Building TopicTrend Detection System based on Slow Intelligence

Building Topic/Trend Detection System based on Slow Intelligence Chia-Chun Shih & Ting-Chun Peng Institute for Information Industry Taipei, Taiwan Presented at DMS’ 10 special session on Slow Intelligence Systems

Agenda • • Introduction Topic/Trend Detection System with Slow Intelligence Conclusion 2

Introduction

Introduction Blog Posts Twitter Posts Facebook Users • Social media is prevailing • Social media is a reflection of real-world – An experiment from HP Social Computing Lab shows: • Twitter-rate time series can accurately predict box-office movie sales with Adjusted R 2 = 0. 973 (amazing!!) • The emerging market for Social Media Monitoring Service – E. g. , Nielsen Buzzmetrics, Radian 6 4

Introduction (cont’d) • Topic Detection and Tracking (TDT) – Initiated by DARPA at 1996 – discover the topical structure in unsegmented streams of news reporting as it appears across multiple media – Tasks: • • • Topic Detection Topic Tracking First Story Detection Story Segmentation Link Detection 5

Introduction (cont’d) • Slow Intelligence provides a software development framework for systems with insufficient computing resources to gradually adapt to environments to handle complexities Environment Knowledge-based Controller Problem 1 Enumerator 2 Adaptor 3 Eliminator 4 Concentrator Solution Slow Intelligence System 6

Introduction (cont’d) • In this paper, we propose a design of online topic/trend detection system for Social Media with the advantages of Slow Intelligence. • Four complexities of designing online topic/trend detection systems are identified, along with corresponding Slow Intelligence solutions. 7

Topic/Trend Detection System

Topic/Trend Detection System • Objective – Detect current hot topics and to predict future hot topics based on data collected from Social Media • Three components – Crawler & Extractor: Collect data and extract information from Social Media – Topic Extractor: Detect hot topics from a set of text documents – Trend Detector: Detect trends (future hot topics) based on currently available data Social Media Crawler & Extractor Topic Extractor Trend Detector Current Hot topics Future Hot topics 9

Topic/Trend Detection System (cont’d) • Crawler & Extractor Social Media User’s Keywords of Interests Web Crawler HTML documents Information Extractor Text documents Web data DB Topic Extractor * Extract articles and metadata (title, author, content, etc) from semistructured web content Crawler & Extractor 10

Topic/Trend Detection System (cont’d) • Topic Extractor Web data DB Topic Word Extraction • Apply TF-IDF scheme to generate Top-N topic words for each document Topic Word Clustering • Apply clustering algorithm to cluster topic words into topic groups. The topic groups are treated as “topics” Current topics Hot topic extraction Current Hot topics • Apply aging theory to find hot topics Topic Extractor 11

Topic/Trend Detection System (cont’d) • Trend Detector Current topics Trend Estimation Algorithms Topic Trend (Future Hot Topics) Trend Detector • The Trend Estimation Algorithm is a black box now, however, it will “find its way” when Slow Intelligence is involved in the system 12

Topic/Trend Detection System with Slow Intelligence

T/TD System with Slow Intelligence • Four complexities of designing online topic/trend detection systems • 1. It is unlikely to collect all web data based on limited amount of computing resources. The system needs to develop data collection strategies which can concentrate limited resources on collecting important web data. Crawler & Extractor 14

T/TD System with Slow Intelligence (cont’d) • 2. Many computation methods are available for estimating trends. If parameter settings are also taken into account, there are too many combinations to choose. Furthermore, Internet is a changing environment, which means current best solution may not perform well in the future. The system needs to automatically (or at least quasi-automatically) find best solution from many alternatives in a changing environment. Trend Detector 15

T/TD System with Slow Intelligence (cont’d) • 3. The crawler needs to revisit websites to collect upto-date data in hourly or daily intervals. Each site has different amount of to-be-update data and different policy to restrict frequent access, which are unknown beforehand. The system needs to find feasible data collection schedule based on past experience. Crawler & Extractor 16

T/TD System with Slow Intelligence (cont’d) • 4. Any changes in web pages may disrupt Extractors. It needs automatic repair mechanism for Extractors if many websites are being monitored. The repair mechanism needs to detect errors of Extractors, find alternatives, and choose the best solution from alternatives to fix the disrupted Extractors. Crawler & Extractor 17

T/TD System with Slow Intelligence (cont’d) 1. SIS to help restrict the range of data collection Knowledge of data Knowledge of algorithm 18

T/TD System with Slow Intelligence (cont’d) 2. SIS to help select and adapt trend detection algorithms 19

T/TD System with Slow Intelligence (cont’d) 3. SIS to help scheduling Crawler 20

T/TD System with Slow Intelligence (cont’d) 4. SIS to help adapt Extractors 21

Conclusion

Conclusion • An online trend detection system requires careful resource allocation and automatic algorithm adaptation to process huge size of heterogeneous data. • This research adopts Slow Intelligence, which provides a framework for systems with insufficient computing resources to gradually adapt to environments, to response the challenges. • Four Slow Intelligence subsystems are proposed, and each subsystem targets a challenge in designing online topic/trend detection systems. 23

If you have any questions, please e-mail us chiachun@iii. org. tw (Chia-Chun Shih) markpeng@iii. org. tw (Ting-Chun Peng)