All provided recommenders are subclasses of PredictRecommender, which uses rating predictions to make recommendations. It supports both deterministic and stochastic item selection policies.
A recommender that returns a random item from the list of unconsumed items foreach online user. It is a useful baseline for calibrating lower-bounds of recommender performance.
A recommender that uses historical ratings to make global rankings of items and recommend items based on items with highest overall popularity. This is a useful baseline for measuring the benefits of personalization.
A recommender that is instantiated with a dense_rating_function, which provides the true ratings of the users for all items. It is a useful baseline for calibrating upper-bounds of recommender performance.
KNNRecommender neighborhood based collaborative filtering algorithm. The class supports both user and item based collaborative filtering. In an user_based KNN recommender, user features are stacked and pairwise similarity metrics between users are measured. An online user is thus recommended an item that was highly rated by a similar user. Conversely in an item_based KNN recommender, item features are stacked and pairwise similarity metrics between items are measured. An online user is thus recommended an item that is highly similar to other highly rated items of the user.
It is wrapper for the LibFM recommender. See www.libfm.org for implementation details. We built a pip installable python package ypyfm based on this C++ implementation, that might be of interest in it's own right.
At each step of the simulation the LibFM recommender re-trains a matrix factorization model. It computes rating predictions as the inner product of user and item factors plus bias terms.
AutoRec is an autoencoder framework for collaborative filtering proposed by this paper. It can be seen as a non-linear generalization of factorization models. We adapted a publicly available implementation
Cfnade is neural autoregressive architecture for collaborative filtering proposed by this paper. We adapted a publicly available implementation.
Llorma is a generalization of low rank matrix factorization techniques based on this (paper)[http://jmlr.org/papers/v17/14-301.html]. The LLORMA algorithm approximates the observed rating matrix as a weighted sum of low-rank matrices which are limited to a local region of the observed matrix. We adapted a publicly available implementation.
SLIM is a sparse linear recommendation model based on this paper. For an user i it models the predicted rating of an unseen item i as a weighted average of the ratings of items previously rated by user u.
EASE predicts ratings bases on item-item similarity model based on this paper. Assuming that the historical data contains N users and M items in a NxM rating matrix X. The model computes a MxM self-similarity matrix B. Unseen ratings are predicted as XB.