This vignette is an introduction to the R package recometrics for evaluating recommender systems built with implicit-feedback data, assuming that the recommendation models are based on low-rank matrix factorization (example such packages: cmfrec, rsparse, recosystem, among many others), or assuming that it is possible to compute a user-item score as a dot product of user and item factors/components/attributes.
Historically, many models for recommender systems were designed by approaching the problem as regression or rating prediction, by taking as input a matrix \(\mathbf{X}_{ui}\) denoting user likes and dislikes of items in a scale (e.g. users giving a 1-to-5 star rating to different movies), and evaluating such models by seeing how well they predict these ratings on hold-out data.
In many cases, it is impossible or very expensive to obtain such data, but one has instead so called “implicit-feedback” records: that is, observed logs of user interactions with items (e.g. number of times that a user played each song in a music service), which do not signal dislikes in the same way as a 1-star rating would, but can still be used for building and evaluating recommender systems.
In the latter case, the problem is approached more as ranking or classification instead of regression, with the models being evaluated not by how well they perform at predicting ratings, but by how good they are at scoring the observed interactions higher than the non-observed interactions for each user, using metrics more typical of information retrieval.
Generating a ranked list of items for each user according to their
predicted score and comparing such lists against hold-out data can
nevertheless be very slow (might even be slower than fitting the model
itself), and this is where recometrics
comes in: it
provides efficient routines for calculating many implicit-feedback
recommendation quality metrics, which exploit multi-threading, SIMD
instructions, and efficient sorted search procedures.
The perhaps most common approach towards building a recommendation model is by trying to approximate the matrix \(\mathbf{X}_{mn}\) as the product of two lower-dimensional matrices \(\mathbf{A}_{mk}\) and \(\mathbf{B}_{nk}\) (with \(k \ll m\) and \(k \ll n\)), representing latent user and item factors/components, respectively (which are the model parameters to estimate) - i.e. \[ \mathbf{X} \approx \mathbf{A} \mathbf{B}^T \] In the explicit-feedback setting (e.g. movie ratings), this is typically done by trying to minimize squared errors with respect to the observed entries in \(\mathbf{X}\), while in implicit-feedback settings this is typically done by turning the \(\mathbf{X}\) matrix into a binary matrix which has a one if the observation is observed and a zero if not, using the actual values (e.g. number of times that a song was played) instead as weights for the positive entries, thereby looking at all entries rather than just the observed (non-zero) values - e.g.: \[ \min_{\mathbf{A}, \mathbf{B}} \sum_{u=1}^{m} \sum_{i=1}^{n} x_{ui} (I_{x_{ui}>0} - \mathbf{a}_u \cdot \mathbf{b}_i)^2 \]
The recommendations for a given user are then produced by calculating the full products between that user vector \(\mathbf{a}_u\) and the \(\mathbf{B}\) matrix, sorting these predicted scores in descending order.
For a better overview of implicit-feedback matrix factorization, see the paper Hu, Yifan, Yehuda Koren, and Chris Volinsky. “Collaborative filtering for implicit feedback datasets.” 2008 Eighth IEEE International Conference on Data Mining. Ieee, 2008.
Such matrix factorization models are commonly evaluated by setting aside a small amount of users as hold-out for evaluation, fitting a model to all the remaining users and items. Then, from the evaluation users, a fraction of their interactions data is set as a hold-out test set, while their latent factors are computed using the rest of the data and the previously fitted model from the other users.
Then, top-K recommendations for each user are produced, discarding the non-hold-out items with which their latent factors were just determined, and these top-K lists are compared against the hold-out test items, seeing how well they do at ranking them near the top vs. how they rank the remainder of the items.
This package can be used to calculate many recommendation quality metrics given the user and item factors and the train-test data split that was used, including:
P@K (“precision-at-k”): this is the most intuitive metric. It calculates the proportion of the top-K recommendations that include items from the test set for a given user - i.e. \[ P@K = \frac{1}{k} \sum_{i=1}^k \begin{cases} 1, & r_i \in \mathcal{T}\\ 0, & \text{otherwise} \end{cases} \] Where \(r_i\) is the item ranked at position \(i\) by the model (sorting the predicted scores in descending order, after excluding the items in the training data for that user), and \(\mathcal{T}\) is the set of items that are in the test set for that user.
Note that some papers and libraries define \(P@K\) differently, see the second version below.
TP@K (truncated \(P@K\)): same calculation as \(P@K\), but will instead divide by the minimum between \(k\) and the number of test items: \[ TP@K = \frac{1}{\min\{k, |\mathcal{T}|\}} \sum_{i=1}^k \begin{cases} 1, & r_i \in \mathcal{T}\\ 0, & \text{otherwise} \end{cases} \]
The “truncated” prefix is a non-standard nomenclature introduced here to differentiate it from the other \(P@K\) metric.
R@K (“recall-at-k”): while \(P@K\) offers an intuitive metric that captures what a recommender system aims at being good at, it does not capture the fact that, the more test items there are, the higher the chances that they will be included in the top-K recommendations. Recall instead looks at what proportion of the test items would have been retrieved with the top-K recommended list: \[ R@K = \frac{1}{|\mathcal{T}|} \sum_{i=1}^k \begin{cases} 1, & r_i \in \mathcal{T}\\ 0, & \text{otherwise} \end{cases} \]
AP@K (“average precision-at-k”): precision and recall look at all the items in the top-K equally, whereas one might want to take into account also the ranking within this top-K list, for which this metric comes in handy. “Average Precision” tries to reflect the precisions that would be obtained at different recalls: \[ AP@K = \frac{1}{|\mathcal{T}|} \sum_{i=1}^k \begin{cases} P@i, & r_i \in \mathcal{T}\\ 0, & \text{otherwise} \end{cases} \] \(AP@K\) is a metric which to some degree considers precision, recall, and rank within top-K. Intuitively, it tries to approximate the are under a precision-recall tradeoff curve. Its average across users is typically called “MAP@K” or “Mean Average Precision”.
Important: many authors define \(AP@K\) differently, such as dividing by the minimum between \(k\) and \(|\mathcal{T}|\) instead, or as the average for P@1..P@K (either as-is or stopping the calculation after already retrieving all test items). See below for the other version.
TAP@K (truncated \(AP@K\)): a truncated version of the \(AP@K\) metric, which will instead divide it by the minimum between \(k\) and the number of test items. Just like for \(TP@K\), the “truncated” prefix is a non-standard nomenclature used here to differentiate it from the other more typical \(AP@K\).
NDCG@K (“normalized discounted cumulative gain at k”): while the earlier metrics look at just the presence of an item in the test set, these items might not all be as good, with some of them having higher observed values than others. NDCG aims at judging these values, but discounted according to the rank in the top-K list. First it calculates the unstandardized discounted cumulative gain: \[ DCG@K = \sum_{i=1}^{k} \frac{C_{r_i}}{log_2 (1+i)} \] Where \(C_{r_i}\) indicates the observed interaction value in the test data for item \(r_i\), and is zero if the item was not in the test data. The DCG@K metric is then standardized by dividing it by the maximum achievable DCG@K for the test data: \[ NDCG@K = \frac{DCG@K}{\max DCG@K} \]
Unlike the other metrics, NDCG can handle data which contains “dislikes” in the form of negative values. If there are no negative values in the test data, it will be bounded between zero and one.
Hit@K (from which “Hit Rate” is calculated): this is a simpler yes/no metric that looks at whether any of the top-K recommended items were in the test set for a given user: \[ Hit@K = \max_{i=1..K} \begin{cases} 1, & r_i \in \mathcal{T}\\ 0, & \text{otherwise} \end{cases} \] The average of this metric across users is typically called “Hit Rate”.
RR@K (“reciprocal rank at k”, from which “MRR” or “mean reciprocal rank” is calculated): this metric only looks at the rank of the first recommended item that is in the test set, and outputs its inverse: \[ RR@K = \max_{i=1..K} \frac{1}{i} \:\:\:\text{s.t.}\:\:\: r_i \in \mathcal{T} \] The average of this metric across users is typically called “Mean Reciprocal Rank”.
ROC AUC (“area under the receiver-operating characteristic curve”): see the Wikipedia entry for details. While the metrics above only looked at the top-K recommended items, this metric looks at the full ranking of items instead, and produces a standardized number between zero and one in which 0.5 denotes random predictions.
PR AUC (“area under the precision-recall curve”): while ROC AUC provides an overview of the overall ranking, one is typically only interested in how well it retrieves test items within top ranks, and for this the area under the precision-recall curve can do a better job at judging rankings, albeit the metric itself is not standardized and its minimum does not go as low as zero.
The metric is calculated using the fast but not-so-precise rectangular method, whose formula corresponds to the AP@K metric with K=N. Some papers and libraries call this the average of this metric the “MAP” or “Mean Average Precision” instead (without the “@K”).
(For more details about the metrics, see the package
documentation: ?calc.reco.metrics
)
NOT covered by this package:
Metrics that look at the rareness of the items recommended (to evaluate so-called “serendipity”).
Metrics that look at “discoverability”.
Metrics that take into account the diversity of the ranked lists.
Now a practical example using the library cmfrec and the MovieLens100K data, taken from the recommenderlab package.
Note that this is an explicit-feedback dataset about movie ratings. Here it will be converted to implicit-feedback by setting movies with a rating of 4 and 5 stars as the positive (observed) data, while the others will be set as negative (unobserved).
This section will load the MovieLens100K data and filter out observations with a rating of less than 4 stars in order to have something that resembles implicit feedback.
library(Matrix)
library(MatrixExtra)
library(data.table)
library(kableExtra)
library(recommenderlab)
library(cmfrec)
library(recometrics)
data(MovieLense)
<- MovieLense@data
X_raw
### Converting it to implicit-feedback
<- as.coo.matrix(filterSparse(X_raw, function(x) x >= 4))
X_implicit str(X_implicit)
#> Formal class 'dgTMatrix' [package "Matrix"] with 6 slots
#> ..@ i : int [1:55024] 0 1 4 5 9 15 16 17 20 22 ...
#> ..@ j : int [1:55024] 0 0 0 0 0 0 0 0 0 0 ...
#> ..@ Dim : int [1:2] 943 1664
#> ..@ Dimnames:List of 2
#> .. ..$ : chr [1:943] "1" "2" "3" "4" ...
#> .. ..$ : chr [1:1664] "Toy Story (1995)" "GoldenEye (1995)" "Four Rooms (1995)" "Get Shorty (1995)" ...
#> ..@ x : num [1:55024] 5 4 4 4 4 5 4 5 5 5 ...
#> ..@ factors : list()
Now leaving aside a random sample of 100 users for model evaluation, for whom 30% of the data will be left as a hold-out test set.
<- create.reco.train.test(
reco_split
X_implicit,users_test_fraction = NULL,
max_test_users = 100,
items_test_fraction = 0.3,
seed = 123
)<- reco_split$X_train ## Train data for test users
X_train <- reco_split$X_test ## Test data for test users
X_test <- reco_split$X_rem ## Data to fit the model
X_rem <- reco_split$users_test ## IDs of the test users users_test
In order to determine if a personalized recommendation model is bringing value or not, it’s logical to compare such model against the simplest possible ways of making recommendations, such as:
This section creates such baselines to compare against.
### Random recommendations (random latent factors)
set.seed(123)
<- matrix(rnorm(nrow(X_test) * 5), nrow=5)
UserFactors_random <- matrix(rnorm(ncol(X_test) * 5), nrow=5)
ItemFactors_random
### Non-personalized recommendations
<- cmfrec::MostPopular(as.coo.matrix(X_rem), implicit=TRUE)
model_baseline <- model_baseline$matrices$item_bias item_biases
This section will fit a few models in order to have different ranked lists to evaluate:
All of these models are taken from the cmfrec
package -
see its documentation for more details about the models.
Important: for the explicit-feedback models, it’s not possible to use the same train-test split strategy as for the implicit-feedback variants, as the training data contains only 4 and 5 stars, which does not signal any dislikes and thus puts these models at a disadvantage. As such, here the user factors will be obtained from the full data (train+test), which gives them a quite unfair advantage compared to the other models.
In theory, one could also split the full ratings data, and filter out
low-star ratings in the test set only, but that would still distort a
bit the metrics for implicit-feedback models. Alternatively, one could
adjust the WRMF model to take low-star ratings as more negative entries
with higher weight (e.g. giving them a value of -1 and a weight of 5
minus rating), which is supported by e.g. cmfrec
. Note
however that the only metric in this package that can accomodate such a
scenatio (implicit feedback plus dislikes) is the \(NDCG@K\) metric.
### Typical implicit-feedback ALS model
### a.k.a. "WRMF" (weighted regularized matrix factorization)
<- cmfrec::CMF_implicit(as.coo.matrix(X_rem), k=10, verbose=FALSE)
model_wrmf <- t(cmfrec::factors(model_wrmf, X_train))
UserFactors_wrmf
### As a comparison, this is the typical explicit-feedback model,
### implemented by software such as Spark,
### and called "Weighted-Lambda-Regularized Matrix Factorization".
### Note that it determines the user factors using the train+test data.
<- cmfrec::CMF(as.coo.matrix(X_raw[-users_test, ]),
model_wlr lambda=0.1, scale_lam=TRUE,
user_bias=FALSE, item_bias=FALSE,
k=10, verbose=FALSE)
<- t(cmfrec::factors(model_wlr, as.csr.matrix(X_raw)[users_test,]))
UserFactors_wlr
### This is a different explicit-feedback model which
### uses the same regularization for each user and item
### (as opposed to the "weighted-lambda" model) and which
### adds "implicit features", which are a binarized version
### of the input data, but without weights.
### Note that it determines the user factors using the train+test data.
<- cmfrec::CMF(as.coo.matrix(X_raw[-users_test, ]),
model_hybrid lambda=20, scale_lam=FALSE,
user_bias=FALSE, item_bias=FALSE,
add_implicit_features=TRUE,
k=10, verbose=FALSE)
<- t(cmfrec::factors(model_hybrid, as.csr.matrix(X_raw)[users_test,])) UserFactors_hybrid
The MovieLens100K data used here comes with metadata/attributes about the users (gender, occupation, age, among others) and the items (genre and year of release), which so far have not been incorporated into these models.
One simple way of adding this secondary information into the same
WRMF model is through the concept of “Collective Matrix Factorization”,
which does so by also factorizing the side information matrices, but
using the same user/item factors - see the documentation of
cmfrec
for more details about this approach.
As well, one typical trick in the explicit-feedback variant is to add a fixed bias/intercept for each user and item, which is also possible to do in the WRMF model by making some slight modifications to the optimization procedure.
This section will fit additional variations of the WRMF model to compare against:
cmfrec
does not provide this option for implicit-feedback
models, but it offers a lot of flexibility in what kind of objective to
optimize in its main function (CMF
), which can mimic a
variety of models (e.g. the “Weighted-Lambda” or the “WRMF”) depending
on the parameters - here, the model will be fit by manually re-creating
the WRMF variant (missing-as-zero, binarized matrix, positive entries
weighted by the actual values plus one as in the original paper), which
in turn requires manually creating the binarized X
matrix
and the weights.First processing the data as required for the new models:
### Processing user side information
<- as.data.table(MovieLenseUser)[-users_test, ]
U <- mean(U$age)
mean_age <- sd(U$age)
sd_age <- levels(U$occupation)
levels_occ ::restore_old_matrix_behavior()
MatrixExtra<- function(U, mean_age,sd_age, levels_occ) {
process.U `:=`(
U[, id = NULL,
age = (age - mean_age) / sd_age,
sex = as.numeric(sex == "M"),
occupation = factor(occupation, levels_occ),
zipcode = NULL
)]<- Matrix::sparse.model.matrix(~.-1, data=U)
U <- as.coo.matrix(U)
U return(U)
}<- process.U(U, mean_age,sd_age, levels_occ)
U <- as.data.table(MovieLenseUser)[users_test, ]
U_train <- process.U(U_train, mean_age,sd_age, levels_occ)
U_train
### Processing item side information
<- as.data.table(MovieLenseMeta)
I <- mean(I$year, na.rm=TRUE)
mean_year <- sd(I$year, na.rm=TRUE)
sd_year
I[is.na(year), year := mean_year
`:=`(
][, title = NULL,
year = (year - mean_year) / sd_year,
url = NULL
)]<- as.coo.matrix(I)
I
### Manually re-creating a binarized matrix and weights
### that will mimic the WRMF model
<- as.coo.matrix(mapSparse(X_rem, function(x) rep(1, length(x))))
X_rem_ones <- as.coo.matrix(mapSparse(X_rem, function(x) x+1))
W_rem <- as.coo.matrix(mapSparse(X_train, function(x) rep(1, length(x))))
X_train_ones <- as.coo.matrix(mapSparse(X_train, function(x) x+1)) W_train
Now fitting the models:
### WRMF model, but with item biases/intercepts
<- cmfrec::CMF(X_rem_ones, weight=W_rem, NA_as_zero=TRUE,
model_bwrmf lambda=1, scale_lam=FALSE,
center=FALSE, user_bias=FALSE, item_bias=TRUE,
k=10, verbose=FALSE)
<- t(cmfrec::factors(model_bwrmf, X_train_ones, weight=W_train))
UserFactors_bwrmf
### Collective WRMF model (taking user and item attributes)
<- cmfrec::CMF_implicit(as.coo.matrix(X_rem), U=U, I=I,
model_cwrmf NA_as_zero_user=TRUE, NA_as_zero_item=TRUE,
center_U=TRUE, center_I=TRUE,
lambda=0.1,
k=10, verbose=FALSE)
<- t(cmfrec::factors(model_cwrmf, X_train, U=U_train))
UserFactors_cwrmf
### Collective WRMF plus item biases/intercepts
<- cmfrec::CMF(X_rem_ones, weight=W_rem, NA_as_zero=TRUE,
model_bcwrmf U=U, I=I, center_U=FALSE, center_I=FALSE,
NA_as_zero_user=TRUE, NA_as_zero_item=TRUE,
lambda=0.1, scale_lam=FALSE,
center=FALSE, user_bias=FALSE, item_bias=TRUE,
k=10, verbose=FALSE)
<- t(cmfrec::factors(model_bcwrmf, X_train_ones,
UserFactors_bcwrmf weight=W_train, U=U_train))
Finally, calculating recommendation quality metrics for all these models:
<- 5 ## Top-K recommendations to evaluate
k
### Baselines
<- calc.reco.metrics(X_train, X_test,
metrics_random A=UserFactors_random,
B=ItemFactors_random,
k=k, all_metrics=TRUE)
<- calc.reco.metrics(X_train, X_test,
metrics_baseline A=NULL, B=NULL,
item_biases=item_biases,
k=k, all_metrics=TRUE)
### Simple models
<- calc.reco.metrics(X_train, X_test,
metrics_wrmf A=UserFactors_wrmf,
B=model_wrmf$matrices$B,
k=k, all_metrics=TRUE)
<- calc.reco.metrics(X_train, X_test,
metrics_wlr A=UserFactors_wlr,
B=model_wlr$matrices$B,
k=k, all_metrics=TRUE)
<- calc.reco.metrics(X_train, X_test,
metrics_hybrid A=UserFactors_hybrid,
B=model_hybrid$matrices$B,
k=k, all_metrics=TRUE)
### More complex models
<- calc.reco.metrics(X_train, X_test,
metrics_bwrmf A=UserFactors_bwrmf,
B=model_bwrmf$matrices$B,
item_biases=model_bwrmf$matrices$item_bias,
k=k, all_metrics=TRUE)
<- calc.reco.metrics(X_train, X_test,
metrics_cwrmf A=UserFactors_cwrmf,
B=model_cwrmf$matrices$B,
k=k, all_metrics=TRUE)
<- calc.reco.metrics(X_train, X_test,
metrics_bcwrmf A=UserFactors_bcwrmf,
B=model_bcwrmf$matrices$B,
item_biases=model_bcwrmf$matrices$item_bias,
k=k, all_metrics=TRUE)
These metrics are by default returned as a data frame, with each user representing a row and each metric a column - example:
%>%
metrics_baseline head(5) %>%
kable() %>%
kable_styling()
p_at_5 | tp_at_5 | r_at_5 | ap_at_5 | tap_at_5 | ndcg_at_5 | hit_at_5 | rr_at_5 | roc_auc | pr_auc | |
---|---|---|---|---|---|---|---|---|---|---|
31 | 0 | 0 | 0.0000000 | 0.0000000 | 0 | 0.000000 | 0 | 0 | 0.8600427 | 0.0203154 |
33 | 0 | 0 | 0.0000000 | 0.0000000 | 0 | 0.000000 | 0 | 0 | 0.8000000 | 0.0200586 |
38 | 0 | 0 | 0.0000000 | 0.0000000 | 0 | 0.000000 | 0 | 0 | 0.7723270 | 0.0684697 |
43 | 1 | 1 | 0.1136364 | 0.1136364 | 1 | 0.826241 | 1 | 1 | 0.9383715 | 0.5255828 |
61 | 0 | 0 | 0.0000000 | 0.0000000 | 0 | 0.000000 | 0 | 0 | 0.8702474 | 0.0067806 |
In order to compare models, one can instead summarize these metrics across users:
<- list(
all_metrics `Random` = metrics_random,
`Non-personalized` = metrics_baseline,
`Weighted-Lambda` = metrics_wlr,
`Hybrid-Explicit` = metrics_hybrid,
`WRMF (a.k.a. iALS)` = metrics_wrmf,
`bWRMF` = metrics_bwrmf,
`CWRMF` = metrics_cwrmf,
`bCWRMF` = metrics_bcwrmf
)<- all_metrics %>%
results lapply(function(df) as.data.table(df)[, lapply(.SD, mean)]) %>%
::rbindlist() %>%
data.tableas.data.frame()
row.names(results) <- names(all_metrics)
%>%
results kable() %>%
kable_styling()
p_at_5 | tp_at_5 | r_at_5 | ap_at_5 | tap_at_5 | ndcg_at_5 | hit_at_5 | rr_at_5 | roc_auc | pr_auc | |
---|---|---|---|---|---|---|---|---|---|---|
Random | 0.012 | 0.0120000 | 0.0024653 | 0.0010162 | 0.0057333 | 0.0110063 | 0.06 | 0.0286667 | 0.4886724 | 0.0159867 |
Non-personalized | 0.198 | 0.1985000 | 0.0538081 | 0.0363141 | 0.1458167 | 0.2026604 | 0.53 | 0.3813333 | 0.8780423 | 0.1264918 |
Weighted-Lambda | 0.036 | 0.0360000 | 0.0080785 | 0.0043460 | 0.0214333 | 0.0366013 | 0.13 | 0.0748333 | 0.7536814 | 0.0522353 |
Hybrid-Explicit | 0.268 | 0.2711667 | 0.0750656 | 0.0508290 | 0.2045333 | 0.2808849 | 0.63 | 0.4601667 | 0.7390052 | 0.1527414 |
WRMF (a.k.a. iALS) | 0.352 | 0.3666667 | 0.1325404 | 0.0975948 | 0.2885889 | 0.3661113 | 0.75 | 0.5573333 | 0.9412953 | 0.2623818 |
bWRMF | 0.342 | 0.3548333 | 0.1194934 | 0.0886808 | 0.2812833 | 0.3529575 | 0.71 | 0.5381667 | 0.9393616 | 0.2548754 |
CWRMF | 0.350 | 0.3646667 | 0.1309673 | 0.0938263 | 0.2875250 | 0.3639270 | 0.77 | 0.5663333 | 0.9423962 | 0.2560516 |
bCWRMF | 0.352 | 0.3653333 | 0.1225364 | 0.0897751 | 0.2847167 | 0.3577057 | 0.73 | 0.5335000 | 0.9409499 | 0.2524747 |
From these metrics, the best-performing model overall seems to be CWRMF (collective version of WRMF or iALS model, which incorporates side information about users and items), but it does not dominate across all metrics.
It is hard to conclude for example whether adding item biases to the WRMF or the CWRMF model is an improvement, as some metrics improve while others deteriorate, and this is where specific properties about the dataset and the desired recommendation goals have to come in mind (e.g. one might decide that \(AP@K\) is simply the most informative metric and make a decision based on it, or perhaps look at more specialized metrics).
To keep in mind: