Variational Autoencoder Model Combining Deep Learning and Probability Statistics and Its Application in Large-scale Data Analysis

A multi-layer generative model is proposed as a means of enhancing the accuracy of large-scale data analysis. This model addresses the problem of limited feature extraction capability and insufficient association with label information in existing topic models. The model is divided into three main modules: text encoding, autoencoder inference, and layer-by-layer learning. The model combines a hierarchical Bayesian model with a deterministic upward random downward network structure. It uses a Poisson Gamma Belief Network as a decoder to capture hierarchical implicit features in text data during text encoding, autoencoder inference, and layer-by-layer learning. Random Gradient Monte Carlo sampling is used for posterior inference to improve the model efficiency. In addition, the Fisher information matrix is used to adaptively adjust the learning rate of different levels and topic parameters, and a layer-by-layer learning strategy is introduced to construct a learning network. Based on this, text data and label information are combined for feature extraction. The results demonstrated that the test error rates of the designed model on the 20News, RCV1, and IMDB datasets were 16.52%, 18.72%, and 11.67%, respectively, all of which were the lowest. Additionally, the testing time was the shortest, at 0.020s, 0.017s, and 0.015s, respectively, indicating a high level of accuracy and efficiency. In addition, the perplexity levels on the 20News, RCV1, and Wiki datasets were 590.23, 953.12, and 982.67, respectively, significantly lower than those of other comparison models. Given this, the designed model has high data analysis and interpretation capabilities and relatively high computational efficiency, which can provide scientific tools for accurately analyzing large-scale data in batches. © 2024 Slovene Society Informatika. All rights reserved.