These datasets are used in the computational experiments of the paper "Simple steps are all you need: Frank-Wolfe and generalized self-concordant functions".
Some of the portfolio instances were taken from Dvurechensky, Pavel, et al., "Generalized Self-Concordant Analysis of Frank-Wolfe algorithms".
Others were generated with the same format and varying dimensions.
Each file is in MAT format, contains a single matrix
W of size
d × p with
d the decision dimension and
p the number of periods.
The logistic regression instances were taken from the LIBSVM datasets and converted using libsvm2csv.
The first columns with non-numeric labels are not features and should be discarded, the outcome is the
target column which may be encoded in different manners (-1 and 1, 1 and 2, etc).
The phase retrieval instances were generated in MAT format as described in the article, they all contain a
W matrix, a
y outcome vector and a
theta vector which is not required.
W has as many rows as the length of
y and columns as the initial input vector.