Rating Prediction with User Business Review

Overview

Use following methods to accomplish the prediction of the restuarant rating

1. Text Data Processing
   • cunt vectorizer : Convert the words into matrix.
   • tf-idf(Term Frequency–Inverse Document Frequency) : Ignore some function words, catch some important words.
2. Feature Extraction and Engineering
   • Use the data which is done by tf-idf to features suitable for modeling.
3. Cross Validation & Hyperparameter Tuning
   • K-fold : Split the data into k equal sized subsamples. Retain one of subsample for testing model and repeat k times.
   • Hyperparameter Tuning : Learning rate, iterators, layers etc.
4. Commercial Recommendation System

Data

• training data
  • total nearly 8000
  • include review_id, business_id, user_id, text, date, stars
• test data
  • total nearly 2000
  • include review_id, business_id, user_id, text, date

Purpose

Use training data to train model, then use the model for test data to predict the restaurant stars

The methods I tried

• texts

  1. Drop the unnecessary column(it must be test many times)
  2. Use nltk.stem library to remove the prefixes, suffixes from a word and and change it to its base form.
  3. Drop the stopword
  4. Use TF-IDF(n_gram: choose numbers of the words, min_df and max_df: filter the result ) to divide the string.

• model

  • discrete - linearSVC and linear regression: convert the matrix into high dimension, the accurancy is higher, MSE is higher because of the result is in integer.
  • continuous - linearSVR and SVR: the result will be out of 5(the maximum of rating), we may need to filter the result to make it range from 1 to 5

Notation

• In tf-idf, the n_gram selecte number of the words. example: Tom is not handsome.
• The continuous training model will be out of the rating range(1 to 5)
  n_gram = 1 -> Tom, is, not, handsome; n_gram = 2 Tom is, is not, not handsome.
• To reduce the RMSE, can make all rating larger than 4.6, convert it to 4.6; all rating smaller than 0.5, convert it to 0.5.

Best Result

• Accuracy 0.5042
• RMSE(Root-Mean-Square Error) 0.8071