網路社群的風行有目共睹，早已成為現代人生活不可或缺的一部份。然而，在網路的屏蔽下，成員可能在社群中展現各種不可預測的行為，使社群管理面臨巨大挑戰。但若能提早發現社群成員的狀態變化，管理者便可預先處理，以維持社群的正面發展。上述需求雖然重要，但在前人研究中卻鮮少提及，顯然需要進一步研究。為了解社群成員未來之發展動向，本研究提出一套線上社群成員之狀態變遷趨勢預測方法。首先，我們運用多種社群指標來描述社群成員的屬性，並透過分群方法歸納出多種典型成員特質，再使以代表社群成員之狀態。為預測成員下一時間點之狀態，我們分別設計了以狀態序列為基礎以及以指標值序列為基礎之預測方法。其次，本研究也提出多種二階式預測架構，以準確性較高的二分類模型進行前置過濾，以降低預測時的誤判狀況。為進一步提升實用性，我們也運用Transformer架構建立Sequence-to-Sequence模型，預測會員未來連續時間點之狀態序列。為驗證提出方法之有效性，本研究自PTT BBS討論區下載實際發文資料進行實驗。實驗結果顯示，針對會員下一時間點之狀態類型預測，準確率可達65.6%。而對於更複雜的狀態序列預測，更可獲得77.3%之平均相似度。上述結果顯示，本研究提出之方法能提供實用的社群成員狀態變遷趨勢預測，可做為社群管理之重要參考依據。

The popularity of online communities is obvious to all, and it has become an indispensable part of modern life. However, under the protection of the Internet, members may exhibit various unpredictable behaviors, posing huge challenges to managers. However, if changes can be detected early, they can be dealt with in advance to maintain the positive development of the community. While these needs are important, they have been seldom mentioned in previous studies and clearly require further research. In order to understand the future development trend of community members, this study proposes a set of methods for predicting status change trends of online community members. First, we use a variety of indicators to describe member attributes, and then use the grouping method to summarize a variety of typical characteristics and use them to represent the status of community members. In order to predict the status of members at the next time point, we designed prediction methods based on state sequences and feature value sequences. Secondly, this study proposes a set of second-order prediction models to reduce misjudgments during prediction through a more accurate two-class classification model. To further improve practicality, we also use the Transformer architecture to build a Sequence-to-Sequence model to predict the future status sequence of members. In order to verify the effectiveness of the proposed method, this study downloaded actual published documents from PTT BBS for experiments. The results show that the accuracy of predicting the status of members at the next point in time can reach 65.6%. For more complex future status sequence predictions, an average sequence similarity of 77.3% can be obtained. The results show that the method proposed in this study can provide an acceptable prediction of the status change trends of online community members, which can be used as an important reference for community management.

第一章 緒論	1
第二章 文獻探討與相關技術	4
2.1 社會網路分析 (Social Network Analysis)	4
2.2 使用深度學習技術建立預測模型	5
2.2.1 長短期記憶(Long Short-Term Memory, LSTM)	5
2.2.2 變形器(Transformer)	6
2.3 萊文斯坦距離 (Levenshtein distance)	7
第三章 社群成員狀態變遷趨勢預測方法	9
3.1 資料集的建立與狀態標籤的產生	9
3.2 成員單一未來狀態之預測 (Members’ Future State Prediction)	13
3.2.1 以狀態序列為基礎之預測 (State sequence based Prediction)	13
3.2.2 以指標值序列為基礎之預測 (Feature-value sequence based prediction)	16
3.3 二階式預測架構 (2-Level Prediction Framework)	17
3.3.1 以模型融合為基礎之預測架構 (Prediction Framework Based on Model Fusion)	17
3.3.2 Category Split二階式預測架構	18
3.3.3 Passive/Active Split二階式預測架構	19
3.3.4 State Change Checking二階式預測架構	20
3.4 未來狀態序列之預測 (Next State Sequence Prediction)	21
第四章 實驗結果與討論	23
4.1 實驗設定	23
4.2 單一未來狀態預測之效能驗證	23
4.2.1 以狀態序列為基礎之預測方法實驗結果	23
4.2.2 以指標值序列為基礎之預測方法實驗結果	25
4.2.3 二階式狀態預測架構之效能	26
4.3 社群成員未來狀態序列預測之效能驗證	30
第五章 結論與未來工作	34
參考文獻	35
附錄	37

圖目錄
圖2-1: LSTM記憶單元	6
圖2-2: Transformer架構(Vaswani et al., 2017)	7
圖2-3: 注意力機制運算過程	7
圖3-1: 使用Elbow法與Silhouette法決定群數之過程	11
圖3-2: 以狀態序列為基礎之預測模型架構(以order 3為例)	16
圖3-3: 指標值序列為基礎之預測模型架構(以order 3為例)	17
圖3-4: 複合式會員未來狀態預測架構	18
圖3-5: 以MLP建立複合式狀態預測模型	18
圖3-6: Category Split二階式預測架構	19
圖3-7: Passive/Active Split二階式預測架構	20
圖3-8: State Change Checking二階式預測架構	21
圖3-9: Seq2Seq Model之模型架構(以order 10lookahead 7為例)	22
圖4-1: 各類資料之數量之比較	24
圖4-2: 資料統計期間社群成員各種狀態變遷樣式之比例統計	33

表目錄
表3-1: 用以描述社群成員特質之七項指標	10
表3-2: 透過七項指標將社群成員活動轉換為指標值向量序列集合	10
表3-3: 依照七項社群指標產生之資料集範例	11
表3-4: 蒐集PTT政黑板13個月資料，將會員特質進行分群之結果	12
表3-5: 社群成員狀態序列生成流程之虛擬碼	13
表3-6: 將使用者在各週之指標值序列資料轉換為狀態變遷樣式集	14
表3-7: 各狀態分群之樣本數	14
表3-8: 社群成員特質之主狀態再細分為子狀態	15
表3-9: 以指標值向量序列為prefix之樣式	16
表3-10: 建立複合模型所使用之資料集	18
表3-11: 產生未來狀態序列預測所需之資料集	21
表4-1: PTT 政黑板在13個月期間各項社群指標之前10名指標值(週平均)	23
表4-2: 狀態序列資料集之各類別資料數及比例(order=3)	24
表4-3: 單一未來狀態預測之各類別資料配比(Train, Validation, Test)	25
表4-4: 社群成員之單一未來狀態預測結果與比較	25
表4-5: 指標值序列預測方法結果之混淆矩陣、召回率及精確率(order=7)	26
表4-6: 利用不同學習方法建立Compound Model之準確率	26
表4-7: Category Split架構之各層模型資料配比(Train, Validation, Test)	27
表4-8: Category Split架構之各層模型準確率(以指標值序列為輸入)	28
表4-9: Passive/Active之各層模型資料配比(Train, Validation, Test)	28
表4-10: Passive/Active Split架構之各層模型準確率(以指標值序列為輸入)	29
表4-11: 二分類之資料配比(Train, Validation, Test)	29
表4-12: 第一階段二元分類預測結果	29
表4-13: 單一未來狀態之各種預測方法準確率比較	30
表4-14: 各種Seq2Seq塑模方法配合不同lookahead之平均序列相似度	31
表4-15: 各種不同order/ lookahead組合之預測平均相似度	31
表4-16: 以Transformer架構進行O10/L7 之Seq2Seq預測	32
附表1: 分群之平方距離和變化	37
附表2: 子狀態之群心	38
附表3: 單一未來狀態預測模型之混淆矩陣	39
附表4: 二階預測架構各層模型之交叉驗證準確率變化圖	42

1.Asim, Y. et al., (2019), "A trust model for analysis of trust, influence and their relationship in social network," Telematics and Informatics 36 (2019) 94-116.
2.BBC News (2016), "How can Facebook fix its fake news problem?", http://www.bbc.com/news/technology-37974306 
3.Berger, B., et al., (2021), " Levenshtein Distance, Sequence Comparison and Biological Database Search," IEEE Transactions on Information Theory, Volume 67, Issue 6.
4.Chang, Jau-Shien, & Wang, Guo-Hao. (2021). "Formulating the dynamics of an online community and identifying the influential members using the animal flocking model". Electronic Commerce Research and Applications.
5.Franks, Henry, Griffiths Nathan, & Anand Singh Sarabjot. (2014). "Learning agent influence in MAS with complex". Auton Agent Multi-Agent Syst, 28:836–866.
6.Huang Y. & Ling Li. (2013). "Analysis of user influence in social network based on behavior and relationship". Proceedings of 2013 2nd International Conference on Measurement, Information and Control.
7.Hutchinson, A., " Internal Documents Show Facebook Usage Among Young Users is in Steep Decline," https://www.socialmediatoday.com/news/internal-documents-show-facebook-usage-among-young-users-is-in-steep-declin/607708/, 2021/10/5.
8.Karatas, A. and Sahin, S., (2021)" A Review on Predicting Evolution of Communities," International Conference on Engineering Technologies (ICENTE'21) Konya, Turkey, November 18-20, 2021
9.Khatua, A., et al. (2019), "Tweeting in Support of LGBT?: A Deep Learning Approach," Proceedings of the ACM India Joint International Conference on Data Science and Management of Data, pp. 342-345.
10.Liang, G., et al. (2016), "Rumor Identification in Microblogging Systems Based on User's Behavior," IEEE Trans. on Computational Social Systems, Volume: 2, Issue: 3, 2016, pp.99-108.
11.Liu., Y., et al., (2014), "Modeling What Friendship Patterns on Facebook Reveal About Personality and Social Capital," ACM Transactions on Computer-Human Interaction, Vol. 21, No. 3, Article 17, May 2014. pp. 16-36.
12.Liu, X., et al., (2019), "A novel negative feedback information dissemination model based on online social network," Physica A 513 (2019) 371-389.
13.Liu, W. and Xia, Z, (2021) "Detect Network Water Army based on Large-scale E-commerce Reviews," 2021 IEEE 2nd International Conference on Information Technology, Big Data and Artificial Intelligence (ICIBA).
14.Revelle, M., Domeniconi, C., and Gelman, B., (2021) "Group-Node Attention for Community Evolution Prediction," 2021 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining.
15.Salve, A., Mori, P., Guidi, B., Ricci, L., and Pietro, R. D., "Predicting Influential Users in Online Social Network Groups," ACM Transactions on Knowledge Discovery from Data, Volume 15, Issue 3, Article No. 35, pp 1–50.
16.Shafiq, Zubair M., Ilyas U. Muhammad, Liu X. Alex, & Radha Hayder. (2013). "Identifying Leaders and Followers in Online Social Networks". IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, 31(9), 618-628.
17.Sheikhahmadi, A., et al., (2017), "Identification of influential users by neighbors in online social networks," Physica A 486 (2017) 517-534.
18.Subbian, Karthik, Aggarwal Charu, & Srivastava Jaideep. (2016). "Mining Influencers Using Information Flows in Social Streams". ACM Transactions on Knowledge Discovery from Data, 10(3), pp 1–28.
19.Suler, John, "The Online Disinhibition Effect," Cyber Psychology & Behavior. 2004, 7 (3): 321–326.
20.Soyusiawaty, D. and Rahmawanto, F., (2018), "Similarity Detector on the Student Assignment Document Using Levenshtein Distance Method," 2018 International Seminar on Research of Information Technology and Intelligent Systems (ISRITI).
21.Vaswani, Ashish, Shazeer Noam, Parmar Niki, Uszkoreit Jakob, Jones Llion, Gomez N. Aidan, Polosukhin Illia. (2017). "Attention is all you need". arXiv:1706.03762.
22.Wibowo, N., S., Mahardika, R., and Kusrini, K. (2020). Twitter Data Analysis Using Machine Learning to Evaluate Community Compliance in Preventing the Spread of Covid-19. 2020 2nd International Conference on Cybernetics and Intelligent System.
23.Wolfsfeld, G., et al., (2013)"The Social Media and the Arab Spring: Politics Always Comes First," The International Journal of Press/Politics 18(2), March 2013.
24.Zubiaga, A., Liakata M, Procter R, Wong Sak Hoi G, Tolmie P (2016) Analysing How People Orient to and Spread Rumours in Social Media by Looking at Conversational Threads. PLoS ONE 11(3).