The client is an eCommerce platform for industrial products and office supplies. They rely heavily on digital marketing and advertising campaigns as they simplify direct selling to their customers.
However, the client was not able to derive insights to improve their investment decisions and sales. For instance, in Google Ads, the client would invest in search keywords, location, and device type levels to drive more clicks and ultimately more sales. But they couldn’t get deep insights into the past data about the campaign performance and buying patterns. Hence, it was tough to ascertain and maximize returns on their investment and redefine the campaigns for better results.
The client partnered with Persistent, a premier digital consultancy, to build and deliver an ML-model to accurately calculate and forecast sales every week. The model considers Google Adwords and Customer Relationship Management (CRM) metrics like email clicks, calls, quotes, SIC, and more.
Building a scalable, automated data ingestion pipeline to train the model
Various types of data were used to train the model, for example – keyword performance report (clicks, impressions, and other metrics for each keyword), shopping performance report (clicks, impressions, and other metrics for each shopping ad), click performance report (all the clicks from Google Ads), product inventory changes, product historical sales, product historical and current price changes, product hierarchical clickstream data, traffic by different online channels, quote pipeline of product, historical conversions, home page exposure, and more.
All these data were integrated into Google BigQuery. Common attributes were identified among the disparate data types to integrate them accurately. The integrated data set was divided into training and test data for feature engineering.
Continuous training of the ML model
The Persistent team built a neural network regression model that predicts the sales for the upcoming seven days. The model was continuously trained to reduce errors and improve prediction accuracy. Weights were assigned to various input features used during the prediction cycle and the client teams can now perform what-if analysis by changing the input features.
After training the model successfully, the training code was also packaged into a module to facilitate training jobs for future use cases.
An interactive UI application for the client teams
Persistent’s ML specialized team also delivered a light, front-end application to test and demonstrate model functioning and to run different sales prediction cycles by adjusting the input features.
Addressing the security and compliance requirements
The client wanted to enable role-based access to data. Hence, datasets specific to user groups were created and views were authorized only to the relevant groups. Also, based on the client’s infrastructure policy and recommended best practices, the Google Cloud network resources were optimized in a shared VPC setup.
Following are the Google Cloud products that were used in this engagement:
Google Cloud Storage: To store incremental clickstream data.
Google Cloud Composer: The fully managed workflow orchestration service was used to schedule the DataFlow jobs.
Google Cloud DataFlow: It was used to transform and enrich streaming and batch data and execute multiple data transformation pipelines.
Google BigQuery: It stored raw data and used the datasets for analysis and training of ML models. The training data was queried from the BigQuery for each iteration of training.
Google Cloud Machine Learning Engine: The Cloud Machine Learning Engine was used for its services, like – AI Platform Notebooks for developing the ML models and AI Platform Jobs to train the models with different hyperparameters to find the best model.
Firebase Authentication & Hosting: The application access was restricted to only Persistent and the client’s domains.
Optimizing the marketing spend and increasing the sales conversions
The client has been able to automate the ingestion of data. Now the model is continually fed with the latest data for an accurate forecast of their sales, and insights to improve their spend decisions and sales conversions.
Every business that works digitally has a huge repository of data that can be turned into a power asset if the data is consolidated for business analytics. In this case too, the client possessed day-to-day transactional and historical data. We are glad to have worked together with them in optimizing their marketing spend by integrating all their data from various marketing and sales channels and implementing a prediction model to drive their decisions. We consider the moment when the implementation won a well-recognized technology innovation award as the highlight of the whole engagement
Asheesh Sharma, Google Business Unit, Persistent
