Check the output of the outputcolumns https://docs.microsoft.com/en-us/dotnet/machine-learning/resources/tasks in all trainers:
binaryclassification: outputs range from 0 to 1 - wether the given data is or is not something (rows labeled 0 or 1). multiclassclassification: outputs an array which contain ranges of 0 to 1 - wether the given data is or is not something. Rows of the label column are integers 0 - N. (If label number 0 is pizza, label number 1 is hamburger and label number 2 is kebab, output array[0] is wether or not the given data is pizza, raging 0.0 - 1.0, and array[1] 0.0 - 1.0 wether it is being hamburger, array[2] 0.0 - 1.0 wether it is being kebab). Seem like it does not always have to outpu float array though?: https://docs.microsoft.com/en-us/dotnet/machine-learning/tutorials/github-issue-classification. But the predicted value is form PreictedLabel - which seems like it must be in string.
regression: outputs a predicted float - based on trained data where one column (of float) was marked as label - with new data missing the values of that column. Basically predicts missing column values in new data.
clustering: outputs a uint - the id of the nearest cluster, and a float array containing the distances to each cluster where index is cluserid - 1. The metrics given from mlContext.Clustering.Evaluate() contains DaviesBouldinIndex. To find the most suitable numberOfClusters for the data - find value of DaviesBouldinIndex where it is being the lowest.
anomalydetection: outputs a double array with dimension depending on the type of transform used from the TimeSeries namespace, for instance DetectSpike or DetectChangepoint. Alert 0 or 1 specifies if the data point with 1 indicating it is spike/is changepoint, score is the input column of the data (in float), P-Value indicate how likely the point is being an anomaly, between 0 and 1 where towards 0 indicate anomaly, and Martingale value of how "weird" the data point is: https://docs.microsoft.com/en-us/dotnet/machine-learning/tutorials/sales-anomaly-detection
ranking:
recommendation: ouputs a a float score where high values indicate suitable recommandation - trained from 2 column features. Find highest score match between (for instance) customerid 1 with N productids.