Space semantic aware loss function for embedding creation in case of transaction data

Transaction data are the most popular data type of bank domain, they are often represented as sparse vectors with a large number of features. Using sparse vectors in deep learning tasks is computationally inefficient and may lead to overfitting. Аutoencoders are widely applied to extract new useful features in a lower dimensional space. In this paper we propose to use a novel loss function based on the metric that estimates the quality of mapping the semantic structure of the original tabular data to the embedded space. The proposed loss function allows preserving the item relation structure of the original space during the dimension reduction transformation. The obtained results show the improvement of the resulting embedding properties while using the combination of the new loss function and the traditional mean squared error one.