Development of Machine Learning/Prediction

27-09-18 Rongbo Hu 0 comment

A basic understanding of the data analysis of all partners is very important for a successful collaboration of data producing and data processing partners. Because of this, the consortium is planning a Data Analytics Workshop in December 2018. EPFL is organizing this event with support from TUM. The main focus of the data analysis in REACH is the Human Activity Recognition (Fraunhofer) and the Prediction of behavior change (EPFL).

  1. Prediction of Behavior change: In the first step, a behavior change recommender system has to preprocess data then it should monitor the behavior profile and intervention profiles and then check the persuasion profile if it is feasible/effective. After evaluating such data, one can say if its approach has been effective/efficient or not. In the first step, a predictive model should be created based on many different patients and then an estimation profile of the current user must be prepared. Based on the two input maybe the behavior change can be predicted. The data set must include a base line behavior and a post-intervention behavior to generate the model of behavior change. Such a model (if the proper functionality is needed) can be created only after many different cases are studied/inputted to build the model. To make the most accurate prediction possible, a very large data set must be used. As many different people as possible should be measured, the length of the measurements is rather insignificant.
  2. Human Activity Recognition: The goal of Human Activity Recognition Paradigm is to understand the health status of the patients and monitor daily routines and analyze changes. Complexity levels of activities can be divided into four categories high-level, mid-level, low-level and locomotion. Considering all these levels a chart of data can be generated which includes many signal patterns and based on these patterns of activities can be recognized. This is implemented via Activity Recognition Chain Concept. In this concept, the raw sensor data will be preprocessed and segmented into sliding windows and then features are extracted for each sliding window. It is considered to follow the guidelines of the Opportunity Challenge experiments for data collection, labeling procedures, and quality control.


machine learning 3


Leave a reply