DBScan clustered back-feeding

Airbus_emitterAirbus_emitter Member Posts: 6 Learner I
Good morning,

I am new here and also in rapid miner. I am trying to do a clustering using dbscan. 
Currently I get the clustering I want by manually giving more or less weight to my features. But I would like to know some operator that automatically does this function. Depending on the input data, the weight that each of the parameters should take is self-adjusted.

I show the small process that I have set up right now.

In this case I deal with signals from a transmitter. I have 6 parameters. RF_min, Rf_max, PRI_max, PRI_min, PW_max, PW_min. 
I would like to modify the type of clustering it does depending on the dispersion of the entered data. By this I mean, for example, if there are values with RFmin around 1000 and others with 8000, it should not give so much weight to RF. However, if all the values are between 5000 and 6000, it should give more weight, so that the clustering is more suitable.
This is how I choose the weights in the generateAttributes operator:

Regarding the dbscan values, I have tried adjusting the min points, or the epsilon, but I don't like the clustering it produces. So right now I have min points:1, epsilon 0.2.

I show an example of what happens. In this case, PW for example should not have variations and should take 6.84 and 7 in the same cluster and not separate it. But this happens because RF should have less weight to disaggregate it less.

Properly done would be to move the weights of RF, to give less importance to this, and so that it does not create new clusters due to that small variation of RF. In this case, it should take as RF is always between 9464 and 9871, without making new clusters between these values. So I de-weight RF, dividing it now by 1000, and I get this.

Thanks for your help
Sign In or Register to comment.