Download presentation
Presentation is loading. Please wait.
Published byΜέδουσα Μαυρίδης Modified over 6 years ago
1
The Recommendation Click Graph: Properties and Applications
Yufei Xue, Yiqun Liu, Min Zhang, Shaoping Ma, Liyun Ru State Key Laboratory of Intelligent Technology and Systems, Tsinghua National Laboratory for Information Science and Technology, Department of Computer Science and Technology, Tsinghua University.
2
Introduction Query Recommendation The Clicks on Query Recommendation
Widely used in commercial search engines Frequently used by users The Clicks on Query Recommendation Show users’ search intents Reflect users are not satisfied with search results
3
Introduction Recommendation Click Graph: Concept Properties
Applications Optimizing search results Recognizing ambiguous queries
4
Related Work Link Analysis on Web Graphs: HITS PageRank TrustRank …
5
Related Work Types of Web Graphs Examples of Web Graphs Word graph
Session graph URL cover graph URL link graph URL terms graph Examples of Web Graphs HyperLink Graph Query Flow Graph
6
Definition The recommendation click graph is a directed graph Gc =(V, E) , where the set of nodes, V = {q|q appear in the recommendation click log as a source or destination query} E={(qi,qj)| < qi,qj >is a recommendation click pair in the recommendation click log}
7
Assumptions For a recommendation click pair < qi,qj >
Assumption 1: qj describes the user's information needs more precisely than qi, or Assumption 2: qj does not describe the user's information needs more precisely than qi, but the user is interested in qjand wants more information.
8
Properites A Recommendation Click Graph Search log of 31 days
58,334,303 clicks 23,516,620 vertices 31,569,262 directed edges
9
Properites Connected Components 2,668,331 components
71% have only 2 vertices The largest component has 16,298,916 vertices
10
Local Analysis Local Subgraph For query qi in a recommendation click graph, we define a local subgraph of qi as where
11
Local Analysis HITS on Local Subgraph
Apply HITS on a local subgraph of the recommendation click graph. Find the queries with high authority value and low hub value. These queries may satisfy more users and be less ambiguous. Use the search results of these queries to optimizing the search result of original query.
12
Optimizing Search Results
13
Optimizing Search Results
Experiment From the queries with out-degree>8, we randomly selected two groups of queries for experiment. Group 1: The queries with out-degree = 8 or 9 Group 2: Randomly Selected.
14
Finding Ambiguous Queries
Inverse PageRank High inverse PageRank Indicates having many outlinks or pointing to vertices with many outlinks. High inverse PageRank on Recommendation Click Graph indicates strong ambiguity. Experiment on 1010 queries Sort the queries by Inverse PageRank in decending order. Divide the queries in 10 buckets.
15
Finding Ambiguous Queries
16
Finding Ambiguous Queries
17
Finding Ambiguous Queries
18
Conclusions Concept: Recommendation Click Graph
Properties: Similary to Hyperlink Graph Applications HITS: Optimizing Search Results Inverse PageRank: Finding Ambiguous Queries
19
Thanks! Questions?
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.