Deep Lattice Networks and Partial Monotonic Functions

• DOI: 10.48550/arxiv.1709.06680
• Publication Date: 2017-01-01
• Authors: Seungil You; Xin Ding; Kevin Robert Canini; Jan Pfeifer; Maya R. Gupta
• Keywords: Benchmark (surveying); Lattice (music)

We propose learning deep models that are monotonic with respect to a user-specified set of inputs by alternating layers linear embeddings, ensembles lattices, and calibrators (piecewise functions), appropriate constraints for monotonicity, jointly training the resulting network. implement projections new computational graph nodes in TensorFlow use ADAM optimizer batched stochastic gradients. Experiments on benchmark real-world datasets show six-layer lattice networks achieve state-of-the art performance classification regression monotonicity guarantees.