CSCE 470 Lecture 23
Jump to navigation
Jump to search
« previous | Wednesday, October 16, 2013 | next »
Recommenders
- Most popular (overall, in genre)
- Sneaky, paid placement
- Most recent
- Most profitable
- Editorial recommenders - Aggregate (Rotten Tomatoes)
- Random
Applications:
- Music
- Video
- Game
- Books/printed material
- Dating sites
- Facebook friends
- Shopping
- Search engine related stuff
2 Recommender Systems:
- Collaborative Filtering: Essentially, "I will like stuff that people like me like"
- Content-Based Recommenders: "I will like stuff similar to what I like"
Rating Matrix
| User | One Direction | Avengers | Pacific Rim | Django Unchained |
|---|---|---|---|---|
| C | 4 | 4.5 | — | 5 |
| D | — | 5 | 4 | 3.5 |
| B | — | 5 | 5 | 5 |
| Cav | — | 4 | — | 5 |
Big question: How to fill out the blanks in that matrix?
Ignore anomalies, outliers, and changes in taste.
Collaborative approach: Cav was very similar to C, so he might like One Direction
Ratings
Ratings can be explicit:
- user asked to rate
- wrote a review
or implicit
- watched the whole thing and then the next in the series
- clicked on the product or viewed the page
Cold Start Problems
- New Item
- When a new item is added to the matrix and there are no ratings for it, how can it be recommended?
- New User
- When a new user is added and we know very little about that user, what can we recommend?
- Amateurs tend to rate "blandly", but grow into connoisseurs with experience