type
status
date
slug
summary
tags
category
icon
password
Created time
Feb 1, 2025 03:26 PM
Link
区域等
Status
 
When I first learned machine learning a decade ago, causal reasoning felt like a missing puzzle piece. As I applied data techniques to manufacturing, one question kept resurfacing: How do we distinguish true causes from mere correlations?
This curiosity led me through six learning phases:
  1. Building vocabulary - Starting with terms like “causal discovery” and “DoWhy”
  1. Historical foundations - Studying Judea Pearl’s “The Book of Why”
  1. Expert conversations - Learning from:
      • Matheus Facure (Python causal inference author)
      • Aleksander Molak (causal AI researcher)
      • Microsoft’s causal team (Chang Zhang, Greg Lewis, Amit Sharma)
  1. Industry mapping - Tracking investments from:
      • Tech giants: Microsoft, Huawei (PCIC conference hosts)
      • Startups: UK’s CausalLens, Germany’s ExplainData, US-based Geminus
  1. Tool experimentation - Testing:
      • Algorithms: PC, GES, Notears
      • Frameworks: DoWhy, EconML
  1. Competition validation - Applying skills through causal inference challenges
 

Key insights

- Causal understanding requires combining statistical methods with domain knowledge - Current tools feel like prototype hammers when we need surgical instruments - The field’s growth stalled as AI research shifted to LLMs
 

Here is what I did:

Building Vocabulary

My learning began with key terms: - Causality vs correlation - Causal discovery methods - Inference frameworks - Tools like DoWhy

Historical Foundations

I discovered Judea Pearl’s “The Book of Why” - the Rosetta Stone for causal reasoning. This masterpiece explains: - The ladder of causation (seeing vs doing vs imagining) - Counterfactual thinking - Bayesian networks’ role in causal analysis

Expanding Concepts

The book helped me grasp: 1. do-calculus fundamentals 2. Structural causal models 3. Mediation vs confounding 4. Pearl’s “causal revolution” framework

Conversations That Shaped My Understanding

These practitioners gave me crucial reality checks: - Matheus Facure (Built Python causal inference ecosystem) - Aleksander Molak (Pioneer in causal AI architectures) - Guo (data specialist) - Dr. Li (Shanghai causal AI group leader)
Notable Researchers I Followed: - Microsoft’s Causal Frontiers Team: - Chang Zhang (Enterprise causal adoption) - Greg Lewis (Econometric causal models) - Amit Sharma (Recommender systems causality)

Where the Money Flows: Corporate vs Startup Playbooks

Corporate Investments: - Microsoft’s end-to-end causal stack - Huawei’s annual PCIC challenge (Pacific Causal Inference Conference)
Startup Innovators: 1. CausalLens (UK) - Enterprise decision engines 2. ExplainData (Germany) - Automated causal reporting 3. Geminus (US) - Physics-informed digital twins

Toolbox Essentials

After investigation in people and companies. I started to looking into these methods/frameworks/tools.
Key Algorithms: - PC (constraint-based discovery) - GES (score-based structure learning) - LiNGAM (non-Gaussian orientation) - Notears (memorable name, abandoned author) - GOLEM (accuracy-focused optimization) - DECI (neural causal discovery)
Practical Frameworks: - DoWhy (Python causal graph analysis) - EconML (Microsoft’s heterogeneous effects)
Overall I feel the huge diversity of tools and how people using them differently (not like everybody is using pytorch now). Some people even told me they never use these frameworks and writing all the methods by themselves.
At the mean time I feel frustrated in finding the correct way of using these tools. Almost all the tutorials are outdated and I can hardly find any answers if there’s a bug.
At the end when I’m running the algorithms on some random datasets, I see extremely low performance (For only a few thousands of records, some methods runs few hours on a M1 mac).
 

Learning by Doing: Why Competitions Work

Like mastering machine learning basics through Kaggle, causal discovery clicked for me when I entered Huawei’s PCIC Challenge. The competition’s industrial manufacturing focus matched my daily work challenges.
I attacked it through: 1. Reverse-engineering winning solutions from past years 2. Implementing conditional independence tests from scratch 3. Benchmarking NOTEARS against traditional PC algorithms 4. Visualizing confounding paths in equipment failure data
The hands-on trial/error exposed gaps in my understanding of collider bias and optimal transport theory. Waking up at 3am to check leaderboard updates became a ritual. When my hybrid approach landed in the top 15%, I finally grasped how causal assumptions shape real-world model performance.

Summary

After a long time learning, my initial sky-high expectations made the subsequent disappointment that much more profound when reading The Book of Why by Judea Pearl. While the fundamental concept of causality remains brilliant, lack of a complete theory and tools/frameworks stopped it from becoming meaningful. It stayed like a 3 year old baby and didn’t grow up.
And it will not grow up considering the AI research shift to LLMs.
For any theory, it takes years of working and collaboration of university-giants-startups. It’s like how a baby grows. It needs father, month, teacher and more to mature.
With LLMs becoming more accessible, “how people learn” will become so different from what people imagine. Will our children even learn to code? Will they even learn history? Will they feel meaningless when learning and taking exams that AI can do 100x better?