Download presentation
Presentation is loading. Please wait.
Published byGwendolyn Kelly Modified over 8 years ago
1
Innovation Team of Recommender System(ITRS) Collaborative Competitive Filtering : Learning Recommender Using Context of User Choice Keynote: Zhi-qiang You Alibaba Research Center For Complexity Sciences 2014.12.26 SIGIR.11(citation 60) author : Shuang Hong Yang (Georgia Tech)
2
Overview Innovation Team of Recommender System(ITRS) 1 Overview 背景介绍问题描述模型及学习数据集及结果表现
3
Innovation Team of Recommender System(ITRS) 2 背景描述 Collaborative Filtering 推荐算法存在的问题 1.Only the binary events of user actions, but totally disregard the context 2.Data sparseness lead to overfitting 3.The interaction process between users and the recommender systems not fully-exploited
4
Innovation Team of Recommender System(ITRS) 3 Context User actions (u, i*) are action dyads, while the dyads {(u, i)} for all i O, and are contextual dyads Set of users, U Set of items, I Each user interaction has an offer set O and a decision set D Each user interaction is stored as a tuple (u, O, D) where D is a subset of O Notation: Traditional method: trade these contextual as,missing data Now: This context is informative which should be exploited
5
Innovation Team of Recommender System(ITRS) 4 Context example User U1 prefer Die Hard ( 虎胆龙威, 主演 Bruce Willis) given a generic set of movies only tells us that the user appreciates action movies. While, a preference for Die Hard over Terminator( 终结者, 主演阿诺德 · 施瓦辛格 ) or Rocky( 洛奇 ,主演史泰龙 ) suggests that the user might favor Bruce willis over other action heros. In other words, the context of user choice is vital when estimating user preferences.
6
Innovation Team of Recommender System(ITRS) 5 Maechanism underlying user choice behaviors Collaborative filtering: “collaboration effects” that similar items get similar responses from similar users The effect of Collaboration Another mechanism governs users’ behavior? The effect of competition Items turn to compete with each other for the attention of users, user u will pick the best item i* (the one with highest utility) when confronted by the set of alternatives O A man with penchant for action movies by Arnold Schwarzenegger. Given the choice between Sleepless in Seattle and Die Hard,he will likely choose the latter. And afforded the choice between the oeuvres of Schwarzenegger, Diesel or willis, he more likely to choose Schwarzenegger. So local competition effect essential for recommender model
7
Innovation Team of Recommender System(ITRS) 6 Problem formulation Set of users, U Set of items, I Each user interaction has an offer set O and a decision set D, D={i*}, D is not empty and contains exactly one item i*, i.e. i* as a decision. Each user interaction is stored as a tuple (u, O, D) where D is a subset of O Collaborative filtering Given observations of dyadic responses {(u, i, y ui )} with each y ui being an observed response( e.g. user’s rating to an item, or indication of whether user u took an action on item i) Collaborative filtering explores the notion of “collaboration effect”, i.e. similar users have similar preferences to similar items.
8
Innovation Team of Recommender System(ITRS) 7 Collaborative filtering 1. Neighborhood models Based on propagating the observations of responses among items or users that are considered as neighbors. The model first defines a similarity measure between items / users. Then, unseen response between user u and item I is approximated based on the response of neighboring users or items. 2.Latent factor models Learn predictive latent factors to estimate the missing dyadic responses. The basic idea is to associate latent factors, for each user u and for each item i.
9
Innovation Team of Recommender System(ITRS) 8 Latent factor models
10
Innovation Team of Recommender System(ITRS) 9 Overfitting CF usually performs poorly on binary response. For the aforementioned interaction process, the response y ui is typically a binary event indicating whether or not item i was accepted by the user u. with the non-action dyads being ignored, the responses are exclusively positive observations. We will obtain an overly-optimistic estimator that biases toward positive responses and predicts positive for almost all incoming dyads.
11
Innovation Team of Recommender System(ITRS) 10 Collaborative competitive filtering CCF for learning recommender models by modeling users’ choice behavior in their interactions with the recommender system. Employ a multiplicative latent factor model to characterize the dyadic utility function. CCF encodes the collaboration effect and completion effect. In practice, a user u could make different decisions when facing different contexts O t Local Optimality of User Choice Each item has a potential revenue to the user which is r ui Users also consider the opportunity cost (OC) when deciding potential revenue OC is what the user gives up for making a given decision OC is c ui = max( i’ | i’ in O \ i) Profit is π ui = r ui – c ui A user interaction is an opportunity give and take process User is given a set of opportunities User makes a decision to select one of the many opportunities Each opportunity comes with some revenue (utility or relevance)
12
Innovation Team of Recommender System(ITRS) 11 Collaborative competitive filtering Local optimality constraint Each item in the decision set has a revenue higher than those not in the decision set Problem becomes intractable with only this constraint, no unique solution
13
Innovation Team of Recommender System(ITRS) 12 Surrogate objectives Softmax model Assume the utility function consists of two components is a deterministic function characterizing the intrinsic interest of user u to item i. for which use the latent factor model to quantify is a stochastic error term reflecting the uncertainty and complexness of the choice process. Assume the error term is an independently and identically distributed Weibull variable.
14
Innovation Team of Recommender System(ITRS) 13 Softmax model
15
Innovation Team of Recommender System(ITRS) 14 Hinge model
16
Innovation Team of Recommender System(ITRS) 15 Learning algorithm
17
Innovation Team of Recommender System(ITRS) 16 Extensions
18
Innovation Team of Recommender System(ITRS) 17 DataSets Dyadic response data 1.Social network data. 2.Netflix 5 star data
19
Innovation Team of Recommender System(ITRS) 18 Evaluation
20
Innovation Team of Recommender System(ITRS) 19 最近的想法 1.CCF in BPR 2.APRIORI to more accurate Neighbors or niche Items 3.Matrix Factorazation with Random Forests
21
Innovation Team of Recommender System(ITRS) Alibaba Research Center For Complexity Sciences
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.