System and method for anomaly detection in dynamically evolving data using random neural network decomposition

A computer system, comprising:

• a) a preparation module configured to:
• receive input data in the form of a matrix A of size m×n and rank k, wherein m represents a plurality of multidimensional data points (MDDPs) and wherein n represents a dimension of each MDDP, construct iteratively i dictionaries Di, wherein each Di is a matrix Bi=Ai′Ri of dimension ki×n, wherein Ai′ is the transpose of matrix Ai, wherein ki is a rank of matrix Ai, wherein Ri is a Gaussian distributed random matrix of dimension mi×ki, wherein each m, is smaller than an iteration wherein mi−1 is an immediately preceding iteration, and wherein each Di is constructed by applying to Ai a multi-layer feedforward artificial neural network made of one hidden layer NN(Ai) such that ∥Bi−NN(Ai)∥ is minimized with respect to parameters of the neural network, and to concatenate all dictionaries Di to construct a dictionary D;
• and
• b) an anomaly detection system including an anomaly detection engine configured to use dictionary D to classify a MDDP as an anomaly or as normal, wherein a detected anomaly is indicative of an undesirable event,
• whereby a reduction in m enhances the performance of the computer system in both processing and storage terms.

A method, comprising:

• a) receiving input data in the form of a matrix A of size m×n and rank k, wherein m represents a plurality of multidimensional data points (MDDPs) and wherein n represents a dimension of each MDDP;
• b) constructing iteratively i dictionaries Di, wherein each Di is a matrix Bi=Ai′Ri of dimension ki×n, wherein Ai′ is the transpose of matrix Ai, wherein ki is a rank of matrix Ai, wherein Ri is a Gaussian distributed random matrix of dimension mi×ki, wherein each mi is smaller than an iteration mi−1, wherein mi−1 is an immediately preceding iteration, and wherein each Di is constructed by applying to Ai a multi-layer feedforward artificial neural network made of one hidden layer (NN(Ai)) such that ∥Bi−NN(Ai)∥ is minimized with respect to parameters of the neural network;
• c) concatenating all dictionaries Di to construct a dictionary D;
• and
• d) using dictionary D to classify a MDDP as an anomaly or as normal, wherein a detected anomaly is indicative of an undesirable event,
• whereby a reduction in m enhances the performance of a computer performing the method in both processing and storage terms.