The first step would be to generate the training data set using emoticons only and assign positive and negative labels to the tweets. During the testing phase, we will extract features from the tweets (e.g., emoticons, positive adjectives, negative adjectives etc) and apply machine learning classifiers such as SVMs, K-Nearest Neigbors, Random Forest and the like. I will use Python's scikit-learn library to implement this project. The output metrics such as recall, precision, F-measure, and accuracy will be extensively discussed. Moreover, the code will be fully documented so that you understand the solution. I look forward to discussing further details with you on chat!