With an old challenge
Training your model to get the detection of spam. Your data has many benefits to some degree, so you plant many hours of work to resume 50/50. Now you are satisfied because he was able to deal with the inequalities of the class. What if I have told you that 60/40 may not be enough, but the best?
For many machines for the division of a machine, one class conditions without other classes. This reduces learning [1] and may participate in trained models [2]. The most commonly used methods are subject to simple law: to find a way to provide all the same classes. Usually, this is done in simple ways as providing more value from smaller-level examples (restrictions), to remove multiple examples of class (not focus), or including very little class conditions.
The validity of these methods usually chat, so the Miori work and empirical indicates which solution is effective in terms of your request [3]. However, there is a hidden idea that can be discussed often and that is often easily taken: repatriation and a good idea? To some extent, these methods apply, so the answer is yes. But we should fully Returning our details? To make it easier, let us take a binary problem of binary separation. Should we return our training data to have 50% of each class? Intuition says yes, as well as the intuition guidelines up to date. In this case, intuition is wrong. For accurate reasons.
What does it mean by “inequality”?
Before we give up how much and why 50% is not the full training for unequalities in binary division, let's explain the right prices. We call ni₀ Number of conditions of one paragraph (usually, a small category), and ni₁ those other class. In this way, the total number of data conditions for training ni=ni+ni₁. Quantity Tongue Today Training Training,
ρ⁽ᵗʳᵃⁱⁿ⁾ = ni₀ /ni .
Proof that 50% underground
The first proof comes from a royal activity in random forests. Malov and participants set up full training inequality, ρ⁽ᵒᵖᵗ⁾, in 20 datasets [4]. They find its value vary in trouble to trouble, but conclude that it is above or less ρ⁽ᵒᵖᵗ⁾ = 43%. This means that, according to their tests, you want most of the smaller grade grants. This is not a perfect story. If you want to aim for the correct models, don't stop here and straight back ρ⁽ᵗʳᵃⁱⁿ⁾ to 43%.
In fact, this year, the Thiri work is Pezzololi et al. [5]It has shown that appropriate inequality is not a valid universe in all applications. Not 50% and is not 43%. It is evident, not perfect inequality varies. Never little sometimes can be smaller than 50% (as kamolovs and estimated participants), and more than 50%. A certain amount of ρ◇ It will depend on the details of each specific partition problem. Another way to find ρ⁽ᵒᵖᵗ⁾ Training model in several values ρ⁽ᵗʳᵃⁱⁿ⁾, and estimate related performance. This is possible by example looks like this:
Although straight patterns decided ρ◇ Is not clear, it seems that when data is full compared to the size of the model, malicity is smaller than 50%, as in Valov test. However, some of the many things – that the unusual conditions are relieving, that the dynamics of unlimited training – how to set the correct training amount, how much performance is lost when a person is lost when a person is lost. ρ⁽ᵒᵖᵗ⁾.
Why is the perfect balance doesn't always be the best
As we mentioned, the answer is actually accurate: As different classes with different structures, there is no reason why both classes will handle similar information. In fact, Pezzololi's group has proven that they are usually unable. Therefore, reduce the excellent organization boundary we may need many scenarios more than another. Pozzicoli's work, in the Mongoli of Aomaly's acquisition, provides a simple and understanding example.
Let's think that information is from the distribution of PUussian multivarian, and that we write the labels all the points on the right of the restriction. In 2D, it looked like this:
The cooked line is our decision-making border, and the points on the right side of the decision is ni₀ anomalies. Let's now restart our data to ρ⁽ᵗʳᵃⁱⁿ⁾ = 0.5. To do so, we need many Anomons. As the Anomomies are not rare, what we can find most closely near the tin. Just eye eyes, this situation is clearly clear:
Anomalies, yellow, tied at the decision border, so they are more informed about its position than blue points. This may not think that it's better for the little class points. Only the other hand, the Anomalies cover one side of the decision boundary, so that one has a small class point, it would be easier to invest in many distances, to cover the other side of the religions. As a result of these two competitive effects, ρ⁽ᵒᵖᵗ⁾ Usually it is not 50%, and its exact value depends on the problem.
The root cause asymmetry
Pezzixo's vision shows that perfect imbalance often differ from 50%, because different classes have different buildings. However, they only analyze one source of diversity between classes, that is, different behavior. However, as it is for example shown by Sarao-Mannelli and Coauntry [6]There are many results, such as the presence of small groups within classes, which can produce the same effect. The consistency of the largest result of the effects determine the diversity between classes, which tells us what is the full imbalance of our specific problem. Until we have a sense of charge all asymmetry sources in detail together (including how the model structures apply), we will not know the unequal of data training in advance.
Having a Key & you can do differently
If so far and re-register your binary data to 50%, you've been doing well, but you may not do it very well. Although we do not have a vision you can tell us what relevant training is the right training, now you know it may not have 50 percent. The good news is to be on the way: Theorists' machine reading deals with this article diligently. Meanwhile, you can imagine ρ⁽ᵗʳᵃⁱⁿ⁾ As a hyperparameter you can use premature, just as any other hyperparameter, cross your data effectively. So before running the following model training, ask yourself: Is you just 50/59 really? Try to switch your class imperfection – your model performance can surprise you.
Progress
[1] E. Francizi, Um. Baity-Jei, and A. Lucchi, an idealized analysis of the ability to study under the Foxince (2023), CML 2023
[2] K. Ghosh, C. Bellinger, R. Corizzo, P. Branco, b. Krawczyk, and n. japkowicz, an inequalities problem in a shallow learning (2024), Machine reading, 113(7), 4845-4901
[3] E. Loffredo, M. Prestore, S. Cocco Nor. Monasson, Restoring Balance: OverSampled under / Data Data Data (2024), CML 2024
[4] F. Arxiv Print Arxiv: 2207.04631
[5] FS Pezzompli, V. ROS, FP LANDES AND. Bainty-Jei, Class Headfall in Anomaly receiving: Learn from the properly suspected model (2025). Aistats 2025
[6] S. Sarao-Mannelli, F. Gerace, N. Rostamzadeh and L. Saglietti, bias-inductucing geometries: Fairbiedness Implement Model (2022), Arxiv Print Arxiv: 2205.15935
