Automating Block Type Classification using Data Science

Hudl IQ's exclusive line battles data provides tremendous detail on the interactions between the offensive and defensive lines. Via this data, we now know who each player engaged with, where they were on the field when this happened, and when/where the engagement both starts and ends. This in turn allows for the creation of metrics and analysis on vital areas of game planning and prep such as run tendencies, while also allowing deeper analysis- such as this look at pockets using our free TB12 data.

Run blocking is a key component of offensive line play. Linemen block defensive players in a manner coordinated to help create space for ball carriers to exploit. Depending on the run concept being used for a play, each player will have a different assignment. While the possible assignments are nearly unlimited, they can be categorized into one of several different block “types.” By identifying these block types at the individual player level on every running play, we can both begin to understand if certain players perform certain tasks better than others while also creating the building blocks for identifying run schemes and ultimately play calls themselves.

Our line battles data is intended to help the analysis of both individual player performance and team schemes and tendencies. The first step is the identification and assignment of those block types. In this article, we’ll go over how we managed it.

To model, or not to model: that is the question

As a data scientist, it is tempting to think that complex statistical models can solve everything. However, they are not always the answer. Sometimes simplicity and explainability are paramount.

Our aim is to attribute one of seven different block types (Base, Drive, Reach, Pull, Slice, Double Team, Second Level) to each offensive line player on each running play. In data science terms, this means training a classifier with seven different outcomes. While there are no issues with this statistically (many classification algorithms can successfully handle predictions for this many outcomes), modeling this data successfully comes with its own challenges.

Broadly speaking, there are two categories of machine learning models. The first category, supervised learning models, use labeled data to train the model. We don’t currently have labels for block type and collecting the number of labels we’d need would take a lot of time and resources.
The second category, unsupervised models, use unlabeled data and instead rely upon the model finding patterns in the data to split the observations into groups. Unsupervised models are unlikely to converge to a solution that aligns to exactly what we want to capture, especially from the standpoint of creating groupings that would be recognizable to football experts. While the model could be fine-tuned and iterated to seek the specific classifications needed, this process adds complexity, consumes time, and does not guarantee success. We can achieve a better solution much more efficiently through other methods.

Enter Stage Left: Subject Matter Experts

With the help of Subject Matter Experts (SMEs) we can approach the problem in a more efficient and quicker way.

Our SMEs led by Head of Football Analysis Matt Edwards collated an initial list of block types to identify. From their experience playing and/or coaching line play at a high level, they know exactly what to look for on film to identify each type. Combining this knowledge with our line engagement data, we can use these football definitions of each block type to derive similarly defined situations within our data. Following this, we can use a sprinkle of more advanced statistical methods to refine those definitions further.

Let’s take a look at how this works, taking reach blocks as an example.

Matt’s definition of reach block is a block where the offensive player is attempting to get outside leverage on a defender. Within our tracking and engagement data, reach blocks are identifiable by the large amount of lateral movement by the lineman between snap and initial engagement with a defender.

As we have the start and end coordinates of the engagement, alongside a blocker’s snap location, we can easily combine this information to quantify lateral movement for a player within a play. There are a couple of different ways we could do this, for example, movement from snap to the start of the engagement, or movement from the snap to the end of the engagement.

Determining which definition creates results which best align with our football definition of reach blocks requires a little trial and error.

To examine the relationship between a proposed definition within the data and instances we have pre-selected as reach blocks, we can use eCDF (empirical cumulative distribution function) and ROC (Receiver Operating Characteristic) curves.

On an eCDF plot, we have the variable of interest on the x-axis, cumulative proportion on the y-axis and separate curves for each group we are interested in (in this case either a reach block or not a reach block). For a variable to be a good option for distinguishing between the given groups, we want there to be clear separation between the curves and ideally a single particular value on the x-axis that provides the most separation. The charts below illustrate this analysis by comparing the accuracy of using distance to either engagement start or engagement end as the heuristic for determining if a reach block occurred.