An improved CNN with Bi-LSTM model for Parameter reduction and Handling imbalanced data

Abstract

This paper presents improvements in a deep learning model combining CNN and Bi-LSTM to address two key issues: imbalanced data and computational complexity. To tackle the problem of imbalanced data, techniques such as SMOTE, undersampling, and class weight adjustments are applied, enhancing accuracy for minority classes in the dataset. Experimental results on the UCI Student Performance dataset demonstrate the model's effectiveness in predicting student academic performance. Meanwhile, to reduce computational complexity, Depthwise Separable Convolutions are employed to decrease the number of model parameters. Results are presented through the air quality prediction task for PM2.5 in Ho Chi Minh City, showing the efficiency in saving computational resources without sacrificing prediction performance.