Sport climbing is a relatively recent sport, and many new climbing routes are continuously built all around the world. Climbers are interested to climb routes which are: challenging, but within their capabilities; enjoyable; suitable for training; and safe. We would like to first understand the preferences of the climbers by monitoring their behaviour, and then to build a recommender system for supporting them in choosing suitable climbing routes.
Modeling a climber’s behavior will be made by using both explicit feedback, e.g., by leveraging data coming from mobile climbing applications (s.a. Vertical Life), and by using implicit feedback, which can come from data acquired by accelerometers placed on the wall or on the body of the climber. By combining these two data sources, we aim at building a climber’s profile and based on such a profile, to generate recommendations.
Aiming at the overall goal of building a climbing routes recommender system, we have now developed machine learning models to predict whether the climber would agree with the “official” difficulty level of the climbing routes. Usually, climbing route difficulty is given by the route setter or a person who initially built the. However, climbers might have different opinions about the grade of a route, hence in disagreement with the route setter. In our application scenario, climbers manually insert their grades for routes that they tried, by using a mobile application (`Vertical-Life’). We use a range of additional data in order to model why they disagree with the setter. We have created a dataset of indoor and outdoor lead climbs, where the official grade and the climbers’ opinions are stored. We have then implemented two ML models to predict the grade difficulty given by the climbers: linear regression and matrix factorization. For the linear regression model, we constructed features of the climber-route interaction which describe how and when the climber deviates from the setter’s grade, and for matrix factorization, we modelled the grade prediction as a special type of rating prediction problem. We have then used singular value decomposition with normalization taking into consideration the route setter’s grade. We show that the models’ predictions are closer to the climbers’ opinions than a baseline model.
In the future, we would like to develop a full system that could be also able to identify climbers in a climbing gym, detect what type of activities they perform, and measure their performance. This system could be used to enrich the climber’s profile and to complete the design of the recommender. Last, but not least, we would like to incorporate the proposed solution into the Vertical-Life app.