The project was developed as a part of the national-level data science challenge at Phillips HackAbout and judged in the top 5. Proposed a 2-level combination of classifier and regression neural network for imbalanced datasets, wherein samples of certain class outnumber the sample of other class.
Since the dataset was imbalanced, our initial predictions for length of stay were highly biased towards the mean.
To overcome the problem of the imbalanced dataset, we divided the dataset into two classes - Near mean (majority class) and Far mean (Minority class). We tried to take into account the known prior information about the skewed class distribution by the use of ensembles. Since it's easier to determine the boundary between the majority and minority classes, we first identify the class to which a data-point belongs to. Since now we have knowledge about the location of the data point in the class distribution, we can now improve now perform regression on the data-point.
The first module is a classifier, which tries to identify to which class the data point belongs to (i.e. Near-mean or Far-mean).
The second module of the ensemble is a set of regression neural networks that are trained on different classes of the dataset. Based on the class-prediction from the classifier, one of the regression neural networks is chosen from the set and used to make predict length of stay of myocardial infarction.
We attained an accuracy score of 87% on the predicted length of stay.