Predicted the Total UPDRS score based on 6000 records from the Parkinson’s telemonitoring dataset using various regression techniques. Achieved a minimum MAE of 0.415 using PCA with a multi-layered perceptron model. Determined an optimum threshold value of 15 for motor UPDRS score discriminated by dysphonia measurements
Implemented an LSTM model with shifting and down sampling the data to achieve an F1 score of 0.868, a signification improvement over random forest baseline which gave an F1 score of 0.39. Compared results with techniques such as SMOTE and weighted loss.
Performed sentiment analysis on a crowd sourced movie review dataset with doc2vec model for vectorization pre-trained on IMDB movie review dataset. Compared results against other vectorization techniques such as count vectorization and word2vec models