2014 IEEE Fourth International Conference on Big Data and Cloud Computing

Item Recommendation Using Collaborative Filtering

in Mobile Social Games: A Case Study

Zhaojie Tao∗ , Ming Cheung† , James She‡ Ringo Lam§

HKUST-NIE Social Media Lab nxTomo Games Limited & Tsinghua University
Hong Kong University of Science & Technology, Hong Kong § [email protected]
∗ [email protected], † [email protected], ‡ [email protected]

Abstract—This paper evaluates the performance of collabo- the items that rated by other users. The hybrid approach is
rative filtering in mobile social game. The evaluation involves a combination of content-based approach and collaborative
both user-based and item-based collaborative filtering on game filtering approach. This paper focuses on the collaborative
items for in-app purchases, and including 4 different social filtering approach by investigating a mobile social game, called
information available in the game. Based on the operational Barcode Footballer.
data from a mobile social game, Barcode Footballer, with more
than 100k users and 50k purchasing history, it is concluded
Barcode Footballer is a football game, it is currently
that both user-based and item-based collaborative filtering have launched by nxTomo, which is a Hong Kong based company.
much higher precision than random recommendation, while user- In this game, each user can create and manage a football team,
based approach with friendship as similar relationship has better shown in Fig. 1.(a) As a mobile social game, it includes 4
performance than original approach. This paper also proposes social elements, i.e., Friendship, Location, League, and Level.
a hybrid method to improve the performance of user-based The game have more than 1 million friendships, users can
friendship approach. The results can be applied to mobile social find out new friends in Friends Searching interface. Users can
games to recommend highly needed items to users so that the change the football team with better footballers and can join
monetization can be enhanced. matches and leagues in order to rise to higher levels and enter
Keywords—Recommendation; collaborative filtering; in-app better leagues. Besides, the game includes over 100 locations
purchases; game items; social mobile game in Hong Kong, which indicates the places that users registered
in the game. In addition to the normal matches with default
players, users can also join the matches with other users, which
I. I NTRODUCTION are based on the users location, so that users can play the game
with other nearby users. As a free download game, it was
Mobile social games are based on social networks and mo- designed with in-app purchase function. Users can purchase
bile technologies, which integrate social features as gameplay items from the item store, shown in Fig. 1.(b), the items in
elements and suitable to play on mobile devices. Mobile social this game can be classified into three types, namely basic
game industry has grown rapidly and altered the game industry items, football shirts and team logos. Items can be purchased
deeply. According to a research by SuperData, there were 77.9 by using game point or club fund, which can be purchased in
million social network players in US by the end of 2012 and app or raise in game respectively. Since in-app purchase can
the market is 6.2 billion USD, which is expected to reach generate great revenue for game companies, how to encourage
8.64 billion USD in 2014. Since mobile social game is a new users purchase virtual goods is one of the most important
developing industry, the ways to design a successful mobile question in monetization. This paper focuses on recommending
social game is still a challenge for all the game companies. game items, which is a popular in-app purchasing virtual good
One of the analytic frameworks in game development is ARM in mobile social games. It investigates the user data from
(Acquisition, Retention, Monetization). Acquisition is to attract Barcode Footballer and tries to compare the performance of
users to play the game, Retention focus to keep users in collaborative filtering with different conditions.
the game, and Monetization is how operators gain revenue. To the best of our knowledge, there is no previous in-
One of the common approach of monetization is using in- vestigation on real datasets to evaluate the performance of
app purchase. According to research by Distimo, almost all collaborative filtering on game item for in-app purchases. The
revenue in free apps are from in-app purchases, which indicates contributions of this paper can be summarized as follows:
the importance of in-app purchases. In a mobile monetization
report released by Swrve, that half of mobile game money 1) Investigate a social mobile game with 100k+ users,
comes from 0.15% of players. As a result, the ways to Barcode Footballer;
monetize more paid users is a prefered challenge for game 2) Evaluate the performance of collaborative filtering in
companies, one possible solution is using recommendation. mobile social game in-app purchases;
At present, recommender systems can be classified into 3) Propose a modified user-based collaborative filtering
three types, namely content-based recommendations, collabo- as well as a hybrid method for item recommendation.
rative filtering recommendations, and hybrid recommendations
[1]. In content-based approach, items of the user is estimated The paper is organized as follows: Section II describes
by the similarity among items. In collaborative filtering ap- the previous works, Section III introduced the methodology
proach, the utility of items for a user is predicted based on in game item recommendation, Section IV describe the details

978-1-4799-6719-3/14 $31.00 © 2014 IEEE 293

DOI 10.1109/BDCloud.2014.73
 (a)
Fig. 2. Find out Relevant Items in Item-based Scenario.

evaluated in mobile game recommendation. The result shows

that item-based collaborative filtering with combined explicit
and implicit data has the best performance (largest AUC
in ROC curve), and in item-based approach, using Pearson
Correlation calculate the similarity has better performance,
while in user-based approach, Euclidean Distance is better.
In [9], they discuss how recommendation algorithm could be
evaluated in order to select the best algorithms from a set of
candidates.

(b)
III. M ETHODOLOGY
Fig. 1. (a) Team Management Interface of Barcode Footballer. (b) Item A. Item-based Collaborative Filtering
Purchase Interface of Barcode Footballer.
In item-based scenario, for each item, recommendation is
made according to the items of the buyers. For instance, for an
of dataset, and the evaluation results with analysis, Section V item A, collect all the items of item A buyers, shown in Fig. 2.
proposes a combined method, Section VI is the conclusion. Firstly count the repeated times of each item, and then sort all
the items, and choose the top-k items as the most related items
of item A. In other words, it reveals that the users who have
II. R ELATED W ORKS
purchased item A also purchased topk items. After doing this
Using an in-app purchase model in mobile apps is a for all the items, a result that include most connected items of
more effective method of converting casual users into paying each item can be get.
customers. It is a different approach from the traditional
payment where users pay and download, which the users Algorithm 1 Find out most related items for an item
may be disappointed by the experience after purchasing [2, Input: Item buyers’ purchased items, {P ocketu };
3]. In-app purchasing provides a recurring revenue stream for Output: Item list, {Itemi }; Times of items be purchased,
game developers, this enhances the importance of analyzing {Counti };
the in-app purchasing pattern and find out how to increase the 1: for each P ocketx of user x in {P ocketu } do
monetization [4]. In [5], the authors analyze the in-app pur- 2: for each Itemy in P ocketx do
chasing patterns in a social game called PuppyRed. The items 3: if Itemy in {Itemi } then
can be purchased by coin (real money) or bean (obtained in 4: Countx ++;
game), which is similar with Barcode Footballer. The authors 5: else
classified the game items into 3 types, which distinguish items 6: Add Itemy to {Itemi };
according to use items on the role, space and the gameplay. The 7: end if
result shows that high spenders like to purchase items by coins 8: end for
for the role and space, while low spenders like to purchase 9: end for
items by bean for the gameplay. In [6], the authors analyze 10: Sort {Itemi } according to {Counti };
the users purchasing pattern and suggest how to distinguishing 11: return {Itemi };
money spenders and non-spenders.
In addition to purchasing pattern, recommendation is also
an important method to enhance revenue [7]. As previous The Algorithm 1 describes how to find out what other items
works show, collaborative filtering recommendation can be do the users most likely to buy when they purchase an item.
classified into two types, namely user-based approach and After using the Algorithm 1 for every items of the game, each
item-based approach. In [8], both of this two approaches are item can get a list of most related items.
To make the recommendation, the data of purchasing

294
Fig. 3. Find out Relevant Items in User-based Scenario. Fig. 4. Distribution of users based on how many items they purchased.

record is analyzed. If a user buy an item, a list of most related of all the friends of the user A (not include items of the user
items of that item can be listed quickly, and then choose top- A), form an item library. Then count the number of repeated
N items from the list, where variable N is the number of purchases of each item, in other words, how many friends
recommendations. have bought that item. After that, sort the items according to
the repeated times and choose top-N items to recommend.
B. Original User-based Collaborative Filtering
In user-based scenario, recommendation will be made IV. E XPERIMENT
based on similar users, in terms of users’ purchasing records.
Firstly, setting up item rating tables, where the rating is A. Dataset
obtained from purchasing records. Secondly, using the rat- The dataset used in this project is provided by nxTomo,
ing table to calculate the similarity between each two users it consist of the information of users, including friendship,
in terms of purchased items. Finlly, recommend the top-N levels, leagues, locations, etc. Besides, it also includes the item
purchased items from top-k similar users, shown in Fig. 3. purchasing history and a record of items in each users pocket.
The algorithm of user-based collaborative filtering is shown in The data is collected from 7 August 2013 to 16 December
Algorithm 2, for each user, a list of items can be generated, 2013, for a more receivable evaluation result, the dataset of
and then choose top-N items from the list according to the purchasing record is split into two parts according to the
recommendation amount requirement. purchasing date. The earlier one is used for analysis the
relationship between items, which is used in making recom-
Algorithm 2 Make user-based collborative filtering mendations. The later one is used for evaluate the results of
Input: Similar users list, {U }; Item buyers’ purchased items, recommendations. The two parts contains almost equal number
{P ocketu }; of records.
Output: Item list, {Itemi }; Times of items be purchased, Statistically, the dataset contains the information of
{Counti }; 197,604 users, and 52,587 purchasing records, which include
1: for each user x in {U } do 7,323 purchased users. The real distribution of how many items
2: for each Itemy in P ocketx do each user has is shown in Fig. 4. Most users purchased items
3: if Itemy in {Itemi } then amount is under 5, approximately 6,000 users. On the contrast,
4: Countx ++; users rarely have over 15 items.
5: else
6: Add Itemy to {Itemi }; B. Social Information
7: end if
8: end for In user-based recommendation with social information, first
9: end for similar users should be find out based on the users profiles. In
10: Sort {Itemi } according to {Counti }; this paper, 4 types of social relationships has been discussed,
11: return {Itemi }; respectively are friendships, locations, leagues and levels.
1) Friendship: The given dataset contains 1,615,586 pairs
C. User-based Collaborative Filtering with Social Information of friends, which means each user has approximately 8 friends
on average. It has been proved that friendship indicates a
In user-based collaborative filtering, similarity between certain similarity, and friends tend to have many things in
users is identified by users’ purchased items. In this section, common.
a novel method which using users’ social information rather
than using purchasing records as similarity is proposed. For 2) Location: Location is the place that the users registered
example, it is believed that two friends have a higher similarity in the game, which includes 248 places in Hong Kong. Since
[10], so they will probably have common items. Base on this, the regional functions differ from each place, it can been infer
for instance, choose a user A, and then collect all the items that users have same location tend to have similar living and

295
 (a) (a)

(b) (b)
Fig. 5. Evaluation results with friendship and league (a) Precision (b) Recall. Fig. 6. Evaluation results with level and location (a) Precision (b) Recall.

working environment, so location is also chosen as a type of Fig. 5 and Fig.6 shows, all the curves of collaborative filtering
relationship. have higher precision than random and hottest curves, where
user-based approach has better performance than item-based
3) League: As a sport game, there are several leagues in approach.
the game. Users have to attend league matches in order to In Fig.5, user-based collaborative filtering using friendship
join higher leagues. In Barcode Footballer, besides game point, and league have slightly better performance than original user-
users can also purchase some items by ”club fund”, which can based collaborative filtering. In Fig.6, the performances of
be only gained from leagues. user-based approach with level and location are laid between
item-based approach and original user-based approach.
4) Level: Similar with league, users in same level will have
In item-based and original user-based collaborative filter-
similarity. Meanwhile, sometimes games will make limitation
ing, the similarities between items or users are calculated
in games so that only users with lower levels cannot purchase
from purchasing records, which can cost a long processing
higher level items.
time when the data size is large. On the contrast, user-based
approach with social information calculate the similarities
C. Results using users’ profiles only, this approach is more efficient than
the original method. On the other hand, user-based approaches
To verify the recommendations, a random recommendation with both friendship and league as social relationship have bet-
and hottest recommendation are made as references. The ter performances than original user-based approach. However,
hottest recommendation curve is drawn by sorting the items league information needs to be filtered from a large data of
according to how many times each item is purchased. Making users’ profiles, while friendship is a separate data table. In
recommendations by using the above 4 types of relationship terms of efficiency, it is better to use friendship as similar
respectively, the evaluation result of precision and recall curves relationship.
of collaborative filtering are shown in Fig. 5 and Fig. 6.
In k-NN based collaborative filtering, it has been proved
V. H YBRID M ETHOD
that higher k values will have higher precision, but this can also
cause higher complexity of calculation. In this experiment, the In the experiment, it is found that the standard deviation of
k value is 55 for both item-based and user-based approach. As social information user-based approach is higher than original

296
 VI. C ONCLUSION
In this paper, three types of collaborative filtering rec-
ommendation methods have been discussed and evaluated
in mobile social game items recommendation. As a result,
both the user-based method and item-based method has a
better performance than random recommendation. However,
by making comparisons between these methods, user-based
recommendation with friendship as relationship is more accu-
rate and efficient than the item-based and original user-based
recommendation. To cope with the decreasing precision when
users have a smaller number of friends, a combine method
is proposed, which increases the precision at lower number
of friends without decreases precision at higher number of
friends.
Fig. 7. User-based approach with friendship as similar relationship. For practical application, in item-based approach, since the
gameplay and user habit will not change dramatically, it can
be inferred that after an analysis of a certain period of data,
the connections between items will keep stable and will not
change after adding new data. In user-based approach, since
the purchasing records and profiles (i.e. Levels, friendships,
etc.) of users will change frequently, it is less efficient to make
user-based approach in this scenario, because the system needs
to be updated frequently. In hybrid method, both purchasing
history and friendship are needed for analysis, it can have
better precision and stability but have lower efficiency. In this
case, game companies need to trade off between the needed
precision and system capacity.

ACKNOWLEDGMENT
This work is supported by HKUST-NIE Social Media Lab.,
HKUST and nxTomo Games Ltd.
Fig. 8. Hybrid approach.
R EFERENCES
[1] Adomavicius, G., Tuzhilin, A., ”Toward the next generation of recom-
mender systems: A survey of the state-of-the-art and possible extensions.”
IEEE Transactions on Knowledge and Data Engineering, pp. 374-749,
approach, which means the performances between different 2005.
users have higher variances. Since frinedship have the best [2] Viennot, N., Garcia, E., Nieh J., ”A Measurement Study of Google Play.”
performance among all 4 social relationships in social infor- The 2014 ACM international conference on Measurement and modeling
mation user-based approach, performance of using friendship of computer systems., 2014.
is analyzed, as it is shown in Fig. 7, the precision is much [3] Lettner, F., Clemens, H., ”Automated and unsupervised user interaction
logging as basis for usability evaluation of mobile applications.” Pro-
lower for those users who have less than 5 friends, while the ceedings of the 10th International Conference on Advances in Mobile
precision is much higher than average when users have more Computing & Multimedia., 2012.
than 100 friends. [4] Yamakami, T., ”A two-layer ecosystem evolution model: lessons from
To solve this problem, a hybrid method by combining stages of mobile data services.” Proceedings of the International on
original user-based approach and social information user- Management of Emergent Digital EcoSystems. , 2010.
based approach is proposed. As it is descripted in Section [5] Wohn, D. Y., Na, E., ”Virtual item purchasing patterns in a social game:
III, a similar user list can be obtained from original user- differences between high and low spenders.” Proceedings of the 2012
iConference., 2012.
based approach, and a friends list can be derived from social
[6] Wohn, D. Y., ”Spending real money: purchasing patterns of virtual goods
information user-based approach. Firstly, sort the users’ friends in an online social game.” Proceedings of the SIGCHI Conference on
list according to the similar users list. Secondly, combine the Human Factors in Computing Systems., 2014.
friends list and similar users list, and then recommend the top- [7] Dutting, P., Henzinger, M., Weber, I., ”Maximizing revenue from strate-
N items from the top-k users. The precision of hybrid approach gic recommendations under decaying trust.” Proceedings of the 21st ACM
is shown in Fig. 8, where k value is 55. International Conference on Information and Knowledge Management.,
As Fig. 8 shows, users with lower friendships using the 2012.
hybrid method can have a much higher precision than the [8] Leroux, B. Dhoedt, B., Demeester, P., ”Performance characterization
of game recommendation algorithms on online social network sites.”
friendship-based approach. Besides, the overall precision of all Journal of Computer Science and Techonology, 2012.
users is also higher. For the users who have over 100 friends,
[9] Herlocker, J., ”Evaluating Collaborative Filtering Recommender Sys-
since the k value smaller than 100, the recommendation is tems.” ACM Transactions on Information Systems, pp 5-53, 2004.
actually made from users’ friends list, the precision does not [10] Drucker, Susan, J., Gumpert, Gary; Cohen, M. Howard, ”Social Media.”
drop. Regulating Convergence. Communication Law, 2010.

297