On Jul 16, 2007, at 05:03 AM, Toon Verwaest wrote:
>>> It is true that we don't have information, but in the case that the
>>> system would try to put such classes together with classes in our
>>> system, this should probably fail since we don't want to find
>>> software
>>> components in libraries. This is where the "Fixating" tab comes in
>>> handy, were we can exclude results library classes. If we do this,
>>> even while they were clustered together, we exclude those results.
>>
>> That was what I would tell you indirectly: It is the best simple
>> solution...
>> I understand perfectly the following of your mail, and I know that a
>> clustering algorithm is not a kind of magic which could give us an
>> ideal solution without any user's intervention!!! The objective was
>> simply pointing out this point which does not occurs only in the case
>> of -System under analysis and Platform library-. Ok I will not discus
>> now the problem of inter-systems collaboration and system's
>> plugins...
>
> I just extended the program as such, as defined in the semi-automatic
> method of Koschke, that certain elements and relations can be ignored
> from results of plugins. The change is that, now after ignoring
> classes (or relationships, which already works but is not yet
> integrated in UI), (parts of) the plugins get recalculated ignoring
> the removed relations or full classes and all relations from/to these
> classes. This instead of just making a subset of the cluster without
> the ignored element. For example when applying SCC to moosified moose,
> if we would originally remove MooseModel and FAMIXClass, which are
> part of a SCC cluster of 125 classes, the new cluster would have been
> 123 classes (125 - 2...) while in the new version, it rechecks the
> graph, and it will find that only about 5 small subgraphs of the
> original
> graph is are still strongly connected.
>
excellent :-)
> The downside... removing links, fixating clusters or removing classes
> forces you to recalculate the results of the plugins (at least
> partly).
nothing is gratuitous (gratos) :-)
>
> Toon
P.S.: writing/working at 05:00 AM is a very good sign :-) , and that
leads me to tell you best wishes and very nice day :)
Hani
thank you very much for your detailed answer.
On Jul 13, 2007, at 19:52 PM, Toon Verwaest wrote:
> On (13/07/07 18:23), Hani ABDEEN wrote:
>> From: Hani ABDEEN <hani.abdeen(a)univ-savoie.fr>
>> To: Toon Verwaest <tverwaes(a)infogroep.be>
>> Subject: Re: [Moose-dev] Class clustering (Bug:Exceptions)
>> Date: Fri, 13 Jul 2007 18:23:04 +0200
>>
>> Hi,
>> thank you,
>> yes It works correctly now :-) and I try to understand and to
>> discover what we can do with this plugin.
>> Just somethings that I would point out:
>> - I have seen that when the clustering algorithm take place, all the
>> classes, the packages, and the namespaces in the model under-
>> consideration will be considered in the same way by the algorithm.
>> What about the classes and the packages which do not really belong to
>> the system that we would analyze it (stubs entities)? Such entities
>> are used (accessed) by the considered system, and we have this
>> information in moose model, but we have not any information about
>> what these entities access?. may be they access our system, and may
>> be not!!
>
> It is true that we don't have information, but in the case that the
> system would try to put such classes together with classes in our
> system, this should probably fail since we don't want to find software
> components in libraries. This is where the "Fixating" tab comes in
> handy, were we can exclude results library classes. If we do this,
> even while they were clustered together, we exclude those results.
That was what I would tell you indirectly: It is the best simple
solution...
I understand perfectly the following of your mail, and I know that a
clustering algorithm is not a kind of magic which could give us an
ideal solution without any user's intervention!!! The objective was
simply pointing out this point which does not occurs only in the case
of -System under analysis and Platform library-. Ok I will not discus
now the problem of inter-systems collaboration and system's plugins...
>
> Important is to think that most of the algorithms, just present the
> basic results without automatically combining them, to the user,
> who can then manually validate clustering information before merging
> it. The reason why this is done like this, is because a lot of the
> basic algorithms (for example the "By Reference") will definitaly
> cause false positives because of the way the algorithm is adapted.
>
> Next to that, the ones which do already combine results, like the
> hierarchical clustering based on delta-ic and interface, can
> automatically be filtered for "false positives", be automatically
> removing nodes in the hierarchies which only add information about
> "ignored classes", which we can ignore in the "fixating tab".
>
> But well, it might be useful to already provide a checkbox which
> automatically exludes these classes, so the user doesn't have to find
> out which ones are part of the subject system and which ones aren't..
>
>>
>> - this is a very small thing that may be it does not cause any
>> problem now, but it could do that latter: According to your
>> implementation of FAMIXClass>>accessedByClasses and
>> FAMIXClass>>allClassAccesses, a class could be accessed be it self.
>>
>
> I guess you are right, that in some situations this could lead to a
> problem. I will move the addition of the class itself to the
> algorithms that use it.
>
> One of the basic things you can do with the program now, is using the
> different visualizations to analyse user-defined clusters. In other
> words how packages are organized in the system. Like this you can see
> for example, that the package "Kernel-Objects" contains "green
> classes" which are "ghostclasses" or classes which probably should be
> in the cluster, but are not. Of course since it is a core package and
> since the class causing the effect is the Object-class it might
> not be that relevant, but it shows how easy you can find things like
> this.
again thank you very much for this details, and before all, for this
useful plugin.
Note that if I interest on this plugin is because that the next step
of my work will be around clustering algorithms and software graphs.
So surely such plugin to Moose will be very useful for me.
bests,
Hani
>
> regards,
> Toon
Hi,
for one of the visualizations of clusters (DirectedHierarchical), it
would be nice to show classes in a class hierarchy, and annotate that
hierarchy with "accesses" between classes. However I don't seem to be
able to just show the hierarchy (treeLayout) and then add the
annotations (also edges), without the annotations being used to
calculate the tree... Is there a way to tell the ViewRenderer not to
use some "localEdges:" for the current layout?
thanks,
Toon
Hi,
I have published my work up until now on classcluster detection in the
scg store under the highly original bundle-name "Clustering". Like
this you can play around a bit to see what it can do etc.
Some notes:
- In order to find clusters, the clustering plugin needs to be
installed, by running
"SCG.Moose.AbstractEntity initializeAllMofDescriptions".
I am sure there is a clean way to do this automatically when loading
packages, but since I have no clue how to install that... this is
how it works for now :)
Next if you open the moose browser, rightclick on a model, "Find
Clusters in Model".
- To get better results, make sure you merge instance and class part
of classes when generating models
- I have included Adrians dendrogram clustering, but have written a
wrapper around it to be able to browse the hierarchy as such that
"double information" is avoided. For example if a b and c are all at
distance 1 of each other, their combination a,b,c has threshold 1
too, so it is not useful to see a,b,c with left a,b with distance 1
and right c, so this info just gets pruned away.
Currently the Interface and DeltaIC plugins are hierarchical and the
reasons prefixed by * are hierarchical reasons (are possibly
browsable).
Adrian: currently I have to do a bit of hacking to get similar
results, but it might be cleaner if the algorithms would support
items laying at Infinity+ from each other (if clustered their
threshold is also Infinity+), so those results can be filtered
easier...
- I know parts of the code (especially the UI and the sorting of
reasons) should be cleaned up a bit... but I don't have time for
that for the moment, and like this at least you can already see some
of the possibilities that it creates
Best regards,
Toon
hi!
I have a question concerning the Moose browser.
As you know, I managed to import models conformed to a personnal metamodel
in Moose ( my personnal metamodel is not FAMIX). In myn metamodel, I
decided to define my own entities Class, Attibute and Method.
When I imported my models (that contains my personnal classes, my
personnal attibutes and personnal methods) in Moose, something wrong
annoyed me:
when I selected "all classes" in the browser, Moose gave me the
possibility to generate Mondrian views as BluePrintComplexity,
MethodInvocations and so on... Of course, since my entities are not
conformed to FAMIX, the visualisation cannot be applied to my model (they
lead into a bug except from UML diagrams that work very well :) )
So, here is my question:
I don't understand why Moose proposes Mondrian visualisations (such as
BlueprintComplexity, Duplcation complexity, Method invocation ans so on)
to my personnal classes knowing that they are not conformed to the Famix
meta model. Where can I fix it? I had a look on MooseUI package but I
didn't find what I wanted :(
I hope that my question was clear:)
thank you for your help
sara
Gentlefolks,
I'm looking for applications which would benefit from maze resolution. The
picture attached can give you an impression. It has 4 collections of
objects and 4 collections of relations (all can be manipulated).
The large center area holds multiple documents (code, text, multi media).
Documents, objects and relations can be moved out of the way if desired.
Resting the mouse over something tells you what it is.
When selecting a relation it highlights the objects which it can see.
Maze resolution can guarantee that no other non-empty relation exists from
top left to bottom right and also not from top right to bottom left. So
there can be nothing in the picture (nothing in the image!) which is
hidden from you. Objects can be relocated and can actively help you to
find another suitable location.
I post in the hope to receiving application ideas from other people, i.e.
what to put into the 8 selection areas. If the picture can be populated
with things that make sense then a prototype would make sense (the example
below could be a start).
Thank you in advance for your feedback, all appreciated.
Cheers
Klaus
P.S. an example: top-left classes, top-left-right #category, top-right
system categories, bottom-left methods, top-left-bottom {#allSelectors.
#class->#allSelectors}, bottom-right methods, bottom-left-right #messages.
Then bottom-right shows the methods which are sent but not implemented
(foreigners) by the top-left classes themselves. And top-right can
optionally show the to bottom-right corresponding foreign system
categories with bottom-right-top #methodClass->#category.
P.P.S. thank you Doru for introducing me to your thesis and to
Moose+Mondrian.
Hi,
I am the student working with Gabriela on adopting the Koschke
clustering algorithms to OO. Of the basic algorithms I have almost
implemented all of them (type-based is not really possible, and
have no idea yet if same-expression is adoptable using famix).
Most of the algorithms can be adopted quite straightforward, but
sometimes using them automatically as was the case, is out of the
question. For that reason I have build a browsable interface around
the basic steps of the clustering so the user has a bigger input.
It's main task is helping the user to wade through the mass of
information all the clustering techniques provide, as fast as possible.
Most of the algorithms aren't based on distance metrics, but
there are a few which are. I hadn't come around to implementing the
hierarchical clustering on top of the basic steps yet, so those
algorithms come quite in handy. I will check how I can incorporate
them, thanks for that Andrian.
Further questions are welcome...
Best regards,
Toon
------------------------------
Message: 9
Date: Sun, 1 Jul 2007 21:34:15 +0200
From: St?phane Ducasse <stephane.ducasse(a)univ-savoie.fr>
Subject: [Moose-dev] distribution map in 6 lines?
To: Moose-dev Bugs <moose-dev(a)iam.unibe.ch>
Message-ID: <EA9BA9AA-35AF-4157-B64F-77C5E5E20381(a)univ-savoie.fr>
Content-Type: text/plain; charset=US-ASCII; delsp=yes; format=flowed
Andrea where can I find the distribution map code you show me (how
to write a distribution map in 6 lines of code
------------------------------
Hi,
Presuming you were referring to me, it's Adrian and not Andrea :)
The code is in my bundle MoosLiDevelopment, but because it is only 3 lines
i've extracted it below:
| engine serializedCollection |
engine := ClusterEngine with: aCollection using: distanceBlock.
engine hierarchicalClusteringShowProgressUsing: #averageLinkage.
serializedCollection := engine dendrogram collect: #yourself.
where:
aCollection: is your collection of elements you want to cluster
distanceBlock: is a block that returns the distance between any 2 elements
of your collection (eg. Euclidean distance for numerical proprieties of
elements)
serializedCollection := the result of the clustering, the serialized version
of the collection
Sorry for the late answer, but i was away for a few days.
Ciao,
Adi
--
Adrian DOZSA
Politehnica University of Timisoara
Computer Science Department
mail: adi.dozsa(a)gmail.com
web: http://adi.dozsa.googlepages.com/
