> ## Documentation Index
> Fetch the complete documentation index at: https://kumo.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# majority_sampling_ratio

#### `majority_sampling_ratio: <list[float]>` (Optional)

## Description

A list of majority sampling ratios for AutoML to explore.

The `majority_sampling_ratio` parameter controls undersampling of the majority class in binary classification tasks.\
It specifies how many majority-class examples to keep **per minority-class example** during training.

#### **In other words:**

For every example in the minority class, we sample `majority_sampling_ratio` examples from the majority class.\
This parameter must be **greater than 0**.

#### **Behavior**

* If the dataset’s actual **majority-to-minority ratio** is **greater** than the specified `majority_sampling_ratio`, undersampling is applied to reduce the imbalance.
* If the dataset’s actual ratio is **less than or equal to** the specified ratio, the parameter has **no effect** (i.e., all data are used).

If row weights are also present for binary classification (from SDK/custom training-table workflows or from [weight\_col](/reference/weight_col)), both signals are used together according to `weight_mode`:

* `weight_mode=sample`: row and class weights are combined for sampling.
* `weight_mode=weighted_loss`: row and class weights are combined for loss weighting.
* `weight_mode=None`: invalid when `majority_sampling_ratio` is set.

**Example 1: Undersampling applied**

Suppose your dataset has:

* Majority-class examples: 10,000
* Minority-class examples: 100\
  → Actual ratio = **100:1**

If you set:

```
majority_sampling_ratio = 20
```

Then for each minority-class example, we keep 20 majority-class examples.\
Resulting sampled data:

* Majority-class examples kept: 100 × 20 = 2,000
* Minority-class examples: 100\
  → Resulting ratio = **20:1**

Undersampling is applied because the actual ratio (100) is **greater** than the desired ratio (20).

**Example 2: No effect (ignored)**

Using the same dataset (10,000 majority, 100 minority → 100:1 ratio), if you set:

```
majority_sampling_ratio = 150
```

Then the desired ratio (150:1) is **larger** than the dataset’s actual ratio (100:1).\
Since the dataset is already *less imbalanced* than the target, no undersampling occurs.\
All majority examples are kept, and this setting is ignored.

**Summary table**

| Dataset Majority:Minority | `majority_sampling_ratio` | Action Taken                    | Resulting Ratio |
| :------------------------ | :------------------------ | :------------------------------ | :-------------- |
| 100:1                     | 20                        | Undersample majority            | 20:1            |
| 100:1                     | 50                        | Undersample majority            | 50:1            |
| 100:1                     | 100                       | No change (equal ratio)         | 100:1           |
| 100:1                     | 120                       | Ignored (ratio already smaller) | 100:1           |

### Supported Task Types

* Binary Classification

### Default Values

| run\_mode | Default Value |
| --------- | ------------- |
| FAST      | None          |
| NORMAL    | None          |
| BEST      | None          |