Bike2Vec: Vector Embedding Representations of Road Cycling Riders and Races

TL;DR

Bike2Vec introduces a vector-embedding framework to represent professional road cycling riders and races using historical results. Rider and race embeddings interact via a dot-product predictor, $y$, passed through a sigmoid and trained with binary cross-entropy, using embeddings of dimension $D=5$ optimized with the Adam optimizer at learning rate $lr=0.001$ for $100$ epochs. The resulting embeddings reveal meaningful structure: race embeddings correlate with terrain/climbing profiles, while rider embeddings cluster into roles like sprinters and climbers, with similarity in embedding space reflecting known pairings of riders. This representation enables downstream tasks such as talent identification and race outcome prediction without manual feature engineering, and the authors suggest extensions to time-varying embeddings and route-feature integrations for future work.

Abstract

Vector embeddings have been successfully applied in several domains to obtain effective representations of non-numeric data which can then be used in various downstream tasks. We present a novel application of vector embeddings in professional road cycling by demonstrating a method to learn representations for riders and races based on historical results. We use unsupervised learning techniques to validate that the resultant embeddings capture interesting features of riders and races. These embeddings could be used for downstream prediction tasks such as early talent identification and race outcome prediction.

Figures (2)

• Figure 1: Visualization of race embeddings coloured by race profile score.
• Figure 2: Visualization of rider embeddings coloured by cluster.