Causal inference and model explainability analysis of a real world sales dataset

arXiv:2512.12605 · cs.LG · Submitted 2025-12-14 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Causal inference and model explainability analysis of a real world sales dataset".

Jane: Here's the gist of the paper. The gist The authors aim to provide a recipe for applying model interpretability and causal inference for deriving sales insights by reviewing existing literature and applying it to a real-world retail dataset, finding that an inherently explainable model has a lower variance of SHAP values, and showing that including multiple confounders through a double machine learning approach allows us to get the correct sign of causal effect.

Tom: Model Interpretability Model interpretability can be at 2 levels: local and global, where local interpretability is a feature-dependent prediction, while global interpretability measures are independent of x<ref:2512.12605#pg4>.

Jane: Whether to use local or global interpretability measures depends on the application<ref:2512.12605#pg4>.

Tom: Another classification of interpretability is into post-hoc and inherently interpretable modeling, where post-hoc interpretability is more commonly used when a model optimizes accuracy and then invokes necessary tools to explain predictions<ref:2512.12605#pg4>.

Title and authors: Tom: Pranav Gupta and Nithin Surendran are the authors here, and the title really tells you they're tackling a problem where you need both understanding *and* cause-and-effect in sales data.

Jane: It’s about moving past just knowing what happened to figuring out what would happen if we changed something, which is the whole point of causal inference.

Lu: The paper sets up the context by pointing out that while most retail models lack interpretability and can't discover causal links, they are collecting tons of data from marketing, supply chain, and customer experience divisions.

Meng: So they’re starting with this big problem: you have these complex systems driving sales metrics but you can’t trust the predictions or find the real drivers.

Lalam: They introduce the idea of a recipe for applying both interpretability and causal inference together to derive those sales insights from a real-world dataset.

The paper's summary: Tom: The core finding they report is that an inherently explainable model has lower variance in its SHAP values compared to some other models they tested, even when comparing it against models like the explainable boosting regressor.

Jane: That’s interesting because usually, you trade off accuracy for interpretability, so finding a way to keep the explanation stable while keeping the model useful is important.

Lu: They also show that using a double machine learning approach, by including multiple confounders—things that can mess up the relationship—they can get the correct sign of the causal effect.

Meng: Including more confounders sounds like it adds complexity to training, but if it helps you get the right direction for a cause-and-effect analysis, that’s valuable for decision making.

Lalam: They specifically mention single-link hierarchical clustering on features to find redundancy between things like customer perception and lack of seller responsiveness, which helps them control for those variables.

The paper's improvements: Tom: One major suggestion they make is that we need to move beyond just using Shapley values when we want to discover causal relationships because the authors suggest they aren't entirely reliable for that purpose.

Jane: So what’s the fix then? They point toward integrating SHAP values into other explanation techniques, like c-SHAP, which sounds like a way to ground those explanations in actual causal knowledge.

Lu: They also suggest expanding the work by using computationally efficient methods for calculating SHAP values and for estimating DAGs, which is Directed Acyclic Graphs.

Meng: I’m thinking about deployment—if we want this to be production ready, we need methods that are fast enough to run on live data without taking too much time or memory.

Lalam: They also suggest using single-link hierarchical clustering on features to identify observable confounding, which helps select the right variables for their causal inference analysis.

Conclusion: Tom: So, to wrap up, the paper "Causal inference and model explainability analysis of a real world sales dataset" shows that inherently explainable models have lower variance in SHAP values, and by using multiple confounders via double machine learning, you can get the correct sign of causal effect.

Jane: It suggests that for high-stakes decisions in retail, we should prioritize inherently explainable models when interpretability is critical because they provide more stable explanations.

Lu: The next steps involve integrating SHAP into causal inference methods and using more efficient ways to calculate those values and estimate the DAGs.

Meng: For us at the startup side, it means we need to be careful about which features we control for when we try to isolate a cause-and-effect relationship in our models.

Lalam: I'm also sure that this work on understanding how these tools interact is a solid foundation for how we can build better analytical systems for sales insights.

Pranav Gupta, Nithin Surendran

Lowe’s Companies Inc

cs.LG

Submitted: 2025-12-14

Updated: 2026-10-16

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 75/100

The gist: The gist The authors aim to provide a recipe for applying model interpretability and causal inference for deriving sales insights by reviewing existing literature and applying it to a real-world

Key concepts

Model Interpretability: This involves understanding *why* a predictive model makes certain predictions. It can be local (explaining one specific prediction) or global (understanding the model's overall behavior). The paper focuses on SHAP values, which are a tool used to explain feature contributions in complex models.
Causal Inference: This helps answer 'what if' questions by determining if a change in one variable actually *causes* a change in sales. Double machine learning is used here; it trains one model to handle confounding factors and then another model to estimate the true causal effect of the feature of interest.
Double Machine Learning: This is a technique for observational causal inference that uses two separate machine learning models sequentially. The first model helps remove the influence of confounders from the feature being studied, and the second model estimates how changing that feature truly affects sales, ensuring a more reliable causal sign.
SHAP Values: Shapley Additive Explanations is a method rooted in cooperative game theory used to explain predictions. The authors found that inherently explainable models produce SHAP values with less variance than other models, suggesting they offer more reliable explanations for critical decisions.

Terminology

Causal Inference — 3× in this paper · explained in 5 episodes
Interpretability — 3× in this paper · explained in 15 episodes
Explainability — 1× in this paper · explained in 4 episodes
SHAP Values — 1× in this paper · explained in 4 episodes

Summary

The gist The authors aim to provide a recipe for applying model interpretability and causal inference for deriving sales insights by reviewing existing literature and applying it to a real-world retail dataset, finding that an inherently explainable model has a lower variance of SHAP values, and showing that including multiple confounders through a double machine learning approach allows us to get the correct sign of causal effect.

Model Interpretability

Model interpretability can be at 2 levels: local and global, where local interpretability is a feature-dependent prediction, while global interpretability measures are independent of x<ref:2512.12605#pg4>. Whether to use local or global interpretability measures depends on the application<ref:2512.12605#pg4>. Another classification of interpretability is into post-hoc and inherently interpretable modeling, where post-hoc interpretability is more commonly used when a model optimizes accuracy and then invokes necessary tools to explain predictions<ref:2512.12605#pg4>. For example, Shapley Additive Explanations (SHAP) are model agnostic, grounded in cooperative game theory<ref:2512.12605#pg4>. The paper focuses on SHAP because of its comprehensive capabilities and additive nature<ref:2512.12605#pg4>.

Causal Inference

Causal inference helps in achieving the objective of understanding counterfactuals, ensuring that the final metric changes “because” of the change in the variable of interest<ref:2512.12605#pg5>. Double machine learning is a flexible tool for observational causal inference applications, which uses confounders to first train a model to deconfound the feature of interest, and then trains another machine learning model to estimate the average causal effect of changing that feature<ref:2512.12605#pg4>. This effect would be measured as a slope between the feature of interest and a target, e.g., sales<ref:2512.12605#pg4>. Double machine learning is not suitable for cases with unobserved confounders<ref:2512.12605#pg4>.

Experimental Findings on Interpretability

The authors trained multiple tree-based regression models between the features F1 to F7 and Y, reporting loss as the mean absolute error divided by the average of the target Y, called mean scaled absolute error (MSAE)<ref:2512.12605#pg5>. Although the explainable boosting regressor has a worse MSAE when compared to other “black-box” models, SHAP values for the inherently explainable regressor have less variance<ref:2512.12605#pg7>. Furthermore, the qualitative features of the SHAP scatter plots do not change significantly across the 3 models<ref:2512.12605#pg7>.

Experimental Findings on Causal Inference

The authors performed single-link hierarchical clustering on the features to measure feature redundancy, finding that there is redundancy between F5 (negative customer perception) and F6 (lack of seller responsiveness), and they have some redundancy with F4 (out-of-stock) and F2 (underperforming products)<ref:2512.12605#pg8>. Controlling for more confounding variables leads to the correct sign of causal effect, i.e., higher the unresponsiveness F6, lower the sales Y<ref:2512.12605#pg9>. By controlling just F5, they got the wrong sign of the causal effect (+ve slope instead of the expected-ve slope) in the left plot<ref:2512.12605#pg9>.

Conclusion and Further Work

The paper concludes that Shapley values are not reliable when it comes to discovering causal relationships<ref:2512.12605#pg8>. Further work could involve the integration of SHAP values into causal inference, for example, c-SHAP<ref:2512.12605#pg8>, and expanding to more features while using computationally efficient methods for SHAP values and DAG estimation<ref:2512.12605#pg8>. The findings in this paper generalize to other use cases in retail too<ref:2512.12605#pg4>.

--- Page 1 ---

Causal inference and model explainability tools for retail <ref:2512.12605#pg2> Pranav Gupta Lowe’s Companies Inc Mooresville, NC, USA pranav.gupta@lowes.com Nithin Surendran Lowe’s Companies Inc Mooresville, NC, USA nithin.surendran01@lowes.com (corresponding) Abstract Most major retailers today have multiple divisions focused on various aspects, such as marketing, supply chain, online customer experience, store customer experience, employee productivity, and vendor fulfillment. They also regularly collect data corresponding to all these aspects as dashboards and weekly/monthly/quarterly reports<ref:2512.12605#pg2>. Although several machine learning and statistical techniques have been in place to analyze and predict key metrics, such models typically lack interpretability<ref:2512.12605#pg2>. Moreover, such techniques also do not allow the validation or discovery of causal links<ref:2512.12605#pg2>. In this paper, we aim to provide a recipe for applying model interpretability and causal inference for deriving sales insights<ref:2512.12605#pg2>. In this paper, we review the existing literature on causal inference and interpretability in the context of problems in e-commerce and retail, and apply them to a real-world dataset<ref:2512.12605#pg2>. We find that an inherently explainable model has a lower variance of SHAP values, and show that including multiple confounders through a double machine learning approach allows us to get the correct sign of causal effect<ref:2512.12605#pg2>. 1 Introduction Large-scale retail enterprises have complex ecosystems that involve marketing, supply chain management, online and in-store customer experiences, employee management, and vendor relations<ref:2512.12605#pg4>. The interplay among these divisions directly influences the overarching success metric of retail enterprises– sales<ref:2512.12605#pg4>. Hence, retailers regularly collect extensive data across these divisions to drive decision-making processes<ref:2512.12605#pg4>. Executives in such enterprises are not only concerned with the “factuals,” i.e., the data for the current reality, but also the “counterfactuals,”[19, 23] i.e., what if they invested in marketing? What if they invested more in vendor management?<ref:2512.12605#pg4>. Retailers frequently use several machine learning models to analyze and predict key performance metrics<ref:2512.12605#pg4>. However, these models often lack interpretability, making it difficult for stakeholders to understand the underlying mechanisms driving predictions<ref:2512.12605#pg4>. Furthermore, existing techniques fall short in facilitating the validation or discovery of causal relationships between different operational aspects and sales metrics<ref:2512.12605#pg4>. Given the complexity of retail enterprises, it is extremely difficult to generate a realistic counterfactual scenario for feeding into a predictive model, even if the model approaches nearly 100% performance during training and evaluation<ref:2512.12605#pg4>. For example, imagine a CEO whose data science team has built a highly accurate model that predicts sales as a function of warehouse inventory, marketing spend, and several other features<ref:2512.12605#pg4>. The CEO wants to understand the intuition behind the model’s decisions, but her team tells her that the model is a deep neural network, whose hidden layer weights and intermediate hidden states cannot be directly mapped to an intuitive understanding<ref:2512.12605#pg4>. While model interpretability is less critical in certain e-commerce applications, in this case it is much more critical because of the high stakes associated with each executive-level decision<ref:2512.12605#pg4>. There is another major problem here<ref:2512.12605#pg5>. Let us say that the model shows that increasing the marketing spend by 5% will lead to a 10% increase in sales<ref:2512.12605#pg5>. However, we do not know what the actual inventory in the “counterfactual” scenario of increased marketing spend<ref:2512.12605#pg5>. Maybe promotions will cause an increase in demand and the demand will reduce the inventory<ref:2512.12605#pg5>. There could be partial redundancy between inventory and marketing spend, or a confounding aspect that drives both inventory and marketing spend<ref:2512.

Improvements for AI systems

improve model interpretability by using inherently explainable models when high stakes are involved, as an inherently explainable model has a lower variance of SHAP values. This allows for more reliable explanations in critical decision-making scenarios.
implement double machine learning to derive the correct sign of causal effect by including multiple confounders through a double machine learning approach. This ensures that the conclusion about whether a change in one variable causes sales is robust against confounding factors.
develop systems capable of predicting accurate counterfactuals by accurately modeling the generative process, as this is needed before feeding counterfactuals into predictive models. This moves beyond simple prediction to understanding how variables interact to create future states.
integrate causal shapley values into existing explanation techniques to address the limitation that Shapley values are not reliable when it comes to discovering causal relationships. This allows for explanations that are grounded in actual causal knowledge rather than just feature importance.
use single-link hierarchical clustering on features to identify observable confounding such as redundancy between features, which can create misleading SHAP values. This structural understanding helps in selecting the correct features for causal inference analysis.

Abstract

Most major retailers today have multiple divisions focused on various aspects, such as marketing, supply chain, online customer experience, store customer experience, employee productivity, and vendor fulfillment. They also regularly collect data corresponding to all these aspects as dashboards and weekly/monthly/quarterly reports. Although several machine learning and statistical techniques have been in place to analyze and predict key metrics, such models typically lack interpretability. Moreover, such techniques also do not allow the validation or discovery of causal links. In this paper, we aim to provide a recipe for applying model interpretability and causal inference for deriving sales insights. In this paper, we review the existing literature on causal inference and interpretability in the context of problems in e-commerce and retail, and apply them to a real-world dataset. We find that an inherently explainable model has a lower variance of SHAP values, and show that including multiple confounders through a double machine learning approach allows us to get the correct sign of causal effect.