Difference between PD profile and CP profile when we manually specify the instances

[Nehagupta90]

Hi

If I have 100 observations/instances and want to create the PD profile of few important features, then what is the difference between the PD profile of these instances using

variab= c("var1","var2", "var3" )
pdp <- model_profile(explainer = explainer, variables = variab)
plot(pdp)

and between the CP profile when we manually specify the instances like

new_observation= data[c(2,7,8,11,15,16,21,24,25,26,30,45,46),]
cp<- ceteris_paribus(explainer, new_observation)
cp_agg <- aggregate_profiles(cp, variables = variab)
plot(cp_agg)

[hbaniecki]

Hi, the diffrerence is that in the second case you create the PD profile based on only a subset of data rows (I see 13 observations), while in the first case you use N = 100 observations to estimate the PD profile.

[Nehagupta90]

Thank you Sir So is it a good idea to provide the profile for specific instances rather than just one profile for all instances.. Let me explain: I predict the Results of students of 100.. Some students/instances are good students/instances, some average and some novice. Now is it a good idea to provide the profile of instances predicted as **good**, another profile for **novice** students and so on? Do you think it is a good idea or just providing the PD profile which averages all the instances? Warm regards

…

On Thu, Jul 29, 2021 at 2:02 PM Hubert Baniecki ***@***.***> wrote: Hi, the diffrerence is that in the second case you create the PD profile based on only a subset of data rows (I see 13 observations), while in the first case you use N = 100 observations to estimate the PD profile. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#444 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AN2ZZ2ODKYWV5EX6K2VFAFTT2E7MFANCNFSM5BERORVA> .

[hbaniecki]

It depends on the explanation that you want to achieve. In model_profile there are groups (group profiles/instances by a variable) and k (cluster profiles/instances) parameters that allow for customization. Actually, we discuss the context of creating explanations in the paper https://arxiv.org/abs/2105.13787, and specifically the context of profiles in https://arxiv.org/abs/2105.12837. Hope it helps

Kind regards

[Nehagupta90]

In model_profile there are groups (group profiles/instances by a variable) and k (cluster profiles/instances) parameters that allow for customization Hi Hubert, could you please the link to this group profiles? Is it allow the same type of instances to be in one group and another type of instances in another group?

…

On Fri, Jul 30, 2021 at 11:34 AM Hubert Baniecki ***@***.***> wrote: It depends on the explanation that you want to achieve. In model_profile there are groups (group profiles/instances by a variable) and k (cluster profiles/instances) parameters that allow for customization. Actually, we discuss the context of creating explanations in the paper https://arxiv.org/abs/2105.13787, and specifically the context of profiles in https://arxiv.org/abs/2105.12837. Hope it helps Kind regards — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#444 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AN2ZZ2PJZUJNBCQOKKIMM33T2JW3PANCNFSM5BERORVA> .

[hbaniecki]

It is all in the documentation https://modeloriented.github.io/DALEX/reference/model_profile.html. Yes, you might want to group observations by a categorical variable.

[Nehagupta90]

It does not work with numeric variables? I have mostly numeric variables in my dataset.

…

On Fri, Jul 30, 2021 at 11:49 AM Hubert Baniecki ***@***.***> wrote: It is all in the documentation https://modeloriented.github.io/DALEX/reference/model_profile.html. Yes, you might want to group observations by a categorical variable. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#444 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AN2ZZ2OPCU23RU6ICNPDC6TT2JYUJANCNFSM5BERORVA> .

[hbaniecki]

Well, the colour is a group determined by a category. You could probably dichotomize a variable. Any specific use case can be customized/programmed, so can the groups in aggregate_profile

[Nehagupta90]

Thanks a lot @hubert for the useful information.

…

On Fri, Jul 30, 2021 at 11:56 AM Hubert Baniecki ***@***.***> wrote: Well, the colour is a group determined by a category. You could probably dichotomize a variable. Any specific use case can be customized/programmed, so can the groups in aggregate_profile — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#444 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AN2ZZ2L2N7PT3K6AHRL6YOLT2JZNTANCNFSM5BERORVA> .