Using Coclustering for Predicting Movie Ratings in Netflix

1
Using Co-clustering for Predicting Movie Ratings
in Netflix

• Tuyen Huynh
• Duy Vu

2
Outline

• Data analysis
• Methods for predicting missing ratings
• Experiment setup
• Results
• Conclusion

3
Data Analysis

• Subset of Netflix dataset
• 12326 movies, 9730 users, 2048499 ratings (1.71
  known ratings)

4
Data Analysis (cont.)
5
Data Analysis (cont.)
6
Data Analysis (cont.)
7
Co-clustering for Missing Value Prediction (MVP)

• Co-clustering can be seen as a matrix
  approximation
• Co-clustering the original matrix to form
  coclusters
• Compute some summary statistics based on these
  coclusters
• Construct an approximation matrix based on these
  summary statistics

Use the approximation matrix for predicting
missing values
8
Co-clustering for MVP (cont.)
Banerjee et. al. A generalized maximum entropy
approach to bregman co-clustering and matrix
approximation.
9
Local prediction based on SVD

• Use a thin SVD of rank k for approximating a
  matrix
• Has been shown that this rank-k matrix is the
  best rank-k approximation matrix of R in
  Frobenius norm (squared error loss)

10
Local prediction based on SVD (cont.)
11
Experiment Setup

• The dataset is split into a training set (90 of
  ratings of each user) and a test set.
• Use squared Euclidean distance and I-divergence
  for the nonnegative rating matrix.
• Use different settings of co-clusters 2x2, 5x5,
  10x10, 15x15, 20x20, 25x25.
• Initialize row and column clusters randomly or by
  using Graclus clustering results.

12
Evaluation Metrics

• Root Mean Squared Error (RMSE)
• Mean Absolute Error (MAE)

13
Results

• Baseline results use the average rating (3.6)
  for predicting missing ratings
• RMSE 1.0834
• MAE 0.90881

14
Results (cont.)
15
Results (cont.)
16
Results (cont.)
17
Results (cont.)
18
Results (cont.)

• Local SVD on the best co-clustering result with
  the setting 20x20 using Euclidean distance and
  scheme 3.

19
Results (cont.)

• Local SVD on the best co-clustering result of
  Graclus.

20
Conclusion

• The co-clustering method with the appropriate
  schemes achieves the significant prediction
  performance.
• The application of local SVD to the best
  co-clusters obtained from co-clustering and
  Graclus is successful.

21

• Thank you!!!