hongooi73/sar

Smart adaptive recommendations

SAR: Smart Adaptive Recommendations

SAR is a practical, rating-free collaborative filtering algorithm for recommendations. It produces explainable results, and is usable on a wide range of problems.

This package provides the following:

• An R interface to the Azure Product Recommendations service, a cloud implementation of SAR. It includes the ability to deploy the backend via the AzureRMR package, as well as a client frontend.

• A standalone R implementation of SAR, for ease of experimentation and familiarisation. The core algorithm is written in C++ and makes use of multithreading and sparse matrices for speed and efficiency.

More information

A detailed description of SAR

Other SAR implementations:

• Python: Microsoft/Recommenders
• Spark: Azure/mmlspark

sar's Issues

cold item modeling does not modify produced similarity matrix as desired

specifying appropriate arguments for catalog_data and catalog_formula have no impact on the returned value of sar(). The issue is an assignment in the main sar.default(). Fix inc.

To reproduce, use the following steps:

data(ms_usage)
data(ms_catalog)

## create some feature variables
ms_catalog$ms <- grepl("microsoft", ms_catalog$name, ignore.case=TRUE)
ms_catalog$surf <- grepl("surface", ms_catalog$name, ignore.case=TRUE)

## create several different SAR models
mod0 <- sar(ms_usage, support_threshold=25)
mod1 <- sar(ms_usage, support_threshold=25, catalog_data=ms_catalog, catalog_formula=f, cold_item_model=NULL)
mod2 <- sar(ms_usage, support_threshold=25, catalog_data=ms_catalog, catalog_formula=f, cold_item_model=NULL,
            cold_to_cold=TRUE)
mod3 <- sar(ms_usage, support_threshold=25, catalog_data=ms_catalog, catalog_formula=f, cold_item_model="lm")
mod4 <- sar(ms_usage, support_threshold=25, catalog_data=ms_catalog, catalog_formula=f, cold_item_model="lm",
            cold_to_cold=TRUE)

## all the similarity matrices are the same
all(mod0$sim_mat == mod1$sim_mat)
all(mod0$sim_mat == mod2$sim_mat)
all(mod0$sim_mat == mod3$sim_mat)
all(mod0$sim_mat == mod4$sim_mat)

## do one comparison by hand:
sim_onlywarm = SAR:::make_similarity(ms_usage$user, factor(ms_usage$item), ms_usage$time, 25, by_user=TRUE, similarity='jaccard')
## augment with cold items to mirror model1
f <- reformulate(names(ms_catalog)[-(1:2)])
sim_withcold1 <- SAR:::get_cold_similarity(cold_item_model=NULL, sim_matrix=sim_onlywarm, catalog_formula=f, catalog_data=ms_catalog, cold_to_cold=FALSE,
                                          similarity='jaccard')
## this is still TRUE
all(mod1$sim_mat == sim_onlywarm)
## it SHOULD be FALSE:
all(sim_withcold1 == sim_onlywarm)