Temporal Topic Inference for Trend Prediction

Publicly available social data has been adoptedwidely to explore language of crowds and leverage themin real world problem predictions. In microblogs, usersextensively share information about their moods, topics ofinterests, and social events which provide ideal data resourcefor many applications. We also study footprints of socialproblems in Twitter data. Hidden topics identified fromTwitter content are utilized to predict crime trend. Since ourproblem has a sequential order, extracting meaningful patternsinvolves temporal analysis. Prediction model requiresto address information evolution, in which data are morerelated when they are close in time rather than further apart. The study has been presented into two steps: firstly, a temporaltopic detection model is introduced to infer predictivehidden topics. The model builds a dynamic vocabulary todetect emerged topics. Topics are compared over time to havediversity and novelty in each time consideration. Secondly, apredictive model is proposed which utilizes identified temporaltopics to predict crime trend in prospective timeframe. The model does not suffer from lack of available learningexamples. Learning examples are annotated with knowledgeinferred from the trend. The experiments have revealed, temporal topic detection outperforms static topic modelingwhen dealing with sequential data. Topics are more diversewhen are inferred in different time slices. In general, theresults indicate temporal topics have a strong correlationwith crime index changes. Predictability is high in somespecific crime types and could be variant depending on theincidents. The study provides insight into the correlation oflanguage and real world problems and impacts of social datain providing predictive indicators.