What does it mean when i have 0 true positives and false positives?

pvds90

Hi guys,

I'm running a couple of clustering algorithms on a set with a target label. I'm measuring precision and recall, but both are 0% (zero false positives, zero true positives) for the agglomerative clustering algorithm. What does this say in general? I've never had something like that. With every measure type, it gives me 1126 TN & 1126 FN. I already downsampled the label to equal numbers, no difference. 

Thanks in advance.

jacobcybulski

I am not sure what you are trying to achieve as we have no control over the process of clustering data, it would be a surprise if the labels and clusters aligned. If you wanted to check if the labels follow some natural segmentation, perhaps you could try to analyse the contingency table of label vs cluster classes? If this is so, try a different clustering method, e. g k-means or DBSCAN. 

jacobcybulski

It means that your model does a "sink test" which means it classifies everything as negative, regardless of anything. So it is a bad model.

pvds90

@jacobcybulski
Do you have any tips how to get better performance? 

jacobcybulski

It is not clear to me how you get precision and recall from agglomerative clustering, especially that you already have some label. I do not think clustering is likely to generate your labels and it would be very hard to align your labels with clusters. Perhaps you are using a cluster variable as a label and then build a classification model? There is too many unknowns.

pvds90

I'm using agglomerative clustering and flatten clustering after that with k=2 (because i have 1 labeled cluster). It works when i have another target label, but with one target label i only get 0 percent scores. 

jacobcybulski

I am not sure what you are trying to achieve as we have no control over the process of clustering data, it would be a surprise if the labels and clusters aligned. If you wanted to check if the labels follow some natural segmentation, perhaps you could try to analyse the contingency table of label vs cluster classes? If this is so, try a different clustering method, e. g k-means or DBSCAN. 

pvds90

Thanks!