Recommender Systems and Matrix Factorization

Recommender Systems

• Reasons for using recommender systems include finding all good items, finding good items in context, and helping others.
• The regularization term in the optimization function of an SGD-trained MF model is used to avoid unbound values in the w and h vectors and to prevent overfitting.

Model-Based Collaborative Filtering

• A common choice for regularization is Tikhonov regularization.
• The mutual proximity approach can effectively reduce hubness in recommender systems.

Memory-Based Collaborative Filtering

• In case of binary ratings, no user bias factor is required when computing memory-based CF.
• User-based CF (in the memory-based variant) tends to scale poorly with the number of users and items.

Information Retrieval

• The IDF monotonicity assumption states that terms that appear in only a few documents of the corpus are more important than terms that appear in many documents.
• The use of a logarithmic term when computing TF and IDF values is motivated by Zipf’s law about the frequency of word occurrences in English texts.
• The TF monotonicity assumption states that terms that appear frequently in a document are more important than terms that appear infrequently.

Content-Based Filtering

• Content-based filtering is well-suited to recommend “long tail” items because content features are less affected by popularity biases than user ratings.

Topic Modeling

• The topic modeling approaches pLSA and LDA assume conditional independence of terms and documents, given a topic z.
• LSA is not a probabilistic variant of LDA.

Context Awareness

• Context acquisition can be explicit, implicit, or inferring.
• The user background describes general, rather static information about the user, e.g., knowledge or demographics.
• The user intent is not the purpose the content creator had in mind when creating the item.