Platforms This year we received 505 valid submissions, 35% more than any past year. We were able to accept only 80 of these, representing an accept rate of about 16%. All accepted papers will appear in the proceedings as full-length publications and will also be presented as both a talk and a poster for interactive discussion. We will send the following additional information to you shortly: 1. Details with the session and duration allocated for your oral presentation (long talk or spotlight talk), and the session for your poster. 2. Information on the procedure and timing for submitting your camera-ready version. 3. Information on registering for WSDM 2017. The program committee worked very hard to thoroughly review all the submitted papers and to provide action points to improve your paper. All papers were reviewed by at least three program committee members (and 90% of papers were reviewed by four), and by a senior PC member to oversee discussion amongst the reviewers and provide an overall recommendation for the paper. We trust that you will take their suggestions into account when producing the final version of your paper. Please note a couple of procedural items: 1. Per the call for papers, if you must modify the authorship of your paper, please get in touch with the program chairs first. 2. As WSDM is a forum for presenting and discussing current research in web search and data mining, at least one author must complete the regular conference registration and commit to present the paper at the conference. If you have multiple papers accepted, it’s fine for a single person to present more than one. Once again, congratulations -- and we look forward to seeing you in Cambridge, England! The reviews for your paper are below. Best regards, Andrew Tomkins and Min Zhang WSDM 2017 Program Co-Chairs ----------------------- REVIEW 1 --------------------- PAPER: 238 TITLE: Bartering Books to Beers: A Recommender System for Exchange Platforms AUTHORS: Jérémie Rappaz, Maria-Luiza Vladarean, Julian Mcauley and Michele Catasta OVERALL EVALUATION: -4 (Reject: I think this paper should be rejected) REVIEWER'S CONFIDENCE: 2 (I have passing familiarity with this area) Rank of the paper: 2 (Top 50% in my batch) ----------- Strengths of the paper ----------- 1. This paper introduced a new approach to recommending trades in the context of online bartering platforms. 2. The proposed model incorporated different kinds of external information to improve the recommendations. 3. Experiments shows the effectiveness of the proposed model. ----------- Weaknesses of the paper ----------- 1. This paper spent large space to analyze the data sets, while the proposed model is not clear and how to derive the solutions. Overall, it is not easy to follow. 2. The subscripts in Eq (1) are not correct. 3. The baseline of MF is rather weak, and the authors should consider some more state-of-the-art baselines. 4. Why you titled the paper as “Bartering Books to Beers”, I think the authors conducted the experiments on sole data source. ----------- Review ----------- ee the detailed comments of strong and weak points. ----------------------- REVIEW 2 --------------------- PAPER: 238 TITLE: Bartering Books to Beers: A Recommender System for Exchange Platforms AUTHORS: Jérémie Rappaz, Maria-Luiza Vladarean, Julian Mcauley and Michele Catasta OVERALL EVALUATION: 3 (Accept: I think this paper should be accepted) REVIEWER'S CONFIDENCE: 2 (I have passing familiarity with this area) Rank of the paper: 4 (Top paper in my batch) ----------- Strengths of the paper ----------- - relevant and interesting problem - clear motivation and presentation - novel approach and insightful evaluation ----------- Weaknesses of the paper ----------- - unclear separation of validation and test data - lack of significance validation ----------- Review ----------- The authors propose a latent factor approach for recommending items in bartering platforms. Rather than relying on users’ explicitly defined wish and giveaway lists, the proposed approach models users’ preferences using their past transactions, as well as social and temporal biases that are seemingly prevalent in such platforms. Recommendation for bartering platforms is an interesting problem. The paper is very well written and organized. The proposed approach is clearly motivated based upon observations drawn from real bartering transactions and is competently derived. The relevant literature is adequately described and the approach is novel to the best of my knowledge. The conducted evaluation is generally sound and insightful and demonstrates the benefits of the various components of the approach. Still, the authors should clarify whether the reported results use separate data points from those used for tuning the hyperparameters of the various approaches. In addition, the reported improvements should be checked for statistical significance. ----------------------- REVIEW 3 --------------------- PAPER: 238 TITLE: Bartering Books to Beers: A Recommender System for Exchange Platforms AUTHORS: Jérémie Rappaz, Maria-Luiza Vladarean, Julian Mcauley and Michele Catasta OVERALL EVALUATION: 3 (Accept: I think this paper should be accepted) REVIEWER'S CONFIDENCE: 4 (I have worked on this problem/area) Rank of the paper: 3 (Top 25% in my batch) ----------- Strengths of the paper ----------- Interesting problem Good selection and analysis of data sets Competent application of techniques to get decent improvements ----------- Weaknesses of the paper ----------- Techniques are fairly standard Minor gaps in treatment of related work ----------- Review ----------- This paper describes a recommendation system for swapping platforms where users can exchanges items they do not need anymore for other items they want. There are a number of such commercial platforms on the web, but the amount of research on these systems is still limited. In terms of previous work, the following two early works should be cited: Haddawy et al: Balanced Matching of Buyers and Sellers in E-Marketplaces: The Barter Trade Exchange Model, 2003 Mathieu: Match-Making in Bartering Scenarios, MS Thesis, University of New Brunswick, 2005 In the second paragraph of the intro: "However, ... platforms ... lack mechanisms to recommend trades". Note that while these specific systems do not have such mechanisms, there are others that do. In particular, barterquest.com has (or at least used to have) a matching engine that recommends trades (see also US Patent 8,645,203). Concerning the discussion of CSEM in 3.1, I am not convinced that it is fundamental to CSEM that items can be recommended to only one user at a time. This may be an assumption, but it does not seem to be central to the technique. Also, I fund the discussion in the next to last paragraph in 3.1 confusing. It seems that whether we model the problem as a weighted directed graph (to find matching or cycles in the graph), is independent of how we determine the edges and their weights. We could just use the have and want lists to create the edges, but we could also use more advanced recommendation techniques to determine the edges and weights. This discussion seems to conflate these two choices. Please rewrite or clarify. I would like to see more discussion of the "comparable value" issue in section 3, and also the issue of focusing on specific domains such as beer and books. While these are reasonable assumptions for the particular platforms that are studied, they do limit the applicability somewhat. In particular, books and beer are easier because we can exactly determine which book or beer people are talking about -- or can users in general ask for "any romance novel" or "any sweet Belgian style beer"? There are other platforms that allow people to exchange different goods services --say, a bicycle in return for painting a garage -- that do not follow these assumptions. Also, the title seems a little misleading since the system studies bartering beers for beers and books for books. Or am I missing something? Still, a nice and well-written paper. ----------------------- REVIEW 4 --------------------- PAPER: 238 TITLE: Bartering Books to Beers: A Recommender System for Exchange Platforms AUTHORS: Jérémie Rappaz, Maria-Luiza Vladarean, Julian Mcauley and Michele Catasta OVERALL EVALUATION: 3 (Accept: I think this paper should be accepted) REVIEWER'S CONFIDENCE: 3 (I have read key papers in this area) Rank of the paper: 4 (Top paper in my batch) ----------- Strengths of the paper ----------- -The paper propose a novel and interesting use of collaborative filtering in the problem of online bartering exchanges -It presents very interesting insights into how current methods that restrict bartering to users' explicit wish-lists are limited by the small number of eligible swapping pairs -The experiments are well-designed and show non-trivial gains due to each of the model attributes (temporal dynamics, social bias and bidirectionality). ----------- Weaknesses of the paper ----------- -It would have been nice to also compare the AUC performance of the paper's method against existing methods like BVEM on a dataset where the number of eligible swapping pairs is high ----------- Review ----------- The paper proposes the use of collaborative filtering to capture implicit user-item preferences for the problem of online bartering. By not restricting trades to confirm to existing user wishlists, the authors resolve the issue of sparsity of eligible swapping pairs in many real-world bartering exchanges. In particular, the paper provides swap recommendations between a pair of users and items using a model based on matrix factorization. The model uses implicit feedback signals from transaction history. and explicitly factors in a social bias term, a temporal dynamics term, and the effect of bidirectionality. This is a very nicely written paper that proposes a novel way of looking at online-bartering exchanges from a collaborative-filtering angle. The specific enhancements to basic matrix factorization for this applications - in particular the social bias and the reciprocal interest terms - make intuitive sense and seem to be quite effective in terms of improving the model performance on three different real-world datasets. The experiments seemed very well designed, and show non-trivial gains due to each of the model attributes (temporal dynamics, social bias and bidirectionality). A few comments -It would be nice to also compare the AUC performance of the paper's method against existing methods like BVEM on a dataset (or a subset of the data) where the number of eligible swapping pairs is high, and check whether the gains still hold. -In terms of the reciprocal interest's aggregate function f, instead of using the mean of the two directional terms, perhaps it might make sense to use the the minimum instead? Does strong preference from one side really offset weak preference in the other direction, or do you need both preferences to be relatively high. ------------------------- METAREVIEW ------------------------ PAPER: 238 TITLE: Bartering Books to Beers: A Recommender System for Exchange Platforms The paper is well written and organized, and describes usage of collaborative filtering techniques in the bartering domain. Much of the the space is devoted to explaining the unique challenges that bartering poses, and to describing the datasets. Overall, I agree with the majority of reviewers that this is an interesting work that raises a novel application of collaborative filtering. ================================================== RecSys 2016 (Reject) ================================================== We regret to inform you that your submission: Paper#: 122 Title: SwapIt: a Recommender System for Bartering Platforms has not been accepted for publication at the ACM RecSys 2016 conference. As usual, RecSys is a very competitive forum. This year, RecSys received a total of 294 submissions and accepted 29 long papers (acceptance rate 18,2%) and 22 short papers (acceptance rate 20%). The review process was extremely selective due to the high number of submissions. The Program Committee worked very hard to ensure that every paper got at least three reviews. All papers, particularly those at the borderline, were also reviewed and thoroughly discussed by a Senior Program Committee member and by the PC Chairs. Please find the reviews for your paper attached to this e-mail. We trust that you will find the reviews helpful to further improve your work. Please consider submitting your paper to the RecSys 2016 poster session (http://recsys.acm.org/recsys16/call/#content-tab-1-5-tab) or to one of the workshops (http://recsys.acm.org/recsys16/workshops). Thank you again for submitting to RecSys 2016, we regret that your submission was not accepted but we certainly look forward to seeing you at the conference in Boston! Conference registration is open already - https://recsys.acm.org/recsys16/registration. Best regards, Jill Freyne and Pablo Castells ACM RecSys 2016 Program Chairs ----------------------- REVIEW 1 --------------------- PAPER: 122 TITLE: SwapIt: a Recommender System for Bartering Platforms AUTHORS: Jérémie Rappaz, Maria-Luiza Vladarean, Julian McAuley and Michele Catasta OVERALL EVALUATION: 0 (borderline paper) REVIEWER'S CONFIDENCE: 4 (high) Relevance for RecSys: 4 (good) Novelty: 3 (fair) Technical quality: 3 (fair) Significance: 3 (fair) Presentation and readability: 3 (fair) ----------- Review ----------- This paper adopts Matrix Factorization and Factorization Machines to model users’ latent preferences for items, and then arithmetic mean to aggregate two users’ preferences for quantifying their reciprocal interest. The developed method is applied to bartering platforms to give recommendation. Experiment shows that the variation based on Factorization Machines performs better in most scenarios. The idea is interesting, and the paper is well written and easy to follow. As this paper is submitted to “application” category, it may be acceptable they mainly use existing techniques to solve a critical recommendation problem. My main concern is about its aggregation function and evaluation. 1. They used arithmetic mean to obtain cross-preference score, by which “a strong preference from one user compensates for a potentially weaker one coming from the other”. But this is questionable for swapping items, because the swapping only occurs when both users’ preferences for each other’s items are strong. In my view, the harmonic mean sounds more reasonable and relevant because it is high when both input values need to be high. Though they mention harmonic mean performs worse than arithmetic mean, no experimental results are given to support this claim. Moreover, they state in future work that “we will experiment aggregating schemes other than the Pythagorean means”, but in my opinion, it is worth experimenting more aggregating schemes in this paper. 2. The evaluation metrics for baselines and SWAPIT are difference (in Section 6.1), so it is unfair to put the results together to compare. It is neither clear how AUC_SWAPIT was used to measure the method’s recommendation accuracy for bother users who swap, and how the baselines ALS and FM were used to solve the bartering problem. 3. It should be more interesting to compare the developed method SWAPIT with the state-of-the-art method BVEM [23] in the experiment. 4. They emphasized the importance of providing serendipitous recommendation in the bartering scenario, but the “serendipity” was not formally measured in their experiment. More clarifications are needed in the following places. 1. In section 1, “Also, previous approaches do not yield recommendations that are predictive for the observed transactions, possibly suggesting that the users are guided by criteria other than the ones considered there.” Please be more specific of what criteria guide users. 2. It is hard to comprehend table 4, as the parameter belta was not introduced. 3. Table 3 is not cited. 4. In Section 4, the paragraph starting with “To better understand how a matching based on listed preferences would perform in terms of generating recommendations, xxx.” should better be accompanied by a figure as example of the described bipartite graph. 5. Section 5.1., “A second method used for modeling user preferences is Factorization Machines (FMs) [17]. This approach allows us to incorporate not only user-item interaction information, but also numerical and categorical item features, thus alleviating the cold-start problem.” What numerical and categorical item features were incorporated? Please be more specific. Presentation: 1. The section 2 is a bit lengthy. The related work on TTCC and CSEM could be reduced, as they are not very relevant to the focus of this work. In my view, this section could be divided into two subsections, one is about related work on bartering, and another is about latent preference modeling (ALS, FM, and BPR). 2. English error: “xxx, we add randomly selected a negative examples with a label of -1” (page 6, in section 6.2). ----------------------- REVIEW 2 --------------------- PAPER: 122 TITLE: SwapIt: a Recommender System for Bartering Platforms AUTHORS: Jérémie Rappaz, Maria-Luiza Vladarean, Julian McAuley and Michele Catasta OVERALL EVALUATION: 0 (borderline paper) REVIEWER'S CONFIDENCE: 3 (medium) Relevance for RecSys: 4 (good) Novelty: 4 (good) Technical quality: 4 (good) Significance: 4 (good) Presentation and readability: 4 (good) ----------- Review ----------- This paper presents a new approach and set of data sets for recommending exchange partners in online bartering platforms. The work is (to my knowledge) novel. The evaluation and algorithm are unclear, lacking some detail: in particular, how exactly was the input data partitioned for cross-validation? Were exchange tuples split, were users split, or some other mechanism? What is the relationship between the authors' proposed *algorithms* and *metrics*? It looks like ALS and FM are applied with two different metrics (baseline and SwapIt), but in Section 5 the authors talk about their approach, which seems to be applying FM on top of a particular preparation of the input data? The relationship between all these items should be clearer. I appreciate the authors' commitment to reproducibility in the form of publishing their code and describing algorithm parameters. I do not see anything about sharing the data sets - I would like to see that if relevant legal requirements permit it. ----------------------- REVIEW 3 --------------------- PAPER: 122 TITLE: SwapIt: a Recommender System for Bartering Platforms AUTHORS: Jérémie Rappaz, Maria-Luiza Vladarean, Julian McAuley and Michele Catasta OVERALL EVALUATION: 0 (borderline paper) REVIEWER'S CONFIDENCE: 4 (high) Relevance for RecSys: 4 (good) Novelty: 4 (good) Technical quality: 4 (good) Significance: 4 (good) Presentation and readability: 2 (poor) ----------- Review ----------- This paper presents a method for bartering-based recommendation. Specifically, the authors developed the model based on Matrix Factorization and Factorization Machines. Overall, the problem is interesting. Also, the authors have provided comprehensive experiments on real-world data to show the performances. However, the writing needs to improve substantially. The paper is poorly organized and there are various grammar errors.