samoa-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gianmarco De Francisci Morales <g...@apache.org>
Subject Re: HoeffdingTree and VerticalHoeffdingTree Classifiers run on KDD Cup 99 Data Set
Date Wed, 16 Sep 2015 12:45:38 GMT
Hi,

Thanks for sharing the dataset.

Yes, the tree consists only of internal nodes (split) and leaves (learning).
Where to call the function is a good question.
Given that we are dealing with unbounded streams, in theory there is no
"end of the stream", so there is no final form of the tree.
Of course this is not true when you are running offline on a fixed dataset.
The class ContentEvent has a method public boolean isLastEvent() which
should allow to catch the last event in a data stream from a file.
This is implemented in InstanceContentEvent, and used
in PrequentialSourceProcessor (though its use is not very consistent
throughout the codebase).
So one way to do it could be to catch this event and call the dumping
function if isLastEvent is true.

Thanks for debugging the NPE.
If I am not mistaken, there should not be a case where the leaf is null,
i.e., one should always be able to sort an instance to a leaf.
Even if the instance is sparse, the missing values of the sparse instance
are defined.
Do you know why that instance fails to be sorted to its correct leaf?

Cheers,


--
Gianmarco

On 15 September 2015 at 15:50, Davut Deniz Yavuz <
davut.deniz.yavuz@gmail.com> wrote:

> Hi,
>
> We have implemented a recursive method to dump the tree by implementing
> the describeSubTree methods in  ActiveLearningNode and SplitNode. (We
> checked the tree and it only consists of these types of nodes.) We print
> the tree by calling the describeSubTree function of the root of the tree.
>
> Unfortunately, we couldn’t decide where to call the
> treeRoot.describeSubTree function. Currently, we are calling it at the end
> of the attemptToSplit
> <https://github.com/apache/incubator-samoa/blob/9b178f63152e5b4c262e0f3ed28e77667832fc98/samoa-api/src/main/java/org/apache/samoa/learners/classifiers/trees/ModelAggregatorProcessor.java#L463>
function.
> Therefore, the code prints a couple of trees and we are not sure if it
> prints the final form of the tree. Which place would be more appropriate to
> call this function to print just the final form of the tree?
>
> --
>
> We debugged the NullPointerException and we’ve found out that the source
> of the problem is in the filterInstanceToLeaf method in SplitNode class.
> The method returns new FoundNode(null, this, childIndex)
> <https://github.com/apache/incubator-samoa/blob/master/samoa-api/src/main/java/org/apache/samoa/learners/classifiers/trees/SplitNode.java#L56>
> when the child is null. This method is called from ModelAggregatorProcessor
> by trainOnInstanceImpl(Instance inst)
> <https://github.com/apache/incubator-samoa/blob/9b178f63152e5b4c262e0f3ed28e77667832fc98/samoa-api/src/main/java/org/apache/samoa/learners/classifiers/trees/ModelAggregatorProcessor.java#L425>.
> Thus, a node returning null when the getNode function of it called is
> produced in the 394. line:
>
> FoundNode foundNode = this.treeRoot.filterInstanceToLeaf(inst, null, -1)
> <https://github.com/apache/incubator-samoa/blob/master/samoa-api/src/main/java/org/apache/samoa/learners/classifiers/trees/ModelAggregatorProcessor.java#L394>.
>
>
> This node, then, is added to foundNodeSet by the line
> this.foundNodeSet.add(foundNode)
> <https://github.com/apache/incubator-samoa/blob/master/samoa-api/src/main/java/org/apache/samoa/learners/classifiers/trees/ModelAggregatorProcessor.java#L452>.
> Because of this, while looping
> <https://github.com/apache/incubator-samoa/blob/master/samoa-api/src/main/java/org/apache/samoa/learners/classifiers/trees/ModelAggregatorProcessor.java#L140>over
> the elements in the foundNodeSet, the getNode() function of the leafNode
> node returns a null node and it causes a NullPointerException.
>
> To solve the problem we checked how moa handles the null node, however, to
> our knowledge, there is not an alike structure like foundNodeSet in moa.
>
> --
>
> The KDD Cup 99 Dataset can be downloaded from here
> <https://drive.google.com/file/d/0B_XfRoyTW4lJWFF5TWttR0VLTnhCbERrcEtjY2thWGtLUlZv/view?usp=sharing>
> .
>
> Respectfully,
> Davut Deniz Yavuz, Ercan Ozturk
>
> 2015-09-11 9:57 GMT+03:00 Gianmarco De Francisci Morales <gdfm@apache.org>
> :
>
>> Sure, the ticket is SAMOA-44
>> <https://issues.apache.org/jira/browse/SAMOA-44>.
>>
>> Arinto had started the work on model dumping, I don't know what's the
>> status there.
>> But it should be straightforward to implement a recursive method.
>>
>> If you could post the dataset somewhere where it is possible to download
>> it, it would be great.
>> If you want to take a stab at debugging what's going on and provide a
>> patch, it would be even better.
>>
>> Cheers,
>>
>> --
>> Gianmarco
>>
>> On 10 September 2015 at 08:49, Ercan Öztürk <e.ozturk111@gmail.com>
>> wrote:
>>
>>> Hi,
>>>
>>> Thank you very much for your quick response.
>>>
>>> We were using an older version of SAMOA. I've updated the code now (The
>>> last commit is currently "SAMOA-29: Excluding the samoa-storm.properties at
>>> compile time and including at test") and after building the code with "mvn
>>> package" the new command we use to run SAMOA is
>>>
>>> local target/SAMOA-Local-0.4.0-incubating-SNAPSHOT.jar
>>> "PrequentialEvaluation -i -1 -f 41920 -l
>>> (org.apache.samoa.learners.classifiers.trees.VerticalHoeffdingTree -p 4) -s
>>> (org.apache.samoa.moa.streams.ArffFileStream -f kddcup_full.arff)"
>>>
>>> The console output when the command is run:
>>>
>>> bin/samoa
>>> Deploying to LOCAL
>>> Command line string =  PrequentialEvaluation -i -1 -f 41920 -l
>>> (org.apache.samoa.learners.classifiers.trees.VerticalHoeffdingTree -p 4) -s
>>> (org.apache.samoa.moa.streams.ArffFileStream -f kddcup_full.arff)
>>> 2015-09-09 15:56:30,036 [main] INFO  org.apache.samoa.LocalDoTask
>>> (LocalDoTask.java:80) - Successfully instantiating
>>> org.apache.samoa.tasks.PrequentialEvaluation
>>> 2015-09-09 15:56:31,221 [main] INFO
>>>  org.apache.samoa.evaluation.EvaluatorProcessor
>>> (EvaluatorProcessor.java:83) - 1 seconds for 41920 instances
>>> 2015-09-09 15:56:31,227 [main] INFO
>>>  org.apache.samoa.evaluation.EvaluatorProcessor
>>> (EvaluatorProcessor.java:169) - evaluation instances = 41,920
>>> classified instances = 41,920
>>> classifications correct (percent) = 99.988
>>> Kappa Statistic (percent) = -0.002
>>> Kappa Temporal Statistic (percent) = 28.571
>>> Exception in thread "main" java.lang.NullPointerException
>>> at
>>> org.apache.samoa.learners.classifiers.trees.ModelAggregatorProcessor.process(ModelAggregatorProcessor.java:142)
>>> at
>>> org.apache.samoa.topology.impl.SimpleProcessingItem.processEvent(SimpleProcessingItem.java:84)
>>> at org.apache.samoa.topology.impl.SimpleStream.put(SimpleStream.java:72)
>>> at org.apache.samoa.topology.impl.SimpleStream.put(SimpleStream.java:61)
>>> at
>>> org.apache.samoa.learners.classifiers.trees.FilterProcessor.process(FilterProcessor.java:93)
>>> at
>>> org.apache.samoa.topology.impl.SimpleProcessingItem.processEvent(SimpleProcessingItem.java:84)
>>> at org.apache.samoa.topology.impl.SimpleStream.put(SimpleStream.java:72)
>>> at org.apache.samoa.topology.impl.SimpleStream.put(SimpleStream.java:61)
>>> at
>>> org.apache.samoa.topology.LocalEntranceProcessingItem.injectNextEvent(LocalEntranceProcessingItem.java:45)
>>> at
>>> org.apache.samoa.topology.LocalEntranceProcessingItem.startSendingEvents(LocalEntranceProcessingItem.java:63)
>>> at
>>> org.apache.samoa.topology.impl.SimpleTopology.run(SimpleTopology.java:44)
>>> at
>>> org.apache.samoa.topology.impl.SimpleEngine.submitTopology(SimpleEngine.java:33)
>>> at org.apache.samoa.LocalDoTask.main(LocalDoTask.java:88)
>>>
>>>
>>> We would be very appreciated if you could send us the link for the
>>> ticket so we can follow the updates on the issue.
>>>
>>> Yes, we would like to dump the model so that we can see the rules of the
>>> model and have a better understanding of it.
>>>
>>> The method body of describeSubtree() in Node.java is currently empty. Is
>>> there any work done on it that we can use as a starting point?
>>>
>>> If you need the data set to investigate the issue, I can send it via any
>>> suitable channel, please let me know.
>>>
>>> Respectfully,
>>> Ercan Ozturk
>>>
>>> 2015-09-09 15:11 GMT+03:00 Gianmarco De Francisci Morales <
>>> gdfm@apache.org>:
>>>
>>>> Hi,
>>>>
>>>> Thanks for reporting the bug.
>>>> I'm not sure what is causing the issue.
>>>> Are you using the master version of SAMOA?
>>>> My line 145 of ModelAggregator is:
>>>>               this.sendToAttributeStream(abce[i]);
>>>>
>>>> From what you say it seems that the problem is a bit above, and
>>>> leafNode is null.
>>>> However, by construction there should always be a leaf node.
>>>>
>>>> As a workaround your solution is fine, but I guess there is some other
>>>> underlying problem with the code, which might cause some loss in accuracy.
>>>> We should investigate this issue further, I'll open a ticket.
>>>>
>>>> Regarding fetching the content of the model, we had some prototype
>>>> model dumper code (Arinto had started it), but I guess it's not working
>>>> anymore. See the describeSubtree() method in Node.java.
>>>> So unfortunately you need to do it yourself. However, the good thing is
>>>> that the tree model is in a single place in ModelAggregator, so it should
>>>> be relatively easy to walk the tree, starting from the root node.
>>>> Do you want to dump the model to a text representation for human
>>>> inspection?
>>>>
>>>> Cheers,
>>>>
>>>>
>>>> --
>>>> Gianmarco
>>>>
>>>> On 7 September 2015 at 18:23, Gianmarco De Francisci Morales <
>>>> gdfm@apache.org> wrote:
>>>>
>>>>> Forwarding to the @dev list.
>>>>> --
>>>>> Gianmarco
>>>>>
>>>>> ---------- Forwarded message ----------
>>>>> From: Ercan Öztürk <e.ozturk111@gmail.com>
>>>>> Date: 7 September 2015 at 16:57
>>>>> Subject: HoeffdingTree and VerticalHoeffdingTree Classifiers run on
>>>>> KDD Cup 99 Data Set
>>>>> To: gdfm@apache.org
>>>>>
>>>>>
>>>>> Hi Mr. Morales and Mr. Bifet,
>>>>>
>>>>> We are a couple of undergrad students from TOBB University. As a data
>>>>> mining class project, we decided to run HoeffdingTree classifier-in moa
and
>>>>> VerticalHoeffdingTree classifier-in samoa on KDD Cup 99 data set (couldn't
>>>>> attach the data set to this mail due to the size limitations of the Apache
>>>>> mail server) and present the results in our project report.
>>>>>
>>>>> We were able to run HoeffdingTree Algorithm on the KDD Cup 99 (both on
>>>>> kddcup_full.arff, kddcup_10_percent.arff) data set.
>>>>> VerticalHoeffdingTree classifier also works fine on
>>>>> kddcup_10_percent.arff. However, when we try to run the
>>>>> VerticalHoeffdingTree classifier on kddcup_full.arff, we got the
>>>>> following error:
>>>>>
>>>>> The command we use to run SAMOA Local:
>>>>>
>>>>> bin/samoa local target/SAMOA-Local-0.3.0-SNAPSHOT.jar
>>>>> "PrequentialEvaluation -i -1 -f 41920 -l
>>>>> (com.yahoo.labs.samoa.learners.classifiers.trees.VerticalHoeffdingTree
-p
>>>>> 4) -s (com.yahoo.labs.samoa.moa.streams.ArffFileStream -f kddcup_full.arff)"
>>>>>
>>>>> The console output of samoa:
>>>>>
>>>>> bin/samoa
>>>>>
>>>>> Deploying to LOCAL
>>>>>
>>>>> Command line string =  PrequentialEvaluation -i -1 -f 41920 -l
>>>>> (com.yahoo.labs.samoa.learners.classifiers.trees.VerticalHoeffdingTree
-p
>>>>> 4) -s (com.yahoo.labs.samoa.moa.streams.ArffFileStream -f kddcup_full.arff)
>>>>>
>>>>> 2015-09-01 22:22:16,160 [main] INFO  com.yahoo.labs.samoa.LocalDoTask
>>>>> (LocalDoTask.java:79) - Successfully instantiating
>>>>> com.yahoo.labs.samoa.tasks.PrequentialEvaluation
>>>>>
>>>>> 2015-09-01 22:22:17,741 [main] INFO
>>>>>  com.yahoo.labs.samoa.evaluation.EvaluatorProcessor
>>>>> (EvaluatorProcessor.java:86) - 1 seconds for 41920 instances
>>>>>
>>>>> 2015-09-01 22:22:17,760 [main] INFO
>>>>>  com.yahoo.labs.samoa.evaluation.EvaluatorProcessor
>>>>> (EvaluatorProcessor.java:172) - evaluation instances = 41,920
>>>>>
>>>>> classified instances = 41,920
>>>>>
>>>>> classifications correct (percent) = 99.988
>>>>>
>>>>> Kappa Statistic (percent) = -0.002
>>>>>
>>>>> Kappa Temporal Statistic (percent) = 28.571
>>>>>
>>>>> Exception in thread "main" java.lang.NullPointerException
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.learners.classifiers.trees.ModelAggregatorProcessor.process(ModelAggregatorProcessor.java:145)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.topology.impl.SimpleProcessingItem.processEvent(SimpleProcessingItem.java:84)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.topology.impl.SimpleStream.put(SimpleStream.java:71)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.topology.impl.SimpleStream.put(SimpleStream.java:60)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.learners.classifiers.trees.FilterProcessor.process(FilterProcessor.java:95)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.topology.impl.SimpleProcessingItem.processEvent(SimpleProcessingItem.java:84)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.topology.impl.SimpleStream.put(SimpleStream.java:71)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.topology.impl.SimpleStream.put(SimpleStream.java:60)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.topology.LocalEntranceProcessingItem.injectNextEvent(LocalEntranceProcessingItem.java:46)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.topology.LocalEntranceProcessingItem.startSendingEvents(LocalEntranceProcessingItem.java:66)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.topology.impl.SimpleTopology.run(SimpleTopology.java:42)
>>>>>
>>>>> at
>>>>> com.yahoo.labs.samoa.topology.impl.SimpleEngine.submitTopology(SimpleEngine.java:33)
>>>>>
>>>>> at com.yahoo.labs.samoa.LocalDoTask.main(LocalDoTask.java:87)
>>>>>
>>>>>
>>>>> We were able to track down the problem to the first instance that
>>>>> causes it; the instance is on the 76426th line in kddcup_full.arff.
>>>>> The instance is as follows:
>>>>>
>>>>>
>>>>> 1,tcp,smtp,SF,2252,331,0,0,0,0,0,1,0,0,0,0,0,0,0,0,0,0,1,7,0,0,0,0,1,0,1,5,216,1,0,0.2,0.01,0,0,0,0,normal
>>>>>
>>>>> We haven’t noticed any differences between the problematic instance
>>>>> and the other instances. Could you lead us to the root of the problem
and
>>>>> could you help us on how to overcome this problem?
>>>>>
>>>>> As a workaround we’ve made the following addition to
>>>>> ModelAggregatorProcessor.java
>>>>>
>>>>> if (leafNode == null)
>>>>>
>>>>>         return false;
>>>>>
>>>>> after the line
>>>>>
>>>>> ActiveLearningNode leafNode = (ActiveLearningNode) foundNode.getNode();
>>>>>
>>>>> Now, also VeriticalHoeffdingTree Classifier works fine on kddcup_full.arff.
>>>>> Is this solution acceptable for the problem, what do you think?
>>>>>
>>>>>
>>>>> Besides, we were wondering how we could fetch model contents such as
>>>>> visiting nodes and node content etc.
>>>>>
>>>>> Thanks for your help,
>>>>>
>>>>>
>>>>> Respectfully,
>>>>>
>>>>> Ercan Ozturk, Davut Deniz Yavuz, Gozde Boztepe, Sezin Gurkan
>>>>>
>>>>>
>>>>
>>>
>>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message