Mutual Information Graph for High-Dimensional Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The analysis of categorical data in high-dimensional spaces is challenging due to the lack of a natural distance metric, leading to difficulties in identifying anomalies, as most statistical learning algorithms rely on distance metrics that become less effective with increasing dimensionality, a problem known as the curse of dimensionality.
Innovation Solution
A probabilistic approach is employed using a mutual information graph and clique tree analysis to determine normality ratings for entities by constructing a mutual information graph and determining clique trees based on attributes with strong mutual dependence, allowing for anomaly detection and classification without relying on distance metrics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If distance metrics are used for anomaly detection in high-dimensional categorical data, then the approach works well in low-dimensional numerical data, but the effectiveness deteriorates as dimensionality increases due to the curse of dimensionality
Solution Approach 1:
The patent replaces distance-based mechanical metrics with information-theoretic measures (mutual information, entropy) to detect anomalies. Instead of measuring geometric distance in high-dimensional space, the system uses probabilistic dependencies between attributes to identify anomalous entities, thereby substituting the flawed mechanical approach with a robust information-theoretic framework.
Solution Approach 2:
The patent transforms the problem from geometric space to probabilistic space by changing the fundamental parameters used for comparison. Rather than using distance metrics in attribute space, the system employs probability distributions and mutual information scores, fundamentally changing the parameter space in which anomaly detection occurs to avoid the curse of dimensionality.
2Ease of manufacture
If rule-based or ground truth-based classification approaches are used, then predefined event classes can be detected, but the system becomes fragile and cannot detect subtle anomalies that do not form cohesive classes
Solution Approach 1:
The patent enables the system to learn anomaly patterns directly from the data without requiring external rule definitions or ground truth labels. The mutual information-based approach automatically discovers dependencies and anomalies inherent in the data structure, making the system self-sufficient and eliminating the fragility associated with manual rule-based approaches.
Solution Approach 2:
The patent pre-computes mutual information scores and probability distributions for all attribute pairs before anomaly detection. This preliminary analysis of data dependencies allows the system to efficiently evaluate new entities against learned patterns without requiring real-time rule evaluation, improving both robustness and performance.
3Measurement precision
If human inspection is used to identify anomalies, then accurate detection is possible, but the process becomes time-consuming and inefficient
Solution Approach 1:
The patent replaces manual human inspection with an automated information-theoretic computation system. By using pre-computed mutual information scores and probability distributions, the system can rapidly evaluate entities for anomalies without requiring human time and effort, thereby achieving both accuracy and high productivity.
Data Source
AI summary
A method executable via operation of configured processing circuitry may include constructing a mutual information graph for categorical data with respect to observed attributes of a plurality of entities described in terms of respective ones of the observed attributes by the categorical data, determining a clique tree correlating attributes having at least a threshold level of mutual dependence among the observed attributes, and determining a normality rating for an entity relative to the plurality of entities based on the clique tree.