Re: [DL-Learner discussion] dl-learner and noise

SourceForge Headquarters 1320 Columbia Street Suite 310 San Diego, CA 92101 +1 (858) 422-6466

Hello Lorenz,

Thank you very much for all the time you spent on this! This is very 
helpful.

Best regards,
Céline

Le 29-12-2014 16:03, Lorenz Bühmann a écrit :
> Hello Céline,
> 
>  let me try to explain why things do not work as you would expect.
> 
>  First of all, please always keep in mind that DL-Learner is searching
> for OWL class expressions that
> 
>  1. cover as many positive examples as possible
>  2. cover as less negative examples as possible
> 
>  This search space is quite large and that's why, assuming that a
> perfect solution exists (this is not always the case despite the
> trivial solution), it might be the case that DL-Learner needs much
> more time than the default 10 seconds.
> 
>  According to your examples:
> 
>  1. has6culture.conf
>  There is a default limit of n=5 for cardinalities, that's why your
> solution is never tested. We are working on a more intelligent
> solution. For now, you can increase the limit by
> 
>  op.type = "rho"
>  op.cardinalityLimit = 10
> 
>  Based on the latest source code, I get
> 
>  solutions:
>  1: hasCulture min 6 Thing (pred. acc.: 100.00%, F-measure: 100.00%)
>  2: hasCulture min 6 (not WaterBody) (pred. acc.: 100.00%, F-measure:
> 100.00%)
>  3: hasCulture min 6 (not Volcano) (pred. acc.: 100.00%, F-measure:
> 100.00%)
>  4: hasCulture min 6 (not View) (pred. acc.: 100.00%, F-measure:
> 100.00%)
>  5: hasCulture min 6 (not Urban) (pred. acc.: 100.00%, F-measure:
> 100.00%)
>  6: hasCulture min 6 (not TriumphalArch) (pred. acc.: 100.00%,
> F-measure: 100.00%)
>  7: hasCulture min 6 (not Tower) (pred. acc.: 100.00%, F-measure:
> 100.00%)
>  8: hasCulture min 6 (not Stadium) (pred. acc.: 100.00%, F-measure:
> 100.00%)
>  9: hasCulture min 6 (not QualityEnvironment) (pred. acc.: 100.00%,
> F-measure: 100.00%)
>  10: hasCulture min 6 (not Amphitheatre) (pred. acc.: 100.00%,
> F-measure: 100.00%)
> 
>  2. hasOldTownOrShopping.conf
>  You have to increase the runtime.
> 
>  alg.maxExecutionTimeInSeconds = 600
> 
>  Moreover, it could help to disable some OWL constructs like negation
> and universal restrictions.
>  op.type = "rho"
>  op.useNegation = false //disable negation (not)
>  op.useAllConstructor = false //disable universal restrictions (only)
> 
>  The output is
> 
>  solutions:
>  1: (hasActivity some OldTown or hasEnvironment some Shopping) (pred.
> acc.: 100.00%, F-measure: 100.00%)
>  2: (hasActivity some OldTown or hasEnvironment some Urban) (pred.
> acc.: 97.50%, F-measure: 98.04%)
>  3: (hasActivity some Promenade or hasEnvironment some Urban) (pred.
> acc.: 92.50%, F-measure: 94.34%)
>  4: (hasActivity some Promenade or hasEnvironment some Shopping)
> (pred. acc.: 92.50%, F-measure: 94.34%)
>  5: hasActivity min 2 (Excursion or Lazing) (pred. acc.: 70.00%,
> F-measure: 80.65%)
>  6: hasActivity min 2 (Excursion or WaterActivity) (pred. acc.:
> 68.75%, F-measure: 80.00%)
>  7: hasActivity min 2 (Excursion or Relaxation) (pred. acc.: 68.75%,
> F-measure: 80.00%)
>  8: hasActivity min 2 (Excursion or Nightlife) (pred. acc.: 68.75%,
> F-measure: 80.00%)
>  9: hasActivity min 2 (Excursion or NaturalEnvironment) (pred. acc.:
> 68.75%, F-measure: 80.00%)
>  10: hasActivity min 2 (Environment or Excursion) (pred. acc.: 68.75%,
> F-measure: 80.00%)
>   
> 
>  3. hasBathingMidSummer.conf
>  The same holds for your 3rd example - this is really complex and far
> from easy to learn. Increase the runtime and disable some OWL
> constructs. Additionally, if you assume to get longer descriptions,
> you can set a parameter for the search heuristic like
> 
>  h.type ="celoe_heuristic"
>  h.expansionPenaltyFactor = 0.02
> 
>  The algorithm is still running and the output so far is
> 
>  (hasEnvironment some Bathing and hasWeather some (avgTemperatureC
> some double[>= 21.8563] and precipitationMm some double[<= 33.65835]
> and concernMonth some hasSeason some MidSummer))
> 
>  which was found after 1min 43s 32ms.
> 
>  I attached all config files.
> 
>  Kind regards,
>  Lorenz
> 
>> Hello Lorenz,
>> 
>> Thanks a lot, that helps.
>> 
>> I have another problem. I have an ontology describing holiday’s
>> destinations. If I want to learn some simple definitions like
>> "hasActivity some Animation" (or any "hasObjectProperty some
>> Concept") it works well. But if my definitions are more complicated,
>> it does not work. There is no noise in my examples.
>> 
>> I want to learn:
>> 1. "hasCulture min 6 Culture"
>> 2. "(hasActivity some OldTown) or (hasEnvironment some Shopping)"
>> 3. "(has Activity some Bathing) and (hasWeather min 2 (
>> (avgTemperature some double [>=23.0]) and (precipitationMm some
>> double[<=70.0]) and (concernMonth some hasSeason some MidSummer)))"
>> 
>> I get:
>> 
>> 1. hasCulture min 5 ((not Archaeology) and (not Architecture))
>> (pred. acc.: 97,50%, F-measure: 94,74%)
>> 2. hasActivity only (not TowedWatersport) (pred. acc.: 68,75%,
>> F-measure: 80,00%)
>> 3. (hasActivity some Excursion and hasEnvironment some
>> WaterActivity) (pred. acc.: 68,75%, F-measure: 73,12%)
>> 
>> Could you explain why I do not get the good definitions (or at
>> least some definitions with 100% of accuracy)? Are my definitions
>> too complicated? Did I miss something?
>> Thank you in advance! Sorry for disturbing you again.
>> 
>> Best regards,
>> Céline
>> 
>> Le 18-12-2014 18:40, Lorenz Bühmann a écrit :
>> Hello Céline,
>> 
>>  there is a noise parameter called "noisePercentage" for the
>> learning
>> algorithm CELOE, so you could simply define a noise value like 20%.
>> 
>> This will allow to return solutions in which 20% of the positive
>> examples are not covered by the solution.
>> 
>>  Your conf file should be extended by the line:
>> 
>>  alg.noisePercentage = "20"
>> 
>>  Hope this helps.
>> 
>>  Kind regards,
>>  Lorenz
>> 
>> Hello,
>> 
>> I have an ontology describing 10000 films. I have built a conf file
>> 
>> with positive and negative examples for “American films”. I
>> want
>> to learn that an American film “isFromCountry some _US”. When I
>> 
>> run DL-Learner, it works well. Now, if I add a little bit of noise
>> (changing one positive example into negative example and vice
>> versa), it does not work anymore. “isFromCountry some Country”
>> is learned (acc=42.86%). I have 10000 films, so this noise
>> represents only 0.02% of error.
>> Is it possible to get the good definition with some noise in my
>> examples by running DL-Learner with some parameters I don’t know
>> about or does DL-Learner just not handle noise?
>> 
>> Thanks in advance for your answer. I attached the files if you want
>> 
>> to make some tests. I use dllearner-1.0-beta-3 on Windows.
>> 
>> Best regards,
>> Céline
>> 
>> 
> ------------------------------------------------------------------------------
>> 
>> Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server
>> from Actuate! Instantly Supercharge Your Business Reports and
>> Dashboards
>> with Interactivity, Sharing, Native Excel Exports, App Integration
>> &
>> more
>> Get technology previously reserved for billion-dollar corporations,
>> 
>> FREE
>> 
>> 
> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk
>> [1]
>> [1]
>> 
>> _______________________________________________
>> dl-learner-discussion mailing list
>> dl-...@li...
>> https://lists.sourceforge.net/lists/listinfo/dl-learner-discussion
>> [2]
>> [2]
>> 
>> Links:
>> ------
>> [1]
>> 
> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&amp;iu=/4140/ostg.clktrk
>> [3]
>> [2]
>> https://lists.sourceforge.net/lists/listinfo/dl-learner-discussion
>> [2]
>> 
>> 
> ------------------------------------------------------------------------------
>> 
>> Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server
>> from Actuate! Instantly Supercharge Your Business Reports and
>> Dashboards
>> with Interactivity, Sharing, Native Excel Exports, App Integration
>> & more
>> Get technology previously reserved for billion-dollar corporations,
>> FREE
>> 
> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk
>> [1]
>> 
>> _______________________________________________
>> dl-learner-discussion mailing list
>> dl-...@li...
>> https://lists.sourceforge.net/lists/listinfo/dl-learner-discussion
>> [2]
> 
> ------------------------------------------------------------------------------
> Download BIRT iHub F-Type - The Free Enterprise-Grade BIRT Server
> from Actuate! Instantly Supercharge Your Business Reports and
> Dashboards
> with Interactivity, Sharing, Native Excel Exports, App Integration &
> more
> Get technology previously reserved for billion-dollar corporations,
> FREE
> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&iu=/4140/ostg.clktrk
> [1]
> 
> _______________________________________________
> dl-learner-discussion mailing list
> dl-...@li...
> https://lists.sourceforge.net/lists/listinfo/dl-learner-discussion [2]
> 
> 
> 
> Links:
> ------
> [1]
> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&amp;iu=/4140/ostg.clktrk
> [2] https://lists.sourceforge.net/lists/listinfo/dl-learner-discussion
> [3]
> http://pubads.g.doubleclick.net/gampad/clk?id=164703151&amp;amp;iu=/4140/ostg.clktrk
> 
> ------------------------------------------------------------------------------
> Dive into the World of Parallel Programming! The Go Parallel Website,
> sponsored by Intel and developed in partnership with Slashdot Media, is 
> your
> hub for all things parallel software development, from weekly thought
> leadership blogs to news, videos, case studies, tutorials and more. 
> Take a
> look and join the conversation now. http://goparallel.sourceforge.net
> 
> _______________________________________________
> dl-learner-discussion mailing list
> dl-...@li...
> https://lists.sourceforge.net/lists/listinfo/dl-learner-discussion