Font Size: a A A

Study On Outlier Detection For Suspicious Financial Behavior Recognition

Posted on:2008-09-20Degree:DoctorType:Dissertation
Country:ChinaCandidate:J TangFull Text:PDF
GTID:1118360215492272Subject:Computer applications
Abstract/Summary:PDF Full Text Request
This dissertation mainly discusses the theoretic system for suspicious financial behavior recognition and surveillance in order to provide a theoretic guidance that is short of in the anti-money laundering data analysis and discrimination area. The research object is applying machine learning and system complexity theory and methodology, to find out those outliers distinctly deviated from the normal behavior pattern among the enormous volume and complex financial transaction records. In the last we can realize an automatic recognition and monitoring system on those suspicious transactions related to money laundering and fraud through high efficient algorithms.Financial system is such a complex system that most researches are using simplified models under linear and rigorous assumptions, which result in both high omission rate and false positive rate. Our research introduces chaotic analyzing methodology based on complexity theory, which could be applied to the analysis of financial time series with exterior stochastic performances output by an intrinsic dynamic system. The major work and contributions are listed as following:(1) Chaotic attributes analysis of financial transactionsChaotic behaviors are seemingly stochastic time series generated from deterministic systems. Chaotic systems, which are predictable in short, term while unpredictable in long term. We utilize phase space reconstruction theory to analyze the chaotic attributes of financial time series. We employed mutual information method to get the time delay, false nearest neighborhood method to get the optimal embedding dimension, and track tracing method to get the largest Lyapunov exponent. The experiments on real world financial data show that there is a limit correlation dimension and a positive largest Lyapunov exponent, which characterize the time series with chaotic model features.(2) Generation mechanism recognition of financial behavior based on chaotic methodology The short-term predictable features of chaotic data provide a new solution for subtle different outlier detection from chaotic background. We reconstruct phase space from chaotic background signals based on Takens theorem, and utilize RBF neural network to construct a prediction model for normal financial transaction. Considering the essential different mechanisms between the normal and the suspicious behaviors, a distinctive deviation would be detected out through the prediction system. Simulation experiments on ideal chaotic time series and real financial data give promising results in outlier detection.(3) Feature extraction and similarity measurement for financial dataA novel feature extraction and similarity measurement method is put forward based on RBF neural network one-step deviation prediction, which is different from traditional time series researches. The method converts time series similarity to feature vectors similarity comparison, while the vectors have considered both the continuous data and discrete data. A feature extraction method on time domain power distance rule is also provided in order to reduce dimension number.(4)An outlier classifier based on One-Class SVMConsidering the enormous data volume and lack of training sets, we construct an outlier classifier using One-Class SVM based on statistic learning theory. An RBF kernel function joined with HVDM distance measurement for heterogeneous data sets structure is provided to train the SVM. Experiments on simulated data and real data show promising results.A set of intelligent discriminative system is composed through the above 4-step inter-coupling processes, which could be applied in complex financial transaction behavior analysis. The engineering system could also be extended in areas such as signal processing, crisis warning, health census, financial auditing, e-commerce, etc.
Keywords/Search Tags:Data Mining, Machine Learning, Outlier Detection, Anti-Money Laundering, Chaotic Time Series Analysis, Feature Extraction, Support Vector Machine
PDF Full Text Request
Related items