The dataset I want to cluster consists of ~1000 samples and 10 features, which have different scales and ranges (negative, positive, both). Using scipy.stats.normaltest() I found that none of the features are normally-distributed (all p-values < 1e-4, small enough to reject the null hypothesis that the data are taken from a normal distribution). But all of the distance measures that I'm aware of assume normally-distributed data (I was using Mahalanobis until I realized how non-uniform the data was). What distance measures would one use in this situation? Or is this where one simply has to normalize every feature and hope that that doesn't introduce bias?
Why do you think all distances would assume normal (which btw. is not the same as uniform) data?
Consider Euclidean distance. In many physical applications this distance makes perfect sense, because it is "as the crow flies". Manhattan distance makes a lot of sense when movement is constrained to two axes that cannot be used at the same time. These are completely appropriate for non-normal distributed data.
Related
This question already has answers here:
Cluster analysis in R: determine the optimal number of clusters
(8 answers)
Closed 3 years ago.
I am going to build a K-means clustering model for outlier detection. For that, I need to identify the best number of clusters needs to be selected.
For now, I have tried to do this using Elbow Method. I plotted the sum of squared error vs. the number of clusters(k) but, I got a graph like below which makes confusion to identify the elbow point.
I need to know, why do I get a graph like this and how do I identify the optimal number of clusters.
K-means is not suitable for outlier detection. This keeps popping up here all the time.
K-means is conceptualized for "pure" data, with no false points. All measurements are supposed to come from the data, and only vary by some Gaussian measurement error. Occasionally this may yield some more extreme values, but even these are real measurements, from the real clusters, and should be explained not removed.
K-means itself is known to not work well on noisy data where data points do not belong to the clusters
It tends to split large real clusters in two, and then points right in the middle of the real cluster will have a large distance to the k-means centers
It tends to put outliers into their own clusters (because that reduces SSQ), and then the actual outliers will have a small distance, even 0.
Rather use an actual outlier detection algorithm such as Local Outlier Factor, kNN, LOOP etc. instead that were conceptualized with noisy data in mind.
Remember that the Elbow Method doesn't just 'give' the best value of k, since the best value of k is up to interpretation.
The theory behind the Elbow Method is that we in tandem both want to minimize some error function (i.e. sum of squared errors) while also picking a low value of k.
The Elbow Method thus suggests that a good value of k would lie in a point on the plot that resembles an elbow. That is the error is small, but doesn't decrease drastically when k increases locally.
In your plot you could argue that both k=3 and k=6 resembles elbows. By picking k=3 you'd have picked a small k, and we see that k=4, and k=5 doesn't do much better in minimizing the error. Same goes with k=6.
Suppose that I have already found the eps for all density. I applied the methodology from here http://ijiset.com/v1s4/IJISET_V1_I4_48.pdf
If you don't mind, please open page 5 and see at Proposed Algorithm section. At step 10.1, the paper tells us to calculate the number of objects in eps-neighborhood.
What does eps represent actually? It is a radius to draw a circle right? So, why the radius is so small, smaller than distances between two objects? If so, the MinPts will be 0 forever.
Yes, if used with Euclidean distance, then it is a radius.
It is not infinitely small (it does not tend to 0). It's just supposed to be small compared to the data set extends, but the authors could have named it "r" instead.
Use the original paper to understand the algorithm, not some indian journal variant of it.
In Euclidean distance, it is the radius. Selection of Eps is a little difficult.
This problem is related to model selection, i.e., the selection of a particular model and its corresponding parametrization. In the case of k-means (which requires from the user the number of clusters as input) there is a plethora of measures in the literature that can help in the selection of the best number of clusters, for instance: silhouette, c-index, dunn, davies-bouldin. These measures are the so-called relative validity criteria.
In the case of Density-based clustering algorithms, there are some measures too, for instance: CDbw and DBCV.
I have two datasets (tracks) with points in x/y which represent GPS positions. I want to analyze the distance between both tracks. The points are not necessary in sync, but having the same frequency, as shown in this little excerpt (each track consists of 1000+ points):
Example Picture
Due to being not in sync I can't just compare the two points which are closest to each other. And since the path is not exactly the same I can't sync the tracks. It might be a solution interpolating a curve for each dataset and then calculating the integral in between. Since the tracks are much longer than shown in the example I can't just use regression functions like polyfit.
How can this be done or are there other/better strategies for analyzing (mean/mean square...) the distance?
am304's answer is by far the easiest, and probably the way to go.
However, I'd like to add a few other ways to do this, which are much more complicated, but could greatly enhance accuracy depending on your use case.
And if it's not for you, then it could be useful for anyone else passing by.
Method 1
Pros: fast, easy
Cons: method is overly optimistic about the smoothness of the tracks
Determine the B-spline representation for both tracks. You then have a parametric relation for both tracks:
The distance between both tracks is then the average of the function
for all applicable t, which is computed through the following integral:
Method 2
Pros: closest to the "physics" of the situation
Cons: hard to get right, specific to the situation and thus non-reusable
Use the equations of motion of whatever was following that track to derive a transition matrix for any arbitrary time step t. When possible, also come up with an appropriate noise model.
Use a Kalman filter to re-sample both tracks to some equally-spaced time vector, which is preferably different from the time vector of both track 1 and track 2.
Compute the distances between the x,y pairs thus computed, and take the average.
Method 3
Pros: fast, easy
Cons: method is overly optimistic about the smoothness of the tracks. Both fits are biased.
Fit a space curve through track 1
Compute the distances of all points in track 2 to this space curve.
Repeat 1 and 2, but vice versa.
Take the average of all these distances.
Method 4
Pros: fast, easy
Cons: method is overly optimistic about the smoothness of the tracks. Fit will be of lesser quality due to inherently larger noise terms.
Fit a space curve to the union of both tracks. That is, treat points from track 1 and track 2 as a single data set, through which to fit a space curve.
Compute the perpendicular residuals of both tracks with respect to this space curve.
compute the average all these distances.
Remarks
Note that all methods here use the flat-Earth assumption. If the tracks are truly long and cover a non-negligible portion of the Earth's surface, you'll have to compute distances via the Haversine formula rather than a mere Pythagorean root. The Kalman filter is less sensitive to this, provided your equations of motion take care of a spherical Earth.
If you have an elevation model of the region of interest, use that. Of course depending on the area, you'd be surprised how much of a difference that makes compared to a smooth Earth.
Is the x/y data logged as a function of time? If so, you can resample one or both datasets to have to same sample time vector using the resample function for timeseries. You'll have to convert your data to a timeseries object first, but it's worth it. Once both data sets are resampled to the same time vector, you simply subtract one from the other.
I have a set of weighted features for machine learning. I'd like to reduce the feature set and just use those with a very large or very small weight.
So given below image of sorted weights, I'd only like to use the features that have weights above the higher or below the lower yellow line.
What I'm looking for is some kind of slope change detection so I can discard all the features until the first/last slope coefficient increase/decrease.
While I (think I) know how to code this myself (with first and second numerical derivatives), I'm interested in any established methods. Perhaps there's some statistic or index that computes something like that, or anything I can use from SciPy?
Edit:
At the moment, I'm using 1.8*positive.std() as positive and 1.8*negative.std() as negative threshold (fast and simple), but I'm not mathematician enough to determine how robust this is. I don't think it is, though. ⍨
If the data are (approximately) Gaussian distributed, then just using a multiple
of the standard deviation is sensible.
If you are worried about heavier tails, then you may want to base your analysis on order
statistics.
Since you've plotted it, I'll assume you're willing to sort all of the
data.
Let N be the number of data points in your sample.
Let x[i] be the i'th value in the sorted list of values.
Then 0.5( x[int( 0.8413*N)]-x[int(0.1587*N)]) is an estimate of the standard deviation
which is more robust against outliers. This estimate of the std can be used as you
indicated above. (The magic numbers above are the fraction of data that are
less than [mean+1sigma] and [mean-1sigma] respectively).
There are also conditions where just keeping the highest 10% and lowest 10% would be
sensible as well; and these cutoffs are easily computed if you have the sorted data
on hand.
These are somewhat ad hoc approaches based on the content of your question.
The general sense of what you're trying to do is (a form of) anomaly detection,
and you can probably do a better job of it if you're careful in defining/estimating
what the shape of the distribution is near the middle, so that you can tell when
the features are getting anomalous.
I have a dataset consisting of a large collection of points in three dimensional euclidian space. In this collection of points, i am trying to find the point that is nearest to the area with the highest density of points.
So my problem consists of two steps:
1: Determine where density of the distribution of points is at its highest
2: Determine which point is nearest to the point found in 1
Point 2 i can manage, but i'm not sure how to solve point 1. I know there are a lot of functions for density estimation in Matlab, but i'm not sure which one would be the most suitable, or straightforward to use.
Does anyone know?
My command of statistics is a little bit rusty, but as far as i can tell, this type of problem calls for multivariate analysis. Someone suggested i use multivariate kernel density estimation, but i'm not really sure if that's the best solution.
Density is a measure of mass per unit volume. On the assumption that your points all have the same mass then you are, I suppose, trying to measure the number of points per unit volume. So one approach is to divide your subset of Euclidean space into lots of little unit volumes (let's call them voxels like everyone does) and count how many points there are in each one. The voxel with the most points is where the density of points is at its highest. This is, of course, numerical integration of a sort. If your points were distributed according to some analytic function (and I guess they are not) you could solve the problem with pencil and paper.
You might make this approach as sophisticated as you like, perhaps initially dividing your space into 2 x 2 x 2 voxels, then choosing the voxel with most points and sub-dividing that in turn until your criteria are satisfied.
I hope this will get you started on your point 1; you seem to be OK with point 2 so I'll stop now.
EDIT
It looks as if triplequad might be what you are looking for.