{
  "id": 16172067,
  "title": "Investigating Feature Leakage and Proxy Variables in a Clash of Clans ML Model",
  "url": "https://urgent.news/2026/10/23/investigating-feature-leakage-and-proxy-variables-in-a-clash-of-clans",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-10-23T08:46:12.000Z",
  "source": {
    "name": "HackerNoon",
    "slug": "hackernoon",
    "url": "https://hackernoon.com/investigating-feature-leakage-and-proxy-variables-in-a-clash-of-clans-ml-model?source=rss"
  },
  "original_language": "en",
  "account": "Welcome to another ML article. In this piece, I want to discuss a problem I encountered while working on my Clash of Clans ML lab. During data preparation for the machine learning task, I observed something peculiar. Multiple trophy-related features seemed to possess an excessive amount of predictive information about the target variable. At first glance, this may seem beneficial, as strong predictive features are typically desirable in a machine learning model. However, something about these features piqued my interest. When a feature exhibits overly high predictive power, questions arise: why is this the case? Could the feature contain sensitive information the model shouldn't have access to? Is the feature subtly encoding the target variable? Or is it simply a particularly strong signal that the model should be allowed to utilize? This leads us to the concept of data leakage. Data leakage occurs when information that should not be accessible to the model inadvertently enters the training dataset. Consequently, the model makes predictions based on data it would not have access to during real-world predictions. This can result in a model that demonstrates exceptional performance during evaluation but fails to perform as well when the leaked information is no longer available. Another possibility exists, though: the feature might be functioning as a proxy. Unlike leakage, a proxy feature does not directly contain the target variable. Instead, it is strongly correlated with it, providing the model with a powerful shortcut for making predictions. The challenge lies in distinguishing between a highly predictive feature that is merely useful information and one that has become a proxy for information the model should not have access to. To determine this, I conducted an investigation into this specific case from my Clash of Clans dataset: the trophies feature.",
  "summary": "Can a predictive feature look like data leakage? A Clash of Clans ML case study on proxies, feature audits, domain knowledge, and model performance.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}