CSCE 470 Lecture 26

« previous | Friday, October 25, 2013 | next »

Quiz 3 is coming up:

• Clustering
• Classification
• Maybe Recommenders

Content-Based Recommenders

Suppose each movie title had other information associated with it (actors, script, etc.)

Create a "profile" for user that represents what the user likes

Recommend items similar to what the user likes

Attacking Recommenders

"Average" Attack

Target specific user(s)

• Create shills
• mimic ratings of target user(s)
• promote other item(s) with high ratings

Project Chaos

Undermine whole system

• Create shills
• random ratings
• separate similar items

"Pushing" / "Nuking"

Torgets multiple users or ("popular" recommenders)

• Create shills
• pick a single target item
• push by rating target item well
• nuke by rating target item poorly

Artificial Clustering

Correlate unrelated movies (target cluster-based recommenders)

• Guess what? create shills
• Rate unrelated target items similarly

Recommender Comparison

In reality, companies use a combination of both to overcome weaknesses of each.

Collaborative Filtering

Pros

• can model anything that has ratings
• can recommend things outside of domain: e.g. You like action films. People who like these movies also like this comedy movie.

Cons

• Can't recommend anything that has no ratings (new item problem)

Content-Based

Pros

• Solo!! (not dependent on the crowd)
• works for small user base
• no new item problem

Cons

• Restricted to recommending things only like what you already like

Using Classification

Two classes: recommended, not recommended

Training data: stuff user likes and doesn't like

Associate "likes" and "don't like"s with features