Topological data analysis applied to multimodal recommendation

Abstract

[en] This project incorporates Topological Data Analysis, TDA, into a multimodal collaborative Recommender System. To do so, it conducts a comprehensive investigation into the tools that topology provides us, and how they can be used to describe data. Then, they are incorporated into a state-of-the-art multimodal Recommender System in many ways: (i) collecting data about product images and textual descriptions of products to enhance the information that the Recommender System receives, and (ii) extracting information about the item graph and utilizing it both to feed the neural network and to prune the graph, aiming to increase speed without losing performance. The analysis has been performed on three well-known datasets and with different collaborative models. The results vary depending on the characteristics of the data studied. Generally speaking, we have found that with smaller datasets the performance increases. We have also seen that the information extracted from images seems to be more useful than the information derived from text descriptions. The changes in the network architecture have not been fruitful, and the run time reduction had major impact on the result. In summary, we observe that the topological data is useful, and we see an improvement in the performance on specific steps of the implementation and with smaller datasets. It can be used in many ways, and this requires a careful preliminary study to evaluate where it can have the best and most meaningful impact.