2021 06-02-tabnet

1. TabNet inputs raw tabular data without any preprocessing and is trained using
gradient descent-based optimization.
2. TabNet uses sequential attention to choose which features to reason from at each
decision step.
-> interpretability and better learning as the learning capacity is used for the most salient features.
3. TabNet outperforms or is on par with other tabular learning models on various
datasets.
4. We show performance improvements by using unsupervised pre-training to
predict masked features.
Contribution

Feature selection broadly refers to judiciously picking a subset of
features based on their usefulness for prediction.
Feature selection & Tree-based learning
Their prominent strength is efficient picking of global features with the
most statistical information gain.
To improve the performance of standard DTs(Decision Trees), one
common approach is ensembling to reduce variance.

Random Forest : random subsets of data with randomly selected features to grow
many trees.
XGBoost & LightGBM : tree-based ensemble. (다음 기회에 ..)
Integration of DNNs into DTs ->
Self-supervied learning -> BERT
Feature selection & Tree-based learning

 TabNet is based on such functionality and it outperforms DTs while reaping their benefits by
careful design which
(i) uses sparse instance-wise feature selection learned from data
(ii) constructs a sequential multi-step architecture, where each step contributes to a
portion of the decision based on the selected features
(iii) improves the learning capacity via nonlinear processing of the selected features
(iv) mimics ensembling via higher dimensions and more steps.
TabNet

𝑴 𝒊 ⋅ 𝒇
𝑴 𝒊
𝑴 𝒊
𝒂 𝒊 − 𝟏
𝑷 𝒊 − 𝟏
𝒅 𝒊
𝒂[𝒊]
BN : Ghost Batch Normalization
GLU : Gated Linear Unit
𝜼

Sparsemax
The idea is to set the probabilities of the smallest values of z to zero and keep only probabilities of the highest
values of z, but still keep the function differentiable to ensure successful application of backpropagation
algorithm.
출처 : https://towardsdatascience.com/what-is-sparsemax-f84c136624e4
𝑎 𝑖 − 1 ∶ 𝑀 𝑖 = 𝑠𝑝𝑎𝑟𝑠𝑒𝑚𝑎𝑥 𝑃 𝑖 − 1 ⋅ ℎ𝑖 𝑎 𝑖 − 1 Note, 𝑗=1
𝐷
𝑀 𝑖 𝑏,𝑗 = 1

 We pass the same 𝐷-dimensional features 𝑓 ∈ ℝ𝐵×𝐷
to each decision step, where 𝐵 is the batch size.
 𝑴 𝒊 ∈ ℝ𝐵×𝐷
for soft selection of the salient features. (𝑴 𝒊 ⋅ 𝒇)
 𝑃 𝑖 is the prior scale term, denoting how much a particular feature has been used previously
 𝑃 𝑖 = 𝑗=1
𝑖
(𝛾 − 𝑴 𝒋 )
 𝛾 : relaxation parameter
 𝛾 = 1, feature is enforced to be used only at one decision step and as 𝛾 increase, more flexibility is provided to use a
featue at multiple decision steps.
 𝑃 0 is initialized as all ones, 1𝐵×𝐷
, without any prior on the masked features.
𝐿𝑠𝑝𝑎𝑟𝑠𝑒 = 𝑖=1
𝑁𝑠𝑡𝑒𝑝𝑠
𝑏=1
𝐵
𝑗=1
𝐷 − 𝑴𝒃,𝒋 𝒊 log 𝑴𝒃,𝒋 𝒊 +𝜖
𝑁𝑠𝑡𝑒𝑝𝑠⋅𝐵
Feature Selection

 If 𝑀𝑏,𝑗 𝑖 = 0, then 𝑗𝑡ℎ
feature of the 𝑏𝑡ℎ
sample should have no contribution to the decision.
 If 𝑓𝑖 were a linear function, the coefficient 𝑀𝑏,𝑗 𝑖 would correspond to the feature importance of 𝑓𝑏,𝑗.
 We simply propose 𝜂𝑏 𝑖 = 𝑐=1
𝑁𝑑
𝑅𝑒𝐿𝑈 𝑑𝑏,𝑐 𝑖 to denote the aggregate decision contribution at 𝑖𝑡ℎ
decision step for the 𝑏𝑡ℎ
sample.
 If 𝑑𝑏,𝑐 𝑖 < 0, then all features at 𝑖𝑡ℎ
decision step should have 0 contribution to the overall decision.
 We propose aggregate feature importance mask,
𝑀𝑎𝑔𝑔−𝑏,𝑗 =
𝑖=1
𝜂𝑏 𝑖 𝑀𝑏,𝑗 𝑖 /
𝑗=1
𝐷
𝑖=1
𝜂𝑏 𝑖 𝑀𝑏,𝑗 𝑖
Interpretability

Tabular self-supervised learning
Propose the task of prediction of missing
feature columns from the others.
binary mask : 𝑆 ∈ 0,1 𝐵×𝐷
encoder inputs : 1 − 𝑆 ⋅ 𝑓
->
decoder outputs : 𝑆 ⋅ 𝑓
𝑃 0 = (1 − 𝑆)
reconstructure loss:
𝑏=1
𝐵
𝑗=1
𝐷 𝑓𝑏,𝑗 − 𝑓𝑏,𝑗 ⋅ 𝑆𝑏,𝑗
𝑏=1
𝐵
𝑓𝑏,𝑗 − 1/𝐵 𝑏=1
𝐵
𝑓𝑏,𝑗
2
2

Instant-wise feature selection (AUC)
The datasets are constructed in such a way that only a subset of the features determine the output.

Performance on real-world datasets

2021 06-02-tabnet

More Related Content

What's hot (20)

Similar to 2021 06-02-tabnet (20)

More from JAEMINJEONG5 (20)

Recently uploaded (20)

2021 06-02-tabnet