Customer Review Analytics using Subjective Loss Function for Conceptual-based Learning

AbstractDeep neural networks (DNNs) are currently among the most commonly used machine learning methods in content understanding such as computer vision and natural language understanding (NLU). One of the best characteristics of these methods is their modular design – the ability to change the connectivity patterns of layers, try different activation functions, inject different statistical approaches such as normalization and dropout in the network, and many other actions – in every aspect of deep learning networks. While the majority of deep learning applications simply use cross-entropy, L1, and L2 losses, subjective loss function can actually result in impressive performance improvement. In addition, architecting the last layer of DNNs – referred to as the prediction layer – according to the needs of the application increases the discriminative power of the DNNs. This paper aims to investigate how particular choices of loss functions and prediction layer architecture affect deep neural networks and their learning dynamics, as well as the robustness of various effects. Furthermore a real-life application to measure customer loyalty called Deep Net Promoter Score (DeepNPS) from online product reviews is also proposed. The results are promising for learning more latent features and matching the customer feedback with the NPS score.