IFT6390_Text-Generation

Final year project for IFT 6390- Fundamentals of Machine learning Team Members:
Kuruba Vijaya Lakshmi
Saadaoui Houda
Binulal Narayanan

In this project, we use different Data cleaning strategies, Imputation techniques and Machine Learning models to classify protein sequences as AMP or non-AMP.

Approach

1. Exploratory data analysis
2. Based on data analysis we found 2 challenges:

• Imbalanced dataset: No of non-AMPs are more way higher than AMP.
• the distribution of the sequence lengths varies a lot between the two classes.

3. We have Tried 4 different options with dataset to address Imbalance in sequence lengths

• Original dataset

• Dataset 1: We reconstructed Non Amp sequences such that we ignored the samples whose sequence legths are not in AMP sequences.

• Dataset 2: Dataset in which sequence lengths >125 is ignored, based on fact that 95 percentile of AMP sequence length is 125.

• Dataset 3 : Considered only Amplify dataset since the samples and sequence lengths are balanced

4. For Data samples imbalance , Tried Oversampling and Undersampling Techniques.
5. Models :

• 1. Random forest classifier + Oversampling + Dataset1
• 2. Random forest classifier + Oversampling + Dataset2
• 3. Random forest classifier + Undersampling + Dataset1
• 4. SVM Classifier + No Sampling +Cost penalised+ Original dataset
• 5. Neural networks + Dataset 3

6. Results

This repositary has 2 colab files.

1. IFT_6390_Final_Project_1_.ipynb - It contains Models option1 - option 5
2. IFT_6390_Final_Project_1_supp.ipynb - It contains different data strategy (data split on different ratios) modelled with Random forest