We implemented the well-known matrix factorization algorithm as proposed by Simon Funk.
Matrix factorization methods are used in recommender systems to derive a set of latent factors, from the original user-item rating matrix, to characterize both users and items by these vectors of user and item factors. The user-item interaction is modeled as the inner product of the latent factors space. Accordingly each item i will be associated with a vector factor V_i, and each user u is associated with a vector factors U_u. An approximation of the rating of a user \(u\) on an item \(i\) can be derived as the inner product of their item and user factor vectors.
The rrecsys package utilizes a stochastic gradient descent optimization technique for computing the item and user factors. The U(user) and V(item) factor matrices are cropped to k features. Each feature is trained until convergence.
For the Rating Prediction task, to train a model with this algorithm, it is required to define an additional argument, k the number of user/item factors.
data("ml100k") d <- defineData(ml100k) e <- evalModel(d, folds = 2) mf_model <- evalPred(e, "funk", k = 10, steps = 100, regCoef = 0.0001, learningRate = 0.001, biases = F) mf_model
For the Item Recommendation task, to provide item recommendations, it is required to define two additional arguments, positiveThreshold the threshold for “positive” ratings, and the topN the number of recommended items.
data("ml100k") d <- defineData(ml100k) e <- evalModel(d, folds = 2) mf <- evalRec(e, "funk", k = 10, steps = 100, regCoef = 0.0001, learningRate = 0.001, biases = F, positiveThreshold = 3, topN = 3) mf
The k default value is 10. The positiveThreshold default value is 3. The topN default value is 3. The learningRate default value is 0.001. The regCoef default value is 0.0001. The steps default value is 10.