Decentralized Classification in Sensor Networks via Sparse Representation and Constrained Fractional Programming

This paper investigates the problem of decentralized classification algorithm in sensor networks, i.e., the data about the agriculture are sensitive (or private) stored in the sensor network nodes without sharing or collecting them together. The contributions of this paper are: i) two new classification algorithms are proposed based on the sparse representation and constrained fractional programming. One is for the centralized environment while the other is for the decentralized environment, where the decentralized network node is able to process its own data to extract useful information by implementing some local computation, communication, and storage operations. Especially, by collaborating with its neighbors, it can achieve the same classification result compared with that of the centralized algorithm; ii) to reduce the redundance of the original data, we form a new classification strategy by combining the sparsity transform with the classifier; iii) to improve the robustness of the classifiers, we construct a constrained fractional programming to enforce the discriminant ability of the classifier so that the transformed coefficient vector should be closer to the class center of itself but being far away from centers of other class; iv) to handle the proposed centralized/decentralized classification problems, we decouple the constrained fraction via the Dinkelbach algorithm and alternating minimization. Especially, we cast the decentralized classification problem as a set of subproblems with consensus constraints so that it can be solved in the network environment via the inherent information communication and local computation ability of alternating direction method of multiplier. Finally, numerical examples are provided to verify the effectiveness of the proposed algorithms.