Fenton Zhang, Data Scientist, Customer & LCM Analytics
Organization: Air Canada
Air Canada is the largest provider of scheduled passenger services in the Canadian market as well as in the Canada-U.S. transborder and the international markets to and from Canada. We carried more than 37 million customers in 2022, to 185 direct destination world-wide. This represents close to 1,000 flights daily.The company grew to close to 36,000 active employees in 2022.
Aeroplan’s membership has already exceeded its 7M active members target. The Customer & Loyalty Analytics team has the privilege to aim at better understanding these customers and Aeroplan Member behaviours so that we can serve them better!
Best Acceleration Use Case
Best Data Democratization Program
Best ROI Story
The Customer and Loyalty Analytics Team works closely with our Marketing Team to design and launch marketing campaigns. We know that increased relevancy and personalization increase marketing effectiveness.
Therefore, one of the C&L Analytics mandates is to produce signals through predictive models, customer segmentations, flags, or recommender systems, that can be leveraged in our marketing campaigns. Signals can be quite intensive to develop, from the business problem understanding, to data understanding, data preparation, modeling, evaluation, and deployment of scoring for accessibility in campaign deployment.
Our objective was to reduce the time to build an MVP of a signal, so that we can have access to more signals and increase our collective customer intelligence.
To solve for the data understanding and data preparation steps, we have created a Customer 360 solution in Snowflake, which consists in pre-crunching hundreds of features at the customer level, ready to use for analytics or BI purposes.
The idea is to democratize the intelligence that we build about our customers. Any logic that a data scientist would run in their code every single time is being crunched once and everyone can benefit from it, instead of reinventing the wheel for every project. This also allows for better governance and for having a single source of truth. We leveraged Dataiku’s visual flows, AutoML & custom ML to connect Customer 360 to Dataiku so that we can easily and quickly fetch multiple potential features and train new predictive models, customer segmentations, and recommender systems in hours instead of week/months.
This solution allows us to quickly get to an MVP, on which we can decide to invest more time (or not) thus increasing performance and accounting for diminishing returns.
Once models are created, all scores are fed back to the Customer 360 storage table, leveraging MLOps, then showcased as profiling variables in the standard profile and democratized to the Marketing community.
Our Marketing Stakeholders now have more signals that they can play with in their marketing campaigns.
We also have more profiling variables available to enrich our standard profiles, which help increase relevancy of our marketing communications.
By creating signals more quickly, Data Scientists can create more signals which add even more value.
Business Area Enhanced: Marketing/Sales/Customer Relationship Management
Use Case Stage: Built & Functional
80% of the total effort to build a predictive model taking weeks/months is now reduced to hours of marginal efforts. This enables us to quickly evaluate the quality of the model and decide whether incremental effort is needed to improve it.
By having quicker access to a wider variety of signals, we can further enrich standard profiles. Marketing stakeholders now have access to more profiling variables in the standard profile, which helps increase the relevancy of our marketing communications.
Opportunity cost saving: Data Scientists now have more time to invest in more added value projects, such as building more models, whose insights feed back into a more accurate customer understanding. A virtuous flywheel!
Value Brought by Dataiku:
Dataiku enabled the orchestration of the processes, allowing us to automate, create the flow and operationalize our models in an efficient manner.