forum.alglib.net

ALGLIB's Multinomial Logistic Regression suite is outstanding, but I am now looking to do weighted multinomial logistic regression. (i.e. where each datapoint has an associated weight. Or another way - each "count" may not be an integer). So I am looking to manually edit the source code for ALGLIB to create a multinomial logistic regression training functon that accepts weights for each datapoint.

On the User Guide page for "Multinomial Logistic Regression" (http://www.alglib.net/dataanalysis/logit.php), there is a note that says "[Multinomial Logistic Regression] is a special case of a neural network, that is, a network with one linear layer and with SOFTMAX normalization."

Indeed, looking into the ALGLIB source code for the multinomial logistic regression training function "mnltrainh", I see that it already uses a (neural) network which is an instance of class multilayerperceptron, and the class multilayerperceptron has member variable called weights.

Can anybody shed some light? What is the meaning of the "weights" of a Neural Network (multilayerperceptron)?

More specifically: Do the "weights" of a Neural Network (in ALGLIB) have the same interpretation as the "weights" of datapoints that I am referring to in multinomial logistic regression? Do you think it would be just as easy as literally setting "network.weights = my_weights_per_datapoint" within mnltrainh ?