Machine Learning Application: Personalized Venues Recommendation Model

Abstract

With the development of smart devices and the convenient of social media platform, people are now more open in sharing their photos while traveling. Geo-tagged data from these photos and user’s additional multisource data such as weather combined, could help us in modelling user’s preferences profile in details. Therefore, exploiting these geo-tagged data, a personalized recommendation model is built to provide recommendations based on personal interested analysis. Retrieving geo data from Gowalla dataset and context data from weather API, we posed the recommendation model as a two-stages architect problem where we first generate a candidate lists (matching candidates) and ranking the list follow user’s modelled preferences afterwards. The dataset was formally clustered into venue centres, then transformed into a multi-class classification problem in the matching process using supported vector machine algorithm, and gradient boosting regressor is deployed to rank the generated list follow preference scoring. Evaluating the model by benchmarking with the one-stage architect by accuracy metrics and ranking metric, the model shows its improvement in handling mitigating cold-start problem and mining long-tail data.