---
tags: Draft
---

# Results from BlockScience

##

Goals:
- Show that we have a semi-supervised ML pipeline for detecting sybil attack
- Explain what that entails (options)
- White box algorithms
- Feature Engineering
- Sensitivity vs Specificity
- Labels and GIGO
- Simple models vs ensemble models
- Impress with some numbers
- Impress with some visualizations
- Convince the audience that will only get better through continous evaluation

Target Audience:
-

Boundary conditions:
- 5min for presentation, 2min for Q&A
- No doxing
- No sharing of feature specifics

## Outline

- Intro (45s)
- "Hi! It is well known that we had lots of attacks. We from BSci together with a tight collaboration with the Gitcoin Dev Team have deved a Semi Supervised Piepline"
- Semi Supervised Pipeline (3m)
- What is a semi supervised pipeline (45s)
- What are the pieces of the pipeline (1min)
- White box algo (40s)
- Labels and GIGO (20s)
- How it gets better with time (1min)
- Feature Engineering (15s)
- Human evaluation (30s)
- Simple & ensemble models (15s)
- What we have now (20s)
- Mixture of DTs and RFs
- Results (1min)
- Summary stats from EDA (20s)
- Prediction results (20s)
- Predictive features (20s)
- Conclusions (30s)
-
- Questions (2min)

## Content resources