To evaluate the predictions, we use Weighted Categorization Accuracy Metric. In which we assign weight to each class and accuracy is calculated based on the weights attached.

Weighted categorization accuracy can be generally defined as:

![](https://cdn.iisc.talentsprint.com/CDS/Images/Kaggle_competition.gif)

where  y are ground truths,  y_hat are the predicted results, and w  is the weight associated with the ith item in the dataset.

In our test set, we assign each testing item a weight according to its category. We set the weights in consideration of the imbalance of the data distribution to minimize the bias and to improve performance caused by the majority class.