Summary of Recommenders

Two Basic Properties of Ratings

1. Type of Feedback (i.e. implicit vs explicit)

   • Explicit Ratings: a metric explicitly defined by the user, which measures customer affinity
   • Implicit Rating: a metric we use as a proxy for an explicit rating(i.e. a proxy for an actual affinity score)

2. Sparsity

   • A matrix of ratings is always sparse because users only interact with a small % of items
   • Meaning, there is a disproportionally large number of known ratings corresponding the most popular items

Two Basic Properties of Recommenders

1. Level of personalization

2. Usage of contextual information

   • Simple content-based and collaborative filtering models are categorized as personalized, non-contextual models

Two Basic Types of Recommenders

1. Content-Based Filtering Models

   • A model using item features (e.g. age, gender, etc) to approximate ratings

2. Collaborative Filtering Models

   • A model using a rating matrix to approximate ratings

User- and Item-Based Neighborhood Methods

• Suppose we have a rating matrix for Spotify of listeners and songs

• An item-based recommender would recommend songs based on other similar songs

  • If I enjoy country and jazz, I may be recommended a song that is similar to other jazz songs I’ve enjoyed
  • Or, I may be recommended a song that is similar to other country songs I’ve enjoyed

• A user-based recommender would recommend songs based on other similar listeners

  • If I enjoy country and jazz, I may be recommended a song enjoyed by other listeners who enjoy country and jazz

• User-based and item-based are structurally similar, where the only difference is in the nature of the similarity measure

  • As a result, we should expect the recommendations to be somewhat similar between the two over time
  • Item-based can have some benefit in terms of scale (Amazon invented it for this reason), since they can return similar recommendations over time
  • To find the right one for the problem at hand involves experimentation

Illustrating Latent Factor Models

• Suppose we have a rating matrix of customers and movies
• A latent factor model (with number of factors learned = 2) could create factors for whether a movie is a blockbuster or not, and whether the movie is designated for children or not

Properties of Content-Based Filtering Recommenders

• User independence (advantage):

  • Can make decent recommendations for new users
  • This is because content-based models don’t learn about user-specific behavior
  • Instead, they learn about feature-specific behavior (using basic classification/regression models)

• More feature engineering (disadvantage):

  • Most content-based features require more feature engineering compared to only retrieving ratings for a rating matrix (for collaborative filtering)

• Trivial recommendations (disadvantage):

  • Can often produce trivial recommendations without any novelty
  • Since, content-based models converge towards the average customer

Properties of Collaborative Filtering Recommenders

• Less feature engineering (advantage):

  • Capable of making recommendations based only on patterns/similarities available in the rating matrix

• Involves transfer learning across users and items (advantage):

  • If most customers who like hamburgers also like hot dogs, we can confidently recommend hot dogs to a customer who’s only ever interacted with hamburgers (and rated them highly)

• Sparsity (disadvantage):

  • Difficult to build reliable rating predictions if the rating matrix is too sparse

• Cold-Start Problem (disadvantage)

  • Unable to handle new users or items

Tradeoff between Sparsity and Collaboration

• In general, there is a tradeoff between mitigating sparsity and producing collaborative recommendations when switching between content-based method and collaborative methods

• Content-based methods address issues with sparsity and the cold-start problem

  • They do this by representing items through their meta-features
  • As these are known in advance, recommendations can be computed even for new items for which no collaborative data has been gathered

• Unfortunately, no transfer learning occurs in content-based models

  • Whereas, transfer learning occurs in collaborative models
  • Instead, content-based models for each user are estimated in isolation and do not benefit from data on other users

• Consequently, content-based models perform worse than matrix-factorization models where collaborative information is available and require a large amount of data on each user

  • Rendering them unsuitable for user cold-start

References

• Textbook about Algorithmic Marketing
• Paper about Hybrid Contextual Model named LightFM