推荐阅读
- NIPS'22 Why do tree-based models still outperform deep learning on typical tabular data?
- nips'23 When Do Neural Nets Outperform Boosted Trees on Tabular Data?
TabLLM
pmlr2023,
Few-shot Classification of Tabular Data with Large Language Models
方法
使用把tabular数据序列化成文字的方法进行classification。
使用的序列化方法有几个，有人工也有AI生成。
效果
做few shot learning的效果
看上去一般。
TransTab
Learning Transferable Tabular Transformers Across Tables
方法
属于transfer learning的方法。对category、binary和numeric值进行embedding后再进行transformers最后进行classification。
使用场景
原文：
- S(1) Transfer learning . We collect data tables from multiple cancer trials for testing the efficacy
of the same drug on different patients. These tables were designed independently with overlapping
columns. How do we learn ML models for one trial by leveraging tables from all trials?
- S(2) Incremental learning . Additional columns might be added over time. For example, additional
features are collected across different trial phases. How do we update the ML models using tables
from all trial phases?
- S(3) Pretraining+Finetuning . The trial outcome label (e.g., mortality) might not be always available
from all table sources. Can we benefit pretraining on those tables without labels? How do we finetune
the model on the target table with labels?
- S(4) Zero-shot inference . We model the drug efficacy based on our trial records. The next step is to
conduct inference with the model to find patients that can benefit from the drug. However, patient
tables do not share the same columns as trial tables so direct inference is not possible.
效果
具体看原文吧，与当时的baseline比有提升。
MET
Masked Encoding for Tabular Data
tabtransformer
2020年，arxiv，TabTransformer: Tabular Data Modeling Using Contextual Embeddings
方法
transformer无监督训练，mlp监督训练。
原文
we introduce a pre-training procedure to train the Transformer layers using unlabeled data . This is followed by fine-tuning of the pre-trained Transformer layers along with the top MLP layer using the labeled data
效果
跟mlp
跟其他模型
tabnet
2020， arxiv，Google Cloud AI，Attentive Interpretable Tabular Learning, 封装的非常好，都可以当工具包使用了。
方法
跟transformer没关系的。
feature selection用的是17年的某个选择模型，最后agg一下做predict。