Embedding Tree Feature Analysis for Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face challenges in analyzing features associated with entities using embeddings, as they require initial operations like generating pseudo-documents or graphs, and determining feature values from embedding vectors is difficult.
Innovation Solution
A system and method for analyzing features using an embedding tree, which involves receiving a dataset, generating embeddings, determining split criteria based on Bayesian Information Criterion (BIC) scores for Gaussian mixture models, and constructing an embedding tree for anomaly detection and graphical representation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional embedding methods are used to learn information from datasets, then embeddings can represent semantic relationships, but initial operations such as generating pseudo-documents or graphs are required before embeddings can be used
Solution Approach 1:
The patent applies preliminary action by pre-processing the dataset to generate feature-enriched data instances before embedding generation. The system identifies relevant features, extracts their values, and integrates them with embedding vectors in advance, so that when embeddings are generated, all necessary information is already prepared and organized, eliminating the need for subsequent pseudo-document or graph generation steps
Solution Approach 2:
The patent segments the data processing into distinct components: feature identification, feature value extraction, embedding generation, and integration. By dividing the complex task of information learning into these manageable segments, the system can process each aspect independently and efficiently without requiring complex intermediate structures like pseudo-documents
2Loss of information
If embedding vectors are used to represent entities, then semantic relationships can be captured, but it becomes difficult to determine information about the feature values of the feature vector
Solution Approach 1:
The patent introduces an intermediary mechanism that bridges embedding vectors and feature values. The system maintains a mapping structure that connects each embedding vector to its corresponding feature values and identifiers. This intermediary layer allows the system to retrieve and interpret feature value information from embeddings without losing the semantic representation capabilities of the embedding vectors themselves
Solution Approach 2:
The patent addresses the information loss by adding another dimension to the embedding structure. Instead of relying solely on the embedding vector space, the system creates a multi-dimensional representation that includes both the embedding vectors and their associated feature values as separate but connected dimensions, enabling easy access to feature information while preserving semantic relationships
3Loss of time
If embeddings are generated directly from datasets, then processing time is reduced, but the system lacks the ability to perform anomaly detection and visualize embedding clusters effectively
Solution Approach 1:
The patent applies preliminary action by organizing embedding data into a tree structure during the embedding generation process itself. The system pre-computes cluster assignments, identifies anomaly candidates, and structures the embedding space hierarchically before any analysis is performed. This preliminary organization enables both rapid processing and effective anomaly detection without requiring additional time-consuming operations later
Solution Approach 2:
The patent implements nesting by creating a hierarchical embedding tree structure where embedding clusters are nested within parent nodes, and anomalies are identified at multiple levels of the hierarchy. This nested structure allows the system to efficiently traverse and analyze embeddings at different granularities, enabling effective anomaly detection and visualization while maintaining fast processing speeds
Data Source
AI summary
Provided is a system for analyzing features associated with entities using an embedding tree, the system including at least one processor programmed or configured to receive a dataset associated with a plurality of entities, wherein the dataset comprises a plurality of data instances for a plurality of entities. The processor may be programmed or configured to generate at least two embeddings based on the dataset and determine split criteria for partitioning an embedding space of at least one embedding tree associated with the dataset based on feature data associated with an entity and embedding data associated with the at least two embeddings. The processor may be programmed or configured to generate at least one embedding tree having a plurality of nodes based on the split criteria. Methods and computer program products are also provided.