Hi Maheshakya, Thanks for very detailed response. We'll be reusing the cluster diagram we use in data exploration view to visualize clusters. What we're mostly missing is some measures about training and resulting model. I will check the measures you have mentioned. :)
Regards, CD On Thu, May 28, 2015 at 2:17 PM, Maheshakya Wijewardena <mahesha...@wso2.com > wrote: > Hi CD, > > Two of the widely used evaluation metrics are Rand index[1] and mutual > information[2]. In addition, there is Homogeneity, Completeness and > V-measure [3]. One issue with these external indices is that they require > ground truth of cluster assignments. Therefore without the true class > labels, these metrics are not usable. There are several internal indices as > well such as Silhouette Coefficient[4] which do not need ground truth. Some > of those methods are discussed here[5][6][7]. I think the more useful > scenario will be to use internal indices since having ground truth cluster > labels is not always the case. > > For visualization, only 2D (or maybe 3D) plots can be used despite there > are large number of features. So available options can be: > > 1. Allowing user to choose 2 or 3 features. > 2. Use PCA based dimensionality reduced (to 2 or 3 components) data - > Here, PCA may need to implemented separately so this option can be quite > tedious. > > It would be nice if the voronoi diagram for the data spread also can be > shown in the same diagram. See [8]. > [1] http://en.wikipedia.org/wiki/Rand_index#Adjusted_Rand_index > [2] http://en.wikipedia.org/wiki/Adjusted_mutual_information > [3] http://aclweb.org/anthology/D/D07/D07-1043.pdf > [4] http://en.wikipedia.org/wiki/Silhouette_%28clustering%29 > [5] > http://stats.stackexchange.com/questions/21807/evaluation-measure-of-clustering-without-having-truth-labels > [6] https://web.njit.edu/~yl473/papers/ICDM10CLU.pdf > [7] http://shonen.naun.org/multimedia/UPress/cc/20-463.pdf > [8] http://www.naftaliharris.com/blog/visualizing-k-means-clustering/ > > Best regards. > > On Thu, May 28, 2015 at 12:24 PM, CD Athuraliya <chathur...@wso2.com> > wrote: > >> Hi Maheshakya, >> >> We'll be adding cluster diagram in model summary for clustering >> algorithms. Please suggest if there exist any other useful evaluation >> metrics. >> >> Thanks >> >> On Thu, May 28, 2015 at 11:58 AM, Maheshakya Wijewardena < >> mahesha...@wso2.com> wrote: >> >>> Nice. >>> >>> Adding up to charts for classification, I think we need some >>> visualization method for clustering as well since there's nothing to show >>> after clustering models are trained. Maybe chart with respect to two >>> selected attributes. >>> >>> On Thu, May 28, 2015 at 11:46 AM, CD Athuraliya <chathur...@wso2.com> >>> wrote: >>> >>>> Hi all, >>>> >>>> Residual plot has been added for numerical prediction algorithms. Using >>>> standard chart types as much as possible is better IMO. It will reduce user >>>> confusion in understanding visualizations. I think we need to look for some >>>> standard chart types for classification algorithms (both binary and >>>> multiclass) as well [1]. >>>> >>>> [1] http://oobaloo.co.uk/visualising-classifier-results-with-ggplot2 >>>> >>>> Thanks >>>> >>>> On Wed, May 27, 2015 at 5:38 AM, Srinath Perera <srin...@wso2.com> >>>> wrote: >>>> >>>>> +1 shall we try those? >>>>> On 26 May 2015 22:52, "Upul Bandara" <u...@wso2.com> wrote: >>>>> >>>>>> +1 for residual plots. >>>>>> >>>>>> Though I haven't used it myself Residual Plot is a useful diagnostic >>>>>> tool for regression models. >>>>>> Especially, non-linearity in regression models can be easily >>>>>> identified using it. >>>>>> >>>>>> "An Introduction to Statistical Learning" book [1] ( page 92-96) >>>>>> contains some useful information about residual plots. >>>>>> >>>>>> [1]. http://www-bcf.usc.edu/~gareth/ISL/ISLR%20Fourth%20Printing.pdf >>>>>> >>>>>> On Tue, May 26, 2015 at 8:47 PM, Supun Sethunga <sup...@wso2.com> >>>>>> wrote: >>>>>> >>>>>>> Hi CD, >>>>>>> >>>>>>> As it pops up in the offline discussion as well, IMHO, for >>>>>>> classifications, this plot may not be the best option. But for >>>>>>> regression, >>>>>>> we can actually use this plot but with a slight modification, that is >>>>>>> taking the difference of the predicted and actual (rather than the >>>>>>> values >>>>>>> it self), and plot that, against a predictor variable (just like its >>>>>>> been >>>>>>> done atm). We can also add a third variable (categorical feature) to >>>>>>> color >>>>>>> the points. This is a standard plot (AKA Residual plot) which is usually >>>>>>> use to evaluate regression models. >>>>>>> >>>>>>> One other thing we can try out is, doing the same for classification >>>>>>> as well. i.e: Taking the difference between the actual probability (o >>>>>>> or 1) >>>>>>> and the predicted probability, and plot that, and see whether it gives a >>>>>>> better overall picture. Not sure how will it come out though :) If it >>>>>>> comes >>>>>>> right, then any point lies above 0.5 (or the threshold we used) is >>>>>>> wrongly >>>>>>> classified, and hence we can get a rough idea, on for which values of >>>>>>> x-axis feature, does the points get wrongly classified. I mean, we >>>>>>> should >>>>>>> be able to see any pattern, if there exists. >>>>>>> >>>>>>> Thanks, >>>>>>> Supun >>>>>>> >>>>>>> On Tue, May 26, 2015 at 6:08 PM, CD Athuraliya <chathur...@wso2.com> >>>>>>> wrote: >>>>>>> >>>>>>>> Hi, >>>>>>>> >>>>>>>> Plotting predicted and actual values against a feature doesn't look >>>>>>>> very intuitive, specially for non-probabilistic models. Please check >>>>>>>> the >>>>>>>> attachments. Any thoughts on making this visualization better? >>>>>>>> >>>>>>>> Thanks >>>>>>>> >>>>>>>> On Fri, May 22, 2015 at 3:27 PM, Srinath Perera <srin...@wso2.com> >>>>>>>> wrote: >>>>>>>> >>>>>>>>> yes, rerun using a random sample from test data is OK. >>>>>>>>> >>>>>>>>> --Srinath >>>>>>>>> >>>>>>>>> On Fri, May 22, 2015 at 2:28 PM, CD Athuraliya < >>>>>>>>> chathur...@wso2.com> wrote: >>>>>>>>> >>>>>>>>>> Hi Srinath, >>>>>>>>>> >>>>>>>>>> Still that random sample will not correspond to predicted vs. >>>>>>>>>> actual values in test results. Given that there is no mapping between >>>>>>>>>> random sample data points and test result points. One thing we can >>>>>>>>>> do is >>>>>>>>>> running test separately (using the same model) for sampled data for >>>>>>>>>> the >>>>>>>>>> sole purpose of visualization. Any other options? >>>>>>>>>> >>>>>>>>>> On Fri, May 22, 2015 at 2:06 PM, Srinath Perera <srin...@wso2.com >>>>>>>>>> > wrote: >>>>>>>>>> >>>>>>>>>>> Hi CD, >>>>>>>>>>> >>>>>>>>>>> Can we take a random sample from the test data and use that for >>>>>>>>>>> this process? >>>>>>>>>>> >>>>>>>>>>> --Srianth >>>>>>>>>>> >>>>>>>>>>> On Fri, May 22, 2015 at 12:00 PM, CD Athuraliya < >>>>>>>>>>> chathur...@wso2.com> wrote: >>>>>>>>>>> >>>>>>>>>>>> Hi all, >>>>>>>>>>>> >>>>>>>>>>>> To implement $subject in ML we need all feature values of the >>>>>>>>>>>> dataset against predicted and actual values for test data. But >>>>>>>>>>>> Spark only >>>>>>>>>>>> returns predicted and actual values as test results. Right now we >>>>>>>>>>>> use >>>>>>>>>>>> random 10,000 data rows for other visualizations and we cannot use >>>>>>>>>>>> same >>>>>>>>>>>> data for this visualization since that random 10,000 data does not >>>>>>>>>>>> correspond to test data (test data is a subtracted from dataset >>>>>>>>>>>> according >>>>>>>>>>>> to the train data fraction at model building stage). >>>>>>>>>>>> >>>>>>>>>>>> One option is to persist test data at testing stage, but it can >>>>>>>>>>>> be too large for some datasets according to train data fraction. >>>>>>>>>>>> Appreciate >>>>>>>>>>>> if you can give your comments on this. >>>>>>>>>>>> >>>>>>>>>>>> Thanks, >>>>>>>>>>>> CD >>>>>>>>>>>> >>>>>>>>>>>> -- >>>>>>>>>>>> *CD Athuraliya* >>>>>>>>>>>> Software Engineer >>>>>>>>>>>> WSO2, Inc. >>>>>>>>>>>> lean . enterprise . middleware >>>>>>>>>>>> Mobile: +94 716288847 <94716288847> >>>>>>>>>>>> LinkedIn <http://lk.linkedin.com/in/cdathuraliya> | Twitter >>>>>>>>>>>> <https://twitter.com/cdathuraliya> | Blog >>>>>>>>>>>> <http://cdathuraliya.tumblr.com/> >>>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> -- >>>>>>>>>>> ============================ >>>>>>>>>>> Blog: http://srinathsview.blogspot.com twitter:@srinath_perera >>>>>>>>>>> Site: http://people.apache.org/~hemapani/ >>>>>>>>>>> Photos: http://www.flickr.com/photos/hemapani/ >>>>>>>>>>> Phone: 0772360902 >>>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> -- >>>>>>>>>> *CD Athuraliya* >>>>>>>>>> Software Engineer >>>>>>>>>> WSO2, Inc. >>>>>>>>>> lean . enterprise . middleware >>>>>>>>>> Mobile: +94 716288847 <94716288847> >>>>>>>>>> LinkedIn <http://lk.linkedin.com/in/cdathuraliya> | Twitter >>>>>>>>>> <https://twitter.com/cdathuraliya> | Blog >>>>>>>>>> <http://cdathuraliya.tumblr.com/> >>>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> >>>>>>>>> -- >>>>>>>>> ============================ >>>>>>>>> Blog: http://srinathsview.blogspot.com twitter:@srinath_perera >>>>>>>>> Site: http://people.apache.org/~hemapani/ >>>>>>>>> Photos: http://www.flickr.com/photos/hemapani/ >>>>>>>>> Phone: 0772360902 >>>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> -- >>>>>>>> *CD Athuraliya* >>>>>>>> Software Engineer >>>>>>>> WSO2, Inc. >>>>>>>> lean . enterprise . middleware >>>>>>>> Mobile: +94 716288847 <94716288847> >>>>>>>> LinkedIn <http://lk.linkedin.com/in/cdathuraliya> | Twitter >>>>>>>> <https://twitter.com/cdathuraliya> | Blog >>>>>>>> <http://cdathuraliya.tumblr.com/> >>>>>>>> >>>>>>> >>>>>>> >>>>>>> >>>>>>> -- >>>>>>> *Supun Sethunga* >>>>>>> Software Engineer >>>>>>> WSO2, Inc. >>>>>>> http://wso2.com/ >>>>>>> lean | enterprise | middleware >>>>>>> Mobile : +94 716546324 >>>>>>> >>>>>> >>>>>> >>>>>> >>>>>> -- >>>>>> Upul Bandara, >>>>>> Associate Technical Lead, WSO2, Inc., >>>>>> Mob: +94 715 468 345. >>>>>> >>>>> >>>> >>>> >>>> -- >>>> *CD Athuraliya* >>>> Software Engineer >>>> WSO2, Inc. >>>> lean . enterprise . middleware >>>> Mobile: +94 716288847 <94716288847> >>>> LinkedIn <http://lk.linkedin.com/in/cdathuraliya> | Twitter >>>> <https://twitter.com/cdathuraliya> | Blog >>>> <http://cdathuraliya.tumblr.com/> >>>> >>> >>> >>> >>> -- >>> Pruthuvi Maheshakya Wijewardena >>> Software Engineer >>> WSO2 Lanka (Pvt) Ltd >>> Email: mahesha...@wso2.com >>> Mobile: +94711228855 >>> >>> >>> >> >> >> -- >> *CD Athuraliya* >> Software Engineer >> WSO2, Inc. >> lean . enterprise . middleware >> Mobile: +94 716288847 <94716288847> >> LinkedIn <http://lk.linkedin.com/in/cdathuraliya> | Twitter >> <https://twitter.com/cdathuraliya> | Blog >> <http://cdathuraliya.tumblr.com/> >> > > > > -- > Pruthuvi Maheshakya Wijewardena > Software Engineer > WSO2 Lanka (Pvt) Ltd > Email: mahesha...@wso2.com > Mobile: +94711228855 > > > -- *CD Athuraliya* Software Engineer WSO2, Inc. lean . enterprise . middleware Mobile: +94 716288847 <94716288847> LinkedIn <http://lk.linkedin.com/in/cdathuraliya> | Twitter <https://twitter.com/cdathuraliya> | Blog <http://cdathuraliya.tumblr.com/>
_______________________________________________ Dev mailing list Dev@wso2.org http://wso2.org/cgi-bin/mailman/listinfo/dev