Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Overview of Anomaly Detection in Time Series Data LÊ VĂN QUỐC ANH Outline Introduction Anomaly detection approaches Classification based Nearest Neighbor Based Predictive Window-Based Disk Aware Discord Discovery And others approaches Comments Conclusion References 2 Introduction Time series data problems: Similarity search Classification Clustering Motif discovery Anomaly/novelty detection Visualization * [Keogh] 3 Introduction Time series data problems: Similarity search Classification Clustering Motif discovery Anomaly/novelty detection Visualization * [Keogh] 4 Problem Definition Anomaly/novelty detection refers to the problem of finding patterns in data that do not conform to expected behavior 5 Problem Definition (cont.) Finding discords in large scale time series [V. Chandola] 6 Applications Intrusion detection for cyber-security Fraud detection for credit cards Fault detection in safety critical systems Industrial damage detection Medical and public health anomaly detection Stock market analysis … 7 Very simple technique: Match the data to known patterns 8 Existing anomaly detection techniques Classification based Nearest Neighbor Based Predictive Window-Based Disk Aware Discord Discovery And others techniques 9 Classification based approaches Learn a model from a set of labeled data instances and then, classify a test instance into one of the classes using the learnt model Operate in two phases: training phase: learning from trainning data testing phase: test instance as normal or anomalous Assumption: A classifier that can distinguish between normal and anomalous classes can be learnt in the given feature space. 10 Classification based approaches(cont.) 11 Classification based approaches(cont.) Some techniques: Neural Networks based Bayesian Networks based Support Vector Machines based Rule based 12 Classification based approaches(cont.) Advantages: can distinguish between instances belonging to different classes testing phase is fast Disadvantages: have to assign a label to each test instance rely on availability of accurate labels for various normal classes 13 Nearest Neighbor Based Assumption: Normal data instances occur in dense neighborhoods, while anomalies occur far from their closest neighbors. require a distance defined between two data instances 14 Nearest Neighbor Based(cont.) 15 Nearest Neighbor Based(cont.) Advantages: purely data driven Disadvantages: if the data has normal instances that do not have enough close neighbors or if the data has anomalies that have enough close neighbors, the technique fails to label them correctly performance greatly relies on a distance measure defining distance measures between instances can be challenging when the data is complex 16 Predictive techniques Forecast the next observation in the time series, using the statistical model and the time series observed so far, and compare the forecasted observation with the actual observation to determine if an anomaly has occurred. Some techniques: Regression, Auto Regression ARMA, ARIMA, SVR (Support Vector Regression) 17 Predictive techniques(cont.) Advantages: provide a statistically justifiable solution for anomaly detection if the assumptions regarding the underlying data distribution hold true Disadvantages: rely on the assumption that the data is generated from a particular distribution 18 Window-Based Extract fixed length (w) windows from a test time series, and assign an anomaly score to each window. The per-window scores are then aggregated to obtain the anomaly score for the test time series. Some proposed techniques: HOT SAX AWDD WAT 19 HOT SAX [Eamonn Keogh,Jessica Lin, Ada Fu] Finding the most unusual time series subsequence discord Improve BFDD algorithm (Brute Force Discord Discovery) with heristic ordering Use SAX for discretization 20 Brute Force Algorithm 21 Heuristic Discord Discovery 22 The two data structures for Inner and Outer heuristics [Keogh] 23 AWDD technique M. Chuah, F. Fu (2006) AWDD - Adaptive Window Based Discord Discovery Apply for ECG time series 24 AWDD technique(cont.) Advantages: use adaptive rather than fixed windows Disadvantages: deal only with ECG datasets 25 WAT technique Y. Bu et al (2006) WAT - Wavelet and Augmented Trie Employs Haar wavelet transform and symbol word mapping orderly on raw time series to build prefix tree for Inner and Outer loop heuristic can view a subsequence in different resolutions the first symbol of each word gives us the lowest resolution for each subsequence 26 WAT technique(cont.) Advantages: require 2 parameter (1 intuitive parameter) better performance than HOT SAX Disadvantages: assume the coefficients are in Gaussian distribution assume that the data reside in main memory 27 DADD technique DADD - Disk Aware Discord Discovery (2008) [Yankov, Keogh and Rebbapragada] Finding unusual time series in terabyte sized datasets on secondary memory Algorithm has two phases: Phase 1: a candidate selection phase given a threshold r , finds a set of all discords at distance at least r from their nearest neighbor Phase 2: a discord refinement phase remove all false discords from the candidate set 28 A candidate selection phase procedure [C]=DC Selection(S, r) in: S: disk resident data set of time series r: discord defining range out: C: list of discord candidates C = {S1} 1 for i = 2 to |S| do 2 isCandidate = true 3 for ∀Cj ∈ C do 4 if (Dist(Si,Cj) < r) then 5 C = C \ Cj 6 isCandidate = false 7 end if 8 end for if (isCandidate) then C = C ∪ Si end if 9 10 11 12 13 end for 29 A discord refinement phase procedure [C,C.dist]=DC Refinement(S, C, r) in: S: disk resident dataset of time series C: discord candidates set r: discord defining range out: C: list of discords C.dist: list of NN distances to the discords 1 for j = 1 to |C| do 2 C.distj = ∞ 3 end for 4 for ∀Si ∈ S do 5 for ∀Cj ∈ C do 6 if Si == Cj then 7 continue 8 end if 9 d = EarlyAbandon(Si,Cj ,C.distj) 10 if (d < r) then 11 C = C \ Cj 12 13 14 15 16 17 C.dist = C.dist \ C.distj else C.distj = min(C.distj , d) end if end for end for 30 DADD technique (cont.) Advantages: equires only two linear scans of the disk with a tiny buffer of main memory very simple to implement Disadvantages: depend on threshold r 31 Proposed approach Using Vector Quantization for discretization Improve BFDD algorithm with ordering heuristic 32 Using histogram model Codebook s=16 Generation Series Transformation cmdbca ifaj bb minj j a ma I n j m hldf ko phcako ogcbl p occbl h l hnkkk pl cacg k k g j h h…… gkgj lp Series Encoding 1121000000001000 1200010011000000 1000000012001100 1000000011002100 0001010100110010 1010000100100011 …… 33 Similarity measure 1 S HM (q, t ) 1 dis (q, t ) with s f i ,t f i , q i 1 1 f i ,t f i , q dis (q, t ) fi,t fi,q 1 2...s 34 Using multiple resolutions • Codebook (6,60) • Codebook (16,30) 35 For each resolution Start with lowest resolution and a group of all subsequences For each resolution groups which have more than one subsequences are splitted based on a threshold r Stop when have groups with one subsequences or reach the highest resolution 36 Improve BFDD Outer Loop Heuristic: groups which have smallest subsequences count are considered first Inner Loop Heuristic: when ith subsequence is considered in the outer loop, all subsequences in the same group are considered first in the Inner Loop 37 References [1] E. Keogh, J. Lin, W. Fu. HOT SAX: Efficiently Finding the Most Unusual Time Series Subsequence. In Proc. of the 5th IEEE International Conference on Data Mining (ICDM 2005), November 27-30, 2005, pp. 226233. [2] D. Yankov, E. Keogh, U. Rebbapragada, Disk Aware Discord Discovery: Finding Unusual Time Series in Terabyte Sized Datasets, 2008 [3] E. Keogh.Mining Shape and Time Series Databases with Symbolic Representations. Tutorial of the 13rd ACM Interantional Conference on Knowledge Discovery and Data Mining (KDD 2007), August 12-15, 2007. [4] J. Lin, E. Keogh, A. Fu, and H. Van Herle, Approximations to Magic: Finding Unusual Medical Time Series, the 18th IEEE International Symposium on Computer-Based Medical Systems, pp. 329-334, 2005. [5] M. Chuah and F. Fu, ECG anomaly detection via time series analysis, Technical Report LU-CSE-07-001, 2007. 38 References (cont.) [6] V. Megalooikonomou, Q. Wang, G. Li, C. Faloutsos. A Multiresolution Symbolic Representation of Time Series. In Proc. of the 21st International Conference on Data Engineering (ICDE 2005), April 5-8, 2005, pp. 668-679, 2005. [7] V. Chandola, D. Cheboli, and V. Kumar, Detecting Anomalies in a Time Series Database,Technical Report TR 09-004, 2009. [8] Y. Bu, T-W Leung, A. Fu, E. Keogh, J. Pei, and S. Meshkin, WAT: Finding Top-K Discords in Time Series Database, in Proc. of the 2007 SIAM International Conference on Data Mining (SDM'07), Minneapolis, MN, USA, April 26-28, 2007. [9] Q. Wang, V. Megalooikonomou, A dimensionality reduction technique for efficient time series similarity analysis, Information Systems 33, 115– 132, 2008. [10] H. B. Kekre Tanuja K. Sarode, Fast Codebook Search Algorithm for Vector Quantization using Sorting Technique , International Conference on Advances in Computing, Communication and Control (ICAC3’09), 2009. 39 Thank you! 40