Make effective actionable decisions to optimize business KPIs by using causal machine learning

Causal analysis in the machine learning platform Azure Machine Learning Studio can answer causal questions through an end-to-end automation framework.

Translator | Li Rui

Reviewer | Sun Shujuan

In different scenarios, commonly used machine learning modeling techniques may misunderstand the true relationships in the data. Here we seek to change this paradigm to find actionable insights beyond spurious correlations based on estimating causal relationships and measuring treatment effects on target key performance indicator (KPI) outcomes. ​

The motivation of causal machine learning

Assume that the historical data or observation data of a certain product of a certain company in the past year are obtained. If a product loses 5% of its customers, the company's goal is to reduce the churn rate through targeted campaigns. Typically a classic customer churn prediction propensity model (propensity score - covariate churn probability of customer behavior) is built and prescribes discounts or upsells/cross-sells to customers by selecting thresholds. ​

Now, business managers want to predict the effectiveness of customer churn, such as whether the company's customers are retained due to promotions or marketing activities, or the opposite? This requires traditional AB testing standard experiments ,experimentation takes some time and is not feasible and ,costly in some cases. ​

Therefore, we need to think about issues beyond the tendency model. Churn prediction with supervision is useful, but not every time because it lacks recommendations for recommending the next best action in hypothetical situations. The problem of targeting those personalized customers who are able to respond positively to a business's marketing proposition without wasting money on failure cases, thereby taking the next best action/intervention and changing future outcomes (e.g. maximizing retention) is causal inference Lift modeling in . ​

When understanding certain counterfactual questions in the consumer world, such as how would consumer behavior change if retail prices were raised or lowered (what is the impact of price on behavioral patterns)? If a business shows ads to customers, will they buy the product (the impact of advertising on purchasing)? This includes data-driven decision-making through causal modeling. ​

In general, forecasting or forecasting questions focus on how many people will subscribe in the next month, while causal questions ask what will happen if some policy changes (for example, if a How many people will subscribe to the event). ​

Causal analysis will go further. It is designed to infer various aspects of the data generation process. With the help of these aspects, one can infer not only the likelihood of events under static conditions but also the dynamics of events under changing conditions. This ability includes predicting the effects of actions (e.g., treatment or policy decisions), determining the causes of reported events, and assessing responsibility and attribution (e.g., whether event x was necessary or sufficient for event y to occur). ​

When one uses supervised machine learning to predict models using pseudo-correlation patterns, there is an implicit assumption that things will continue as they have been in the past. At the same time, the environment is being actively changed in a way that often breaks these patterns, as a result of decisions made or actions taken based on predicted outcomes. ​

From prediction to decision-making​

For decision-making, you need to find the characteristics that lead to the outcome and estimate how the outcome will change if the characteristics change. Many data science problems are causal problems, and estimating counterfactuals is common in decision-making scenarios. ​

• A/B Experiment: If you change the color of the buttons on your website, will it lead to higher engagement? ​
• Policy Decision: If this treatment/policy is adopted, how will it lead to a change in outcomes? Will this result in healthier patients/more revenue? ​
• Policy Assessment : What changes did the company make in the past or what is known until now, and how did the results change, and did the policies it created help or hinder the product it was trying to change? ​
• Credit Attribution: People buy things because See the ad? Will they buy it? ​

What is causation and causal effect? ​​​

If an action or treatment (T) causes a result (Y) , if and only if the action (T) results in a change in the outcome (Y), holding everything else constant. Causality means that by changing one factor, another factor can be changed. ​

For example: If aspirin relieves a headache, it will occur if and only if aspirin can cause a change in the headache. ​

If marketing can bring about an increase in sales, if and only if marketing activities can bring about a change in sales, then everything else can remain the same. ​

The causal effect is the magnitude of the change in Y with a unit change in T, not the other way around:​

Causal effect = E [Y | do(T=1)] - E [Y | do (T = 0)] (Judea Pearl's Do-Calculus)

Causal inference requires domain knowledge and assumptions and expertise. Microsoft's ALICE research team developed the DoWhy and EconML open source libraries to make people's work and life easier. The first step in any causal analysis is to ask a clear question: ​

• What treatments/actions are you interested in? ​
• What outcomes do you want to consider? ​
• What confounding factors may be associated with outcome and treatment? ​

Causal analysis pipeline: End-to-end causal inference (DECI) based on deep learning (Microsoft patent). ​

Causal discovery-causal identification-causal estimation-causal verification. ​

Responsible Artificial Intelligence Dashboard (Azure Machine Learning Studio): Cause Analysis

This function is based on fitting the model in the model registry explanation, one can explore what might have happened if there was a causal understanding of the same variables. The causal effects of different characteristics can be observed and compared with idiosyncratic effects, and different groups can be observed and what characteristics or policies work best for them. ​

• DECI: Provides a framework for end-to-end causal inference, which can also be used alone for discovery or estimation. ​
• EconML: Provides a variety of causality estimation methods. ​
• DoWhy: Provides multiple identification and verification methods. ​
• ShowWhy: Provides code-free end-to-end causal analysis for causal decision-making in a user-friendly graphical user interface (GUI).

Summary​

Modern machine learning and deep learning algorithms can find complex patterns in data that interpret black-box algorithms, and their interpretations may mean that machine learning algorithms learn from the world To what. ​

When these learned machine learning algorithms are applied to society to make policy decisions such as loan approvals and health insurance policies, the world it understands does not necessarily reflect the world well. What's going on. ​

However, data-driven predictive models are transparent but cannot truly explain. Interpretability requires a causal model (as evidenced by the Table 2 fallacy). Causal models reliably represent some process in the world. Explainable AI should be able to reason to make effective decisions without bias. ​

Original title: ​​Causal Analysis in Azure Machine Learning Studio to answer your Causal questions through an end-to-end automated framework​​ , Author: Hari Hara

The above is the detailed content of Make effective actionable decisions to optimize business KPIs by using causal machine learning. For more information, please follow other related articles on the PHP Chinese website!