We address class imbalance problems. These are classification problems where the target variable is binary, and one class dominates over the other. A central objective in such problems is to identify features that yield models with high precision/recall values, the standard yardsticks for assessing such models. Our features are extracted from the textual data inherent in these problems. We use n-gram frequencies as features and introduce a discrepancy score that measures the efficacy of an n-gram in highlighting the minority class. The frequency counts of n-grams with the highest discrepancy scores are used as features to construct models with the desired metrics.
According to the best practices followed by the services industry, many customer support tickets will get audited and tagged as “contract-compliant” whereas some will be tagged as “over-delivered”. Based on in-field data, we use a random forest classifier and perform a randomized grid search over the model hyperparameters. The model scoring is performed using an 𝐹𝐹𝛽𝛽 scoring function. Our objective is to minimize the follow-up costs by optimizing the recall score while maintaining a base-level precision. The final optimized model achieves an acceptable recall score while staying above the target precision. We validate our feature selection method by comparing our model with one constructed using frequency counts of n-grams chosen randomly.
We propose extensions of our feature extraction method to general classification (binary and multi-class) and regression problems. The discrepancy score is one measure of dissimilarity of distributions and other (more general) measures that we formulate could potentially yield more effective models.