MRec is an academic project. It targets for item recommendation from implicit feedback and rating prediction based recommendation. It supports cross validation and holdout evaluation. It includes several important algorithms tailored for implicit feedback, such as WRMF[1], content-aware collaborative filtering for implicit feedback[3], GWMF (graph regularized weighted matrix factorization). Also it includes some state-of-the-art algorithm for location recommendation, such as IRenMF[4], GeoMF[2]. In the future, we will further implement algorithms including Bayesian personalized ranking based matrix factorization, non-negative matrix factorization, Poisson Matrix Factorization.
The dataset are assumed to be saved in a format file, including columns of users, items and scores (count|rating).
u1 i1 5
u2 i2 4If dataset has been split into train.txt and test.txt, you can load data via
[train, test] = readData('path/to/file', 1); # 1 indicating indexes of users and items start with zeroIf dataset is not split, we can load data via
data = readContent('path/to/file'); # zero_start is defaultafter that, you can manually split data into train and test
[train, test] = split_matrix(data, 'un', 0.8); # mode='un' and ratio=0.8Ratio corresponds to the ratio of training set. There are five types of matrix split methods, by specified by 'mode'.
* 'un' (default) user-oriented split, splitting each user data into 'folds' folds.
* 'in' item-oriented split,
* 'en' entry-oriented split.
* 'u' splits users into 'folds' folds, which is used for cold-start user evaluation
* 'i' splits items into 'folds' folds, which is used for cold-start item evaluation.The main portal is item_recommend for implicit feedback and rating_recommend for explicit feedback. It is very simple to use. The following two parameters are mandatory to these two portals
See test/dataset/test_script.m for some examples of how to use this portal
Three types of running schemes are available, including cross validation and holdout evaluation.
The following command specifies cross validation, where folds means the number of folds. And it returns summary which include recall, precision, ndcg, map, auc, mpr (mean percentile rank), and 'elapsed'. 'elapsed' records training and testing time. There are two rows in summary, corresponding to mean and std of cross validation. For example, summary.item_recall(1,20) is the averaged recall@20 while summary.item_recall(2,20) is the standard deviation of cross validation.
[summary, detail, elapsed] = item_recommend(@(mat) iccf(mat, 'K', 50, 'max_iter', 20), data, 'folds', 5); which can be simplified as
[summary, detail, elapsed] = item_recommend(@iccf, data, 'folds', 5, 'K', 50, 'max_iter', 20);The following command specifies holdout evaluation, specified by the ratio of the testing set. Note that it is possible to specify the ratio of the training set by train_ratio (default=1), but it means that in can not be used separately. In other words, it should always be used together with test_ratio, since it means it will get the training set from the remaining part after excluding the test set. Besides, it should be accompanied by times option, indicating how many times are need for such holdout evaluation with specified test_ratio.
item_recommend(@(mat) iccf(mat, 'K', 50, 'max_iter', 20), data, 'test_ratio', 0.2, 'times', 5);There are also five types of matrix split methods, specified by split_mode, the same as fold_mode.
If you have both training set and test set, using the following commands for training and testing.
item_recommend(@(mat) iccf(mat, 'alpha', 30, 'K', 50, 'max_iter', 20), train, 'test', test);If only dataset is provided, it will return topk recommendation results for each user, where topk is a parameter with default value 200.
Besides, it provides some easy-to-use utility function, including hyperp_search for searching parameters, readContent to read matrix from tuple-based dataset file/ feature file. You can refer script test/dataset/test_script.m for detailed usage.
Here I would like to emphasis the iccf algorithm[3], since it subsumes weighted regularized matrix factorization, and enables taking any feature from both user side and item side into account. Below I will introduce the usage of this algorithm. A lot of options, including dimension K, maximum number of iteration max_iter, the weight of positive preference alpha, could be specified in this algorithm. In order to specify features from user side and item side, it can specify X as user side feature matrix and Y as item size feature matrix. Some examples are listed below
item_recommend(@(mat) iccf(mat, 'K', 50, 'max_iter', 20, 'alpha', 30), data, 'test_ratio', 0.2, 'times', 5);Note that in this case zeros(size(data,1), 0) is the default value of X and zeros(size(data, 2), 0) is the default value of Y, and this is equivalent to WRMF[1].
When item/user features are fed, reg_u and reg_i is used to specify the weight/importance of these features.
item_recommend(@(mat) iccf(mat, 'K', 50, 'max_iter', 20, 'alpha', 30, 'Y', item_featue_matrix, 'reg_i', 100), data, 'test_ratio', 0.2, 'times', 5); It is worth mentioning that piccf is a parallel version of iccf. You can specify the same options as iccf.
When using the iccf algorithm for recommendation from implicit feedback datasets, alpha, reg_u and reg_i are three important parameters affecting the recommendation performance. And these parameters may be re-tuned when changing the dimension of latent space.