Abstract
Social media platforms have today become an important part of people’s lives. People share their daily activities and interests through these platforms and also read about others’ activities. The increasing use of social media has led to rapid growth in the amount of shared information, and consequently has caused the challenge of information overload. This causes difficulties for users in filtering this huge amount of shared data in order to find and extract the information they need. This consumes the user’s time, increases information clutter and reduces their satisfaction. To manage this challenge, there is a need for personalisation services such as recommender systems. One popular social networking site is Twitter; it provides a rich environment for shared information that can help with recommender system research. However, short-text-based recommender systems based on Twitter activity suffer from a lack of reliable data. One solution to this problem is to build more powerful user profiles that reflect the recent interests of the user and consequently provide more accurate recommendations. A prompt way of building such a profile is to gather the user’s recent activities from their Twitter timeline. However, many users do not provide enough up to date data to build such a profile. Several researchers have tried to overcome this problem by enriching the profile via external textual sources such as Wikipedia. However, external sources may not be able to provide valid data that reflect the user’s interests and consequently the recommendation quality can be affected.This research studies short-text social media user modelling by utilising explicit and implicit relationships. This approach aims to build personal profiles through a different method, using tweets from the user’s Twitter network to provide more accurate recommendations. Our method exploits a Twitter user’s explicit and implicit relationships in order to extract other users’ important tweets that will help in building the personal profile. These relationships were identified by our proposed influence algorithm, which is a new way of measuring the influence rule from the user’s perspective rather than the influencers’ perspective. The usefulness of this proposed method is validated by implementing a tweet recommendation service based on the content-based recommender system mechanism, and by performing offline evaluation on a real dataset of 40 users collected from Twitter. The proposed user profiles are compared against other profiles, used as a baseline, based on similarity and distance metrics to check the effectiveness of the proposed method. Our proposed method shows increased performance gains in reflecting the user’s interests by recommending more accurate items.
The main contribution of this work is an influence algorithm and a framework that is able to identify influential friends of a user (people that the user follows) more accurately than other influence and similarity algorithms, and then use them as sources in enriching the user’s profile via their tweets. It also helped us to enrich the profile with tweets that were collected from other sources, which were found via explicit and implicit relationships. As a result, more accurate recommendations can be delivered.
Date of Award | May 2019 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Stylianos Kapetanakis (Supervisor), Roger Evans (Supervisor) & Nikolaos Polatidis (Supervisor) |