Defensive-Packing-Pass-Network-Disruption-Index: The metric that turns pressing into a graph attack

NavjotML’s model scores defenders not just for winning the ball, but for breaking the opponent’s passing structure, especially when the target is a central node.

9 min read • View on GitHub • More from NavjotML

A football pitch drawn as a passing network, with a dense central hub and a defender closing in from the side. Several routes snap and unravel behind the targeted node, showing that the real value of the press is structural damage, not just pressure. — The core idea is simple: a press matters most when it collapses the opponent’s network, not when it merely forces a touch.

Key Takeaways

This repo treats defense as network sabotage, not as a tally of tackles or pressures.
Its real leap is weighting a press by the passing importance of the player it targets.
Freeze-frame context and centrality scores let the model value structural damage instead of raw defensive volume.
The dashboard turns that idea into something analysts can inspect sequence by sequence.

Most defensive metrics ask a blunt question: did the press win the ball? This repo asks a sharper one: who did the press shut down inside the opponent’s passing graph? That shift matters because a press on a hub player can change the shape of an attack even when it does not end the possession.

The project aims to quantify the disruptive impact of defensive actions on opposition passing networks.

Navjot Singh, Author/Maintainer · Project README

The hidden trick: pressing the player who matters most

The repo’s reversal is elegant. Packing usually describes how much ground an attack eliminates. Here, the logic flips. A defender is rewarded for making the opponent’s passing network smaller, flatter, and less useful.

That is why the model cares about centrality. Pressing a fringe player is cheap. Pressing a high-value connector, like a hub with strong PageRank, can damage the entire possession structure.

Why classic pressing stats miss the point

Press success rate, tackles, and interceptions are useful, but they are narrow. They often reward the final event and ignore the chain reaction that came before it. A press can fail to win the ball and still force the opponent into a worse passing shape.

Metric	What it values	What it misses	Best use case
Classic packing	How many opponents an action bypasses	Defensive structure and target importance	Attacking progression
Press success rate	Whether the press ended in a turnover or error	Partial disruption and hidden pressure	Basic defensive summaries
Tackles and interceptions	Visible ball-winning events	The wider network effect of the action	Traditional box score style analysis
Disruption Index	How much a defensive action breaks the passing structure	Some tracking-level nuance and data quality issues	Valuing structural defensive impact

The difference is not cosmetic. In the final third or a compact mid-block, the cost of a failed press can be tiny if it still cuts the opponent’s best options. The repo tries to measure that hidden value instead of throwing it away.

A close-up of three linked network nodes on a tactical board. One node carries the heaviest weight, another acts as a bridge, and a defender pin interrupts the most connected point while nearby routes bend and compress. — The close-up makes the math tactile. A press is not just an action on a player. It is an intervention on the most valuable connection in the network.

How the Disruption Index works

The model blends space and structure. A press is scored not only by where it happens, but by how much of the opponent’s network it disrupts.

df['combined_press_value'] = (0.40 * spatial_packing + 0.35 * lanes_cut + 0.25 * network_importance)

That formula does the editorial work of the project. Spatial packing captures the local defensive squeeze. Lanes cut captures immediate blockage. Network importance adds the missing layer, which is the structural weight of the player being pressured.

This is why the repo feels more like a research instrument than a dashboard toy. It turns a football action into a composite value that reflects both geometry and graph position.

The network engine behind the metric

Under the hood, the repository precomputes network metrics for players across matches. PageRank estimates which nodes matter most in the passing flow. Betweenness picks out the bridges that connect otherwise separate parts of the attack.

That matters because the model is not guessing importance on the fly. It is loading structural context from processed data, then using that context to score defensive actions more intelligently.

import networkx as nx

# passing network centrality
pagerank = nx.pagerank(G, weight='weight')
betweenness = nx.betweenness_centrality(G, weight='weight')

The repository’s data footprint shows the same ambition. It is built around processed match sequences, network metrics, and a Streamlit layer that can explore thousands of defensive moments without turning the experience into a static report.

What the dashboard actually lets you do

The Streamlit app is the research surface. Analysts can inspect press sequences, compare players, review model outputs, and move between tactical maps and diagnostic plots without leaving the same interface.

That design choice matters. The repo is not only proposing a metric. It is packaging a workflow around it, so the idea can be tested against match context instead of staying abstract.

Sequence-level press value for individual actions.
Player and match filters for tactical comparison.
Model diagnostics such as SHAP and ROC outputs.
Pitch visualizations for press maps and defensive patterns.

Where this beats packing, and where it does not

The Disruption Index wins when the question is defensive value. Packing is a great attacking lens, but it is not built to reward a press for destabilizing a key passing hub. This repo is.

Question	Better tool	Why
How much progress did the attack make?	Packing	It measures bypassed opponents directly
How much did the press destabilize the opponent?	Disruption Index	It values the press by network damage
Did the defender win the ball?	Press success rate	It is a simple outcome metric
Which action changed the opponent’s options most?	Disruption Index	It combines space, lanes, and centrality

It still has limits. It depends on freeze-frame availability, event quality, and the assumptions baked into the weighting formula. That is not a flaw to hide. It is the price of making defensive context measurable from event data.

What this repo says about the future of defensive data

The larger lesson is not about one formula. It is about where football analytics is heading. The next useful defensive metrics will care less about raw event counts and more about structure, context, and downstream damage.

Full tracking data will make this cleaner. Until then, NavjotML’s repo is a strong bridge. It shows how far you can get by combining open event data with graph theory and a clear tactical question.