This is the first part in a project that will provide a more sophisiticated recommendation system than is currently found in industry. When different people go to a restaurant they have different priorities. Some people prioritize cost. Others prioritize the atmosphere. Current recommender systems don't take this into account.
I train a new model to estimate the sentiment of each word using the star ratings of the review as whether the sentiment should be positive or negative. In accordance with [1] I include a neutral category in the training.
- Sentiment Analysis with Spacy and Sci-kit learn for the machine learning and tf-idf processing.
This project was inspired by the following papers:
- Koppel, Moshe, and Jonathan Schler. "The importance of neutral examples for learning sentiment."
- http://harshtechtalk.com/get-informative-features-scikit-learn/
- Proof that we can use SVM coefficients to get the relatively important coefficients.