DISCOVERY
OF RANKING FRAUD FOR MOBILE APPS
Abstract—Ranking fraud in the
mobile App market refers to fraudulent or deceptive activities which have a
purpose of bumping up the Apps in the popularity list. Indeed, it becomes more
and more frequent for App developers to use shady means, such as inflating
their Apps’ sales or posting phony App ratings, to commit ranking fraud. While
the importance of preventing ranking fraud has been widely recognized, there is
limited understanding and research in this area. To this end, in this paper, we
provide a holistic view of ranking fraud and propose a ranking fraud detection
system for mobile Apps. Specifically, we first propose to accurately locate the
ranking fraud by mining the active periods, namely leading sessions, of mobile
Apps. Such leading sessions can be leveraged for detecting the local anomaly
instead of global anomaly of App rankings. Furthermore, we investigate three
types of evidences, i.e., ranking based evidences, rating based evidences and
review based evidences, by modeling Apps’ ranking, rating and review behaviors
through statistical hypotheses tests. In addition, we propose an optimization
based aggregation method to integrate all the evidences for fraud detection.
Finally, we evaluate the proposed system with real-world App data collected
from the iOS App Store for a long time period. In the experiments, we validate
the effectiveness of the proposed system, and show the scalability of the
detection algorithm as well as some regularity of ranking fraud activities.
EXISTING SYSTEM
Generally speaking, the
related works of this study can be grouped into three categories. The first
category is about web ranking spam detection. Specifically, the web ranking
spam refers to any deliberate actions which bring to selected webpages an
unjustifiable favorable relevance or importance. For example, Ntoulas et al. have
studied various aspects of content-based spam on the web and presented a number
of heuristic methods for detecting content based spam. Zhou et al. have studied the problem of unsupervised web
ranking spam detection. Specifically, they proposed an efficient online link
spam and term spam detection methods using spamicity. Recently, Spirin and
Han have reported a survey on web spam
detection, which comprehensively introduces the principles and algorithms in
the literature. Indeed, the work of web ranking spam detection is mainly based
on the analysis of ranking principles of search engines, such as PageRank and
query term frequency. This is different from ranking fraud detection for mobile
Apps. The second category is focused on detecting online review spam. For
example, Lim et al] have identified several representative behaviors of review
spammers and model these behaviors to detect the spammers. Wu et al. have studied the problem of detecting hybrid
shilling attacks on rating data. The proposed approach is based on the
semisupervised learning and can be used for trustworthy product recommendation.
Xie et al. have studied the problem of singleton review spam detection.
Specifically, they solved this problem by detecting the co-anomaly patterns in
multiple review based time series. Although some of above approaches can be
used for anomaly detection from historical rating and review records, they are
not able to extract fraud evidences for a given time period (i.e., leading
session). Finally, the third category includes the studies on mobile App
recommendation. For example, Yan and Chen developed a mobile App recommender
system, named Appjoy, which is based on user’s App usage records to build a preference
matrix instead of using explicit user ratings. Also, to solve the sparsity
problem of App usage records, Shi and Ali studied several recommendation models
and proposed a content based collaborative filtering model, named Eigenapp, for
recommending Apps in their website Getjar. In addition, some researchers
studied the problem of exploiting enriched contextual information for mobile
App recommendation. For example, Zhu et al. proposed a uniform framework for
personalized context-aware recommendation, which can integrate both context
independency and dependency assumptions. However, to the best of our knowledge,
none of previous works has studied the problem of ranking fraud detection for
mobile Apps.
PROPOSED SYSTEM:
First, the download
information is an important signature for detecting ranking fraud, since
ranking manipulation is to use so-called “bot farms” or “human water armies” to
inflate the App downloads and ratings in a very short time. However, the
instant download information of each mobile App is often not available for
analysis. In fact, Apple and Google do not provide accurate download
information on any App. Furthermore, the App developers themselves are also reluctant
to release their download information for various reasons. Therefore, in this
paper, we mainly focus on extracting evidences from Apps’ historical ranking,
rating and review records for ranking fraud detection. However, our approach is
scalable for integrating other evidences if available, such as the evidences
based on the download information and App developers’ reputation. Second, the
proposed approach can detect ranking fraud happened in Apps’ historical leading
sessions. However, sometime, we need to detect such ranking fraud from Apps’
current ranking observations. Actually, given the current ranking ra now of an
App a, we can detect ranking fraud for it in two different cases. First, if ra
now > K_, where K_ is the ranking threshold introduced in Definition 1, we
believe a does not involve in ranking fraud, since it is not in a leading
event. Second, if ra now < K_, which means a is in a new leading event e, we
treat this case as a special case that Te end ¼ te now and u2 ¼ 0. Therefore,
such real-time ranking frauds also can be detected by the proposed approach.
Module 1
Leading events
Definition 1 (Leading
Event). Given a ranking threshold K e¼½testart;teend_ and corresponding
rankings of a, which satisfies Ra start _ K _ a start_1 <r , and r a end _ K
_ <r a endþ1 . Moreover, 8t k 2ðt e start ;t e end Þ, we have r a k . Note that we apply a ranking threshold K _ K _ _ which is usually
smaller than K here because K may be very big (e.g., more than 1,000), and the
ranking records beyond K _ (e.g., 300) are not very useful for detecting the
ranking manipulations. Furthermore, we also find that some Apps have several
adjacent leading events which are close to each other and form a leading
session. For example, Fig. 2b shows an example of adjacent leading events of a
given mobile App, which form two leading sessions. Particularly, a leading
event which does not have other nearby neighbors can also be treated as a
special leading session.
Module
2
Leading
Sessions
A leading session s of App
a contains a time range T s ¼½t s start ;t s end _ and n adjacent leading
events fe 1 ; ...;e n g, which satisfies t and there is no other leading session
s s start _ ¼ t e 1 start , t that makes
T _ . Meanwhile, 8i ½1;nÞ, we have ðt e
iþ1 start _ t e I end s end s I ¼ t _ _ T Þ < f, where f is a predefined time
threshold for merging leading events. Intuitively, the leading sessions of a
mobile App represent its periods of popularity, so the ranking manipulation
will only take place in these leading sessions. Therefore, the problem of
detecting ranking fraud is to detect fraudulent leading sessions. Along this
line, the first task is how to mine the leading sessions of a mobile App from
its historical ranking records.
Module
3
Identifying
the Leading Sessions for Mobile APPs
There are two main steps
for mining leading sessions. First, we need to discover leading events from the
App’s historical ranking records. Second, we need to merge adjacent leading
events for constructing leading sessions. Specifically, Algorithm demonstrates the pseudo code of mining
leading sessions for a given App In
Algorithm , we denote each leading event e and session s as tuples <t e
start ;t e end > and <t s start > respectively, where E is the set of
leading events in session s. Specifically, we first extract individual leading
event e for the given App a (i.e., Step 2 to 7) from the beginning time. For
each extracted individual leading event e,we check the time span between e and
the current leading session s to decide whether they belong to the same leading
session based on Definition 2. Particularly, if ðt s e start ;t _ t Þ < f,
will be considered as a new leading session (i.e., Step 8 to 16). Thus, this
algorithm can identify leading events and ses- sions by scanning a’s historical
ranking records only once. S end s end ;E s
CONCLUDING REMARKS
In this paper, we
developed a ranking fraud detection system for mobile Apps. Specifically, we
first showed that ranking fraud happened in leading sessions and provided a method
for mining leading sessions for each App from its
historical ranking
records. Then, we identified ranking based evidences, rating based evidences
and review based evidences for detecting ranking fraud. Moreover, we proposed an
optimization based aggregation method to integrate all the evidences for
evaluating the credibility of leading sessions from mobile Apps. An unique
perspective of this approach is that all the evidences can be modeled by statistical
hypothesis tests, thus it is easy to be extended with other evidences from
domain knowledge to detect ranking fraud. Finally, we validate the proposed
system with extensive experiments on real-world App data collected from the
Apple’s App store. Experimental results showed the effectiveness of the
proposed approach. In the future, we plan to study more effective fraud
evidences and analyze the latent relationship among rating, review and
rankings. Moreover, we will extend our ranking fraud detection approach with
other mobile App related services, such as mobile Apps recommendation, for
enhancing user experience.
REFERENCES
[4] (2012). [Online].
Available: http://venturebeat.com/2012/07/03/ apples-crackdown-on-app-ranking-manipulation/
[5] (2012). [Online].
Available: http://www.ibtimes.com/applethreatens-
crackdown-biggest-app-store-ranking-fra ud-406764
[8] L. Azzopardi, M.
Girolami, and K. V. Risjbergen, “Investigating the relationship between
language model perplexity and ir precision- recall measures,” in Proc. 26th
Int. Conf. Res. Develop. Inform. Retrieval, 2003, pp. 369–370.
[9] D. M. Blei, A. Y. Ng,
and M. I. Jordan, “Latent Dirichlet allocation,” J. Mach. Learn. Res., pp.
993–1022, 2003.
[10] Y. Ge, H. Xiong, C.
Liu, and Z.-H. Zhou, “A taxi driving fraud detection system,” in Proc. IEEE
11th Int. Conf. Data Mining, 2011, pp. 181–190.
[11] D. F. Gleich and
L.-h. Lim, “Rank aggregation via nuclear norm minimization,” in Proc. 17th ACM
SIGKDD Int. Conf. Knowl. Discovery Data Mining, 2011, pp. 60–68.
[12] T. L. Griffiths and
M. Steyvers, “Finding scientific topics,” Proc. Nat. Acad. Sci. USA, vol. 101,
pp. 5228–5235, 2004. [13] G. Heinrich, Parameter estimation for text analysis,
“ Univ. Leipzig, Leipzig, Germany, Tech. Rep., http://faculty.cs.byu.edu/~ringger/ CS601R/papers/Heinrich-GibbsLDA.pdf, 2008.
[14] N. Jindal and B. Liu,
“Opinion spam and analysis,” in Proc. Int. Conf. Web Search Data Mining, 2008,
pp. 219–230.
[15] J. Kivinen and M. K.
Warmuth, “Additive versus exponentiated gradient updates for linear
prediction,” in Proc. 27th Annu. ACM Symp. Theory Comput., 1995, pp. 209–218.
No comments:
Post a Comment