Deep Lattice Networks and Partial Monotonic Functions
This addresses the need for interpretable and reliable models in domains requiring monotonicity guarantees, such as finance or healthcare, though it appears incremental as it builds on existing lattice and calibration techniques.
The paper tackles the problem of learning deep models that are monotonic with respect to user-specified inputs by proposing deep lattice networks with alternating layers and constraints, achieving state-of-the-art performance on benchmark and real-world datasets for classification and regression.
We propose learning deep models that are monotonic with respect to a user-specified set of inputs by alternating layers of linear embeddings, ensembles of lattices, and calibrators (piecewise linear functions), with appropriate constraints for monotonicity, and jointly training the resulting network. We implement the layers and projections with new computational graph nodes in TensorFlow and use the ADAM optimizer and batched stochastic gradients. Experiments on benchmark and real-world datasets show that six-layer monotonic deep lattice networks achieve state-of-the art performance for classification and regression with monotonicity guarantees.