Music listening can be regarded as a social activity, in which people can listen together and make friends with one other. Therefore, social relationships may imply multiple facets of the users, such as their listening behaviors and tastes. In this light, it is considered that social relationships hold abundant valuable information that can be utilized for music recommendation. However, utilizing the information for recommendation could be di cult, because such information is usually sparse. To address this issue, we propose to learn the latent social listening representations by the DeepWalk method, and then integrate the learned representations into Factorization Machines to construct better recommendation models. With the DeepWalk method, user social relationships can be transformed from the sparse and independent and identically distributed (i.i.d.) form into a dense and noni.i.d. form. In addition, the latent representations can also capture the spatial locality among users and items, therefore bene ting the constructed recommendation models.
Underlying almost all recommendation algorithms is an
attempt to model the interaction among users and items.
There have been some studies working on utilizing auxiliary
information for improving recommendations. In [
Figure 1 illustrates the framework of the proposed method.
We transfer the listening history and friends relationship into
a social listening graph. Given the social listening graph,
DeepWalk is used to learn the implicit representation by
seeking the possible path on the graph. The idea is to maximize
the co-occurrence patterns in each generated path so that the
potential distance is modeled. The representation, usually
presented as a vector, can encode the context information
for further processing [
To nd the possible path, we utilize random walk to uniformly sample a series of random vertex from a graph. The primitive graph is constructed on the sole social network. For the speci c application of music recommendation, we propose 3 di erent ways to construct such a social graph: Social graph: Build the graph only based on users' friends, which will enable DeepWalk to detect local community. Listening graph: Build the graph only based on user-item listening matrix, which will enable DeepWalk to identify the association patterns about users and items.
Social Listening graph: Build the graph based on the above two relations including the user-user and user-item matrices, which may hopefully fuse the merits of the preceding two approaches. 3.
Our experiments involve two real-world music datasets { hetrec2011-lastfm-2k and KKBOX-50K. The rst one is a public benchmark dataset derived Last.fm. The second one is collected from a music streaming company KKBOX. Table 2 shows some statistics of the datasets.
We randomly hold out 80% records for each user as the training data. The remaining 20% records of all users are
U1 U2 U3 U4
I1 I2 I3 I4 I5 I6
U1 U1
I2 I3 I2 U4
U4 U2
I5
U4U1 I2
I1
I4
X (d) Latent Representations (a) Given Matrices (b) Built Graphs (c) Random Walks (e) Designed Matrix for FM
Figure 1: Exploring latent social listening representations for music recommendation. Approach Sole User-to-Item Friendship Indexes Random Social Graph Social Graph Listening Graph Social Listening Graph treated as the testing item pool. Instead of using pure precision, we use recall and mean average precision (MAP) as the performance measurements. We repeat the evaluation process 5 times with di erent randomly selected training sets and report the average performance.
Since FM model is already a promising approach of MFbased model, we directly perform all the experiments based on FM, including a) one primitive approach that uses pure user-to-item matrix, b) one baseline approach that directly uses friendship indexes as auxiliary feature c) one testing approach that builds the graph by randomly generated social links and d) three proposed approaches that integrate di erent types of latent representation in correspondingly.
Evaluation results are reported in Table 1.
Social links are much more sparser than the case of userto-item ratings so that direct use of friendship is ine ective.
Besides, combing the social graph with listening graph is able to lter out the inadequate connections. Consequently, we can see that the use of social listening graph leads to about 60% relative improvement to baseline approach (i.e.
Friendship Indexes) for hetrec2011-lastfm-2k dataset and about 20% of improvement for KKBOX-50K dataset in terms of MAP and recall measurements. A randomly generated social graph is also examined for verifying the e ectiveness of social relations in music data. It can be found that the random graph leads to inferior result than the proposed ones
We propose a novel method that bridges unsupervised representation learning for social links and context-aware factorization model for recommendations. According to the experimental results, direct use of friendship indices as features does not perform well. Among the three proposed ways for building the graph, integrating the latent representation learned from the social listening graph achieves the best improvement in both MAP and recall measurements.
In current work, all the connections are considered to have the same weight (i.e. the binary response) and some types of connections are omitted (i.e. the item-to-item connections). Hence, it is possible to earn better performance by, for example, assigning numeric weights to the connections or adding item-to-item connections. We leave these as future work.