Improving PageRank using sports results modeling

How to rank participants of a sports tournament is fundamental importance. While PageRank has been extensively used for this task, the algorithm’s superiority over simpler ranking methods never clearly demonstrated. We address knowledge gap by comparing performance multiple on synthetic datasets where true known and methods’ can be thus quantified standard information filtering metrics. Using results from 18 major leagues, we calibrate state-of-art model, variation classical Bradley–Terry results. identify relevant range parameters under which model reproduces statistical patterns found in analyzed empirical datasets. Our evaluation shows that outperforms benchmark number wins only early when small fraction all games have played yet. Increased randomness data due home team advantage, example, further reduces PageRank’s superiority. propose new variant combines forward backward propagation directed network representing input The method evaluated settings and, sufficiently sport not too random, it also wins. Beyond presented comparison methods, our work paves way designing optimal algorithms data.