AI Briefing[NeurIPS 2024 Series] A B-spline-based Method for Normalizing Tabular Data Inputs - LG AI Research BLOG
Key point
LG AI Research has proposed a new technique that uses B-spline to dynamically normalize tabular data according to its complex distributions.
Details
Tabular Data has the characteristic that the data distribution varies greatly and is heterogeneous across each column. Existing normalization methods such as min-max, standardization, and quantile transformation are not optimized during the neural network training process, and have the limitation of being effective only for specific distributions.
To solve this problem, LG AI Research's DI Lab proposed a learnable numerical input normalization technique based on B-spline. This method defines the normalization process as a single continuous function (curve), and optimizes the curve's parameters simultaneously with neural network training.
The main reasons for choosing B-spline are as follows:
- It can freely adjust the slope of a specific section, allowing the neural network to designate areas that need to be learned as important
- It can generate smooth, continuous curves
- It can maintain the order of the data, ensuring the stability of normalization
The research team dynamically adjusts the normalization transformation according to the learning difficulty of each data sample. Through a mechanism that assigns low importance to easy samples with low training loss and adjusts the curve's slope so that the network can focus more intensively on samples that are difficult to learn, the method maximizes representation learning performance for tabular data.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.