Every information that is recordable but not numerical can be analyzed with ML to answer business questions

Classified by categories:

Text – NLP

Image – CV

Video – CV

Sound – DL

Biometrics – Bioinformatics

Reasons to use ML to understand unstructured data:

Cost reduction and Scalability

Objectivity

Built-in other prediction systems

Problems of using ML to understand unstructured data

Measurement Errors

Interpretation

Y = a + b * D + g(X) + epsilon

Outcome

Y can be generated through ML with error

Treatment:

D can be generated through ML with error

Control:

X can be generated through ML with error

X can be selected by ML with error

Two general approaches

Standard error re-computation (Instrumental Variable)

De-biasing

But how?

Remember we have training data.

Using training data to learn a better mapping on the error structure

We typically say experiments are golden standards for causal inference.

Why?

The estimator is simple

It achieves the sqrt-n efficiency

It is unbiased in small samples

Basically it suits to tell a causal story

It is hard:

Cross-validation cannot be directly applied to hyper-parameter tuning for causal inference models (Athey and Imbens 2016)

Good performance in predicting the propensity score or outcomes cannot be directly translated into good causal performance (Belloni and Chernozhukov 2014)

Regularization in ML will introduce additional biases (Belloni et al. 2016)

What are the core problems in causal inference?

Non-overlapping/balance, unconfoundedness, power, functional-forms etc.

ML will not magically solve these fundamental problems

ML specifically targeting functional form problems

This is a very fast evolving field (compared to other fields in econometrics)

The framework requires:

Some regularity condition

ML estimator to converge at n^-1/4 (which you should realize is slower than n^-1/2 and most parametric models, according to delta method, converge at n^-1/2)

The framework outputs:

N^-1/2 – consistent estimator: means that the distribution of the estimator converges in probability to the true estimator with rate n^-1/2.

Many estimators you learnt in parametric world has this properties.

In frequentist world, converge in probability typically also means asymptotically normal, which means you can easily construct standard errors.

Two sources of biases as described by the authors:

Regularization bias:

The ML model cannot perfectly predict the function fast enough according to existing data.

This is fixed by orthogonality moment conditions (this is why the paper is also called Neyman orthogonality)

Overfitting bias:

The ML model may be biased by the existing data compared to the true distribution

This is fixed by sample splitting (very similar to cross validation, which is the key idea in this paper)

Orthogonality will help us to get rid of b.

The idea is not new. It appears in Robinson (1988) and Frisch-Waugh-Lovell theorem (1930)

More recently, it appears on the authors’ prior paper: Belloni et al. (2014) using lasso

The basic idea is to train two machine learning model and double selecting out the biases. This is called double selection in Belloni et al. (2014).

The authors show that the estimator through this procedure will

Require:

Some regularity condition

ML estimator to converge at n^-1/4 (which you should realize is slower than n^-1/2 and most parametric models, according to delta method, converge at n^-1/2

Output:

N^-1/2 – consistent estimator: means that the distribution of the estimator converges in probability to the true estimator with rate n^-1/2.

The original data generalization process does not need to be a partial linear model in this more generalized proof!

Double/de-biased machine learning (DML)

Correct the bias of ML estimators through Neyman orthogonal score functions + cross-fitting

Newey (1994), Chernozhukov et al. (2018, 2022), Farrell et al. (2020, 2021), Athey et al. (2018), Ellickson et al. (2022), Fan et al. (2022), Gordon et al. (2022) etc.

Valid estimation and inference with multiple experiments

Azevedo et al. (2020), Dasgupta et al. (2015), Athey et al. (2021), Pashley and Bind (2019), etc.

(Conjoint Analysis, Green and Rao 1971)

Experiments on online platforms

Evaluating and optimizing the strategies of a large-scale online platform.

Ye et al. (2022), Zeng et al. (2022), Zhang et al. (2020), Cui et al. (2019, 2020), Feldman et al. (2021), Schwartz et al. (2017), Bojinov et al. (2023), Abadie and Zhao (2023) etc.

We need to observe at least (m+2) conditions or (m+1) treatment conditions:

There are m+2 parameters to be estimated

Why do we choose m+2?

In practice, people typically run single experiment + joint holdout experiment, which gives us m+2 conditions including control conditions

Overlapping conditions:

All treatment conditions must be linked to each other through some path.

For example, if m = 3,

We observe (1,0,0), (1,0,1), (0,1,0). Then we cannot run our analysis since the second treatment condition does not link with any other treatment condition.

However, if we observe (0,1,1) instead of (0,1,0), now the second treatment condition is linked to third one through (0,1,1) and the third one is linked to the first one through (1,0,1).

There is a nice graphical representation of the overlapping conditions

Personalized Pricing and Targeting: Leverage on observational/experimental data.

Lu, H., Simester, D., & Zhu, Y. (2025), Yang, J., Eckles, D., Dhillon, P., & Aral, S. (2024), Dubé, J. P., & Misra, S. (2023), Daljord, Ø. et al. (2023), Simester, D., Timoshenko, A., & Zoumpoulis, S. I. (2020),

Experiments on online platforms: Evaluating and optimizing the strategies of a large-scale online platform.

Urban, G. L., Liberali, G., MacDonald, E., Bordley, R., & Hauser, J. R. (2014), Ye et al. (2022), Zeng et al. (2022), Zhang et al. (2020), Feldman et al. (2021), Tucker C. (2014), Tucker, C., & Zhang, J. (2010), etc.

Policy learning with continuous variables: Policy evaluation and optimization.

Kallus & Zhou(2018), Chernozhukov et al. (2019), Athey et al. (2018), etc.

Double/de-biased machine learning (DML): Correct the bias of a plug-in estimator through Neyman-orthogonal score functions.

Newey (1994), Chernozhukov et al. (2018, 2022), Farrell et al. (2020, 2021), Ellickson et al. (2022), Fan et al. (2022), Ye et al. (2023), etc.

High-Dimensional Causal Estimation + Policy Learning

First framework to recover continuous policy values & personalized policies with high dimensional data of discrete A/B test

Strong Theoretical Guarantees with Explicit Bounds on

Approximation error

Estimation error

Policy regret

Empirical Validation

Large-scale field experiment with ~7.4m of users

Shows practical gains from continuity + personalization in real marketing decisions

Practical Impact

The framework is currently utilized by the platform

GitHub link: https://github.com/ZhiqiZhang1229/Policy-Learning-through-Discrete-Experimentation-Theory-and-Empirical-Evidence

ML to optimize multiple treatments (MAB-type-of paper)

ML to optimize complex MDPs in inventory:

Gijsbrechts, J., Boute, R. N., Van Mieghem, J. A., & Zhang, D. J. (2022). Can Deep Reinforcement Learning Improve Inventory Management? Performance on Lost Sales, Dual-Sourcing, and Multi-Echelon Problems. Manufacturing & Service Operations Management.

Dai, Jim G., and Mark Gluzman. "Queueing network controls via deep reinforcement learning." Stochastic Systems 12.1 (2022): 30-67.

Three stages of DRL research in OM/OR:

Vanilla ML and DRL

ML/DRL for Operations Problems with specific structures

ML/DRL for Operations Problems that select past policies (instead of actions)

Static structural model (BLP, static games, games on networks, etc.)

Core problems:

Subject to functional assumptions

Slow estimation due to numerical integration

Dynamic structural models (Russ, dynamic games, etc)

Core problems:

Cannot handle large data due to slow estimation and optimization

Subject to very strong functional-form assumptions

Dynamic game models are very restrictive on game structures

Counterfactuals are limited by optimization power

ML Fairness and Discrimination: https://fairmlbook.org/

Fairness Definitions and Impossibility Results

Test Fairness in AI

Fix Fairness issues in AI

ML and Labor Economics

Eloundou, Tyna, et al. "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models." arXiv preprint arXiv:2303.10130 (2023).

Acemoglu, Daron, and Pascual Restrepo. "Tasks, automation, and the rise in us wage inequality." Econometrica 90.5 (2022): 1973-2016.

Data Privacy

Transparency in AI

Privacy and Business outcomes:

Goldfarb, Avi, and Catherine E. Tucker. "Privacy regulation and online advertising." Management science 57, no. 1 (2011): 57-71.

Data and ML in IO: Data is now an input into the production and consumption function, which can be priced, traded and incorporated into decisions

Any equilibrium involving algorithms, Data, and other features distinct to AI.

Asker, John, Chaim Fershtman, and Ariel Pakes. "Artificial Intelligence, Algorithm Design and Pricing." In AEA Papers and Proceedings, vol. 112, pp. 452-56. American Economic Association, 2022.

ML as a species: Whether human is only part of the civilization of intelligence?

Butlin, Patrick, et al. "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness." arXiv preprint arXiv:2308.08708 (2023).

Hendrycks, Dan. "Natural selection favors ais over humans." arXiv preprint arXiv:2303.16200 (2023).