TopicSpecific Recommendation An Approach to Greater Prediction Diversity

Topic-Specific Recommendation An Approach to Greater Prediction Diversity and Accuracy CS 345 a Minho Kim Brian Tran

Outline l Motivation l Topic-Specific Recommendation l Comparison to other methods l A specific example l Future

Problems w/ Recommendation l l l Prediction Diversity Improved Accuracy Maximize Long Tail recommendation – Possibly provide recommendations for less popular movies

Topic-Specific Recommendation l l Divide items into different topics (genre) Find similar users within each topic Provide recommendations for each topic (even unseen ones) Recommendations should be: – – more diverse more accurate

Comparison to Other Methods MAE RMSE < 0. 5 Diff Exact Match Topic Specific 0. 89 1. 14 7539 1579 Per-Item Average 0. 98 1. 23 6309 342 STI Pearson 0. 95 1. 22 6791 304 Non. Personalized 0. 89 1. 12 6905 169 Optimal Constant Weight 1. 16 1. 49 5320 1428

A More In-Depth Look… l l In Amazon, we entered the following movies: All were considered dramas

The Results Ours: All were dramas… Amazon’s: One drama, but also comedy/romance!

Futhermore Rotten Tomatoes: Rotten Avg Rating: 3. 4 User’s Rating: 4. 0 Our Predicted Rating: 5. 0 Rotten Tomatoes: Rotten Avg Rating: 3. 5 User’s Rating: 4. 0 Our Predicted Rating: 4. 0 Rotten Tomatoes: Fresh Avg Rating: 3. 6 User’s Rating: 4. 0 Our Predicted Rating: 4. 2 No other methods recommended these movies

Future Possibilities l l Different type of dataset Larger dataset (Netflix) Try it on different topics Handling new items and/or users

Questions?