Discussion Notes

Jan 22

Horting Hatches an Egg: A New Graph-Theoretic Approach to Collaborative Filtering

• focusing on accuracy and learning curve - the Achilles heel of collaborative filtering
• directed graph model: nodes - users, directed edges - predictability
• introduces the idea of making an indirect recommendation
• concepts of horting and predicting are not transitive in that a direct hammock is not implied
• recommendations are not symmetric - ala Siteseer
• evaluation is tricky
• takes care of the effusivity and negation aspects of making recommendation (e.g. making use of left shifting)
• items presented for evaluation to users are partitioned into a "hot set" - to increase commonality in order to recommend better; and a "cold set" - to increase coverage
• incorporates a hierarchical classification and creative links which violate that hierarchical classification
• one of several techniques of the IRA, situation analyzer
• tested against mythical e-commerce site / artificial data

Empirical Analysis of Predictive Algorithms for Collaborative Filtering

• second to Resnick and Varian as the most widely referenced paper in recommender systems - an indication of the prematurity of the field
• a purely statistical model (correlation, Pearson's r, vector similarity, inverse user frequency)
• use default votes to handle sparsity
• an extreme end of the spectrum of methods of evaluating a recommender system
• focused on accuracy of predictions, not efficiency
• distinguishes between memory-based CF (lazy learning, non-parametric) and model-based CF (eager learning, parametric) - involves a lot of preprocessing; incrementality becomes an issue - almost impossible with a neural network or Bayesian network
• item by item recommendations versus ranked lists
• make use of protocols developed for other domains
• makes vivid a very important aspect of recommender systems, that is, human satisfaction cannot be replaced with functions
• do not address incrementality and do not capture the underlying social process
• the other end of the spectrum is going out and doing satisfaction surveys, studies, etc. (HCI) - purely social
• Is there something in between these two extremes that can make CS folks happy???

Eigentaste: A Constant Time Collaborative Filtering Algorithm

• accurate and efficient recommendations to users in constant time
• of course, not really constant time, because of all the work they do up front
• distinguishes between so-called universal queries and user-selected queries
• Universal queries (i.e. every user rates n number of items, called the "gauge" set - arbitrarily chosen?) yield a dense matrix, a solution to sparsity.
• validated their work in the domain of jokes - jester - a domain in which there is minimal variation in thoughts between the different people who are providing evaluations
• not dynamic
• evaluation done once again via a function
• function could be a black box (e.g. a neural network)
• function misses other roles
• One has to realize that in a two moded model (e.g. people and movies), the second mode (e.g. movies) brings the first mode together.
• concept of "affiliation" network -
  • people - primary
  • movies, books, cds, etc. - secondary
• While functional mapping (using a training set and a test set) as means of an evaluation technique is the most widely accepted and easiest in the field of recommender systems, it is also the most shallow.
• Bottom line is that there is a social element to recommender systems