Learning Sparse Graphs for Functional Regression using Graph-induced Operator-valued Kernels
Authors: Akash Saha, Balamurugan Palaniappan
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | An extensive empirical evaluation on both synthetic and real data demonstrates the utility of the proposed learning framework. Our experiments show that simultaneous learning of F along with sparse graphical structure helps in discovering significant relationships among the input functions, and motivates interpretability of such relationships driving the regression problem. |
| Researcher Affiliation | Academia | Akash Saha EMAIL IEOR, IIT Bombay Mumbai, India P. Balamurugan EMAIL IEOR, IIT Bombay Mumbai, India |
| Pseudocode | Yes | Algorithm 1 MCP Regularization of L ... Algorithm 2 Alternating Minimization of J ... Algorithm 3 Sample-based Approximation ... Algorithm 4 Op MINRES(P, b, maxiter) ... Algorithm 5 Op Lanczos Step(P, qk, qk 1, βk) ... Algorithm 6 Sym Ortho(a, b) |
| Open Source Code | Yes | All methods were coded in Python 3.7 and the codes are made public.1 All experiments were run on a Linux box with 182 Gigabytes main memory and 28 CPU cores. 1Codes used for the experiments can be found at https://github.com/akashsaha06/graph-induced OVK. |
| Open Datasets | Yes | We consider 1 minute data of Wyoming ASOS data collected from IEM ASOS One Minute Data (Iowa Environmental Mesonet, 2022). ... This data is available in the Github repo NBA Movement Data (Seward, 2018). |
| Dataset Splits | Yes | The following data splits have been considered: (80/20/20), (160/40/40) and (320/80/80), representing the number of training samples/validation samples/test samples. ... The following random data splits have been considered: (80/20/20) and (472/123/123), representing the number of training samples/validation samples/test samples in small weather data and full weather data settings, respectively. ... A random data split of (233/59/59) representing the number of training samples/validation samples/test samples has been considered. |
| Hardware Specification | Yes | All experiments were run on a Linux box with 182 Gigabytes main memory and 28 CPU cores. |
| Software Dependencies | Yes | All methods were coded in Python 3.7 and the codes are made public. ... The quadratic programs involved in (23) and (15) are solved by using CVXOPT (Andersen et al., 2023). |
| Experiment Setup | Yes | For the MCP-based L learning and D learning in the proposed alternating minimization framework, we use a decaying step-size in the projected gradient descent. The decaying step-size regime involves starting with an initial step-size (e.g. 10^-4) and reducing it by a fixed factor (e.g. 2) after a set of iterations (e.g. 5) continuously till a final step-size (e.g. 10^-9). |