ConvexFit.jl

Fit vectors with convex combinations of data columns

ConvexFit.jl is a lightweight Julia package for fitting vectors with convex combinations of data columns. Notably, the coefficients are always restricted to be nonnegative and sum to one. This restriction arises naturally in circumstances where predictions involving extrapolation are undesirable, for instance, when constructing weights for synthetic controls.

The coefficients for the convex combinations are obtained by solving a constrained optimization problem of the following form:

min_x ||Ax - b||² + λ||x||²
st.   x_i ≥ 0 for all i and Σ_ix_i = 1

where A is a matrix containing the data columns; x is a vector of coefficients on the unit simplex; b is a vector to be fitted by Ax; and λ is a nonnegative regularization parameter. Only the Euclidean norm is supported at this moment.

The optimization problem is solved iteratively with a conditional gradient method, the Frank-Wolfe algorithm, that directly searches solution candidates on the unit simplex. In practice, some extent of regularization is often desired and that can be controlled by the magnitude of λ. The choice of λ depends on the context of the specific problem and is left to be zero by default. If appropriate, one may consider selecting λ based on the leave-one-out cross validation, which is implemented in this package.

Most of the functionality can be accessed by calling convexfit. To fit a vector b with a convex combination of columns in A and regularize the coefficients x with some λ:

using ConvexFit
r = convexfit(A, b, λ)

The results are stored in SolverResult and can be retrieved from the corresponding fields. For example, r.sol gives the optimal x; while r.fit gives the fitted values.

To select an optimal λ based on the leave-one-out cross validation, one may either provide a grid in place of λ to convexfit for exhaustive search or specify a solver that searches the optimal λ over an interval. An example for the latter case that uses a solver from Optim.jl is as follows:

using ConvexFit, Optim
# Specify a solver in a function that takes an objective function as argument
# The returned object must be a tuple of the solver result and the minimizer
function fmin(f::Function)
    r = optimize(f, 0.0, 100.0, Brent(), abs_tol=1e-4, store_trace=true)
    return r, minimizer(r)
end
# Fit b under the optimal λ selected from [0, 100] based on leave-one-out cross validation
r = convexfit(A, b, optim(fmin))

More details can be found in the help mode of Julia REPL.

Jaggi, Martin. 2013. "Revisiting Frank-Wolfe: Projection-Free Sparse Convex Optimization." Proceedings of the 30th International Conference on Machine Learning 28 (1): 427-435.