FRIEND RECOMMENDATION SYSTEM BY USING MACHINE LEARNING

Abstract

In this thesis, we have presented a friend recommendation model that is based on user lifestyle, which is generated using daily activities performed by the user. Existing technology for friend recommendation is based on social networking graph, in which common friend of users decide the friend recommendation score. We believe the recommendation by social graphs is not the befitting real life friend selection for user. Thus, we have presented an approach of for friend recommendation that is based on matching life style and not on social graph. To implement this semantic-based friend recommendation system, we have taken advantage of sensor-rich smartphones to recognize human activities. After this we have used these human activities to create user life document, which represents user’s life style. Likeness between lifestyles of users is used to measure the similarity between users for recommending friends. First, we collected sensor data of user using mobile application and then we performed activity recognition. After finding activities of user we have created a life document of user, from which the algorithm Latent Dirichlet Allocation (LDA) [1] is used to select the life style. We further propose a similar metric algorithm in order to compute the likeness of life styles among users, and measure the impact the life style of users by a friend-matching graph. When user requests for friends, then our algorithm will return people’s list sorted by recommendation score, from which user can choose friends to send request. We have implemented this system for Android based smartphone and its performance has been evaluated on with largescale experimental data. Finally, the results manifest that the recommendations properly reflect the preferences of users in choosing friends. This approach exploits gradient boosting algorithm, Auto-regression Model, Signal Magnitude Area (SMA), tilt angle, standard deviation mean & median.