Elapsed time prediction with Strava activities

Abstract

[en] Machine learning is a discipline that allows the automation and extraction of patterns in prediction tasks, among others. In this project we study the problem of predicting the elapsed time for sport activities taking into consideration the data gathered from the user’s profile and their activities and elapsed times. More concretely, in this project we address the following topics: a protocol for gathering the data of the different activities from Strava users is proposed, data cleaning and curation is considered, and finally, the usage of different supervised learning techniques for predicting the elapsed time duration of an activity are compared and appropriate metrics are established. The project approaches the study of some machine learning methods, such as Elastic Net, Huber Regressor, Regression Trees, and lastly, additional importance is given to the study of deep neural networks (deep learning). Additionally, some metrics are also set about the success of the differents results obtained based on the accepted threshold of the regression values obtained, all of this applied in a case use of the predictors used within a business model. The results obtained show that neural networks allow us to obtain, for a sparse range of activities within a dataset, successful results in a 60% of the predictions made. But the best performance we have managed to get in this first iteration of the investigation,is yielded by the ElasticNet regressor as it has the lowest percentage of error on average. The results obtained in this project also leave a door open for potentially commercialize the investigation and being able to apply it on real case scenarios.