Unveiling Trust: How Explainable AI Elevates Marketing Analytics Credibility
Artificial intelligence (AI) is being used extensively to automate tasks, analyze data, and make predictions that impact our daily lives. However, with such a growing influence, it’s crucial to understand why AI systems make choices they do.
For an AI model to be reliable, it must be interpretable. Unfortunately, many current AI models, known as Black Box Models, lack interpretability. These models are unable to explain the reasoning behind their decisions, leading to potential problems.
One notable example occurred in 2015 when a black box model algorithm used in Google Photos mistakenly labeled black people as “gorillas,” highlighting the dangers of opaque decision-making in AI systems.
Explainable AI aims to solve this problem by providing explanations for AI outputs. It’s like a doctor explaining your symptoms, test reports and why they think you’re sick, rather just giving you medicine without explaining the diagnosis.
XAI helps us in ensuring that the designed AI models are not only accurate but also accountable, ethical, and trustworthy. By doing so, we can ensure that AI continues to benefit us while avoiding the potential for catastrophic consequences.
Explainable AI in Marketing
In today’s marketing era, Artificial Intelligence (AI) and Machine Learning (ML) algorithms are driving forces behind data analysis, predicting customer buying behaviors, and creating personalized campaign strategies. However, the lack of transparency in these ‘black box’ models create lack of confidence in the predictions, leaving marketers with limited understanding of why an algorithm recommends specific campaigns or predicts certain customer behavior.
To address the lack of transparency and accountability in such models, Explainable AI (XAI) provides transparent insights in understandable formats, enabling marketers to understand the reasoning behind the algorithm’s recommendations, identify potential biases and align the recommendations with marketing goals.
The benefits of XAI go beyond transparency, its helps:
- Build trust and credibility: It provides understandable insights into the predictions made by algorithm.
- Mitigating bias: XAI helps identify any Potential biases in data and adjust the model to ensure fairness.
- Enhanced Campaign Performance: It Allows marketers to refine results and devise more personalized strategies.
With XAI, marketers can confidently justify their campaign strategies, build customer trust, and incorporate predictions into their strategies in a more efficient and effective manner. It’s time for marketers to embrace the power of XAI and revolutionize the marketing landscape.
XAI Methods
Explainable AI is not a single technique, it is a much wider term that encompasses many approaches to explain the AI models. Some techniques provide explanations for particular data points (local explanations) while some provide global explanations, some are specific to the underlying models while some work independent of the underlying model (model-agnostic).
SHAP (SHapley Additive exPlanations) is a widely used XAI technique. It is a very versatile tool that can provide explanations at a global as well as local level. It is a model-agnostic technique. It uses Shapley values to calculate contributions of various features towards the output.
Use Case
We are often asked by our clients and partners after seeing the results — what makes Skewb models forecast with over 95–96% accuracy whereas other AI models don’t.
We thought to run an independent study with all types of models to compare the results but most importantly — compare Skewb’s outcomes with other Black Box AI models claiming to be using advanced statistical techniques and Algos.
We analyzed different consumer product categories (Like Daily consumables, Dairy, Bakery, Sweets, Fruits, Vegetables, etc.) and took customer level data containing various features depicting Demographics, Behavior, Purchase Product Category, Advertising data and Sales Channels. This data depicts past campaigns, and past purchases of the customers in different categories. We had to predict the customer response to a future campaign.
We applied several explainable models, a black box model, and Skewb’s own explainable model. The results are as follows.
Performance of different Classification techniques:
We observed that the black box model’s predictive accuracy surpassed that of logistic regression and decision tree models. However, we still lack interpretability regarding which features influence user response to the campaign.
On applying SHAP on the black box model and analyzed the global feature importance, we could quantify the feature importance in determining user response to campaign.
This helped us conclude that the behavior of customer plays a major role in predicting customer reaction, but still we were losing many potential customers who could respond.
At Skewb, we deploy ensemble modeling techniques that are inherently explainable to determine feature responsiveness. As a result, our prescriptive models outperform other techniques. The feature contribution provided by this model is as follows:
Here, we observe that while “behavior” retains the highest contribution, its significance has decreased compared to the black box model, while “sales channel” now holds a substantial contribution. Customers who make purchases both in-store and online are more likely to respond, suggesting that targeting them in these channels is beneficial. Additionally, features such as “complaints,” which were previously considered insignificant by the black box model, are now deemed important.
The high accuracy and interpretability of our model enable us to target customers more effectively, avoiding spending resources on customers or channels that are less likely to yield a response.
It’s very important to remove biases in any analytical model and explain the reasons behind any ups and downs to predict with utmost accuracy. This is our secret sauce of building trust and credibility by delivering explainable outcomes for any problem.
