Using Hybrid Recommender Systems to Predict Customer Needs
Customers are engaging online more frequently, in turn accelerating e-commerce, entertainment, and other real-world platforms. Developing personalized offerings with a better understanding of customers can help companies differentiate themselves and provide customers with exactly what they are looking for. They can also drive innovation, as companies continuously learn and improve from user experiences.
Recommendation systems are information filtering systems that predict user responses to large and complex information spaces, helping to narrow down possibilities. They are part of the future of personalization to optimize product offering and the customer experience. They make multiple, possible touchpoints of a customer journey possible.
In the latest work, Ramin Raziperchikolaei and Young Joo Chung of RIT San Mateo have been researching recommender systems and ways to gather critical insights from past user experiences and behaviors. To develop a better understanding of user demands and item features, they focus on a hybrid recommender system drawing from side information, which includes content such as category and title, and user feedback. Hybrid RS is proposed to solve the cold-start problem where the system is not able to recommend items to users. They use content because cold-start items do not have user feedback.
The key contribution of this paper is proposing an efficient method called Neural Representation for Prediction. Hybrid RS methods mainly use autoencoders to create latent neural representations for users and items. Authors found that existing hybrid RS methods were utilizing such neural representation for regularization, not for prediction. By proposing a model making prediction with this neural representation, NRP is able to achieve faster training with fewer computation cost, while maintaining the recommendation performance.
We here at Rakuten continue to be excited by the potential to support researchers in developing systems that can build capabilities to enhance the customer journey and to nurture the future into reality. There is still a lot of work—this work does not yet represent a system that can be fully implemented in a customer ecosystem. While the model can make better predictions than many other expert models, there are many factors to consider for a system to be more impactful in real-world application. Designing recommender systems is a popular topic with infinite possibilities. While NRP model outperformed many other expert models, there are many unexplored techniques and area like graph network and counterfactual inference that can help better recommendation. At Rakuten, we are continuing to improve the RS to provide better user experience.
Please see the paper for all acknowledgments and further details on the work. In addition, the implementation code is available through the github repo found below. We would like to thank ACM SIGIR who helped bring the attention of this research to a wider audience.
Read the Neural Representations in Hybrid recommender systems: prediction versus regularization here
Check out the github repo here