CHOSUN

CNN을 이용한 SNS 사용자 관심사 카테고리 분류 및 팔로잉 추천 방법

Metadata Downloads
Author(s)
홍택은
Issued Date
2016
Abstract
With the development of various smart devices, an increasing number of people are using SNS, which enables communication and sharing of information regardless of distance and place. Users who used to focus on the functions of forming relationship and communication are now using SNS for the function of sharing information. The existing way of recommending following was mostly based on relationship so there is the limitation of recommending following for users who use SNS for information. Also, as studies using SNS postings mostly suggest ways of categorizing interests and recommending following based on images and texts, it is hard to understand users' intention exactly.
So this paper suggests the way of recommending information-centered following that finds out users' interests by using images and texts and categorizing interests of users. This study secured objectivity by defining DOMZ, which is established as ODP(Open Directory Project), as the standard for category of interest, and by using CNN(Convolutional Neuronal Network), one of the machine learning methods, it studied images and text and divided interests into categories. It defined users who post many postings as information providers, established information database and categorized users with the same interests. Then it measured the similarity of the distance between information providers and users and recommended the information provider with the most similar interests for following.
As a result of categorizing interests of information providers and of users using the method suggested in this paper, the rate of accuracy was 80% and 79.8% respectively, and the overall rate of accuracy was 79.93%, which means a good categorization overall. Categories of interest with low rate of accuracy are thought to be because the meanings were similar and similar data was collected when images and texts were searched. There seems to be a slight correlation between categories. Therefore, it seems that there need to be studies on the selection of categories and standards for making SNS categories, and also studies on establishing proper CNN models to learn and categorize texts that are made up of special characters and emoticons due to the nature of texts of SNS postings.
What can be suggested is to divide users' interests into 16 categories then recommend following in order to recommend more proper following for information-focused SNS users, and by using images and texts together, it was possible to accurately understand the intention of users. Through the way suggested by this paper, it is likely that it will be possible to recommend following that is suitable for SNS users, and categorization of interests will be used in customized services, which recommend contents and information suitable for users' interests, and SNS marketing services.
Alternative Title
Interest Category Classification of SNS User and Following Recommendation Method using CNN
Alternative Author(s)
Taekeun Hong
Affiliation
조선대학교 산업기술융합대학원
Department
산업기술융합대학원 소프트웨어융합공학과
Advisor
신주현
Awarded Date
2017. 2
Table Of Contents
ABSTRACT

Ⅰ. 서론 1
A. 연구 배경 및 목적 1
B. 연구 내용 및 구성 2

Ⅱ. 관련 연구 3
A. 관심사 분류 관련 기존 연구 3
B. Convolutional Neural Network 6
1. 이미지 분류 9
2. 텍스트 분류 11
C. TensorFlow 13
D. DMOZ 15

Ⅲ. 관심사 카테고리 분류를 이용한 팔로잉 추천 방법 17
A. 시스템 구성도 17
B. 정보제공자 데이터베이스 구축 18
1. 관심사 카테고리 정의 19
2. CNN을 이용한 관심사 카테고리 분류 20
a. 이미지에서 관심사 카테고리 분류 20
b. 텍스트에서 관심사 카테고리 분류 25
3. 관심사 카테고리를 이용한 데이터베이스 구축 29
C. 사용자 팔로잉 추천 32
1. 사용자 관심사 카테고리 분류 32
2. 거리 측정을 통한 팔로잉 추천 35
Ⅳ. 실험 및 결과 39
A. 실험 환경 39
B. 데이터 셋 40
C. 실험 평가 및 분석 41
1. 실험 평가 방법 41
2. 실험 결과 분석 42

Ⅴ. 결론 및 제언 47

참고문헌 48
Degree
Master
Publisher
조선대학교 산업기술융합대학원
Citation
홍택은. (2016). CNN을 이용한 SNS 사용자 관심사 카테고리 분류 및 팔로잉 추천 방법
Type
Dissertation
URI
https://oak.chosun.ac.kr/handle/2020.oak/1951
http://chosun.dcollection.net/common/orgView/200000265867
Appears in Collections:
Engineering > Theses(Master)(산업기술창업대학원)
Authorize & License
  • AuthorizeOpen
Files in This Item:

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.