Cluster knowledge-driven vertical federated learning

In industrial scenarios, cross-departmental collaboration is necessary to achieve quality traceability. However, data cannot be shared due to privacy concerns. Vertical Federated Learning (VFL) enables heterogeneous industrial sectors to jointly train models while preserving product privacy. However, existing traditional VFL algorithms only focus on aligning feature benefits and suffer from high communication costs and poor performance. This paper proposes a "Cluster Knowledge-Driven Vertical Federated Learning" (Cluster-VFL) algorithm, which integrates cluster intelligence to optimize heterogeneous distributed environments. In Cluster-VFL, each participant engages in training as an individual within the cluster, taking into account the utilization of non-aligned features. Cluster-VFL promotes model updates by identifying global optimal individuals and transferring global optimal knowledge. Subsequently, this knowledge is merged with the individual optimal knowledge obtained from local training of each participant. We conducted extensive experiments using an open-source diagnostic dataset and a proprietary dataset from Company A. The results unequivocally demonstrate that this algorithm enhances participants' learning abilities, improves their communication efficiency.