How an Australian telco uses digital personalisation to target customers

A major Australian telco was looking to boost retail sales by optimising digital onsite personalisation—which was becoming an increasingly important component of its overall sales revenue.

Its marketing team hoped to use machine learning to target website visitors with relevant content and use advanced insights to A/B test its way to success. Servian brought the marketing team on a journey that delivered the machine learning personalisation infrastructure that integrated with their site and web marketing stack, while upskilling their staff on how to utilise and manage the machine learning system.

the situation

Based on recent experiences, the telco wanted to develop its own capability in machine learning that would be impactful and transparent and could be tuned to make use of its diverse first and third party data sources.

Google Cloud’s data and machine learning infrastructure in combination with Google Marketing Platform provided a compelling value proposition as it simplified data ingestion, development of machine learning systems and provided the channels to personalise and measure the impact.

the framework

The importance of being in control of its own intellectual property was a neat fit for Servian’s six-step ‘ProductionML’ process, which helps to uplift the internal capacity for machine learning capability.

ProductionML is a six-step process which seeks to work collaboratively with customers to build internal machine learning capability with its people while using a first use case as a demonstration of the technical possibilities of both machine learning and cloud technologies.

The six steps are:

  • Step 1 – Business Impact Workshop
  • Step 2 – Model Development
  • Step 3 – Solution Architecture
  • Step 4 – Model: Evolving Dev to Prod
  • Step 5 – Governance
  • Step 6 – Model: Operate and Innovate

the process

Servian delivered the six workshops plus an additional workshop at a pace determined by the customer. The approach was adjusted to explore architectures and machine learning through a series of hypotheses.

Two machine learning models were developed using Jupyter style notebooks and related Python ML libraries to build an intent model. In parallel, a full personalisation pipeline was implemented including 

data ingestion, data transformation, model decision and activation through API to Google Analytics and Google Optimise 360 for measurement.

As a demonstration of the end-to-end capability in production machine learning, a full CI/CD pipeline was deployed onto Google Cloud which included model management and tracking capabilities with Dataflow used as the basis of a Data Pipeline and ML Engine for serving recommendations.

the architecture

platform used

To achieve the outcomes required, the following GCP products were used:

The data was sourced from Google Analytics, which helps provide a baseline for future datasets including DoubleClick and Google Adwords.

the results

Key elements of the design included being able to feed the recommendations back into the A/B testing tool of choice. In addition, in line with meeting their technology preferences, the serverless architecture that was implemented limited the amount of technical infrastructure that the client needed to maintain and minimised the cost.

The telco is now in their optimise phase, using the personalisation machine learning model in experiments with their site content and planning for iterative improvement of the site, content and model in order to drive growth.