Personalizing Forum Search using Multidimensional Random Walks
0 likes910 views
This document describes a method for improving forum search through personalization. It proposes using multidimensional random walks to compute user similarities based on multiple dimensions like co-participation, interactions, topics and profiles. The approach builds a multidimensional heterogeneous graph and executes random walks with weights based on egocentric relations. Key results show this method predicts similar users to answer future questions better than 6 baselines, and can enhance keyword search by re-ranking results based on contributor authority scores from multiple relations.
Personalizing Forum Search using Multidimensional Random Walks
- 1. 1 Personalizing Forum Search using Multidimensional Random Walks Gayatree Ganu Amélie Marian Project: Patient Emotion and stRucture Search USer tool (PERSEUS) NSF Grants IIS-084493, BCS-1027801, Google Research Award GOAL: Improve forum search by leveraging user similarities • Co-participation • Past interactions • Topical similarities • Profile similarities
- 2. 2 Our Approach • Compute user similarities using multidimensional random walks (MRW) • Algorithm Outline: – Build multidimensional heterogeneous graph – Execute rooted MRW with egocentric relation weights – Scores represent similarity incorporating multiple relations Relation A Relation B Relation C r1 r22 b 1 4 c 4 21 4 Amélie Marian – Rutgers University – ICWSM’14
- 3. 3 Key Results • Personalized forum search, predicting similar users will answer questions in future threads – MRW has highest prediction MAP compared to 6 baselines – MRW improves over best baseline by 10% for top-10 similar users, 21% for top-100 • Enhanced keyword search – Re-ranked tf*idf results using the authority of contributors computed over multiple user relations – 5% relevance improvement in MAP@10, judged by human raters Amélie Marian – Rutgers University – ICWSM’14