Hard-Disk-Predictive-Maintenance

Repository of models trained for CS3244 Machine Learning term project. See project paper here.

Logistic Regression (LogisticRegression.ipynb)

• Attributes used: 'smart_12_normalized', 'smart_189_normalized', 'smart_190_normalized', 'smart_193_normalized', 'smart_199_normalized', 'smart_240_normalized', 'smart_242_normalized', 'smart_5_normalized', 'smart_187_normalized', 'smart_188_normalized', 'smart_197_normalized', 'smart_198_normalized'
• Randomly sampled 836 negative examples (non-failed hard disks) because we only have 836 positive examples
• Train: 80%; Test: 20%
• Score: 0.7014925373134329
• Lots of false negatives

AdaBoost

• Attributes used: all except 'date_x','serial_number','model','failure_x','date_actual_fail'
• Used various oversample/undersample methods (SMOTE, SMOTETomek, ADASYN)
• For AdaBoostSMOTETOMEK(biased data).ipynb, negative:positive ratio is 1:5