Many tasks in graph machine learning, such as link prediction and node\nclassification, are typically solved by using representation learning, in which\neach node or edge in the network is encoded via an embedding. Though there\nexists a lot of network embeddings for static graphs, the task becomes much\nmore complicated when the dynamic (i.e. temporal) network is analyzed. In this\npaper, we propose a novel approach for dynamic network representation learning\nbased on Temporal Graph Network by using a highly custom message generating\nfunction by extracting Causal Anonymous Walks. For evaluation, we provide a\nbenchmark pipeline for the evaluation of temporal network embeddings. This work\nprovides the first comprehensive comparison framework for temporal network\nrepresentation learning in every available setting for graph machine learning\nproblems involving node classification and link prediction. The proposed model\noutperforms state-of-the-art baseline models. The work also justifies the\ndifference between them based on evaluation in various transductive/inductive\nedge/node classification tasks. In addition, we show the applicability and\nsuperior performance of our model in the real-world downstream graph machine\nlearning task provided by one of the top European banks, involving credit\nscoring based on transaction data.\n