image_doctor

417 reputation

bio	website
	location
	age
visits	member for	1 year, 2 months
	seen	May 7 at 21:26
stats	profile views	51

	417 reputation	bio	website		visits	member for	1 year, 2 months
17 badges		location			seen	May 7 at 21:26

summary answers questions tags badges favorites bounties reputation activity

110 Actions

suggestions reviews revisions comments badges posts accepts all

Apr 27	comment	Is it possible to convert 'ow'l to 'arff' file format Some type of mapping is possible, but perhaps it will help if you outline what your goal in doing this might be ?
Apr 26	comment	Analysis of computer software performance It may be helpful to define what the goal of your analysis is.
Apr 18	comment	Positive, Likely Positives and Unknowns in Random Forests Is there a reason to choose 'Likely Negative' as the correct imputation from 'Unlabeled'? Have you thought about trying some form of cluster analysis on the large unlabeled set?
Apr 18	comment	Is the number 20 magic? In general no, in specific situations maybe. Do you have the references and what are your goals ?
Apr 16	comment	Which performance measure for unbalanced binary classification without an 'active' class? When you say a method invariant to the class distribution ... learning algorithms are explicitly dependent on the class distribution. If your samples have large variations in distribution, perhaps you could consider aggregating samples or using a sample technique that balanced the classes. But classifiers trained this way will show poorer performance on highly unbalanced samples. Otherwise you might want to consider algorithms which are more tolerant to sample variation, or if your samples are bimodal, train a separate classifier for each mode.
Apr 16	comment	How many attributes to select for classification Unfortunately there is no general answer to your problem, as it explicitly depends on the specific data you have which make up your M observations of your k features. There are two basic approaches, filter methods and wrapper methods. Filter methods select features based on searches in combinatorial spaces based on metrics. Wrapper methods incorporate a classifier/s of choice into the feature selection process. Wrapper methods are more specific to the problem at hand but may be computationally expensive. A variety of metrics and classifiers are available to plug into the two approaches.
Apr 15	comment	How to reduce dimensionality of vector? Maybe you can shed some light on what you want to achieve with this reduction in dimensionality and that might help to find some relevant solutions :)
Apr 15	comment	How to reduce dimensionality of vector? Sadly, PCA won't work for you either, not if you want a 1D result.
Apr 15	comment	How to reduce dimensionality of vector? Method 1, even if constrained to +ve only values, is still not unique: For instance, vectors {2,1} and {1,2} map to the same value under the Norm operation. PCA maps a vector in one N dimensional space to another vector in another N dimensional space. You can select the most important dimensions as a form of dimensionality reduction, but you have no assurance of a unique mapping. Once you start adding a set of values, e.g. Norm, weighted sum , together you can easily find you have outcomes with the same value.
Apr 15	comment	How to reduce dimensionality of vector? Method 1 does not give a unique mapping. Method 4 will not give a unique mapping if you reduce the dimensionality to 1.
Apr 15	comment	Is ARIMA better in comparision with Neural Networks? Thank you, I understand your comments a little better now. By "ignore" you mean that NN give outliers no special significance. In general I agree, though I imagine you could add a weighting term to the instances that you felt might be suspect. I think I have more of a philosophical challenge when it comes to identifying outliers, without extra information I'm not sure there is any reason to suspect 5 to be an outlier in the series, unless you make assumptions not supported by the data from an empirical standpoint.
Apr 15	comment	Which performance measure for unbalanced binary classification without an 'active' class? The cost function would normally be derived from the application domain dependent on how "expensive" making false positive and false negative errors are. In the example you give, you are declaring an error in mislabeling a B instance to be four times as damaging as mislabeling an instance of A. What goal are you trying to achieve ? Other techniques are over sampling the minority class, under sampling the majority class, adding noise to the to minority class instances to make new samples, estimating the distribution of the minority class and sampling from that.
Apr 15	comment	Which performance measure for unbalanced binary classification without an 'active' class? This is where the cost function for errors comes in. If your errors have EQUAL weights then the classifier that predicts the majority class in that case is the more accurate classifier. If the errors do not have equal weights, then you should weight the errors appropriately using a suitable cost function :)
Apr 15	comment	Which performance measure for unbalanced binary classification without an 'active' class? I'm not sure I understand the question, the accuracy on one dataset, say breast cancer prediction, and the accuracy on another dataset, say horse racing, are related in what way? Or are you saying you are sampling the same underlying distribution and the samples have varying ratios of class A and B?
Apr 15	comment	Which performance measure for unbalanced binary classification without an 'active' class? Your instances may have equal weights, do your errors have equal weights ? If they do, then accuracy is a valid measure of performance.
Apr 15	comment	Which performance measure for unbalanced binary classification without an 'active' class? Do you have a cost function ?
Apr 13	comment	Is ARIMA better in comparision with Neural Networks? can you say a little more about the way in which Neural Networks ignore outliers?
Apr 8	comment	Confusion matrixes with percentages rather than number of instances? A percentage without an absolute magnitude or a single magnitude without an overall measure of scale both lack part of the picture, so in that respect neither is an ideal solution. I don't see anything wrong with percentages if they convey better what it is you wish to say or interpret.
Mar 23	awarded	Yearling
Feb 19	comment	Effective way to speed-up my mass-univariate problem (i.e. a large set of per-point optimization problems) This may depend on the specific local behaviour of your observed function.