以卷積與遞迴類神經網路為基礎之動態社群網路預測

A Dynamic Social Network Prediction Based on Convolution and Recurrent Neural Networks

黃士綸(Shih-Lun Huang)

指導教授 ： 王英宏

淡江大學/工學院/資訊工程學系碩士班/碩士(2022年)

https://doi.org/10.6846/TKU.2022.00760

摘要

隨著網路日益普及的發展，人與人之間已經一種網路社會，稱為社群網絡。透過社群網路的互動，可以了解用戶的交友圈，並且推薦哪些用戶適合相加為好友。我們把社群網路以圖形的方式呈現出來，將一定的時間點作為固定間隔，並劃分為動態社群網絡，以用來觀察連結之間的演化趨勢，來預測未來會有哪些連結出現或消失在圖中。這就是所謂的動態網路連結預測(Dynamic network link prediction)。我們提出了一種新的基於CNN+LSTM的模型，C3D-LSTM，用於動態網路的連結預測。C3D-LSTM結合了Conv3D和LSTM，將資料分群後，以3D立體的形式，來訓練資料，並且跟GRU、LSTM、Conv2D-LSTM做比較。實驗結果顯示C3D-LSTM有相當高的AUC。最後在對模型進行參數調整，以達到最佳的預測結果。

關鍵字

動態社群網路 ； 長短期記憶 ； 動態網路連結預測

並列摘要

With the growing popularity of the Internet, an online society has been formed between people and people, which called social network. Through social network interaction, we can understand users' circle of friend-making and recommend which users are suitable to be added as friends. We present social networks graphically, using time-step to fix the interval, and divide it into dynamic network, to observe evolutionary trend between links, then predict which links will appear or disappear from the graph in the future. This is called Dynamic Network Link Prediction (DNLP). We propose a new model based on CNN+LSTM, C3D-LSTM, which predict for dynamic networks. C3D-LSTM combined Conv3D and LSTM. We separate the data in 3D form, and train the dataset, then comparing with GRU, LSTM, Conv2D-LSTM to see which one is better. In the result, C3D-LSTM had the best performance.

並列關鍵字

Dynamic Social Network ； Long Short-Term Memory ； Dynamic Social Network Prediction

參考文獻

[1] E. Adar and L. Adamic, “Friends and neighbors on the Web,” in Social Networks,

Volume 25, Issue 3, Pages 211-230, 2003.

[2] C. Aggarwal, W. Yu, W. Cheng, H. Chen and W. Wang, "Link Prediction with Spatial

and Temporal Consistency in Dynamic Networks", in International Joint Conference on

Artificial Intelligence (IJCAI), Pages 3343-3349, August 2017