The goal of this project was to train a few models to classify customer sentiment given a product review, which was trained on Amazon Reviews here. This dataset is extremely huge, which helped out in the training phase.
As is known, models have difficulty learning meaningful patterns from raw, unprocessed data. Thus, I cleaned the data and ensured the data was balanced so that the learning has no problems. The following steps were taken to clean the data:
- Lowercase all the reviews
- Remove all HTML tags and URLs
- Remove non-English alphanumeric characters and extra spaces
- Expand contractions (won't became will not)
- Remove stop words
- Perform lemmatization
The average character length of the reviews were reduced from 309 to 183 characters (down by 40%).
Next, I randomly sampled 100k instances of each class (positive and negative sentiment) to be the final processed dataset. By doing this, the model will be fitted on a high-quality dataset that is well-balanced.
So far the reviews are still in their textual format. By utilizing Sklearn, I was able to extract features from the reviews using TF-IDF, which is a value that is intended to reflect how important a word is to a dataset. This step converted each review from its textual format into an array of floats.
A total of 4 models were trained using Sklearn: Perceptron, Linear SVM, Logistic Regression, and Multinomial Naive Bayes. The following are the prediction results against the testing set after training the models:
| Model | Precision | Recall | Accuracy | F1 score |
|---|---|---|---|---|
| Perceptron | 85% | 89% | 85% | 84 |
| Multinomial Naive Bayes | 87% | 87% | 87% | 87 |
| Linear SVC | 89% | 90% | 89% | 89 |
| Logistic Regression | 90% | 90% | 90% | 90 |