Amazon Movies and TV Ratings Recommendation Systems

Summary

In this project, the user ratings from a subset of Amazon reviews are used to train various recommendation systems by evaluating various algorithms with hyperparameter optimization. The models return a list of recommended items based on the reviewer's previous actions.

Data

The Movies_and_TV ratings data was retrieved from Recommender Systems and Personalization Datasets. This includes (item, user, rating, timestamp) tuples. A subset of the data was used in the analysis due to computation constraints due to sparcity.

Modeling

Model Training Process

Models were trained using:

• surprise library
• A ranking model using tensorflow-recommenders
• Collaborative filtering models using tensorflow
• scipy library
• A popularity recommender model

For the initial training of the models using surprise:

• The default parameters of BaselineOnly, KNNBaseline, KNNBasic, KNNWithMeans, KNNWithZScore, CoClustering, SVD, SVDpp, NMF and NormalPredictor were evaluated using the cross_validate method using 3-fold cross validation to determine which algorithm yielded the lowest RMSE errors.
• Then hyperparameter optimization was performed to find the best parameters for the four model types with the lowest RMSE errors (SVDpp, SVD, BaselineOnly and KNNBaseline).

For the construction of the ranking recommender model:

• The data was loaded in a tensorflow dataset and partitioned into train/test sets.
• A vocabulary was generated to map the feature values to embedding vectors.
• The model used an embedding_dimension=64 and contained multiple stacked dense layers with activation='relu'.
• Raw features were used as input to compute_loss and MeanSquaredError was used for the loss metric.
• The model was fit for 20 epochs and evaluated.
• The ranking model was tested by computing predictions for items and ranking by the predictions made.
• The model was saved and exported for serving purpose in TensorFLow Lite.

For the training of the SVD based models using SciPy:

• A rating matrix with items and reviewers was constructed.
• The parameters for the model were U, sigma, Vt were tested in randomized_svd and TruncatedSVDTruncatedSVD.
• A diagonal matrix was constructed in SVD.
• Then the ratings were predicted and RMSE was calculated.

For the construction of the popularity recommender model:

• A recommendation score was created by counting each reviewer for each unique item.
• This score was sorted and a recommendation rank was created based on scoring.
• Predictions were then calculated for various reviewers.