Introduction to Data Analysis Final Exam
Problem
§ You have learned supervised learning and unsupervised learning in this class.
§ What if we have a limited number of labeled data for supervised learning tasks and addi tional unlabeled data?
• You are given a limited number of labeled data and additional unlabeled data in this exam.
• The goal is to improve your model’s accuracy as much as possible using a given dataset consi sting of labeled and unlabeled data.
Strategies
§ You can improve your supervised learning task model’s accuracy using additional unlabel ed data.
§ We suggest two simple strategies for this problem.
• Self-Training
• Co-Training
§ You can follow these strategies, or you can apply your own method.
Self-Training
§ Train your classifier with labeled data.
§ Use the trained model to predict labels for the unlabeled data.
• A confidence score can be used for predicting labels.
Score(x) > threshold The model's prediction for x can be regarded as a confident label. Score(x) < threshold The model’s prediction for x cannot be regarded as a confident label.
§ Retrain the model with the pseudo and labeled datasets tog ether.
Co-Training
§ Split features into two exclusive feature sets.
§ Train two different classifiers using two different feature sets.
§ Get pseudo-labels from unlabeled data.
§ Expand labeled data using pseudo-labels from the different view models.
Libraries
For the library, you can use the library used in the previous HW (i.e., seaborn, matplotlib, pandas, NumPy), sklearn, xgboost, kmodes, and mvlearn.
• “sklearn.semi_supervised” and “mvlearn.semi_supervised” are allowed.
