Anomaly Detection in Time-Series Data

Conformal LSTM for Anomaly Detection in Julia

Project Overview

This project implements a high-performance LSTM (Long Short-Term Memory) network in Julia to detect anomalies in streaming time-series data. Utilizing the Numenta Anomaly Benchmark (NAB), I developed a system that doesn’t just predict values, but provides a statistically sound “safety margin” using Conformal Prediction.

View Project on GitHub Open in Colab Dataset (GitHub - Numenta Anomaly Benchmark)

---

Methodology

1. High-Performance Modeling (Julia & Flux)

Leveraging the Julia programming language for its C-like speed, I implemented the LSTM architecture using the Flux.jl library. This allows for rapid training and inference on large-scale AWS Cloudwatch and traffic datasets.

2. Conformal Prediction Framework

Most AI models give a “guess.” This project uses Conformal Prediction to generate a dynamic threshold based on historical error distribution. If a new data point falls outside this mathematically rigorous interval, it is flagged as an anomaly with a specific confidence level (e.g., 95%).

3. Supervised vs. Unsupervised Pipelines

• Supervised: Trained on known anomaly patterns to recognize specific signatures.
• Unsupervised: A forecasting-based approach that identifies “surprises” in data it has never seen before.

---

Evaluation Results

The model was evaluated against the Numenta Anomaly Benchmark (NAB). Below are the top-performing categories for both the supervised and unsupervised pipelines.

Supervised

data_sup = FileAttachment("supervised_metrics.csv").csv({ typed: true })

data_sup_top = data_sup
  .slice()
  .sort((a, b) => b["F1 Score (%)"] - a["F1 Score (%)"])
  .slice(0, 5)

Inputs.table(data_sup_top, {
  layout: "auto",
  columns: [
    "filename", "Precision (%)", "Recall (%)", "F1 Score (%)"
  ],
  width: {
    filename: 280,
    "Precision (%)": 100,
    "Recall (%)": 100,
    "F1 Score (%)": 100
  },
  header: {
    filename: "Dataset"
  }
})

Unsupervised

data_unsup = FileAttachment("unsupervised_metrics.csv").csv({ typed: true })

data_unsup_top = data_unsup
  .slice()
  .sort((a, b) => b["F1 Score (%)"] - a["F1 Score (%)"])
  .slice(0, 5)

Inputs.table(data_unsup_top, {
  layout: "auto",
  columns: [
    "filename", "Precision (%)", "Recall (%)", "F1 Score (%)"
  ],
  width: {
    filename: 280,
    "Precision (%)": 100,
    "Recall (%)": 100,
    "F1 Score (%)": 100
  },
  header: {
    filename: "Dataset"
  }
})

Full Experimental Results

The sections below contain the complete results.

Supervised Full Data

Inputs.table(data_sup, {
  required: false,
  rows: 10,
  sort: "F1 Score (%)", 
  reverse: true,
  layout: "auto"
})

Unsupervised Full Data

Inputs.table(data_unsup, {
  required: false,
  rows: 10,
  sort: "F1 Score (%)", 
  reverse: true,
  layout: "auto"
})

---

Complete Analysis Notebook

To view the full Julia implementation, mathematical derivations for the conformal thresholds, and performance logs, please visit the technical notebook page.

View Technical Implementation - Julia Notebook

TipProject Documentation

The complete methodology, data analysis, and model performance metrics are detailed in the full technical report.

View Final Report