India’s biggest sports gaming platform Dream11 is personalising its fantasy sports experience for users with artificial intelligence (AI) and machine learning (ML) capabilities in collaboration with software company Databricks.
Dream11, operating under Mumbai-based parent startup Dream Sports is a unicorn with over 150 million users playing an array of sports like cricket, football, hockey, kabaddi, handball, basketball, volleyball, rugby, futsal and baseball online.
The gaming app revolves around participants forming virtual teams of real players and competing based on the statistical performance of those players in actual games. It also has partnerships with several national and international sports bodies.
Dream11’s assistant vice president for data science, Aditya Prasad told a recent Databricks conference that the company's objective has been to personalise the app and help users discover relevant products.
“Personalisation was not just a need of the hour, but a basic expectation of our users,” he said.
Dream11 has struggled to manage 120 million requests per minute(rpm), over 10 million concurrent users and peak data collection volume of 25 terabytes per day.
The firm has also seen a surge in multiple user touchpoints and conversation rates on the app.
Managing the user experience and ensuring efficient operation was a big challenge with high revisit rates of users, Prasad said.
The firm also lacked the ability to develop recommendation engines to personalise each user journey, offering customised contests and content efficiently.
“We needed a reliable AI platform that could provide contextual recommendations in inferencing contests and products in real-time,” he added.
The firm decided to leverage Databricks in crafting personalised, contextual, and timely engagement campaigns to drive higher app monetisation and retention.
Personalisation
Dream11 has followed multiple personalisation approaches running several different algorithms ranging from recency frequency monitoring (RFM), collaborative filtering, factorisation machines, LearningtoRank, deep neural networks(DNNs) and contextual bandits.
It has drawn unique benefits out of these approaches powered with Spark and TensorFlow on Databricks, Prasad said.
For instance, through RFM, the platform can keep track of its users’ activity and monetary values to help Dream 11 curate user-centric campaigns, he added.
"These algorithms are adapted into our workflows, which gets user concurrency of 10 million plus… they are all powered with GPUs for training and influencing processes,” Prasad explained.
It has chosen server-side inference rather than user/client or edge inferencing. “This has given us the versatility to experiment,” he added.
The company's personalisation model architecture with multi-task heads has effectively improved relevance and engagements based on specific historical and real-time in-app user activity, Prasad said.
Moreover, the company has also scaled up its model training process and production inferencing with insights from Databricks.
Prasad said the adoption of Databricks over the past three years has “significantly” helped the firm in achieving new milestones in user concurrency and rpms.
Dream11 can now ramp up more personalised features and decrease time to production, he concluded.
