PDF Publication Title:
Text from PDF Page: 003
Practical Diversified Recommendations on YouTube with Determinantal Point Processes CIKM’18, October 2018, Turin, Italy 2.3 Related Works Summary & Design Choices In summary, many researchers before us have studied how to im- prove diversity in both recommendation and search results. Some researchers deal with several of these diversity notions at the same time. For instance, Vargas et al. [39] addresses coverage and re- dundancy, as well as the size of the recommendation list. We are interested in a technique that works well in practice in a large-scale recommendation system that can be served to hundreds of millions of users per day. The notion of diversity should be flexible enough that it can evolve over time. As a result, we chose not to pursue taxonomic or topic-coverage approaches, as they require some ex- plicit representation of diversity (e.g., an explicit guess at the user’s intent or topic coverage). Instead, we propose an approach using determinantal point pro- cesses (DPPs) [4, 7, 13, 22]. DPP is a set-wise recommendation model that only requires two explicit and natural elements: how good is each item for the user, and how similar are each pair of items. As a result, our focus is on eliminating redundancy. 3 BACKGROUND 3.1 YouTube Homepage Feed Overview and the Need for Diversification The overall structure of the system for generating the video recom- mendations on a user’s YouTube mobile homepage feed is illustrated in Figure 1. The system is comprised of three phases: (1) candidate generation, wherein the feed items are selected from a large cat- alogue, (2) ranking, which orders the feed items, and (3) policy, which enforces business needs such as requiring that some content appear at a specific position on the page. Phases (1) and (2) both make heavy use of deep neural networks [9]. Figure 1: The basic serving scheme. Candidate generation is substantially influenced by the previous behavior of the user on our system, and computes relatively simple measures of how well items match user preferences. For example, co-utility is one measure that is used: if a user enjoyed video A, and many other users who enjoyed A also enjoyed B, then B might be selected in the candidate generation phase. The ranking phase also makes heavy use of user features, but additionally relies on richer item features (such as embeddings of the video in some semantic space). As one might expect, the ranking phase tends to give similar videos similar utility predictions, leading to feeds that have repetitive content and, often, runs of very similar videos. In order to mitigate the redundancy problem, at first, we intro- duced heuristics in the spirit of [32, 45] to the policy layer, such as a requirement that an individual uploader can contribute no more than n items to any user’s feed. While this rule is somewhat effective, our experience is that it interacts quite poorly with the underlying recommendation system. Since the candidate genera- tion and ranking layers are unaware of this heuristic, they make suboptimal predictions by wasting space on items that will never be presented. Furthermore, as the first two layers evolve over time, we need to repeatedly retune the parameters of the heuristics—a task that is expensive and hence in practice is not done with enough frequency to maintain much of the rule’s effectiveness. Finally, the interactions between multiple types of heuristics yields, in practice, a recommendation algorithm that is very hard to understand. The result is a system that is suboptimal and difficult to evolve. 3.2 Definitions To be more precise, let us denote the observed interactions of a user with items in a given feed as a binary vector y, (e.g., y = [0, 1, 0, 1, 1, 0, 0, . . .]), where it is understood that the user typically will not look at the entire feed, but will start at the lower num- bered indices. Our present goal is to maximize the total number of interactions: G′= yui. (1) u ∼Users i ∼Items In order to train models from records of previous interactions, we try to select the parameters of the model to maximize the cumulative gain by reranking the feed items: G= yui, (2) u ∼Users i ∼Items j where j is the new rank that the model assigns to an item. This quantity increases as we rank interactions more highly. (In practice, we minimize jyui instead of maximizing yui , but the two expres- j sions have the same optima.) In the following discussion, we will drop the u subscript for simplicity, although all values should be assumed to differ on a per-user basis Let us further assume we are provided with some black box estimates of y’s quality: qi ≈P(yi =1|featuresofitemi). (3) The obvious ranking policy is to sort the items according to q. Note though that qi is a function of only a single item. If there are many similar items with similar values of qi they will be ranked adjacent to each other, which may lead to the user abandoning the feed. Given that our ultimate goal is to maximize the feed’s total utility, we call two items similar when: P(yi =1,yj =1)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 (Standard Web Page)