Optimizing Recommendations for a Mobile Trading Platform

Onebyzero developed a custom, ML-driven recommendation engine for Kelly, leveraging open-source libraries for greater control and incorporating factors like event popularity, recency, and CLTV. The system uses multiple models and a learning-to-rank algorithm to refine recommendations based on user feedback, with A/B testing to optimize performance and maximize long-term profits.

Kelly is an opinion trading platform where users trade binary contracts based on beliefs about real-world outcomes. Its key challenge is user engagement, driven by a recommendation engine that suggests relevant topics for trading. Recommendations are based on user interests, past trades, event popularity, trade volume, and recency, with a focus on providing real-time suggestions to boost activity and revenue.

Onebyzero built a custom ML-driven recommendation engine for Kelly on AWS, opting for open-source libraries over AWS Personalize for greater control and to incorporate factors like event popularity, recency, and time-to-close. The solution includes multiple models (collaborative filtering, item-based and user-based similarity metrics) and a learning-to-rank algorithm that adjusts factors based on user feedback. An ensemble model combines scores from different models while factoring in Customer Lifetime Value (CLTV) to maximize long-term profits. A/B testing was used to optimize the recommendation approach.