Genetic Variant Classifier : Random Forest Beats Deep Architecture (By a HUGE MARGIN)

Hello Readers!

Welcome to yet another value prediction work! Today, we will be looking at the in-demand dataset , namely Genetic Variant Classifications.

We will look at this dataset and go for it's primary objective, that is classification of  the two lab reports and determining whether they both conflict or not.

The Kernel you may want to look at for more information : Conflicting result classifications

As usual, we will be looking at the dataset with the aim of EDA, Feature Engineering and Predictions

Exploratory Data Analysis

One would like to see what are the Chromosomes vs Class distribution of this data. For that, you can simply use : 

As you can observe in the graph given below, the dataset happens to be heavily biased towards the non- conflicting genes and that too with the CHROM == 2 standing out as the clear bias winner.

Since the incidents where the genes are recorded to be conflicting, we can assume that our classifiers won't be doing much of a great job and we can assume that most of them would take CHROM 2 as their most important feature if we don't feature hash it into something of a lower dimension

Once that is clear, let's go on and check out our feature selections:

Feature Engineering

Applying Feature Hashing and Binary Classification on various classes.

Machine Learning Models

I will be skipping the pure modelling part here for that you guys can visit the kernel from the link above and will just share the interesting detail which I came across recently, namely Feature Importance. 

The feature importance of the Worst Performer (Logistic Classifier) can be shown as : 

From the graph above, we can easily understand that the most important features taken into consideration by our Logistic Regressor are Inversion, Microsatellite and LOW, while the chromosomes are given very little or no importance.

While the Feature importance of the Best Performance (Random Forest Classifier) can be show as : 

Here, we can see the emphasis on Chromosome types.

Deep Learning Model

I tried the following deep learning model for the same prediction (although it failed. Pretty badly)

and well, it came out with a performance of worse than the Decision Tree Classifier (discussed in Kernel itself)

Conclusion

By this, we can conclude that Random Forest Classifier is our best bet, and our deep learning model probably failed because of extensive binary and feature modelling and evident bias.