What 'Similar Champions' Means in Winrate
16 de febrero de 2026
When Winrate says two champions are similar, it does not mean they have the same kit. It means they tend to create value in similar ways across many real games in the same role.
Why This Feature Exists
Most similarity lists are built from class labels or manual judgment. Those can be useful, but they are coarse.
The goal here is more practical: if you perform well on champion A, the "similar" shortlist for champion B should feel like a real adjacent option, not a random champion with the same class tag.
What The Model Learns From Games
Each training row is one player in one match. A typical run uses a large sampled dataset, around 500,000 rows. From each row, the model computes 43 behavioral features. These are signals about how a champion is played, not just what class Riot assigns it.
No single stat decides similarity. Each champion's fingerprint is a blend of all 43 signals, averaged over many games to reduce one-off noise.
To make that concrete, here is an illustrative pattern-level comparison:
This is why Thresh often lands near engage supports instead of pure enchanters: the whole profile is closer, even if some individual stats overlap.
What We Actually Do To Produce "Similar Champions"
After feature extraction, each champion gets one stable fingerprint. We then compare champion fingerprints directly and rank the closest matches.
Before final ranking, we apply a role filter, so support champions are compared against supports, mids against mids, and so on. This avoids many cross-role pairs that look mathematically close but are not useful in real draft decisions.
The product returns the top five nearest neighbors. A higher score means the two champions are closer in overall play pattern.
Similar means "tends to create value in comparable ways in this role." It does not mean the champions are identical, equally strong on the patch, or correct into the same matchup.
Current Examples
These examples are useful because they are about pattern rather than class label. For example, the support cluster around Thresh reflects engage/setup behavior, not just "support" as a broad bucket.
How We Evaluate Whether It Works
The core evaluation question is straightforward: do high-similarity pairs line up with real skill transfer across players?
Raw win rate is not enough for this. Player skill is a confounder. If stronger players happen to play both champions, a pair can look good even when the play patterns are not actually close.
So the evaluation uses relative performance: for each player/champion, it asks whether that player did better or worse than expected on that champion. Then it checks whether players who overperform on champion A also tend to overperform on champion B.
Example: if champion A has a 50% baseline and a player wins at 55% on A, that player is +5 on A. If champion B has a 49% baseline and the same player wins at 53% on B, that player is +4 on B. When this pattern repeats across many shared players, A and B show positive skill-transfer signal.
That gives a cleaner signal of transferable skill. In plain language: if your personal "above baseline" champions overlap with the model's similar pairs, the model is doing useful work.
Current documented result for this version is 74.7% top-10% precision. That means when we look at the highest-ranked similarity pairs, about three out of four show positive skill-transfer signal in evaluation.
This supports the ranking as useful directional signal. It is not a guarantee for a specific matchup, lane state, or draft.
Why We Shipped This Version
Earlier iterations either learned shortcuts (champion identity cues) or added complexity without clear production benefit. The current version gave the best balance: behavior-rich signals, stable output quality, and an evaluation result that holds up.
Why Similarity Is Not Defined As Raw Skill-Transfer Correlation
That seems attractive at first, but it creates practical problems. Coverage drops fast for niche champions because many pairs do not have enough shared-player data. Correlation is also noisy across patch and meta shifts, so pair strength can move for reasons unrelated to stable play pattern. It can also capture community overlap (the same players picking both champions) alongside true transferability. Finally, a pure correlation graph is harder to generalize when fresh data is thin.
Skill transfer is still central. We use it as the validation target, while behavior-based similarity provides a more stable ranking signal.
How To Use It In Product
The main product use is champion pool growth. You can use Similar Champions to find adjacent picks you are likely to learn faster, then pressure-test those options inside Draft Picker. Use similarity to generate candidates, then use Draft Picker context (matchups, comp, bans, patch) to make the final choice.