Image processing through machine learning techniques such as neural networks is a field of study that has been advancing for some years, but still has open workspaces. Likewise, Transfer Learning (TL) is a more recent approach that has been developed as an alternative to advance in the solution of some problems associated with the training of machine learning models. In this order of ideas, this work seeks to combine image processing and TL for the creation of a model that allows identifying and categorizing emotions in images taken from recordings of face-to-face classes at the university level. This study seeks to combine the benefits of image processing techniques with the ability to transfer knowledge from previously trained models, which allows a more effective detection of emotions in a variety of educational contexts, initially, classes and activities developed in these contexts. The results show that it is feasible to transfer emotion classification models generated from training models with a generic dataset and then transfer them to a particular case represented in a smaller dataset of images taken from a face-to-face educational context. Regarding the F Score performance metric, the algorithm that delivered the best results was LinInt: Linear Interpolation between SrcOnly and TgtOnly with 83.7% accuracy for the transferred model, this is compared to other models such as RegularTransferNN - Regular Transfer with Neural Networks and RegularTransferLR - Regular Transfer with Linear Regression by Transfer learning techniques tested, and also to classifiers such as SVM - Support Vector Machine and KNN - K-Nearest Neighbors.

Learning analytics (LA) and educational data mining (EDM) are two complementary approaches to modeling and understanding teaching-learning processes and, in general, data from academic environments. LA is applied to data from various sources, which can vary in format, granularity, and structure. Integrating these data is key to addressing the challenge of scalability in LA, a fundamental aspect. To this end, interoperability, understood as the ability of different systems, devices, or applications to connect, interact, and work together effectively, is crucial and generates the need for specifications for the case of academic information systems and Learning Management Systems. According to this context, the objective of this work was to address through a literature review the following main question: What are the main challenges for modeling architecture to support the interoperability of educational data to apply Learning Analytics? To develop the review, the team used Parsifal, an online tool designed to conduct systematic literature reviews in the context of software engineering. The initial search was done in six databases, deciding to include twenty papers in the final report. The results showed that there are still many open spaces for research and development in terms of the design and use of educational data specifications for the subsequent application of LA, to make the transition from models built on data coming from a single source to the construction of models that report results from the integration of several sources using specifications like Caliper Analytics or Experience API.