Introduction

Cancer is one of the most dangerous disease. As there is cure for some types of cancer it is very costly and all humans cannot afford it. So diagnosing the cancer in early stages itself is the best thing we can do. So this project is about classify the different types of cancer using different machine learning algorithms.

For detailed analysis please check this article

Machine Learning Algorithms Used

• Naive Bayes
• Logistic Regression
• Support Vector Machine
• Random Forest
• Stacking Classifier(NB,SVM,LR)
• Voting Classifier

Dataset Link

Data is used from Kaggle website. Use this link to download data.

About the files

• TFIDF_approaches.ipynb - This file consists code and results of all machine learning algorithms with TFIDF vectorization(unigrams and bigrams) of data
• LR with 4 grams.ipynb - This file consists detailed analyis of Logistic Regression with one hot encoding and response coding with 4 grams
• LR AND LSTM approach.ipynb - This file consists results for two layer LSTM model.
• TFIDF_2000.ipynb - This file consits analysis on only top 2000 features. For best results check this notebook.

Results

BOW VECTORIZER

ALGORITHM USED	TRAIN LOG LOSS	CV LOG LOSS	TEST LOG LOSS	MISCLASSIFIED POINTS(%)
NB(ONE HOT)	0.90	1.27	1.21	39.84
KNN(RESPONSE CODING)	0.705	1.130	1.002	39.47
LR+BALANCING(ONE HOT)	0.614	1.143	1.048	34.77
LR+IMBALANCING(ONE HOT)	0.628	1.185	1.054	36.28
SVM+BALANCING(ONE HOT)	0.739	1.132	1.063	36.47
RF(ONE HOT)	0.703	1.192	1.097	36.84
RF(RESPONSE CODING)	0.052	1.325	1.211	46.8
STACKING CLASSIFIER(NB,LR,SVM)	0.663	1.177	1.081	36.24

Observation: RF(Response Coding) model is overfitted. Considering all business constraints mentioned in this blog, LR+Balancing better than all other models.

TF-IDF VECTORIZER

ALGORITHM USED	TRAIN LOG LOSS	CV LOG LOSS	TEST LOG LOSS	MISCLASSIFIED POINTS(%)
NB(ONE HOT)	0.906	1.229	1.225	41
LR+BALANCING(ONE HOT)	0.57	1.114	1.113	36.84
LR+IMBALANCING(ONE HOT)	0.56	1.13	1.12	36.09
LINEAR SVM(ONE HOT)	0.68	1.15	1.19	37.78
RF(ONE HOT)	0.64	1.12	1.16	37.7
RF(RESPONSE CODING)	0.05	1.38	1.44	48.8
STACKING CLASSIFIER(NB,LR,SVM)	0.63	1.16	1.19	38.90
MAXIMUM VOTING CLASSIFIER(NB,LR,SVM)	0.63	1.16	1.19	40.10

Observation : When compared with BOW vectorizer results of TF-IDF Vectorizer are better and LR is outperforming all other algorithms and RF is overfitting again. So lets see detailed analysis on LR

LR (UNIGRAMS AND BIGRAMS)

ALGORITHM USED	TRAIN LOG LOSS	CV LOG LOSS	TEST LOG LOSS	MISCLASSIFIED POINTS(%)
LR(BIGRAMS)	0.83	1.17	1.19	40.7

LR(4 GRAMS WITH TOP 2000 FEATURES)

ALGORITHM USED	TRAIN LOG LOSS	CV LOG LOSS	TEST LOG LOSS	MISCLASSIFIED POINTS(%)
LR(4 GRAMS)	0.439	0.957	0.982	35.4

LR (4 Grams With top 2000 features) outperformed all other algorithms and log loss value less than 1.0

Feel free to add pull request if you good results on this dataset.

Don't forget to star the repo if you liked the analysis.

Acknowledgment

This case study is part of this course