Sam Torres - When Traffic Gets Weird: ML for Anomalies & Forecasts

Transcript

2. TRAFFIC HAS CHANGED
3. Let’s Get Real for a Second…
4. Use the right tool for the job Unstructured data LLMs Structured data ML
5. Where ML shines • Specialized and built to do specific tasks • Pattern recognition • Anything that requires math • Consistent, reproducible results • More efficient • No “Yes and”
6. But machine learning models are scary…
7. Let’s meet Google Colab
8. What makes Colab special? •Removes setup tasks – start working immediately •Easy to share, and is YOURS •Free to use •Great learning environment •Version control
10. Colab Building Blocks Text cells Add formatted text and documentation using Markdown. Code cells Write and execute your code, usually Python. Click play or Shift+Enter to run. File browser Upload files, access data, and mount Google Drive. Menu bar File operations, runtime settings, and sharing options. Top right shows RAM and disk usage.
11. Let’s use it!
12. My ML Workflow 1. What am I trying to do? 2. What kind of data do I have to get me there? 3. Ask Claude what models may be useful based on this scenario 4. Choose a model 5. Ask Claude to build me a Colab notebook to run this model (or maybe build a notebook for each model recommendation) 6. Troubleshoot (Claude, are you sure?) 7. Ecco! Have cool graphs and look super smart.
13. What’s the Problem Type • Classification • Regression • Clustering • Removing Noise • Anomaly Detection
14. What About My Dataset • Size • Data form (numbers, tables, time series, images) • Seasonality / holidays • Number of dimensions • Data quality
15. • Anomaly detection • Forecasting Let’s see it in action
16. Plot twist! Anomaly detection and forecasting can use the exact same models. The difference is all in the application.
17. Start with these models • Isolation forest • Local outlier factor • Forecast-residual approach
18. Isolation Forest Where it’s awesome • Works well with multiple dimensions (impressions, CTR, & even crawl stats) • Great first sniff test, global weirdness • Fast Less ideal use cases • Identifying more gradual changes • Seasonality (holidays!)
19. Local Outlier Factor (LOF) Where it’s awesome • Identifying weirdness amongst its peers • Keyword data • Patterns that are odd within its own group • A/B test Less ideal use cases • Global awareness • Datasets with lots of disparate performance or dimensions
20. Forecast-Residual Approach Where it’s awesome • Best for time series and seasonality (including holidays) • Gradual and subtle shifts • SERP volatility Less ideal use cases • Small dataset (or time length) • Can be a bit persnickety
22. Finding Anomalies
23. Keep Learning
24. Thank you! /company/graydotco /in/samantha-torres-seo @thegray.company @sam.thegray.co