Experiments for TIT paper "Optimal Rates for Agnostic Distributed Learning".

---

Installation on LINUX

To install FALKON, set the MATLAB shell path to the kernels folder and run:

mex -largeArrayDims ./tri_solve_d.cpp -lmwblas -lmwlapack

mex -largeArrayDims ./inplace_chol.cpp -lmwblas -lmwlapack

To install libsvm for loading libsvm datasets, set the MATLAB shell path to the ./utils/libsvm/matlab folder and just run:

make

Folders

• kernels contains basic functions for kernels and random features.
• methods consists of compared methods.
• utils involves utility tools, including figures, data processing, simulated data generating and so on.
• datasets stores the primal datasets from libsvm or other sources.
• data stores '.mat' files after data processing for datasets.
• results stores intermediate results and outputs.

Utils

• generate_simulated_spline is used to generate simulated dataset with different settings.
• data_preprocess generates the training and testing sets based on given datasets.

Description

• best_parameters_SSL_DKRR includes used hyperparameters for datasets, including data partitions, $\sigma$, $\lambda$ and iteration times.
• test_ssl_unlabeled records the test errors and training times in terms of different number of unlabeled examples.
• test_ssl_unlabeled is used to draw two figures: classification error and training time versus the number of unlabeled examples.
• test_ssl_partitions records the test errors and training times in terms of different number of partitions.
• draw_ssl_partitions is used to draw two figures: classification error and training time versus the number of partitions.
• run_script_SSL_DKRR is the script that conducts experiments and draw visual comparison in terms of different datasets.

How to RUN

• Just run run_script_SSL_DKRR.
• OR run test_ssl_unlabeled and test_ssl_unlabeled for comparion on the number of unlabeled exmaples.
• OR run test_ssl_partitions and draw_ssl_partitions for comparion on the number of partitions.