Multi Interest

Multi-interest learning aims to model users' diverse preferences by representing them with multiple interest vectors, rather than a single vector, significantly improving the accuracy of recommendation systems. Current research focuses on developing sophisticated model architectures, such as multi-tower networks and hierarchical co-networks, that effectively disentangle and learn these diverse interests, often incorporating techniques to address issues like spurious correlations and out-of-distribution generalization. This approach leads to more personalized and accurate recommendations, impacting various applications like micro-video matching and e-commerce, as demonstrated by improved metrics such as recall and gross merchandise value in real-world deployments.