Synthetic Samples Generation for Imbalance Class Distribution with LSTM Recurrent Neural Networks
Biprodip Pal, Arnab Kanti Tarafder, Md. Shahinur Rahman · 2020
Developing a classifier model is difficult to properly classify rare items or minority data as long as the training data distribution is imbalanced. Existing resampling issues to remedy this problem, often causes loss of information or are depended on proper computation of neighborhood information to generate new samples. On the other hand, Long Short Term Memory has been used as successful generative models for time series and sequence generation tasks. This article proposes a resampling approach to synthesize new minority samples leveraging Long Short Term Memory based Deep Learning technique to capture association information from manifolds of minority samples and generates new minority sequences based on the conditional probability on the parent sequences. The obtained experimental results on benchmark imbalance model testing datasets revealed that the proposed approach outperforms baseline classifiers by up to 5% as compared to their performance while trained on imbalanced data and outperforms one of the benchmark imbalance reduction method SMOTE by up to 34% in specific cases.