Modeling the relationship between a human’s stress level and their physiological responses can pave the way for numerous advances in neuropsychology. This work follows suit in tackling such prob lem by adapting a multi-modal framework in determining stress response. In this paper, we modeled the valences and arousals of human subjects as they were put into an impromptu job interview session; visual, acoustic, text, and physiological features were ob tained fromsubjects.Weproposeatemporal-sensitivedeeplearning network framework that consists of a CNN-LSTM and a CNN-CNN module with late fusion. Various combinations of modalities were tested and it was determined that using a combination of acoustic and visual features yielded the best concordance correlation coeffi cient (CCC) for valence–reaching 0.708; on the other hand, fusing all the acoustic, visual, and text features had the highest CCC for arousal, equaling to 0.3692. 

An anonymized recommendation system of anime shows that uses neural network and federated learning was implemented. This is performed without storing user information via federated learning to address anonimity in using services that require personal information. Federated learning uses model averaging to preserve data privacy by not requiring the user to send personal information directly to the recommendation system but instead updates the model on the device level. The paper implemented three (3) collaborative filtering recommendation systems: vanilla using matrix slower rate taking into account the small batch implemented with the limitation on the computing system available for simulation.