PDF Publication Title:
Text from PDF Page: 001
Practical Diversified Recommendations on YouTube with Determinantal Point Processes Mark Wilhelm, Ajith Ramanathan, Alexander Bonomo, Sagar Jain, Ed H. Chi, Jennifer Gillenwater Google Inc. {wilhelm,ajith,bonomo,sagarj,edchi,jengi}@google.com ABSTRACT Many recommendation systems produce result sets with large num- bers of highly similar items. Diversifying these results is often accomplished with heuristics, which are impoverished models of users’ desire for diversity. However, integrating more complex sta- tistical models of diversity into large-scale, mature systems is chal- lenging. Without a good match between the model’s definition of diversity and users’ perception of diversity, the model can easily degrade users’ perception of the recommendations. In this work we present a statistical model of diversity based on determinantal point processes (DPPs). We train this model from examples of user preferences with a simple procedure that can be integrated into large and complex production systems relatively easily. We use an approximate inference algorithm to serve the model at scale, and empirical results on live YouTube homepage traffic show that this model, coupled with a re-ranking algorithm, yields substantial short- and long-term increases in user satisfaction. ACM Reference Format: Mark Wilhelm, Ajith Ramanathan, Alexander Bonomo, Sagar Jain, Ed H. Chi, Jennifer Gillenwater. 2018. Practical Diversified Recommendations on YouTube with Determinantal Point Processes. In Proceedings of International Conference on Information and Knowledge Management (CIKM’18). ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn 1 INTRODUCTION Online recommendation services often present content in the form of a feed—an ordered list of items through which the user browses. Examples include the YouTube mobile homepage feed and the Face- book news feed. The goal is to select and order a set of k items such that the utility of the set is maximized. Often times recommenders do this by ranking based on item quality—assigning each item i a pointwise quality score, qi , and sorting by this score. However, this is sub-optimal as the pointwise estimator ignores correlations between the items. For example, given that a basketball video has already been shown on the page, it may now be less useful to show another basketball video. This is exacerbated by the fact that similar videos tend to have similar quality scores. Unfortunately, even if we build a good set-wise estimator, scoring every possible permutation of the ranked list is prohibitively expensive. In this paper, we apply a particular machine learning model called a determinantal point process (DPP) [4, 13, 22], which is a Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s). CIKM’18, October 2018, Turin, Italy © 2018 Copyright held by the owner/author(s). ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. https://doi.org/10.1145/nnnnnnn.nnnnnnn probabilistic model of repulsion that can be used to diversify sets of recommended items (e.g., lists of videos, books, or search results) [7, 14, 20, 21]. One key aspect of a DPP is that it can efficiently score an entire list of items rather than scoring each item individually, allowing us to better take into account item correlations. Implementing a DPP-based solution in a mature recommenda- tion system is non-trivial. First, the training methods for DPPs are significantly different from those used in typical recommender systems [3, 12, 14, 20, 21, 26, 27]. Second, integrating the DPP op- timization with existing recommenders is complex. One option would be to retool the entire infrastructure in terms of set-wise recommendations, but that would discard the large investment in, and the sophistication of, the existing pointwise estimators. Instead, we use DPPs on top of existing infrastructure as a last-layer model. This allows the various underlying system components to evolve independently. More specifically, for a large-scale recommendation system, we build a DPP using two inputs: 1) pointwise estimators from a deep neural network built for recommendations [9], which gives us a high-precision estimate of item quality qi , and 2) pair- wise item distances Dij computed in a sparse semantic embedding space. (e.g., [19]). From these inputs, we construct a DPP and apply it to the top n items in a feed. Our approach has the advantage of enabling teams of researchers to continue to develop the qi and Dij estimators simultaneously with our development of a set-wise scoring system. We can therefore achieve our diversification goals while leveraging existing investments in a large-scale prediction system. Empirical results on YouTube show substantial short- and long-term increases in user satisfaction. Our contributions are: (1) We offer a simple and effective procedure for set-wise rec- ommendations by leveraging DPPs. We define a parameter- ization and learning algorithm for DPPs that makes use of pointwise quality scores for items and pairwise distances between items. (2) We describe a practical and modular approach which can be applied in the context of latency-sensitive, large-scale recommender systems. (3) We offer both offline and online empirical results verifying that our approach improves recommendation accuracy on top of a mature, large-scale recommender system. The paper is organized as follows. We start with related works in §2. We describe the diversification needs in the current recom- mendation system in §3, defining basic terminology in §3.2. In §4, we briefly review DPPs, then describe our current choice of DPP kernel, work-in-progress on a more complex kernel, and a rank- ing algorithm that makes use of these kernels. Finally, we provide a summary of our online experimental results in §5 and end by offering a few concluding remarks in §6.PDF Image | Practical Diversified Recommendations on YouTube with Determinantal Point Processes
PDF Search Title:
Practical Diversified Recommendations on YouTube with Determinantal Point ProcessesOriginal File Name Searched:
cikm2018.pdfDIY PDF Search: Google It | Yahoo | Bing
Cruise Ship Reviews | Luxury Resort | Jet | Yacht | and Travel Tech More Info
Cruising Review Topics and Articles More Info
Software based on Filemaker for the travel industry More Info
The Burgenstock Resort: Reviews on CruisingReview website... More Info
Resort Reviews: World Class resorts... More Info
The Riffelalp Resort: Reviews on CruisingReview website... More Info
| CONTACT TEL: 608-238-6001 Email: greg@cruisingreview.com | RSS | AMP |