PDF Publication Title:
Text from PDF Page: 002
CIKM’18, October 2018, Turin, Italy Mark Wilhelm, Ajith Ramanathan, Alex Bonomo, et al. 2 RELATED WORK Current recommender research is generally focused on improving the pointwise estimate qi —a prediction of how much a user will enjoy one particular item. This line of research was initially started over 20 years ago with user-based collaborative filtering [34] and item-based collaborative filtering [36], and then refined using ma- trix factorization techniques [19]. In our system, we now obtain these pointwise estimates from deep neural networks, in which a user’s preference features are combined with the the item features to estimate how much the user will enjoy that content [9]. During the course of these refinements, there was also significant study of users’ need for novelty and diversity in the recommen- dation results [16, 24, 29, 39, 41, 43, 45]. Similarly, there has been significant work on diversification in information retrieval systems such as web search [6, 8, 10, 11, 15, 33, 35, 40, 42]. Considering all of this literature, researchers have proposed many notions of diver- sification. Here we briefly summarize and contrast two different perspectives on the purpose of content diversification. 2.1 Diversification to Facilitate Exploration First, diversification is sometimes seen as a way to facilitate explo- ration; showing the user more diverse content will (A) help them discover new topics of interest or (B) help the recommender system discover more about the user. For discovering user intent, there is a thread of work in infor- mation retrieval on using taxonomy to resolve ambiguity in user intent [2, 35]. For instance, IA-Select in [2] uses a taxonomy to cover an ambiguous query, and then aims to maximize the probability that the user will select at least one returned result. Santos et al. [35] estimate how well a ranked result covers an uncovered aspect of the answer for an ambiguous query. Whereas these methods require a problem-specific taxonomy, the solution we present only requires uncalibrated item distances (the calibration is learned as part of the training procedure). For facilitating topic discovery, if a topic contains multiple as- pects, then one can further divide the topic into subtopics, and then make sure that each subtopic is well-covered by the results retrieved [10, 40, 42]. For instance, Dang et al. [10] proposes to return a result list that has per-topic coverage that is proportional to that topic’s popularity. As another example, Perez et al. [32] uses categories of businesses to ensure recommendation results for a local business recommendation problem has sufficient topical coverage. In [23], Kwon and Adomavicius argue that users essen- tially want a multi-criteria rating system, in which they can specify which aspects of the recommendation they want. In contrast to these methods, we are able to learn the appropriate amount of coverage based directly on user behavior. Perhaps it is worth noting that while exploration likely does happen to some extent in all recommenders, imperfect information about user preferences and correlated recommendations are fun- damentally orthogonal problems. Exploration is still needed in the presence of uncorrelated recommendations, and diversification is still needed in the presence of perfect information. Somewhat consistent with the exploration perspective on diver- sity is that it is a secondary product objective. This perspective suggests a fundamental trade-off between diversity and utility, and can be seen in work that focuses on increasing a diversity metric as much as possible, without hurting the utility too much. In recent work that is similar to ours, Chen et al. [7] described the use of DPPs to optimize exploration without hurting the user utility. Their DPP kernel parameterization is different, and our work offers not just offline experiments but also a large-scale online experiment. More importantly, in contrast, we optimize for user utility while increasing diversity using DPP. 2.2 Diversification in Service of Utility A different perspective on diversity, and the one we adopt for this work, is that diversity operates directly in service of utility—by ap- propriately diversifying impressions, one can maximize the feed’s utility. From this perspective diversity is purely about the corre- lation of interactions, and increasing diversity means replacing redundant video impressions with alternatives that a user is more likely to concurrently enjoy. These new videos generally have lower individual scores but lead to a better page overall. Concisely, one way of achieving diversity is avoiding redundancy, which is particularly important for recommender systems [5, 30, 32, 43, 45]. For instance, in their seminal work in 2005, Ziegler et al. [45] minimize the similarity between recommended items using a greedy algorithm with a taxonomy of books. The output is then merged with a non-diversified result list using a diversification factor. In another seminal work in information retrieval, Carbonell and Goldstein [5] propose the maximal marginal relevance (MMR) method. This method involves iteratively selecting one item at a time. The score of an item under consideration is proportional to its relevance minus a penalty term that measures its similarity to previously selected items. Other explicit notions of redundancy are studied in [32], which uses a decay function on pairwise similarities. More recently, Nassif et al. [30] describe an approach using sub- modular optimization to diversify music recommendation. Lin and Bilmes [25] describe a way to use submodular functions to perform document summarization, a task with similar coverage goals as set diversification tasks. Tschiatschek et al. [38] describe an approach using submodular maximization to select sequences of items, while Teo et al. [37] describe using submodular diversification to re-rank top items based on category. Our goals are quite similar in nature, but use a different optimization technique. Additionally we do not take item diversity as an a priori goal; our aim is simply trying to increase the number of positive user interactions by making diver- sity information available to the overall recommendation system. One can imagine iterating on the model presented here to express a personalized notion of diversity. The recommended content feed is also a convenient context for this approach, since (unlike in search) users are typically not looking for a specific item and may interact with multiple items in the course of a session. The notion of redundancy can be further broken up into two separate relevance notions: substitutes and complements. These notions have been employed by several recommender systems [28, 44]. In e-commerce recommendation applications, before the user makes a purchasing decision, it might be more helpful to offer substitutes of candidates under consideration, while complement products might be offered after the user has made a purchase.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 |