Leveraging Machine Learning and Optimization to Efficiently Allocate Personalized Vouchers at Scale for Gojek

Core Concepts

Gojek developed a machine learning-driven, multi-objective solution to efficiently allocate personalized vouchers to millions of customers across multiple geographies, maximizing business objectives while adhering to budget constraints.

Abstract

The article discusses how Gojek built a machine learning-driven voucher allocation engine to serve millions of customers across multiple geographies. The key highlights are:

Identifying the "Persuadables" - customers who make more transactions when targeted with a voucher, as opposed to the "Sure Things", "Lost Causes", and "Do Not Disturbs".
Using a deep learning-based causal inference algorithm to predict the uplift in business objectives and cost for each customer given a voucher. These predictions are then fed into a knapsack optimizer to recommend the optimal voucher allocation to maximize the business objective while adhering to the budget constraint.
Leveraging data transformation tools like dbt to efficiently extract and process thousands of customer features from hundreds of source tables, with extensive data testing and monitoring using tools like elementary.
Employing best practices for scalable code development, such as CI/CD, code testing, and configuration management using Hydra.
Benefiting from Gojek's in-house platforms like Merlin (ML deployment) and Campaign Portal (voucher allocation at scale) to enable the efficient implementation of the voucher allocation solution.

Stats

Gojek serves millions of customers across multiple geographies.
The data science team manages customized hyperparameter configurations for ML models across multitudes of on-demand services.

Quotes

"We use historical data of customers to observe past effects of vouchers on them. This is a typical causal inference problem to measure effect (for example, incremental transactions) of treatment (voucher) for our customers."
"The 'objective' in the problem formulation depends on the business use case (multi-objective)."
"dbt is a wonderful tool!"

Key Insights Distilled From

How Gojek Allocates Personalised Vouchers At Scale

by Praveen Pras... at medium.com 07-17-2024

https://medium.com/gojekengineering/how-gojek-allocates-personalised-vouchers-at-scale-41cad5d6f218

Deeper Inquiries

How does Gojek's voucher allocation solution adapt to changing customer behavior and preferences over time?

Gojek's voucher allocation solution adapts to changing customer behavior and preferences over time through the use of a Machine Learning (ML) driven engine. By categorizing customers into different groups such as Persuadables, Sure Things, Lost Causes, and Do Not Disturbs based on their response to vouchers, Gojek can tailor its voucher allocation strategy to target the most responsive customers. This segmentation allows Gojek to identify patterns in customer behavior and adjust its voucher allocation strategy accordingly. Additionally, Gojek's causal inference algorithm helps in measuring the effect of vouchers on customers, enabling the company to understand how customer behavior changes over time in response to different voucher offerings. By continuously analyzing and updating customer data, Gojek can refine its voucher allocation strategy to better align with changing customer preferences and behaviors.

What are the potential ethical considerations and risks associated with using personalized vouchers and causal inference techniques at scale?

When using personalized vouchers and causal inference techniques at scale, there are several potential ethical considerations and risks to be aware of. One major concern is the privacy of customer data, as personalized vouchers require access to sensitive information about individual customers. Ensuring that customer data is handled securely and in compliance with data protection regulations is crucial to mitigate privacy risks. Additionally, there is a risk of algorithmic bias, where the ML models used for voucher allocation may inadvertently discriminate against certain groups of customers based on factors like demographics or past behavior. It is important to regularly audit and monitor the ML models to detect and address any biases that may arise. Furthermore, there is a risk of unintended consequences, where the use of personalized vouchers could lead to unintended outcomes such as customer manipulation or unfair treatment. Implementing transparency and accountability measures in the voucher allocation process can help mitigate these risks and ensure ethical use of personalized vouchers and causal inference techniques.

How could Gojek's approach to scalable data engineering and ML deployment be applied to other business domains beyond voucher allocation?

Gojek's approach to scalable data engineering and ML deployment can be applied to other business domains beyond voucher allocation by leveraging similar techniques and tools in different contexts. For example, the use of a data build tool like dbt for efficient data transformation can be beneficial in various business domains that require processing large volumes of data from multiple sources. By implementing dbt for data transformation, businesses can streamline their data processing pipelines and ensure data reliability across different domains. Similarly, the use of a machine learning platform like Merlin for ML model deployments can be applied to other domains that rely on predictive analytics or AI-driven solutions. By adopting a standardized ML deployment platform, businesses can accelerate the development and deployment of ML models in various domains. Additionally, practices such as enforcing code testing, tracking test coverage, and implementing CI/CD pipelines can be valuable in ensuring code quality and efficient collaboration in different business domains. Overall, Gojek's approach to scalable data engineering and ML deployment offers a blueprint for implementing advanced data analytics and ML solutions in diverse business contexts beyond voucher allocation.