Data Stream Outlier Detection Algorithm

A Data Stream Outlier
Detection Algorithm Based
On Reverse K Nearest
Neighbors
9/28/2016 HecoSoft Publications 1

Abstract
• A new data stream outlier detection algorithm SODRNN is proposed
based on reverse nearest neighbors. We deal with the sliding window
model, where outlier queries are performed in order to detect
anomalies in the current window. The update of insertion or deletion
only needs one scan of the current window, which improves efficiency.
The capability of queries at arbitrary time on the whole current
window is achieved by Query Manager procedure, which can capture
the phenomenon of concept drift of data stream in time. Results of
experiments conducted on both synthetic and real data sets show that
SODRNN algorithm is both effective and efficient.

Introduction
• In many emerging applications, such as fraud detection, network flow
monitoring, sensor network data management, video tracking[1],etc., data
arrive continuously and unboundedly, but it is either unnecessary or
impractical to store all incoming objects. In this context, an important
challenge is to find the most exceptional objects (called outliers) in the data
stream. Outlier detection is a key issue in traditional data mining. There are
already several kinds of outlier detection algorithms, such as statistic-based
methods, distance-based methods, density-based methods and deviation-
based methods, etc. Traditional methods need to perform multiple scans of
the database, they are not suitable for the data stream environment. It is
very important to detect data stream outliers through only one scan of the
stream.

Cont.
• Outlier detection in data streams has attracted more and more
attention from researchers. A distributed framework to approximate
data distributions coming from a sensor network is presented In [2]. It
is used to report outlying values in the union of the readings coming
from multiple sensors. This algorithm has good efficiency, but is
specifically designed to support sensor networks. In [3], a data stream
outliers detection algorithm based on k-means partitioning DSOKP is
proposed which applies k means clustering on each partition of the
data stream to generate mean reference point set, and subsequently
picks out those potential outliers of each periods according to the
definition of outliers. DSOKP is better in efficiency, but it can not
capture the concept drift information.

PROBLEM DEFINITIONS
• Definition 1: (Data Stream) A data stream DS is a possible infinite series
of objects {…at-N+1,at-N+2,… ,at-1,at,…} where at denotes the object
observed at time t, N is the width of the current window. In the sliding
window model, the window is identified by two sliding endpoints at-
N+1 and at.
• Definition 2: (Hawkins’s definition to outlier)[7] An observation which
deviates so much from other observations as to arouse suspicions that
it was generated by a different mechanism. Hawkins’s definition is
intuitive and it captures the general spirit of an outlier.

Cont.
• Definition 3: (k-distance of an object p) For any positive integer k, the k-
distance of object p, denoted as kdistance(p), is defined as the distance
d(p,o) between p and an object oD such that: (1) for at least k objects oD{p}
it holds that d(p,o) d(p,o) (2) for at most k-1 objects oD{p}it holds that
d(p,od(p,o)
• Definition 4: (k-distance neighborhood of an object p) Given the k-distance
of p, the k-distance neighborhood of p contains every object whose distance
from p is not greater than the k-distance, i.e. Nk-distance(p)(p)={qD{p}|d(p,
q) k-distance(p) } These objects q are called the k-nearest neighbors of p.
• Definition 5: (reverse k nearest neighborhood of p) The reverse k-nearest
neighborhood RNN is an inverse relation which can be defined as:RNNk(p)
={q|qD,pNk(q)}. RNNk(p).

Cont.
• Definition 6: (rknn-outlier) The rknn outlier is the data object in the current
window which has the least number of reverse k nearest neighbors. If object
p has less reverse k nearest neighbors in the current window, from the
definition of RNN, we know that it is because less objects regard object p as
their k nearest neighbors. So the outlierness of object p is much higher than
that of others. The rknn-outlier conforms well to the essence of Hawkins’s
outlier definition. An important character of data stream is concept drift.
Due to data stream evolution, object properties can change over time and,
hence, evaluating an object for outlierness when it arrives, although
meaningful, can be reductive in some contexts and sometimes misleading.
On the contrary, by classifying single objects when a data analysis is
required, data concept drift typical of streams can be captured. To this aim,
it is needed to support queries at arbitrary points-intime, called query
times, which classify the whole population in the current window instead of
the single incoming data stream object.

ALGORITHM
A. Algorithm method
• In this section the algorithm SODRNN, standing for Stream Outlier Detection based
on Reverse k Nearest Neighbors, is described. This algorithm consists of two
procedures: the Stream Manager and the Query Manager. And the entire window
should be allocated in memory. The former procedure receives the incoming data
stream objects and efficiently updates the current window. When new stream
object comes, in order to maintain current window perfectly, it needs only update
the knnlist and rknnlist of the influenced objects in the current window rather
than that of all the data stream objects in the current window. When new coming
object is inserted, it needs only one pass of scan to the current window to find all
objects whose k nearest neighbors are influenced. The update of the knnlists of
the influenced objects in the current window can update their rknnlists at the
same time. The deletion of the expired object needs only update the rknnlists of
the influenced objects in the current window according to its knnlist, and then
update the knnlists of the influenced objects in the current window according to
its rknnlist.

Cont.
B.Algorithm description
• ALGRITHM: SODRNN
• INPUT: DS, current window size N, integer k, query
• time Uquery, number of outlier m
• OUTPUT: m outliers
• METHOD:
• BEGIN
• (1) SM(DS,N,k);
• (2) when (Uquery) QM(m);
• END

Cont.
C. Stream Manager Procedure
• PROCEDURE SM(DS,N,k)
• BEGIN
• (1) FOR each data stream object obj with arrival time t
• DO
• (2) IF the oldest object q of current window expires
• (3)FOR all objects o in q.knnlist DO o.rknnlistdelete(q);
• (4) FOR all objects o in q.rknnlist DO
• o.knnlistdelete(q);
• (5) ENDIF
• (6) remove object q from current window
• (7) object p(obj,t);

Cont.
• obj with p.obj = obj, p.id = t
• (8) FOR all objects o in current window DO
• //update current window
• (9) dist=o.distance(p);
• (10) p.knnlistinsert(o); //the k nearest neighbors of p
• (11) o.rknnlistinsert(p);
• (12) IF dist<=o.k_distance()
• (13) o.knnlistinsert(p);
• (14) p.rknnlistinsert(o); //the reverse k nearest
• neighbors of p
• (15) ENDIF
• (16) ENDFOR
• (17) Insert object p into current window.
• (18) ENDFOR
• END

Outlier Query Management Procedure
• PROCEDURE QM(m)
• BEGIN
• (1) perform a single scan of current window;
• (2) return m objects with minimal RNNk(p) as outliers.
• END

Spatial Analysis
• In the following we use N to denote the size of the current window.
Each node consists of four members: n.obj, n.id, n.knnlist[],
n.rknnlist[], we use d to denote the space required to store an object
n.obj, an integer which stores the identifier of n.obj. For all objects of
the current window, when there is a k nearest neighbor, there must be
a reverse k nearest neighbor correspondingly somewhere in the
current window due to the definition of the reverse k nearest
neighbor. So for the whole current window the total number of the k
nearest neighbors and that of the reverse k nearest neighbors are
equal to Nk. Therefore, the spatial cost of the SODRRN algorithm is
O(N(d+2k+1)).

Temporal analysis
• The stream manager procedure consists of three main stages when
processing a new object. Firstly, update current window and delete the
expired node q. Step(3) updates rknnlists of the objects influenced by
the deletion, whose cost is O(klogk); Step(4) updates knnlists of the
objects influenced by the deletion, whose cost is also O(klogk). This is
because we can find directly the objects whose k nearest
neighborhood are influenced. We need not make a whole scan of the
current window by using the rknnlist. It is noticeable that when the
distance between q and o is less than or equal to o.k_distance() and
o.knnlist has only k objects we should take a knn query to find the
nearest object of o and insert it into o.knnlist, whose cost is
O(klogklogN).

Cont.
• Step(6) has constant cost. So the first main stage has a cost of O(klogklogN).
When k is far less than N, the cost is O(logN). Secondly, the creation of new
node p has constant cost. Thirdly, update the current window and insert
node p into it. Step(8) finds all the objects whose k nearest neighborhood
are changed, it takes a cost of O(N), Step(9) has constant cost.; Step(10) has
a cost of O(logk); Step(11) updates rknnlists of the objects influenced by the
insertion, whose cost is O(logk); Step(13) updates the knnlists of the objects
influenced by the insertion when the distance of p and o is less than
o.k_distance(), whose cost is O(logk); It is noticeable that if there already
exist k-1 objects within o.k_distance() we should delete the farthermost
nearest neighbor r of object o and update r.rknnlist. In this case, the
insertion of p into o.knnlist has a cost of O(log2k); Step(14) has a cost of
O(logk); Step(17) has constant cost. So the third stage has a cost of
O(Nlog2k), When k is far less than N, the cost is O(N). Summarizing, the cost
of the stream manager procedure is O(N).

EXPERIMENTAL RESULTS
A. Experiments on real data sets
• In this paper, in order to verify the validity of the proposed algorithm, we
employed the KDD Cup 1999 benchmark evaluation data sets, which has
been extensively used to evaluate intrusion detection algorithms. The data
consists of network connection records of several intrusions simulated in a
military network environment. As the original data sets is too large, we used
only 50,000 TCP connection records from about one week of data. TCP
connections has been elaborated to construct a data set of 23 numerical
features, which consists of duration, src_bytes, num_file_creations,
num_failed_logins, etc. Intrusion behaviors account for 1% of the entire
data set, which consist of 4 types of intrusion behaviors: DOS(denial of
service),R2L(unauthorized access from a remote machine),
U2R(unauthorized access to local super user privileges),
probing(surveillance and other probing).

Cont.

Experiments on synthetic data sets
• In order to test the efficiency of our algorithm, we used a two
dimensions synthetic data set with 30,000 records which has three
main clusters. We inserted one percent records as outliers that has
certain difference with the three main clusters. The window size N was
set to 10,000. We performed experiments to SODRNN and IncLOF
algorithm respectively. In this experiment, different values of k were
chosen when comparing the time cost of processing a single record. As
we can see from figure 2, the processing time of the two algorithm
both are increased nearly linearly with the increasing of k.. The
increasing speed of SODRNN is slower than that of IncLOF.

Cont.

CONCLUSIONS
• In this paper, a novel data stream outlier detection algorithm SODRNN
is presented. In contrast to IncLOF, this algorithm reduces the number
of scans to only one. Experiments conducted on both synthetic and
real data sets show that the proposed method is efficient and
effective. The future work is to improve the performance of this
algorithm in high-dimensional data stream environment.

REFERENCES
• [1] Han J, Kamber M, Data Mining: concepts and techniques, 2nd edn.
• Morgan Kaufmann,2006:1-29
• [2] [2] Song Xiuyao, Wu Mingxi, Jermaine C. Statistical change
• detection for multi-dimensional data.KDD,California,2007:667-676
• [3] [3] Wu Mingxi, Jermaine C.A Bayesian method for guessing the
• extreme values in a data set.In Proc of
• VLDB,Vienna,Austria,2007:471-482
• [4] [4] Abe N, Zadrozny B, Langford J. Outlier detection by active
• learning. Pennsylvania, KDD,2006:504-509
• [5] [5] Barbara D, Domeniconi C, Rogers J P. Detecting outliers using
• transduction and statistical testing. Pennsylvania,KDD,2006:55-64
• [6] [6] Aggarwal C C, Yu P S. An effective and efficient algorithm for
• high-dimensional outlier detection. The VLDB Journal,2005:211-221
• [7] [7] Lazarevic A,Kumar V.Feature bagging for outlier
• detection.Chicago,KDD, 2005:157-166

Data Stream Outlier Detection Algorithm

More Related Content

What's hot

Viewers also liked

Similar to Data Stream Outlier Detection Algorithm

Recently uploaded

Data Stream Outlier Detection Algorithm