Identification and Visualization of Zombie Enterprise Portraits - Mining Short-time Series Features from the Perspective of Image

Governance of zombie enterprises is an important means to ensure the healthy, sustained development of the economy. Traditional methods such as identifying zombie enterprises based on expert knowledge suffer from incomplete expert database and increasingly complex economic environment. Thus the proposed data-driven system is implemented to not only identify zombie enterprises, but also visually present the enterprise portraits. In this paper, the three-year data of 50000 enterprise is transformed into N×N×3 image-format-matrix (N×N are the number of features). Afterward, Convolutional Neural Network, namely CNN is applied and result is got in one stage instead of fitting the data of each year and voting. It is also proved that CNN can effectively mine the short-time series features of enterprises by reconstruct the data into image-format-matrix. Considering the imbalance of data, Focal-Loss is implemented as the loss function when applying CNN model to the data. Grad-CAM, a model interpretive method in the image domain, is used to explain the CNN network after the fitting is completed. It is found that the model pays too much attention to salient features. Thus Mutual Channel Loss is further implemented to make the model pay attention to those indistinguishable features. At the same time, CBAM attention module is added to pay selective attention to different characteristics of enterprises in different years. The three-year information of 15050 enterprises collected from the State Administration for Industry and Commerce of China is used as the source data. The results show that comparing with other models, our CNN model reached the state of art in the rate of misjudgment and missed judgment.