Introduction To Recommender Systems at Universität Bern

The TF-IDF concept can be used to create a profile of a document/object. A movie could be described as a weighted vector of its tags. These TF-IDF profiles can be combined with ratings to create user profiles, and then matched against future documents.

Advantages:

• Set of users can be small
• Good for finding subsitutes
• Good for finding the right product
• Good explainability

Disadvantages:

• Unlikely to find surprising connections
• Harder to find complements (rather than substitutes)

In the context of Recommender Systems, a prediction represents the anticipated opinion of a certain user for a given item. The predicted rating should have the same numerical scale as the real ratings.

A recommendation, on the other hand, is a list (subset) of all the items. The items in the list are chosen in a way that the user will probably like them the most of all the items in the system.

• Default Voting: insert default voting for unrated items, which corresponds to neutral or somewhat negative preference.
• Preprocessing using Averages: for each user, assign his or her mean rating score to the unobserved items.
• Filterbots: use (intelligent) automated rating agents to rate new items.
• Dimensionality Reduction: use dimensionality reduction techniques like SVD. This may also solve the synonymy problem since the latent representation of two items with different names might be similar.

We can either just pick the useres that have the highest similarity with the active user or we can incrementally add a new user to the neighborhood, calculate the new centroid of this neighborhood and proceed by selecting the next user that is closest to this centroid.

• Recommend the N most frequent items that have been rated in the neighborhood of the active user but not by the active user itself.
• Make use of association rules and choose the ones with the highest confidence.

• User tastes are stable
• For a given domain (e.g. movies): if two users agree on one type of item, thy also agree on other types of items

The lecture slides are somewhat ambivalent regarding this question. In general, k seems to be often somewhere in between 25 and 100 in many domains, but it might be better to start with a value between 20 or 30 and 50. For movie ratings, Herlocker et al. found k = 20 to be a good value.

Yes. For example, if a user always rates things very low that the active user rates very high, we can infer that their tastes are somewhat orthogonal. We could use the bad ratings of this user to predict items that the active user is going to like.

• Pearson Correlation
• Constrained Pearson Correlation
• Spearman Rank Correlation
• Cosine Similarity

In the Constrained Pearson Correlation, we use a neutral value instead of the mean rating of each user. This neutral value should denote neither agreement nor disagreement.

• Item-item similarity is fairly stable.
• We can precompute the item-item similarity (since it is stable)
• Faster recommendations for systems with a lot of users.