=Paper=
{{Paper
|id=Vol-1441/recsys2015_poster6
|storemode=property
|title=
|pdfUrl=https://ceur-ws.org/Vol-1441/recsys2015_poster6.pdf
|volume=Vol-1441
|dblpUrl=https://dblp.org/rec/conf/recsys/ChenCLTY15
}}
====
https://ceur-ws.org/Vol-1441/recsys2015_poster6.pdf
Exploiting Latent Social Listening Representations for
Music Recommendations
Chih-Ming Chen1,2 , Po-Chuan Chien1 , Yu-Ching Lin3 , Ming-Feng Tsai1 , Yi-Hsuan Yang2
1 2 3
Department of Computer Research Center for Information Research Center
Science Technology Innovation KKBOX Inc.
National Chengchi University Academia Sinica Taipei 11501, Taiwan
Taipei 11605, Taiwan Taipei 11564, Taiwan aaronlin@kkbox.com
{99303052, mftsai}@nccu.edu.tw {cmchen, yang}@citi.sinica.edu.tw
ABSTRACT methods [1, 8] fuse the idea of topic modeling with the prob-
Music listening can be regarded as a social activity, in which abilistic matrix factorization on social networks, in order to
people can listen together and make friends with one other. infer useful latent topics for collaborative filtering.
Therefore, social relationships may imply multiple facets
of the users, such as their listening behaviors and tastes. 2. METHODOLOGY
In this light, it is considered that social relationships hold Figure 1 illustrates the framework of the proposed method.
abundant valuable information that can be utilized for mu- We transfer the listening history and friends relationship into
sic recommendation. However, utilizing the information for a social listening graph. Given the social listening graph,
recommendation could be difficult, because such information DeepWalk is used to learn the implicit representation by seek-
is usually sparse. To address this issue, we propose to learn ing the possible path on the graph. The idea is to maximize
the latent social listening representations by the DeepWalk the co-occurrence patterns in each generated path so that the
method, and then integrate the learned representations into potential distance is modeled. The representation, usually
Factorization Machines to construct better recommendation presented as a vector, can encode the context information
models. With the DeepWalk method, user social relation- for further processing [7]. From theoretical perspective, the
ships can be transformed from the sparse and independent social representation learning is a technique of combining the
and identically distributed (i.i.d.) form into a dense and non- recent developments of language modeling and unsupervised
i.i.d. form. In addition, the latent representations can also representation learning. In this work, we use the technique
capture the spatial locality among users and items, therefore to learn the representations on a social listening graph, and
benefiting the constructed recommendation models. then feed the learned representations into Factorization Ma-
chines (FM) [9]. After being processed by DeepWalk, the
Keywords sparse social relations will become dense and in a non-i.i.d.
form, which can be helpful for increasing the connections
Representation Learning, Factorization Machine, Recom-
among users and items. To our best knowledge, this work is
mender System, Social Network, Graph
the first attempt to use deepwalk for music recommendation.
To find the possible path, we utilize random walk to uni-
1. BACKGROUND formly sample a series of random vertex from a graph. The
primitive graph is constructed on the sole social network.
Underlying almost all recommendation algorithms is an For the specific application of music recommendation, we
attempt to model the interaction among users and items. propose 3 different ways to construct such a social graph:
There have been some studies working on utilizing auxiliary • Social graph: Build the graph only based on users’ friends,
information for improving recommendations. In [2, 5], social which will enable DeepWalk to detect local community.
relationships are utilized to densify the ratings of users to
• Listening graph: Build the graph only based on user-item
improve the similarity computation behind the Collabora- listening matrix, which will enable DeepWalk to identify
tive Filtering-based (CF-based) methods. For the Matrix the association patterns about users and items.
Factorization-based (MF-based) methods, some studies focus
• Social Listening graph: Build the graph based on the
on how to incorporate social relations with other attributes above two relations including the user-user and user-item
of users [3, 4, 10] and how to affect the regularization term [6, matrices, which may hopefully fuse the merits of the pre-
11]. In addition, the Collaborative Topic Regression (CTR) ceding two approaches.
3. EXPERIMENTS
Our experiments involve two real-world music datasets –
hetrec2011-lastfm-2k and KKBOX-50K. The first 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.
Copyright is held by the author(s).
We randomly hold out 80% records for each user as the
RecSys 2015 Poster Proceedings, September 16-20, 2015, Austria, Vienna. training data. The remaining 20% records of all users are
I1 I2 I3 I4 I5 I6 Y
I3
I6
U1 1 1 1 - - - I5 U2
I1 I2 U4 U3
U2 - - 1 - 1 1 U1
I2
U1
U1 U4
U3 - - 1 1 1 1
I2 I5
I1 I4
U4 - 1 1 - - - U2
U2 I3
X
User-Item Matrix I3
(d) Latent Representations
U4
U1 U2 U3 U4 U3 U3
I4
U1 - 0 0 1 User User Feature Item Item Feature
U2 - - 0 0 I2
U3 1:1 (U1 ) 1:1 (I1 )
U4
U3 - - - 1 I5 U1 1:1 (U1 ) 2:1 (I2 )
I3
1:1 (U1 ) 3:1 (I3 )
U4 - - - - U4 2:1 (U2 ) 3:1 (I3 )
Relationship Matrix I6 2:1 (U2 ) 5:1 (I5 )
I6
… … … …
(a) Given Matrices (b) Built Graphs (c) Random Walks (e) Designed Matrix for FM
Figure 1: Exploring latent social listening representations for music recommendation.
hetrec2011-lastfm-2k KKBOX-50K
Approach DeepWalk MAP Recall@100 Recall@200 MAP Recall@100 Recall@200
Sole User-to-Item 7 4.993% 29.554% 40.308% 5.167% 4.498% 8.143%
Friendship Indexes 7 5.204% 32.275% 39.111% 2.764% 2.031% 3.971%
Random Social Graph 3 4.993% 30.479% 41.089% 0.311% 0.004% 0.004%
Social Graph 3 5.646% 33.328% 45.153% 5.155% 4.494% 8.038%
Listening Graph 3 8.423% 46.285% 60.486% 6.094% 5.376% 9.658%
Social Listening Graph 3 8.807% 47.962% 62.113% 6.157% 5.446% 9.708%
Table 1: Recommendation Performance
Dataset #User #Item #Record #Relations for building the graph, integrating the latent representation
lastfm-2k 1,892 17,632 92,834 12,717 learned from the social listening graph achieves the best
KKBOX-50K 50,000 200,000 50,000,000 80,000 improvement in both MAP and recall measurements.
In current work, all the connections are considered to have
Table 2: Data Statistics 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 exam-
treated as the testing item pool. Instead of using pure ple, assigning numeric weights to the connections or adding
precision, we use recall and mean average precision (MAP) item-to-item connections. We leave these as future work.
as the performance measurements. We repeat the evaluation
process 5 times with different randomly selected training sets
and report the average performance. 5. REFERENCES
[1] C. Chen, X. Zheng, Y. Wang, F. Hong, and Z. Lin.
Since FM model is already a promising approach of MF- Context-aware collaborative topic regression with social
based model, we directly perform all the experiments based matrix factorization for recommender systems. In Proc.
on FM, including a) one primitive approach that uses pure AAAI, 2014.
user-to-item matrix, b) one baseline approach that directly [2] W.-S. Hwang, S. Li, S.-W. Kim, and K. Lee. Data
uses friendship indexes as auxiliary feature c) one testing imputation using a trust network for recommendation. In
Proc. WWW Companion, 2014.
approach that builds the graph by randomly generated so-
[3] M. Jiang, P. Cui, R. Liu, Q. Yang, F. Wang, W. Zhu, and
cial links and d) three proposed approaches that integrate S. Yang. Social contextual recommendation. In Proc. ACM
different types of latent representation in correspondingly. CIKM, 2012.
Evaluation results are reported in Table 1. [4] X. Liu and K. Aberer. Soco: A social network aided
Social links are much more sparser than the case of user- context-aware recommender system. In Proc. WWW, 2013.
to-item ratings so that direct use of friendship is ineffective. [5] C. Luo, W. Pang, Z. Wang, and C. Lin. Hete-cf:
Besides, combing the social graph with listening graph is Social-based collaborative filtering recommendation using
able to filter out the inadequate connections. Consequently, heterogeneous relations. In Proc. IEEE ICDM, 2014.
we can see that the use of social listening graph leads to [6] H. Ma, D. Zhou, C. Liu, M. R. Lyu, and I. King.
Recommender systems with social regularization. In Proc.
about 60% relative improvement to baseline approach (i.e. ACM WSDM, 2011.
Friendship Indexes) for hetrec2011-lastfm-2k dataset and [7] B. Perozzi, R. Al-Rfou, and S. Skiena. Deepwalk: Online
about 20% of improvement for KKBOX-50K dataset in terms learning of social representations. In Proc. ACM KDD, 2014.
of MAP and recall measurements. A randomly generated [8] S. Purushotham, Y. Liu, and C. J. Kuo. Collaborative topic
social graph is also examined for verifying the effectiveness regression with social matrix factorization for
of social relations in music data. It can be found that the recommendation systems. CoRR, 2012.
random graph leads to inferior result than the proposed ones [9] S. Rendle. Factorization machines with libFM. ACM Trans.
Intell. Syst. Technol., 2012.
[10] X. Yang, H. Steck, and Y. Liu. Circle-based recommendation
4. CONCLUSIONS AND FUTURE WORK in online social networks. In Proc ACM KDD, 2012.
We propose a novel method that bridges unsupervised [11] Q. Yuan, L. Chen, and S. Zhao. Factorization vs.
representation learning for social links and context-aware regularization: Fusing heterogeneous social relationships in
top-n recommendation. In Proc. ACM RecSys, 2011.
factorization model for recommendations. According to the
experimental results, direct use of friendship indices as fea-
tures does not perform well. Among the three proposed ways