Date of Award

December 2016

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Management Science

First Advisor

Atish Sinha

Second Advisor

Huimin Zhao

Committee Members

Stephen France, Xiaojing Yang

Keywords

Friend Recommendation, Health and Fitness, Location Based Service, Recommendation System, Social Networking Site, Text Mining

Abstract

Social networking sites (SNSs) first appeared in the mid-90s. In recent years, however, Web 2.0 technologies have made modern SNSs increasingly popular and easier to use, and social networking has expanded explosively across the web. This brought a massive number of new users. Two of the most popular SNSs, Facebook and Twitter, have reached one billion users and exceeded half billion users, respectively.

Too many new users may cause the cold start problem. Users sign up on a SNS and discover they do not have any friends. Normally, SNSs solve this problem by recommending potential friends. The current major methods for friend recommendations are profile matching and “friends-of-friends.” The profile matching method compares two users’ profiles. This is relatively inflexible because it ignores the changing nature of users. It also requires complete profiles. The friends-of-friends method can only find people who are likely to be previously known to each other and neglects many users who share the same interests. To the best of my knowledge, existing research has not proposed guidelines for building a better recommendation system based on context information (location information) and user-generated content (UGC).

This dissertation consists of three essays. The first essay focuses on location information and then develops a framework for using location to recommend friends--a framework that is not limited to making only known people recommendations but that also adds stranger recommendations. The second essay employs UGC by developing a text analytic framework that discovers users’ interests and personalities and uses this information to recommend friends. The third essay discusses friend recommendations in a certain type of online community – health and fitness social networking sites, physical activities and health status become more important factors in this case.

Essay 1: Location-sensitive Friend Recommendations in Online Social Networks

GPS-embedded smart devices and wearable devices such as smart phones, tablets, smart watches, etc., have significantly increased in recent years. Because of them, users can record their location at anytime and anyplace. SNSs such as Foursquare, Facebook, and Twitter all have developed their own location-based services to collect users’ location check-in data and provide location-sensitive services such as location-based promotions. None of these sites, however, have used location information to make friend recommendations.

In this essay, we investigate a new model to make friend recommendations. This model includes location check-in data as predictors and calculates users’ check-in histories--users’ life patterns--to make friend recommendations. The results of our experiment show that this novel model provides better performance in making friend recommendations.

Essay 2: Novel Friend Recommendations Based on User-generated Contents

More and more users have joined and contributed to SNSs. Users share stories of their daily life (such as having delicious food, enjoying shopping, traveling, hanging out, etc.) and leave comments. This huge amount of UGC could provide rich data for building an accurate, adaptable, effective, and extensible user model that reflects users’ interests, their sentiments about different type of locations, and their personalities. From the computer-supported social matching process, these attributes could influence friend matches. Unfortunately, none of the previous studies in this area have focused on using these extracted meta-text features for friend recommendation systems.

In this study, we develop a text analytic framework and apply it to UGCs on SNSs. By extracting interests and personality features from UGCs, we can make text-based friend recommendations. The results of our experiment show that text features could further improve recommendation performance.

Essay 3: Friend Recommendations in Health/Fitness Social Networking Sites

Thanks to the growing number of wearable devices, online health/fitness communities are becoming more and more popular. This type of social networking sites offers individuals the opportunity to monitor their diet process and motivating them to change their lifestyles. Users can improve their physical activity level and health status by receiving information, advice and supports from their friends in the social networks. Many studies have confirmed that social network structure and the degree of homophily in a network will affect how health behavior and innovations are spread. However, very few studies have focused on the opposite, the impact from users’ daily activities for building friendships in a health/fitness social networking site.

In this study, we track and collect users’ daily activities from Record, a famous online fitness social networking sites. By building an analytic framework, we test and evaluate how people’s daily activities could help friend recommendations. The results of our experiment have shown that by using the helps from these information, friend recommendation systems become more accurate and more precise.

Share

COinS