The objective of this repository is to create prediction on product ratings based on a text given from customer feedback. This repository will be creating a model to predict the rating of product 1-5 stars with 5 being the highest.
To create the model I will be training the model on Amazon reviews. Amazon product reviews have a star rating system so that can be used as the target output prediction.
There are different methods for classifying text as a rating system.
Learns by the joint probability distribution p(x,y) then by using Bayes is transformed to p(y/x) however you are not directly computing p(y/x). In this classifier for text analysis, there is an assumption that the words are independent of one another.
Learns the conditional probability distribution p(y/x) "probability of y given x. This type of model would read the data sequentially to try and figure out the difference in ratings based on examples. In this experiment I tried a number of different models (CNN, Multi Size CNN, and a LSTM).
The input variable in this is the text summary of a review, and the target output is the rating
- The target variable (rating) I converted this to one hot encodings
- Vectorized the text summary, I also removed emojis text
- Calculated the average, max, min of the text summary
- Set a sequence length that is appropriate
- Tested x3 models (single hidden layer, CNN, LSTM)
Out of the three models, I tested. The LSTM model had an accuracy of 0.898. It should be noted minimal fine-tuning of the model was done and the model was only trained on 10,000 of the 500,000 available rows.
