=Paper=
{{Paper
|id=Vol-1618/IFUP_paper_5
|storemode=property
|title=RBPR: Role-based Bayesian Personalized Ranking for Heterogeneous One-Class Collaborative Filtering
|pdfUrl=https://ceur-ws.org/Vol-1618/IFUP_paper_5.pdf
|volume=Vol-1618
|authors= Xiaogang Peng,Yaofeng Chen,Yuchao Duan,Weike Pan,Zhong Ming
|dblpUrl=https://dblp.org/rec/conf/um/PengCDPM16
}}
==RBPR: Role-based Bayesian Personalized Ranking for Heterogeneous One-Class Collaborative Filtering==
https://ceur-ws.org/Vol-1618/IFUP_paper_5.pdf
RBPR: Role-based Bayesian Personalized Ranking for
Heterogeneous One-Class Collaborative Filtering
Xiaogang Peng, Yaofeng Chen, Yuchao Duan, Weike Pan* and Zhong Ming*
College of Computer Science and Software Engineering, Shenzhen University
patrickpeng@126.com, {chenyaofeng,duanyuchao}@email.szu.edu.cn, {panweike,mingz}@szu.edu.cn
*: corresponding author.
ABSTRACT Intelligent recommendation systems and technology have
Heterogeneous one-class collaborative filtering (HOCCF) is played a more and more important role in various real-world
a recently studied important recommendation problem, which applications, with a wide spectrum of entertainment, social
consists of different types of users’ one-class feedback such as and professional services. Some recent work show that one
browses and purchases. In HOCCF, we aim to fully exploit important line of research have gradually transferred from
the heterogenous feedback and learn users’ preferences so collaborative filtering (CF) with numerical ratings to one-
as to make a personalized and ranking-oriented recommen- class CF (OCCF) with homogeneous one-class feedback such
dation for each user. For HOCCF, we can apply existing as purchases [2] and heterogeneous OCCF (HOCCF) with
solutions for OCCF with purchases only such as Bayesian more than one types of one-class feedback such as browses
personalized ranking (BPR) or make use of both browses and purchases [3]. In this paper, we focus on the problem
and purchases such as transfer via joint similarity learning setting of HOCCF, which is very common in real industry
(TJSL). However, BPR may be not very accurate due to scenarios.
the ignorance of browses, and TJSL may be not very effi- The main challenge of HOCCF is the heterogeneity of the
cient due to the mechanism of joint similarity learning and two different types of one-class feedback, since a user’s pref-
base model aggregation. In this paper, we propose a novel erence behind a purchase action may be different from that
perspective for the different types of one-class feedback via of a browse action. In a very recent work [3], a similari-
users’ different roles, i.e., browser and purchaser. Specifi- ty learning algorithm is proposed for this challenge, which
cally, we design a two-stage role-based preference learning aims to combine browses and purchases in a principled way.
framework, i.e., role-based Bayesian personalized ranking The improved performance in [3] shows the complementar-
(RBPR). In RBPR, we first digest the combined one-class ity of browses to the well exploited feedback of purchases
feedback as a browser to find the candidate items that a user in OCCF models [1, 4]. However, the proposed algorithm,
will browse, and then we exploit the purchase feedback to i.e., transfer via joint similarity learning (TJSL) [3], may be
refine the candidate list as a purchaser. Empirical results not efficient enough for large datasets due to the complex
on five public datasets show that our RBPR is an efficien- prediction rule and base model ensemble.
t and accurate recommendation algorithm for HOCCF as In this paper, we interpret the HOCCF problem from a
compared with the state-of-the-art methods such as BPR novel view of users’ roles, i.e., a purchaser (as reflected in
and TJSL. a purchase feedback) is converted from a browser in a se-
quential manner. Based on this perspective, we propose
a two-stage framework, including browser-based preference
CCS Concepts learning and purchaser-based preference learning. Those t-
•Information systems → Personalization; •Human- wo preference learning tasks are connected via a candidate
centered computing → Collaborative filtering; list of items that a user will browse, which is assumed to
contain the potential items that a user will finally purchase.
In each of the two tasks, we apply the seminal work for
Keywords homogeneous one-class feedback, i.e., Bayesian personalized
Role-based Preference Learning; Bayesian Personalized Rank- ranking (BPR) [4], and for this reason, we call our approach
ing; Heterogeneous One-Class Collaborative Filtering role-based BPR (RBPR).
In our empirical studies, we compare our RBPR with the
1. INTRODUCTION state-of-the-art methods of BPR and TJSL using various
ranking-oriented evaluation metrics on five public datasets.
The studies show that our RBPR is able to produce com-
petitive recommendations efficiently. We list our main con-
tributions as follows: (i) we propose a novel and generic
staged role-based preference learning framework, which is a
frustratingly easy, scalable and effective solution for collab-
orative ranking with heterogeneous one-class feedback; and
(ii) we conduct extensive empirical studies and obtain very
. promising results.
Figure 1: An illustration of role-based preference learning for heterogeneous one-class collaborative filtering
(HOCCF), including browser-based preference learning and purchaser-based preference learning.
2. ROLE-BASED BAYESIAN PERSONALIZED Input: Users’ browses B and purchases P.
RANKING Output: Top-K recommended items for each user.
2.1 Problem Definition Step 1. Conduct browser-based preference learning
In HOCCF, we have a set of n users (U), a set of m items via BPR(B ∪ P) as shown in Eq.(1) and obtain 3K
(I), and two different sets of user feedback, e.g., browses B candidate items with highest predicted scores.
and purchases P. Our goal is to find some likely-to-purchase Step 2. Conduct purchaser-based preference learn-
items from unpurchased items for each user. ing via BPR(P) as shown in Eq.(2); predict the
scores on the 3K candidate items and refine the list.
In order to fully exploit heterogeneous feedback in HOC-
CF such as browses and purchases, we propose not to model
those different feedback jointly as a whole as done in a recen- Figure 2: The algorithm of role-based Bayesian per-
t work [3], but separately in a staged manner. Specifically, sonalized ranking (RBPR).
we model different feedback of a typical user via different
roles such as browser and purchaser. From the perspective
of browser and purchaser, in our role-based Bayesian per-
sonalized ranking (RBPR), we have two tasks of preference
learning, including browser-based preference learning and In order to solve this task, we propose to use the purchase
purchaser-based preference learning. We illustrate the main data only to refine the candidate list from the first step.
procedure of our proposed solution in Figure 1. The reason is that the purchase feedback is more helpful in
answering whether a certain item will be bought by a user.
2.2 Browser-based Preference Learning Due to the fact that a user’s purchase feedback are few, we
In the first step, we assume that a typical user is first may not get good results if we only apply the second step,
a browser before he/she is converted to a purchaser. And i.e., only use the purchase feedback to find items that will
thus, in our first task, we focus on answering the question be bought by a user. This phenomenon is also observed in
of “whether a user will browse an item”. our empirical studies.
In order to address this task, we propose to combine the Similarly, we again adopt BPR [4] for model training, but
two types of one-class feedback, i.e., browses and purchas- use purchase feedback P only. Mathematically, we learn the
es, together, and then apply an algorithm for homogeneous model parameters as follows,
one-class feedback such as BPR [4], i.e., BPR(B ∪ P). Math- X X X
ematically, we will solve the following optimization problem, min fuij , (2)
X X X ΘP
u∈U i∈Pu j∈I\Pu
min fuij , (1)
ΘB∪P
u∈U i∈(Bu ∪Pu ) j∈I\(Bu ∪Pu )
where ΘP denotes the model parameters to be learned from
where Bu and Pu are item sets browsed and purchased by the purchase data only.
user u, respectively, fuij is the tentative objective function With the learned model parameters ΘP , we can predict
for a randomly sampled triple (u, i, j), and ΘB∪P denotes the preference of each item i in the candidate list of each user
the set of model parameters to be learned [4]. u, and then re-rank the items in the list. The refined list is
Once we have learned the model parameters, we can gen- expected to better represent the purchase likelihood of a cer-
erate a candidate list of items that a user is likely to browse. tain user, i.e., the recommendation may be more accurate,
Specifically, for a top-K recommended problem, we will gen- which is also verified in our empirical studies. We illustrate
erate 3K items in this step, so that the refinement in next the effect of the difference between those two ranked lists in
step may have more room for improvement. Figure 1.
2.3 Purchaser-based Preference Learning For the optimization problems in the aforementioned two
In the second step, we assume that a user will most like- learning tasks, we can apply stochastic gradient descent to
ly choose an item from the candidate list that he/she has learn the model parameters [4]. We put the two preference
browsed. For this reason, in our second task, we mainly an- learning tasks in one single algorithm in Figure 2 in order
swer the question of “whether a user will purchase an item”. to get a complete picture.
Table 1: Description of the datasets used in the experiments, including the numbers of users, items, purchases
and browses, and the number of purchases in validation and test data. Note that the data of ML100K, ML1M
and Alibaba are from [3], and the statistics of ML10M and Netflix are for the first copy of the three generated
copies of each dataset.
Dataset # user # item # purchase # browse # purchase (validation) # purchase (test)
ML100K 943 1682 9438 45285 − 2153
ML1M 6040 3952 90848 400083 − 45075
Alibaba2015 7475 5257 9290 60659 − 2322
ML10M 71567 10681 309317 4000024 308673 308702
Netflix 480189 17770 4554888 39628846 4556347 4558506
3. EXPERIMENTAL RESULTS For BPR, TJSL and RBPR, we fix the dimension as d =
20 and the learning rate as γ = 0.01. For BPR and TJSL
3.1 Datasets and Evaluation Metrics on ML100K, ML1M and Alibaba2015, we directly use the
In our empirical studies, in order to directly compare our results from [3]. For RBPR on all the datasets and BPR
RBPR with the very recent algorithm for HOCCF, i.e., TJS- on ML10M and Netflix, we search the best tradeoff pa-
L [3]. We first use the three public datasets in [3]1 , including rameter from {0.001, 0.01, 0.1} and iteration number from
MovieLens 100K (ML100K), MovieLens 1M (ML1M) and {100, 500, 1000} via NDCG@15. In order to make the re-
Alibaba2015. The detailed description of those three da- sults consistent and comparable with [3], we run five times
ta can be found in [3]. We also study the performance of of RBPR on ML100K, ML1M and Alibaba2015, and report
our RBPR on two large datasets, including MovieLens 10M the averaged performance. For ML10M and Netflix, we re-
(ML10M)2 and Netflix. port the averaged results on three copies of data.
ML10M is a public data with about 10 million numerical 3.3 Results
ratings in {0.5, 1, 1.5, ..., 4.5, 5}, and Netflix is the dataset
used in the famous $100 Million competition with about 0.1 We report the recommendation performance in Table 2.
billion scores in {1, 2, 3, 4, 5}. For both ML10M and Netflix, We can have the following observations:
we first divide the data into five parts with equal number- • RBPR and TJSL are better than BPR in all cases in-
s of (u, i, rui ) triples, we then take one part and keep the cluding five evaluation metrics and five datasets, which
(u, i) pairs with rui = 5 as purchases for training, take one clearly shows that the feedback browses are useful for
part and keep the (u, i) pairs with rui = 5 as purchases for learning and mining users’ hidden preferences, and RBPR
validation, and take one part and keep the (u, i) pairs with and TJSL are able to make use of users’ heterogeneous
rui = 5 as purchases for test, and finally take the remaining feedback well.
two parts and keep all the (u, i) pairs as browses. We repeat
this procedure for three times in order to obtain three copies • RBPR and TJSL are comparable on three small dataset-
of data. s, e.g., TJSL is the best on ML100K, RBPR is the best
We put the statistics of the datasets in Table 1. on ML1M, and TJSL and RBPR are comparable on
For evaluation, we use five ranking-oriented metrics, in- Alibaba2015.
cluding Precision@5, Recall@5, F1@5, NDCG@5 and 1-call@5.
• TJSL is too slow to generate recommendations on t-
3.2 Baselines and Parameter Settings wo large datasets within 24 hours, while RBPR can
produce significantly better results than BPR, which
Because HOCCF is a relatively new recommendation prob-
shows that our RBPR is a more practical solution re-
lem, very few solutions have been proposed. In our empirical
garding the efficiency.
studies, we thus include the very recent algorithm TJSL [3]
for HOCCF and also the seminal work BPR [4] for OCCF. The overall performance in Table 2 shows that our RBPR
performs the best in making use of the heterogeneous one-
• BPR (Bayesian personalized ranking) is an efficien-
class feedback.
t and accurate recommendation algorithm for homo-
In order to check the performance improvement of our
geneous one-class feedback such as purchases, which
two-stage role-based preference learning solution, we also
mines users’ preferences by assuming that a user prefer-
check the performance of the generated candidate items as
s a purchased item to an unpurchased item.
shown in Figure 1. Specifically, we denote the method for
• TJSL (transfer via joint similarity learning) is the state- generating the candidates as RBPR(Browser) since it is based
of-the-art method for heterogeneous one-class feedback on the role of browser only, and the final recommendation as
such as browses and purchases, which jointly learns the RBPR(Browser,Purchaser). We report the performance on
similarity between a candidate item and a purchased Precision and NDCG in Figure 3 (other metrics are similar),
item, and the similarity between a candidate item and from which we can see that the second stage of candidate re-
a likely-to-purchase item. finement using the purchase data can significantly improve
the performance. The improvement also verifies our main
1
http://www.cse.ust.hk/∼weikep/TL4HOCCF/ assumption that there are usually two separate stages for a
2
http://grouplens.org/datasets/movielens/10m/ user’s shopping action, i.e., browse and purchase.
Table 2: Recommendation performance of RBPR, BPR and TJSL on ML100K, ML1M, Alibaba2015, ML10M
and Netflix using Prec@5, Rec@5, F1@5, NDCG@5 and 1-call@5. The significantly best results are marked
in bold (p value < 0.01). Note that the results of BPR and TJSL on three small datasets are from [3]. We
use “−” to denote the case that the training procedure does not finish within 24 hours.
Dataset Method Prec@5 Rec@5 F1@5 NDCG@5 1-call@5
BPR 0.0552±0.0006 0.1032±0.0019 0.0673±0.0007 0.0874±0.0020 0.2425±0.0034
ML100K TJSL 0.0697±0.0016 0.1393±0.0028 0.0864±0.0019 0.1133±0.0047 0.3033±0.0071
RBPR 0.0654±0.0013 0.1275±0.0048 0.0803±0.0021 0.1058±0.0047 0.2890±0.0047
BPR 0.0928±0.0008 0.0829±0.0002 0.0717±0.0003 0.1121±0.0010 0.3609±0.0018
ML1M TJSL 0.1012±0.0011 0.0968±0.0012 0.0821±0.0009 0.1248±0.0010 0.3961±0.0022
RBPR 0.1086±0.0009 0.1017±0.0015 0.0858±0.0009 0.1327±0.0016 0.4151±0.0055
BPR 0.0050±0.0006 0.0193±0.0026 0.0077±0.0009 0.0138±0.0017 0.0246±0.0031
Alibaba2015 TJSL 0.0071±0.0004 0.0283±0.0016 0.0110±0.0006 0.0200±0.0008 0.0347±0.0017
RBPR 0.0076±0.0005 0.0304±0.0023 0.0118±0.0008 0.0220±0.0013 0.0367±0.0024
BPR 0.0629±0.0002 0.0855±0.0006 0.0603±0.0003 0.0861±0.0004 0.2648±0.0017
ML10M TJSL − − − − −
RBPR 0.0719±0.0013 0.0977±0.0017 0.0690±0.0014 0.0994±0.0020 0.2990±0.0050
BPR 0.0716±0.0007 0.0480±0.0005 0.0446±0.0005 0.0818±0.0011 0.2846±0.0022
Netflix TJSL − − − − −
RBPR 0.0797±0.0002 0.0595±0.0004 0.0527±0.0003 0.0939±0.0003 0.3174±0.0011
0.12 0.15
RBPR(Browser) RBPR(Browser)
RBPR(Browser,Purchaser) RBPR(Browser,Purchaser)
NDCG@5
0.08 0.1
Prec@5
0.04 0.05
0 0
ML100K ML1M Alibaba2015 ML10M Netflix ML100K ML1M Alibaba2015 ML10M Netflix
Dataset Dataset
Figure 3: Recommendation performance of RBPR with different configurations, i.e., RBPR(Browser) for
browser only, and BPR(Browser,Purchaser) for both browser and purchaser.
4. CONCLUSIONS AND FUTURE WORK tion of Guangdong Province No. 2014A030310268 and No.
In this paper, we study an important recommendation 2016A030313038.
problem called heterogeneous one-class collaborative filter-
ing (HOCCF) from a novel perspective of users’ roles. Specif- 6. REFERENCES
ically, we propose a novel role-based preference learning frame- [1] S. Kabbur, X. Ning, and G. Karypis. Fism: Factored
work, i.e., role-based Bayesian personalized ranking (RBPR), item similarity models for top-n recommender systems.
based on a seminal work [4]. Extensive empirical studies In Proceedings of the 19th ACM SIGKDD International
show that our RBPR is more accurate than the seminal work Conference on Knowledge Discovery and Data Mining,
for OCCF, i.e., BPR [4], and a very recent similarity learn- KDD ’13, pages 659–667, 2013.
ing algorithm for HOCCF, i.e., TJSL [3]. Furthermore, our [2] R. Pan, Y. Zhou, B. Cao, N. N. Liu, R. Lukose,
RBPR is very efficient with the inherited merits of BPR [4], M. Scholz, and Q. Yang. One-class collaborative
while TJSL [3] is difficult to produce recommendations on filtering. In Proceedings of the 8th IEEE International
two large datasets. Conference on Data Mining, ICDM ’08, pages 502–511,
For future work, we are interested in extending and ap- 2008.
plying our role-based preference learning framework to oth- [3] W. Pan, M. Liu, and Z. Ming. Transfer learning for
er recommendation settings with more types of users’ roles heterogeneous one-class collaborative filtering. IEEE
such as searcher, browser, purchaser, rater and friends. Intelligent Systems, 2016.
[4] S. Rendle, C. Freudenthaler, Z. Gantner, and
L. Schmidt-Thieme. Bpr: Bayesian personalized
5. ACKNOWLEDGMENT ranking from implicit feedback. In Proceedings of the
We thank the support of Natural Science Foundation of 25th Conference on Uncertainty in Artificial
China (NSFC) No. 61502307 and Natural Science Founda- Intelligence, UAI ’09, pages 452–461, 2009.