Enum PredictionAggregationMethod

Namespace

AiDotNet.Augmentation

Assembly

AiDotNet.dll

Specifies how to combine predictions from multiple augmented versions of the same input.

public enum PredictionAggregationMethod

Fields

GeometricMean = 6

Multiply predictions together (then take the Nth root). Best for probability products.

Use this when predictions represent independent probabilities that should be combined. Less common but useful in specialized scenarios like combining class probabilities.

Max = 2

Take the highest prediction. Useful for object detection confidence scores.

Use this when you want to capture the "best case" prediction. Common in object detection where you want the most confident bounding box.

Mean = 0

Average all predictions together. Best for regression and probability outputs.

This is the most common choice. Works well for:

• Predicting prices, temperatures, or any continuous value
• Classification confidence scores (e.g., "80% confident this is a cat")

Median = 1

Take the middle prediction when sorted. More robust when some predictions are outliers.

Use this when you suspect some augmentations might give bad predictions. For example, if 4 predictions are around 100 but one is 1000, median ignores the outlier.

Min = 3

Take the lowest prediction. Useful for conservative estimates.

Use this when you want the most conservative prediction. For example, when estimating costs where you'd rather underestimate.

Vote = 4

Count votes from each prediction and pick the winner. Best for classification.

Use this for classification tasks. Each augmented prediction "votes" for a class, and the class with the most votes wins. Like a democratic election for predictions.

WeightedMean = 5

Like Mean, but gives more weight to higher-confidence predictions.

Use this when some augmentations produce more reliable predictions than others. Predictions with higher confidence scores contribute more to the final answer.

Remarks

For Beginners: When you make predictions on several variations of an image (flipped, rotated, etc.), you need a way to combine those predictions into one final answer. This enum controls how that combination happens.

Think of it like asking 5 friends to estimate the price of a used car:

• Mean: Take the average of all estimates ($15K + $18K + $16K + $14K + $17K) / 5 = $16K
• Median: Take the middle value when sorted ($14K, $15K, [$16K], $17K, $18K) = $16K
• Vote: If 3 friends say "buy" and 2 say "don't buy", go with "buy"