Learning based spectral clustering for LTE downlink CoMP systems

Coordinated Multi-Point (CoMP) systems appear as advanced promising strategies to improve user throughputs, especially in interference limited regions, at cell edge. Whether CoMP consists in jointly processing data from multiple transmission points, or in smartly coordinating the allocation of resources, CoMP implementation requires significant computation effort, signaling exchange that may impact the bandwidth limited back-hauling resources. To alleviate signaling and computation cost, the solution consists in partitioning the cells into coordinating sets. The study proposes a novel approach to cluster the cells into coordinated sets and to dynamically update it, based on data collected from the field. Clustering algorithm is based on spectral clustering and aims at grouping cells with high ‘similarity’ between them. In the context of Dynamic Point Selection, users are dynamically switched from one cell to another one from the coordinating set, to take profit from macro diversity or load balancing gains. Similarity then captures the likelikood of switching, through a combined distance metric based on geometrical distance and Handover frequency. Analysis is carried out from data mining tools, with input data from the field, in the dense wireless network of Manhattan, in New York. Geographical Information System application completes the interpretation of clustering results, by displaying cluster maps, and proves the capability of clustering thanks to this approach.