Evaluation of Attributed Network Embedding Algorithms for Patent Analytics

Patent analytics is a specialized branch of data analytics where patent documents are analysed to understand behavioural information. Citation network analysis is one of the common techniques to examine the importance of a patent by studying its citations. Typical patent citation network (PCN) will have millions of attributed nodes and edges. Inferencing on such a large network necessitates the use of attributed network embedding (ANE) techniques to bring down the computational requirements by reducing the dimensionality of the network data. Identifying the suitable ANE algorithm for PCN analytics is the purpose of this study. Multiple ANE algorithms are applied on the patent dataset to create low-dimensional embeddings, and these embeddings are used as the input for performing the innovation value prediction using linear regression model. Mean square error (MSE) is calculated between the predicted innovation values and the actual innovation values. MSE values obtained with different ANE algorithms are analysed to identify the most suitable ANE algorithm for patent analytics. GraphSAGE with mean-based aggregator resulted in the least MSE compared to all other ANE algorithms evaluated for patent analytics.