This repository demonstrates a practical approach to detecting anomalies in tabular data by repurposing Transformer-inspired concepts like embeddings and self-attention. Built in C# with a clear Domain-Driven Design (DDD) structure, it focuses on:
- Automatic normalization of numeric fields via order-of-magnitude scaling.
- Hash-based embeddings for discrete fields, transforming categories into short, stable vectors.
- Dimension-wise z‑score analysis to identify which fields drive anomalies.
- Parallel processing for efficient handling of large data sets.
-
Embedding Discrete Fields
Converts categorical values (likeDepartment,JobTitle) into fixed-size continuous vectors using a hashing function (SHA‑256). No training or large embedding models needed. -
Normalization of Numeric Fields
Automatically adjusts each numeric field (e.g.,Salary,Age) by its order of magnitude, preserving outliers without cumbersome min–max configs. -
Dimension-Wise Outlier Detection
After computing a “row embedding,” the code groups rows (e.g., byDepartment) and calculates mean/std. dev. embeddings. Then, z‑scores pinpoint anomalies dimension by dimension. -
Explainability
Once a row is flagged, we re-check each field’s embedding to see which dimension/field contributed most to the anomaly. This mirrors a self-attention–style focus.
- DDD Architecture: Code is organized into Domain, Application, Infrastructure, and Presentation layers.
- Parallel Processing: Uses
Task.Run(...)and concurrency structures for faster execution on large datasets. - Hash & Normalize: A stable, zero-configuration approach for embedding discrete fields and scaling numeric fields.
- Explanatory Logs: Logs detail which field triggered the anomaly, along with dimension-wise z‑scores for transparency.
- Domain: Contains
Employee,EmbeddingVector,IEmbeddingService, normalization, and anomaly detection logic. - Application: Defines commands (e.g.,
ProcessEmployeeDataCommand) and services (EmployeeProcessingService) that drive the domain operations. - Infrastructure: Responsible for data I/O (e.g., CSV loading via
CsvHelper), plus any external resources. - Presentation: Entry point (
Program.cs) and controllers (e.g.,EmployeeController) that wire everything together.
- Clone the Repository
git clone https://github.com/YourUserName/EmbeddingsSelfAttentionAnomaly.git cd EmbeddingsSelfAttentionAnomaly - Update Program.cs or EmployeeController with the path to your CSV (default is "data.csv") - Sample data.csv is in project directory
Load Employees The CsvEmployeeLoader reads your CSV asynchronously, creating a list of Employee objects.
Compute Embeddings The EmployeeProcessingService uses SimpleEmbeddingService to hash discrete fields (department, job title, etc.) and Normalizer to scale numeric fields.
Group & Detect Anomalies Employees are grouped by Department (and then JobTitle) to compute group-level mean and std. dev. Embeddings. Each row’s embedding is checked dimension by dimension using z‑scores.
Log Anomalies The application logs anomalies to the console, detailing which field caused the most significant deviation.
We welcome contributions! Feel free to open an issue or submit a pull request for bug fixes, feature ideas, or expansions—especially if you want to add advanced embedding logic, additional grouping strategies, or performance optimizations.
This project is licensed under the MIT License. You’re free to use, modify, and distribute this code in your own projects.
Vaswani, A. et al. (2017). Attention Is All You Need. Evans, E. (2003). Domain-Driven Design: Tackling Complexity in the Heart of Software. Chandola, V., Banerjee, A., & Kumar, V. (2009). Anomaly Detection: A Survey.
Feel free to open an issue or let me know if you have any comments, questions, or suggestions. Enjoy exploring Transformer-like embeddings for your tabular data anomalies!