Classifying and Grouping Narratives with Convolutional Neural Networks, PCA and t-SNE

Abstract

Each week, the Consumer Financial Protection Bureau (CFPB) receives thousands of consumer complaints about financial products and services. These complaints must be forwarded to the responsible company and posted on the site after 15 days or when the company responds to the complaint, whichever comes first. Published complaints and solutions help consumers solve their problems and also serve as a repository of help for other consumers to avoid or solve problems on their own. Every complaint provides information about the problems people are having, helping them to identify inappropriate practices and allowing them to stop before they become major problems. Culminating in better results for consumers and a better financial market for everyone. Each of the complaints contains information on submission date, company to send the complaint, complaint narrative, among others. However, complaints do not have information on the department to which it should be forwarded. Therefore, in this work, the three approaches to analyze each complaint are: (i) convolutional neural network (CNN) to classify the narratives; (ii) principal components analysis (PCA); and (iii) t-distributed stochastic neighbor embedding (t-SNE) to create a three-dimensional embedding for clustering. Embedding from scratch, Pre-trained Word Vectors (word2Vec) and Global Vectors (GloVe) vectors are used and compared in six different CNNs modeling. The results increase the evidence that pre-trained word vectors is important and that convolutional neural networks and t-SNE can perform remarkably well on real text classification data.