隨著機器學習的日益普及，越來越多的產業引入機器學習來輔助產業發展，也使得它更加融入我們的生活，與之對應的所需技術也層出不窮。然而擴展到更多領域的機器學習都勢必會經歷到的瓶頸，那就是設備資源的限制。以常用在圖形辨識任務的卷積神經網路來說，作為輸入資料的圖片可以自由縮放成訓練所需的尺寸。但是，對於社群網路來說，社群網路圖的圖象化尺寸遠超過一般的圖形資料，並且難以割捨其中的資料內容而無法使用一般的縮放技術，也就不可能以普通的操作進行機器學習的訓練。我們提出了一套應用於動態社群網路分析時所使用的系統。藉由我們提出的階層式群集演算法，並採用多區塊分割法，最後結合 autoencoder 所形成的複合壓縮技術。在確保資料不失真以及壓縮效率上取得最佳平衡。實驗證明，我們所提出的方法能大幅增加神經網路模型能處理的社群網路資料量，並且能減少預測模型的運算負擔，同時也可降低對硬體設備的依賴程度。 ;With the increasing popularity of machine learning, more and more industries have introduced machine learning to assist the development of the industry, which has made it more integrated into our lives, and the corresponding required technologies have also emerged. However, the bottleneck that machine learning is bound to experience is the limitation of equipment resources when expands to more fields. In the case of convolutional neural networks, which is commonly used in graphic recognition tasks, uses the pictures as input data that can be freely scaled to the size required for training. However, for the social network, the image size of the social network graph is much larger than the general graphic data, and it is difficult to discard the data content in it, so it is impossible to use the general scaling technology, and it is impossible to use ordinary operations for machine learning training. We have proposed a system for use in dynamic social network analysis. With the hierarchical clustering algorithm that we proposed, and then using the multi-block partition method, finally combined with the autoencoder, it becomes a composite compression technology formed. The best balance is achieved in ensuring that the data is not distorted and compression efficiency. Experiments show that our proposed method can greatly increase the amount of social network data that the neural network model can process, and can reduce the computational burden of the prediction model, and can also reduce the degree of dependence on hardware devices.
