For modern datascience we need machine learning. Hence for .NET a modern machine learning library would be very welcome, especially one with a good F# interface. It should contain the most used standard algorithms & models like

• clustering: k-means, spectral clustering, kernel k-means, gaussian mixture, ...


• classification & regression: (regularized) generalized linear models , neural networks (classic & deep learning), kernel models (SVM, kernel logistic regression,...), probabilistic (max likelihood) & Bayesian methods/graphical models/gaussian processes, decision trees & random forests,...


• ensemble methods (bagging, boosting,...)


• dim reduction (SVD, kernel SVD, manifold learning, sparse representations (L1 regularization),...)


• probabilistic programming (cf. infer.NET) ?


• ...

Also applications to specific domains would be nice like
- recommendations!!!
- text mining/information retrieval (calculate tf-idf,LSA, LDA,...)
- network analysis
- adaptive DSP/compression/image processing/computer vision
- ...

As a start we could leverage accord.Net and build further on that.

as a side project one could have library for optimization that could be used by the ML lib & directly:
- gradient descient
- linear/quadratic/ convex programming
- other methods like genetic algos/evolutionary computing,simulated annealing,...

Hi,
First, congratulations for this ongoing work! I came to Math.Net after attempting to create my own libraries, so here is an idea which my be of a help.

The idea would be to use Structs to define matrices instead of Classes to improve performance.

As mentioned above, I did a check with a homemade matrix class and another one being a matrix struct, and for matrices up to 5000X5000 structs are circa 10 times faster for matrix multiplication.

I compared Math.Net with MKL against matlab and it appears to be circa 2 times slower for such matrix sizes. This gives some hope that using structs might bring execution times similar to matlab.