Preference Modeling via Sentiment Analysis and Recommender System

Abstract

본 논문에서는 사용자의 선호도를 모델링하는 두 가지 연구 문제 (감정 분석 및 추천)에 대한 최고 성능의 접근 방법을 제안한다. 구체적으로, 감정분석에 대해서는 주어진 목표 개체와 다른 단어 간의 거리를 고려하면서 학습 시간을 감소시킨 컨볼루션 신경망 기반 모델을 설계한다. 또한, 감정 분석 분야에서 적은 양의 데이터와 언어의 다양성 부족 문제를 해결하기 위해 가장 큰 감정 분석 데이터를 공개하여 연구 커뮤니티에 기여한다. 추천 문제에 대해서는 소비 이력이 적은 사용자의 선호도를 더 잘 파악하기 위해 사용자 리뷰 글 정보를 활용하는 추천 시스템을 추천 정확도와 확장성 관점에서 고안한다. 이를 위해 주의 신경망과 컨볼루션 신경망을 기반으로 추천 시스템을 설계하여 두 가지 문제를 해결한다. 또한 사용자가 개체에 가지는 전역적 및 지역적 관심 유지도를 모델링하여 추천 정확도를 높이는 연구를 제안한다. 감정 분석과 추천 문제에 대해 제안한 접근 방법들은 실제 데이터를 기반으로 한 광범위한 실험을 통해 기존 방법론들보다 뛰어난 성능을 보여준다.

The rapid expansion of online services (e.g., social medias and e-commerce services) has substantially increased the amount of user-generated data, which contain users’ preference on entities such as politicians or products. According to the recent report from Facebook, 2.47 billion people use the service and 4 petabytes of data are generated per day. The users’ preference, which is obtained from the abundant data, has notable applications such as public opinion polls by analyzing user-generated texts in social medias. Consequently, modeling users’ preference has been an essential line of research these days. The preference modeling can be categorized into two research problems; preference extraction and future preference prediction. The preference extraction from the user-generated data has been actively researched since a large portion of users does not explicitly provide their preference but only generate unstructured data (e.g., text, audio and video), which contain their preference in an implicit way. The goal of this problem I is to identify ‘how much users prefer entities’ from the data. Sentiment analysis is a representative problem of the preference extraction, and aims to classify users’ preference on entities in text into predefined sentiments (e.g., positive or negative). The prediction of users’ future preference is another major line of research as we can suggest attractive services (e.g., videos in YouTube) to users based on their future preference. This problem aims to predict ‘which entity users will prefer’ based on the observed users’ preference. Recommendation is a fundamental problem of the future preference prediction, and its goal is to predict a user’s future preference based on the preference of other users who are like minded to the target user. The aim of this dissertation is to advance the state-of-the-art approaches in the main research problems (i.e., sentiment analysis and recommendation) of modeling users’ preference. More concretely, for sentiment analysis, we design a neural architecture to utilize the distance between the target entity and other words in text, and also publish the largest English and Korean datasets to address the limited size and language diversity of existing datasets. For recommendation, we invent approaches to supplement users’ preference with service reviews written by users because the observed users’ preference is generally sparse, which degrades the recommendation accuracy. In addition, we develop a recommender system that captures ‘how much the interest of users in entities will sustain in the future’ to better model users’ concept drift over time. The proposed methods show outstanding performances of modeling users’ preference through extensive experiments. In summary, this dissertation proposes the state-of-the-art approaches in both research problems of modeling users’ preference.