Hnot T. Modeling of Internet-marketing strategy using Data Science tools

In the dissertation, a conceptual approach of Internet-marketing strategy formation with the use of mathematical modeling and tools of Data Science is formulated. The relevance of the topic is expressed by the sharp growth of the online market and Internet marketing in the world and Ukraine; rapid development of analytical methods and algorithms for machine learning. These factors allowed the approaching of the issue from an analytical point of view. As a result, such main steps of Internet marketing strategy were modeled using Data Science: analysis of the competitors, customers profiling, content creation, and product promotion. A procedure for factorization was created to deal with highly sparse data. It includes the efficient decomposition of a rating matrix into latent factors of customers and products, using a combination of gradient descent methods, error weighing, and L-BGFS-B algorithm for efficient function optimization. For visual content, an approach, which assumes building a deep representation of images based on multiple trained neural network models was developed. All models should be trained on different tags to encode different information from images. Taking into account limited labeled data availability, we tested different approaches for fine-tuning and described the best one, which produced the best model. Also, among different neural-network architectures, ResNet is the best one for fine-tuning purposes based on training time and accuracy. The effectiveness of this approach in comparison with the “classical” methods is substantiated and confirmed, which improves the quality of recommendation models and allows to take into account the visual content and increase the efficiency of searching for relevant products. The obtained methods, algorithms, and approaches allowed to increase the effectiveness of Internet marketing at different steps using different data sources.