keras-io/structured-data-classification-grn-vsn

Model description

This model is built using two important architectural components proposed by Bryan Lim et al. in Temporal Fusion Transformers (TFT) for Interpretable Multi-horizon Time Series Forecasting called GRN and VSN which are very useful for structured data learning tasks.

1. Gated Residual Networks(GRN): consists of skip connections and gating layers that facilitate information flow efficiently. They have the flexibility to apply non-linear processing only where needed.
   GRNs make use of Gated Linear Units (or GLUs) to suppress the input that are not relevant for a given task.
   The GRN works as follows:
   • It first applies Non-linear ELU tranformation on its inputs
   • It then applies a linear transformation followed by dropout
   • Next it applies GLU and adds the original inputs to the output of the GLU to perform skip (residual) connection
   • Finally, it applies layer normalization and produces its output
2. Variable Selection Networks(VSN): help in carefully selecting the most important features from the input and getting rid of any unnecessary noisy inputs which could harm the model’s performance.
   The VSN works as follows:
   • First, it applies a Gated Residual Network (GRN) to each feature individually.
   • Then it concatenates all features and applies a GRN on the concatenated features, followed by a softmax to produce feature weights
   • It produces a weighted sum of the output of the individual GRN

Note: This model is not based on the whole TFT model described in the mentioned paper on top but only uses its GRN and VSN components demonstrating that GRN and VSNs can be very useful on their own also for structured data learning tasks.

Intended uses

This model can be used for binary classification task to determine whether a person makes over $500K a year or not.

Training and evaluation data

This model was trained using the United States Census Income Dataset provided by the UCI Machine Learning Repository.
The dataset consists of weighted census data containing demographic and employment related variables extracted from 1994 and 1995 Current Population Surveys conducted by the US Census Bureau.
The dataset comprises of ~299K samples with 41 input variables and 1 target variable called income_level
The variable instance_weight is not used as an input for the model so finally the model uses 40 input features containing 7 numerical features and 33 categorical features: