THE DATASET CLASS BALANCER
Intermediate
Difficulty: Medium Tags: Data Science, Sampling, Imbalanced Data
Story
A training pipeline computes inverse-frequency class weights to counteract an imbalanced dataset. A bug divides by the number of classes instead of the number of examples, scaling the loss incorrectly.
The Math
For N examples across C classes, with class count n_c > 0:
w_c = N / (C * n_c)
If a class has zero examples, its weight is defined as 0.
Input Format
counts, a length-C list of non-negative class counts whose sum is N.
Output Format
A length-C list of weights, each to 6 decimal places.
Constraints
1 <= C <= 10^5,0 <= N <= 10^9.- Time limit: 1.0 second.
Example
Input
counts = [5, 3, 2]
Output
[0.666667, 1.111111, 1.666667]
Explanation: N = 10, C = 3, so N/C = 10/3; dividing by each class
count gives 10/3/5 = 0.666667, 10/3/3 = 1.111111, 10/3/2 = 1.666667.
Keywords
data-sciencesamplingimbalanced-data
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.