The model needs to be moved from the benchtop to production. It is here we address the concerns around:
– How do we expose our recommendations to other systems?
– How do we scale the model up to meet the demands of the existing systems?
– How do we keep our model current?
– How do we adapt to the impacts of our interventions on future data?
– How do we measure the impact of the new algorithm?
– How do we roll it out – big bang orf incremental deployment?
– What are the change impacts on the people who use the systems and what to be communicated to them?
Typically, we write the models in Python and bundle them up in Docker Containers to deploy via the client’s favourite container orchestrator (Kubernetes, Amazon Container Services, Swam). For larger-scale problems, we tend to use Amazon Sagemaker as our workhorse. The choice, however, is normally determined by the nature of the problem and the capabilities and preferences of the client.
3. Design the Features
We analyse which existing features are important for understanding the supply, demand and matchmaking problems. We ask the question, what data do we wish we had? What data would transform our understanding of the customer?
1. Orientation
We start by meeting the team, gaining access to the data and getting a feel for the current approach to the problem.
2. Framing the problem
We workshop ideas to bring alignment between the different stakeholders who have an interest in the algorithm so that we all leave with the same definition of what a success algorithm looks like and we have clarity on what we are optimizing for.
4. Truly Understand What Drives the Customer To Choose a Particular Supplier
The essence of our solution is the ability to predict what a given customer will do when presented with a particular product. Will the user read the article? Will they buy this product? Will they enquire about that car?
Customer-obsessed businesses measure the accuracy with which they can predict these events. Through the use of statistics, we can quantify the degree to which we actually understand our audience and we can monitor our improvement over time.
5. Simulate different recommendation algorithms
Now that we can reliably predict the behaviour of our typical customers we can experiment with different matchmaking algorithms to find the one that maximises the value to the business.
Taking an empirical, bottom-up approach to simulating ensures that we capture the nuances surrounding the arrival of different types of customers, taking into account the real-world complexities of finite and variable supply.
We can articulate the value that we will capture when the best algorithm is deployed.
1. Deployment and Monitoring
We need to monitor the impact of the interventions to make sure that the business is following the new process and that people are acting on the predictions
We need to monitor the performance and reliability of the deployment. Is engineering stable? Is the recommender playing nicely with the other systems?
