Abstract: We propose a depression classification model of deep neural network to solve problems of feature redundancy, single feature dimensions and difficulty to determine feature subset in big data of depression related scale．By combining principal component k-means algorithm (PC k-means), we selected features of scale big data without destroying original feature space, and optimized randomness of original algorithm and uncertainty of number of clusters．To enhance diversity of depression recognition dimensions, a deep neural network depression classification model with factor decomposition machine (FM-DNN) was constructed．Analysis and comparison showed that PC k-means could not only effectively select features, but also improve accuracy of depression classification by combining with traditional classifier and FM-DNN．This work provides new ideas and direction for introduction of scale big data analysis into deep learning．