We propose an alternative way to efficiently exploit rating data for collaborative filtering with Factorization Machines (FMs). Our approach partitions user-item matrix into 'slices' which are mutually exclusive with respect to items. The training phase makes direct use of the slice of interest (target slice), while incorporating information from other slices indirectly. FMs represent user-item interactions as feature vectors, and they offer the advantage of easy incorporation of complementary information. We exploit this advantage to integrate information from other auxiliary slices. We demonstrate, using experiments on two benchmark datasets, that improved performance can be achieved, while the time complexity of training can be reduced significantly.
In this paper, we investigate the idea that the ‘right’ data, rather than all data, should be exploited to build an efficient recommender system. We introduce an approach called ‘Slice and Train’ that trains a model on the right slice of a dataset (a ‘sensible’ subset containing the current items that need to be recommended) and exploits the information in the rest of the dataset indirectly. This approach is particularly interesting in scenarios in which recommendations are needed only for a particular subset of the overall item set. An example is an e-commerce website that does not generate recommendations for out-of-season or discontinued products. The obvious advantage of this approach is that models are trained on a much smaller dataset (only the data in the slice), leading to shorter training times.
The ‘Slice and Train’ approach also has, however, a less expected advantage, namely, that it offers highly competitive performance with conventional approaches that make use of the whole dataset, and in some cases even improves the performance. This advantage means that it is helpful to apply ‘Slice and Train’ even in cases in which predictions are needed for all items in the dataset, by training a series of separate models, one for each slice.
Our approach consists of two steps: first the data is partitioned
into a set of slices containing mutually exclusive items. The slices
can be formed by grouping items based on their properties (e.g., the
category of item), availability, item), or by more advanced slicing
methods such as clustering. In the second step the model is trained
using the samples in the slice of interest, i.e., target slice, while
other slices are indirectly being exploited as auxiliary slices. To
efficiently exploit information from auxiliary slices our approach
trains the recommender model using Factorization Machines [
The ‘Slice and Train’ method is implemented with
Factorization Machines [
The ‘Slice and Train’ method creates feature vectors x only for the ratings in the target slice (i.e., the slice of interest). The rating information from the other slices is exploited indirectly by extending the feature vectors that are created for the target slice. Using this method, the accuracy of the recommender is preserved, while the training time is significantly reduced, since the number of samples in the target slice is lower than in the original dataset.
Figure 1 shows how the feature vectors in the ‘Slice and Train’ method are constructed. The feature vectors have a binary part, which reflects the corresponding user and item of a rating, and a real-valued auxiliary part, which is constructed by using the rating information of auxiliary slices.
To understand how the auxiliary features are built, assume that the dataset is divided into m slices fS1; : : : ; Smg. Let us also assume that the items that are rated by user u in slice Sj is represented by sj (u). By extending the feature vector x(u; i) with auxiliary features, we can represent it with the following sparse representation: x(u; i) = f(u; 1); (i; 1); z2(u); : : : ; zm(u)g
| targe{tzslice } | auxilia{rzy slices } where zj (u) is sparse representation of auxiliary features from slice j and is defined as:
zj (u) = f(l; ϕj (u; l)) : l 2 sj (u)g where ϕj (u; l) is a normalization function that defines the value of the auxiliary features. We define ϕj based on the ratings that user u gave to items in slice j and normalize it based on the average value of user ratings as follows: ϕj (u; l) = rj (u; l) rmax r(u) rmin where rj (u; l) indicates the rating of user u to item l in slice j, r(u) indicates the average value of user ratings, and rmax and rmin indicate the maximum and minimum possible values for rating items.
In this work we tested our method on two benchmark datasets
of MovieLens 1M dataset1 and Amazon reviews dataset [
In order to test our approach, the data in target slices are divided
into 75% training and 25% test data. For every experiment 10%
of training data is only used as validation data to tune the
hyperparameters. Factorization Machines can be trained using three
different learning methods [
The experiments are implemented using WrapRec [
Table 1 presents the performance of our sliced training method compared with the situation that the no slicing is done. The experiments are evaluated using RMSE and MAE evaluation metrics. The reported results are the averaged metrics over all slices when each of them is considered as the target slice. The four different setups that are compared are the followings:
FM-ALL: FM applied on the complete dataset.
MF-ALL: MF applied on the complete dataset.
FM-SLICE: FM applied on independent slices. No auxiliary features are used for this setup. (1) (2) (3) FM-SLICE-AUX: FM applied to slices with feature vectors extended by auxiliary features derived form auxiliary slices.
The first two lines of Table 1 report results when all the data are used. The effectiveness of FM as our model is confirmed when we compare FM-ALL baseline with MF-ALL, the Matrix-Factorizationbased method. Comparing the remaining lines yields to some interesting insights. Not surprisingly, when we train the model only on a target slice without using any auxiliary information (FM-SLICE) the performance of the model drops. This drop can be attributed to the smaller number of samples that is being used when training. However, when the auxiliary feature are exploited with the ‘Slice and Train’ method (FM-SLICE-AUX) the performance becomes even better than the situation where the complete data is being used (FM-ALL). We attribute this improvement to the items in the slices being more homogeneous than the dataset as a whole. However, relying on homogeneity alone does not lead to the best performance (FM-SLICE). Instead the best performance is achieved when slicing is combined with auxiliary features. Furthermore, the timecomplexity of training is significantly reduced when the model is trained on a target slice. In our setup the training time of one slice including auxiliary features compared to the training time of complete dataset falls from 7431 ms to 1605 ms for Amazon dataset and from 14569 ms to 6574 ms for the MovieLens dataset. 4.
In this paper, we presented a brief overview of the ‘Slice and Train’ method, which trains a recommender model on a sensible subset of items (target slice) using Factorization Machines, and exploits other information indirectly. This sort of targeted training yields improvements in both performance and complexity.
This work is supported by funding from EU FP7 project under grant agreements no. 610594 (CrowdRec)