That has all changed with the announcement of Clarify, which offers you the ability to detect bias and implement model explainability in a repeatable and scalable manner. SEATTLE–(BUSINESS WIRE)–Today, Amazon Web Services (AWS), an Amazon.com, Inc. company (NASDAQ: AMZN), and the German Bundesliga, Germany’s top national football league, announced three new Bundesliga Match Facts powered by AWS to give fans deeper insights into action on the pitch. The new Match Facts - Most Pressed Player, which highlights how often a player in possession experiences a significant pressure situation throughout a match; Attacking Zones, which shows fans where their favorite team is attacking and which side of the pitch they view as most likely to score from; and Average Positions - Trends, which shows how changes to a team's tactical formation can impact a match's outcome - will … “Every Bundesliga match generates data that can improve play and help fans better understand team strategies, and we are making tremendous strides in leveraging the vast amount of data in our archives and from our league’s current games to develop and roll-out new Match Facts. In the plot, a feature’s SHAP value serves as an arrow that pushes to increase (positive value) or decrease (negative value) the prediction value. Bundesliga Match Facts are generated by gathering and analyzing data from live game video feeds as they’re streamed into AWS. He works on extracting interesting insights from football data using AI/ML for both fans and clubs. Without further ado, let’s dive in! Gabriel’s background is in Mathematics and Machine Learning, but he is additionally pursuing his PhD in Sports Analytics at the University of Tübingen and working on his football coaching license. Attacking Zones: As teams look to exploit defensive weaknesses, approach their opponent’s goal, and ultimately score, Attacking Zones allows fans to see where the teams focus their offense to create those scoring opportunities. Better interpretability leads to better adoption. Unsurprisingly, a strong inverse relationship exists between DistanceToGoal and PressureSum for those match events with a high goal prediction; as the former decreases, the latter rises. The positive and negative impact on the goal prediction value is shown on the x-axis, derived from our SHAP values. The new Bundesliga Match Facts powered by AWS. Although none of our match events were penalties (all having a feature value =1), it must still be included in the Clarify processing job because it was also included in the original XGBoost model training. These advanced statistics help audiences better understand areas like decision-making on the pitch and the probability of a goal for each shot. He works with clients across industries to help them tell stories with data using machine learning. The higher the xGoals value (with all values lying between 0–1), the greater the likelihood of a goal. We can see a negative relationship between the DistanceToGoal and the target variable, with the likelihood of a goal increasing as we get closer to the goal. Zum kommenden 21. It started small. These statistics are delivered to viewers via national and international broadcasters, as … It’s worth noting that for regions that have an increased vertical dispersion of results, we simply have a higher concentration of data points that are overlapping, which gives us a sense of the distribution of the Shapley values per feature. Dit is slechts een greep van alle wedstrijdinzichten die Luuk voor de Bundesliga heeft ontwikkeld. Bundesliga’s new Match Fact, Attacking Zones, from AWS. The primary implications for Bundesliga Match Facts powered by AWS going forward are twofold. Conversely, for angles greater than 25, a player moving at a slow speed towards the goal reduces the likelihood of a goal compared to a player moving at a greater speed. Luuk Figdor is a data scientist in the AWS Professional Services team. In his spare time he likes to learn all about the mind and the intersection between psychology, economics and AI. The objective of this advanced statistic is to show fans where their favourite team is attacking and which side of the pitch they seem to view as most likely to score. Unsurprisingly, few headers were scored with an angle less than 25. As you move further away from the goal, the angle reduces. The Bundesliga xGoals ML model goes beyond previous xGoals models in that it combines shot-at-goal event data with high-precision data obtained from advanced tracking technology with a 25-Hz frame rate. “Amazon SageMaker Clarify brings the power of state-of-the-art explainable AI algorithms to the fingertips of our developers in a matter of minutes and seamlessly integrates with the rest of the Bundesliga Match Facts digital platform—a key part of our long-term strategy of standardizing our ML workflows on Amazon SageMaker,” reports Gabriel Anzer, Data Scientist at Sportec Solutions (STS), a key partner organization of Bundesliga Match Facts powered by AWS. One particularly interesting use case for Clarify is from the Deutsche Fußball Liga (DFL) on Bundesliga Match Facts powered by AWS, with the goal of uncovering interesting insights into the xGoals model predictions. (German-language video) MPEG-4 Video Insights are generated by gathering and analysing data from live game video feeds as they’re streamed into AWS. When we apply Clarify, we can both enhance goal prediction models and contextualize football match events on a per-play basis. Bundesliga Match Facts are generated by gathering and analyzing data … Nearly all goals that are scored close to the goal are hit with an angle greater than 45 degrees. In today’s world where predictions are made by ML algorithms at scale, it’s increasingly important for large tech organizations to be able to explain to their customers why they made a certain decision based on an ML model’s prediction. We need to have consistency between the two feature sets for model training and Clarify processing. The goal probability is calculated in real time for every shot to give viewers insight into the difficulty of a shot and the likelihood of a goal. Theoretical approaches for overcoming this lack of model explainability have undeniably matured in recent years, with one standout framework becoming a crucial tool in the world of explainable AI: SHAP (SHapley Additive Explanations). Echtzeitstatistik mit AWS. Bundesliga Match Facts powered by AWS provides advanced real-time statistics and in-depth insights, generated live from official match data, for Bundesliga matches. For more information on the xGoals training process with the Amazon SageMaker Python SDK and XGBoost hyperparameter optimization, see The tech behind the Bundesliga Match Facts xGoals: How machine learning is driving data-driven insights in soccer. These three new Match Facts join Speed Alert , Average Positions, and xGoals to bring the total number of insights available for Bundesliga fans to six. Outside of work he loves to spend time travelling, trying new cuisines and reading about science and technology. We see how, starting from the bottom and working our way up, the features start to have an ever-increasing impact on the final prediction, with some extreme cases showcasing how AngleToGoal, DistanceToGoalClosest, and DistanceToGoal really have the final say in our XGBoost model’s probability prediction. With the Bundesliga Match Fact xGoals, the DFL can assess the probability of a player scoring a goal when shooting from any position on the field. Every data point in the following plots represents a single attempt at a goal. How Eintracht Frankfurt beat Bayern Munich: AWS match facts analysis 5 weeks ago. We use the open-source SHAP library to plot the SHAP values that are computed inside our processing job. Let’s take a look at Can’s goal in action, brought to life in 2D animation simply by using the positional tracking data of the players at the time of the goal. We simply select a feature and then plot the feature value on the x-axis and the corresponding SHAP value on the y-axis. An elegant feature of the SHAP framework is that it’s both model agnostic and highly scalable, working on both simple linear models and deep, complex neural networks with hundreds of layers. The new Bundesliga Match Facts include Most Pressed Player, Attacking Zones, and Average Position Trends; it will debut during today’s match between RB Leipzig and FC Augsburg.. As suggested by the vertical axis on the right side of the plot, a red data point indicates a higher value of the feature, and a blue data point indicates a lower value. This new Match Fact divides the last third of the pitch into four equally sized Attacking Zones. The following plot shows that relationship for our most important features: When we take a closer look at two of our (less influential) categorical variables, we see that, all other things being equal, a header invariably decreases the likelihood of a goal, whereas a freekick increases it. Like most ML tools, it was missing a way of diving deeper and explaining the results of said models, or investigating training datasets for potential bias. Every time the attacking team enters one of these zones, either by dribbling or with a pass, the ball possession algorithm counts an attack and displays it in the graphic. Knowing respective feature attributions and explaining outcomes helps in model debugging, which in turn results in higher-quality predictions. Voted to be the best goal of the 2019–2020 season by 22% of Bundesliga viewers, Emre Can’s jaw-dropping strike was given a near-zero (3%) chance of going in and, taking into account his great distance from the goal (approximately 30 meters) and at such a flat angle (11.55 degrees), we can see why. This additional match information kicks off with "Average Positions" and "Expected Goals" (xGoals): Average Positions tracks players' average location on the pitch in … We used the area under the ROC curve (AUC) as the objective metric for our training job, and trained the xGoals model on over 40,000 historical shots at goals in the Bundesliga since 2017, using the Amazon SageMaker XGBoost algorithm. Moritz Mücke erläutert, worum es dabei geht: »Grundsätzlich ist es unser Ziel, mit den Bundesliga Match Facts Leistungen von Spielern … Bundesliga Match Facts are generated by gathering and analyzing data from live game video feeds as they’re streamed into AWS. These affects are again reversed for the AngleToGoal: as the pressure increases, we see an increased AngleToGoal decreasing the SHAP value of PressureSum. In 2006, Amazon launched two fairly simple services: computers you could rent by … Through this, over 500 million Bundesliga fans around the world gain more advanced insights into players, teams, and the league, and are delivered a more personalized experience and the next generation of statistics. The dashed lines are those match events in which a goal occurred. As technology for capturing football data has advanced dramatically in recent years, so too have the models that we can use to model this growing mountain of data. It’s reassuring to have our feature interaction plots confirm our preconceived ideas of the game, as well as quantify the various powers at play. Media partners and commentators can now choose which time spans to analyse and then compare those sections of the match, making it easier to identify tactical trends such as whether a team visibly reacts or begins a period of increased pressure after a significant event such as a goal, red card, or substitution. “Together with AWS, we’re delivering a new perspective on what happens on the field and offering a new and engaging way for fans to follow their favourite teams.”, “Expanding our work with Bundesliga means more fans will gain an appreciation for the incredible talent on the field and the decisions made by teams, at the same time as the league differentiates itself through the use of advanced analytics to improve the quality of play,” added Klaus Buerg, General Manager for AWS Germany, Austria, and Switzerland, Amazon Web Services EMEA SARL. When we look at the sixth goal of the game, scored by Leon Bailey, which the model predicted with relative ease, we can see that many of the (key) feature values are exceeding their average, and contributing toward increasing the likelihood of a goal, as reflected in the relatively high xGoals value of 0.36 in the following force plot. “In just one year since the launch of Match Facts, AWS and Bundesliga have created statistics that are giving fans around the world a completely new way to experience the game. But this was clearly not enough to stop Can. The only features working to increase his chances of scoring were the fact that he had very little pressure on him at the time, with only two players in the local vicinity capable of closing him down. Click here to return to Amazon Web Services homepage, The tech behind the Bundesliga Match Facts xGoals: How machine learning is driving data-driven insights in soccer, Begun automating the process of exploring and analyzing goal prediction data at scale, in novel ways, Offered a model explainability and bias platform that can be improved on for the further capture of interesting and significant shot patterns. Crucially, this can be seen as a direct move away from the underlying models being closed boxes for which we can observe the inputs and outputs, but not the internal workings. Upon inspection we see that, even for the two most important features, they have a very minimal effect on changing the SHAP value of PressureSum. More interestingly, however, when comparing the affects that a header or FootShot has on the likelihood of a goal being scored, we see that for any given angle in the range 25–75, a header reduces it. This makes sense; how often is it that you see someone score a goal from the sideline when 40 meters out? We can improve the dependence plots by highlighting the interaction between different features—the additional affect, after we take into account the individual feature effects. Amazon Web Services (AWS), an Amazon.com, company, and the German Bundesliga, Germany’s top national football league, have announced three new Bundesliga Match Facts powered by AWS to give fans deeper insights into action on the pitch. All the arguments in this processor are generic and are related only to your current production environment and the AWS resources at your disposal. The XGBoost model starts its prediction at this baseline, with positive and negative forces that either increase or decrease the prediction. Lack of explainability can often create a barrier for organizations to adopt ML. The new Match Facts – Most Pressed Player, which highlights how often a player in possession experiences a significant pressure situation throughout a match; Attacking Zones, which shows fans where their favourite team is attacking and which side of the pitch they view as most likely to score from; and Average Positions – Trends, which shows how changes to a team’s tactical formation can impact a match’s outcome – will debut during Matchday 21 on February 12th featuring RB Leipzig vs. FC Augsburg. Nick’s background is in Astrophysics and Machine Learning and, despite occasionally following the Bundesliga, he has been a Manchester United fan from an early age! As we can see from both plots, a noticeable divide exists between the impact that AngleToGoal < 25 and AngleToGoal > 25 have on the goal prediction. When we compare this plot across the three seasons (2017–2018, 2018–2019, and 2019–2020), we see little to no change in both the feature importance and their associated SHAP value distribution.