A Recurrent Neural Network with Triplet Loss Approach for Similarity Matching and User Re-Identification

Customer profiling builds one of the core concepts of Marketing, but can be aggravated under certain conditions. Although more personalized devices are used, the awareness for privacy and data protection is increasing. In such cases, or if user accounts and loyalty programs are shared across users, profiling can be complicated. Family or group accounts make it also difficult to address individual behavior. However, re-identification of sequences poses a number of challenging problems. In particular, for the case of excessively high-dimensional data and long sequences, traditional similarity-based approaches to time series clustering or sequence alignment algorithms have their known weaknesses and limitations. The authors develop a novel general purpose solution to the problem of quantifying sequence similarities. The proposed framework combines recurrent neural network architectures, which are well established in the field of natural language processing, with a specific type of error propagation procedure used in image recognition. They apply this methodology on the re-identification of Internet browsing histories.