I'd be grateful if people could help me find an efficient way (probably low memory algorithm) to tackle the following problem.
I need to find the stationary distribution x of a transition matrix P. The transition matrix is an extremely large, extremely sparse matrix, constructed such that all the columns sum to 1. Since the stationary distribution is given by the equation Px = x, then x is simply the eigenvector of P associated with eigenvalue 1.
I'm currently using GNU Octave to both generate the transition matrix, find the stationary distribution, and plot the results. I'm using the function eigs(), which calculates both eigenvalues and eigenvectors, and it is possible to return just one eigenvector, where the eigenvalue is 1 (I actually had to specify 1.1, to prevent an error). Construction of the transition matrix (using a sparse matrix) is fairly quick, but finding the eigenvector gets increasingly slow as I increase the size, and I'm running out of memory before I can examine even moderately sized problems.
My current code is
[v l] = eigs(P, 1, 1.01);
x = v / sum(v);
Given that I know that 1 is the eigenvalue, I'm wondering if there is either a better method to calculate the eigenvector, or a way that makes more efficient use of memory, given that I don't really need an intermediate large dense matrix. I naively tried
n = size(P,1); % number of states
Q = P - speye(n,n);
x = Q\zeros(n,1); % solve (P-I)x = 0
which fails, since Q is singular (by definition).
I would be very grateful if anyone has any ideas on how I should approach this, as it's a calculation I have to perform a great number of times, and I'd like to try it on larger and more complex models if possible.
As background to this problem, I'm solving for the equilibrium distribution of the number of infectives in a cattle herd in a stochastic SIR model. Unfortunately the transition matrix is very large for even moderately sized herds. For example: in an SIR model with an average of 20 individuals (95% of the time the population is between 12 and 28 individuals), P is 21169 by 21169 with 20340 non-zero values (i.e. 0.0005% dense), and uses up 321 Kb (a full matrix of that size would be 3.3 Gb), while for around 50 individuals P uses 3 Mb. x itself should be pretty small. I suspect that eigs() has a dense matrix somewhere, which is causing me to run out of memory, so I should be okay if I can avoid using full matrices.
Power iteration is a standard way to find the dominant eigenvalue of a matrix. You pick a random vector v, then hit it with P repeatedly until you stop seeing it change very much. You want to periodically divide v by sqrt(v^T v) to normalise it.
The rate of convergence here is proportional to the separation between the largest eigenvalue and the second largest eigenvalue. Each iteration takes just a couple of matrix multiplies.
There are fancier-pants ways to do this ("PageRank" is one good thing to search for here) that improve speed for really huge sparse matrices, but I don't know that they're necessary or useful here.
Your approach seems like a good one. However, what you're calling x, is the null space of Q. null(Q) would work if it supported sparse matrices, but it doesn't. There's a bunch of stuff on the web for finding the null space of a sparse matrix. For example:
http://www.mathworks.co.uk/matlabcentral/newsreader/view_thread/249467
http://www.mathworks.com/matlabcentral/fileexchange/42922-null-space-for-sparse-matrix/content/nulls.m
http://www.mathworks.com/matlabcentral/fileexchange/11120-null-space-of-a-sparse-matrix
It seems the best solution is to use the Power Iteration method, as suggested by tmyklebu.
The method is to iterate x = Px; x /= sum(x), until x converges. I'm assuming convergence if the d1 norm between successive iterations is less than 1e-5, as that seems to give good results.
Convergence can take a while, since the largest two eigenvalues are fairly close (the number of iterations needed to converge can vary considerably, from around 200 to 2000 depending on the model used and population sizes, but it gets there in the end). However, the memory requirements are low, and it's very easy to implement.
Related
This problem may seem very different from the normal Matrix Factorization task which is widely used in recommender system.
My problem is described as below:
Given a dense Matrix M
(approximately 55000*200, may contain much negative elements, 0.1< abs(M[i][j]) <1 )
I have to find two matrix A(55000*1400) and B(1400*200), such that:
AB=M
However, we have some knowledge about A. We have another Matrix C, if C[i][j] = 0, then A[i][j] must be zero, otherwise it can be any value(C[i][j] = 1).
In my practice , I use machine learning to solve the problem, my loss function is:
||(A*C)(element-wise product) x B - M ||(2)(L2 norm)
I have tried adagrad,momentum,adadelta and some other optimization method, but the train error is pretty much and is cut down slowly (learning_rate = 0.1)
UP1:
Well, actually I've got a machine with 32GB memory and I only need 2 min for each epoch. I decompose an element in M only if its corresponding element in C is anotated as 1. Practically , I only decompose M[i][j] when C[i][j] = 1, and after I decompose M[i][j], I solve the gradient for M[i][j] to update A[i : ] and B[ : j] at once. So, the batch I used is too small--just contain one element. Also , I have to mention that C is a pretty sparse matrix. For each line in C, there is only 2-3 elements that are anotated as 1.
After struggling with it for nearly half month, I finally got the answer: I should update the matrix A much more quickly, say, update the parameters at a more smaller step. I originally updated every element in A only once per epoch, much less than B. However, after I changed the code to let A be updated at the same speed as B, then surprise happened: it worked pretty well!
Maybe smaller steps will help SGD work better? I don't really believe it mathematically.
I have a few distance functions which return distance between two images , I want to combine these distance into a single distance, using weighted scoring e.g. ax1+bx2+cx3+dx4 etc i want to learn these weights automatically such that my test error is minimised.
For this purpose i have a labeled dataset which has various triplets of images such that (a,b,c) , a has less distance to b than it has to c.
i.e. d(a,b)<d(a,c)
I want to learn such weights so that this ordering of triplets can be as accurate as possible.(i.e. the weighted linear score given is less for a&b and more for a&c).
What sort of machine learning algorithm can be used for the task,and how the desired task can be achieved?
Hopefully I understand your question correctly, but it seems that this could be solved more easily with constrained optimization directly, rather than classical machine learning (the algorithms of which are often implemented via constrained optimization, see e.g. SVMs).
As an example, a possible objective function could be:
argmin_{w} || e ||_2 + lambda || w ||_2
where w is your weight vector (Oh god why is there no latex here), e is the vector of errors (one component per training triplet), lambda is some tunable regularizer constant (could be zero), and your constraints could be:
max{d(I_p,I_r)-d(I_p,I_q),0} <= e_j for jth (p,q,r) in T s.t. d(I_p,I_r) <= d(I_p,I_q)
for the jth constraint, where I_i is image i, T is the training set, and
d(u,v) = sum_{w_i in w} w_i * d_i(u,v)
with d_i being your ith distance function.
Notice that e is measuring how far your chosen weights are from satisfying all the chosen triplets in the training set. If the weights preserve ordering of label j, then d(I_p,I_r)-d(I_p,I_q) < 0 and so e_j = 0. If they don't, then e_j will measure the amount of violation of training label j. Solving the optimization problem would give the best w; i.e. the one with the lowest error.
If you're not familiar with linear/quadratic programming, convex optimization, etc... then start googling :) Many libraries exist for this type of thing.
On the other hand, if you would prefer a machine learning approach, you may be able to adapt some metric learning approaches to your problem.
I developed a image processing program that identifies what a number is given an image of numbers. Each image was 27x27 pixels = 729 pixels. I take each R, G and B value which means I have 2187 variables from each image (+1 for the intercept = total of 2188).
I used the below gradient descent formula:
Repeat {
θj = θj−α/m∑(hθ(x)−y)xj
}
Where θj is the coefficient on variable j; α is the learning rate; hθ(x) is the hypothesis; y is real value and xj is the value of variable j. m is the number of training sets. hθ(x), y are for each training set (i.e. that's what the summation sign is for). Further the hypothesis is defined as:
hθ(x) = 1/(1+ e^-z)
z= θo + θ1X1+θ2X2 +θ3X3...θnXn
With this, and 3000 training images, I was able to train my program in just over an hour and when tested on a cross validation set, it was able to identify the correct image ~ 67% of the time.
I wanted to improve that so I decided to attempt a polynomial of degree 2.
However the number of variables jumps from 2188 to 2,394,766 per image! It takes me an hour just to do 1 step of gradient descent.
So my question is, how is this vast number of variables handled in machine learning? On the one hand, I don't have enough space to even hold that many variables for each training set. On the other hand, I am currently storing 2188 variables per training sample, but I have to perform O(n^2) just to get the values of each variable multiplied by another variable (i.e. the polynomial to degree 2 values).
So any suggestions / advice is greatly appreciated.
try to use some dimensionality reduction first (PCA, kernel PCA, or LDA if you are classifying the images)
vectorize your gradient descent - with most math libraries or in matlab etc. it will run much faster
parallelize the algorithm and then run in on multiple CPUs (but maybe your library for multiplying vectors already supports parallel computations)
Along with Jirka-x1's answer, I would first say that this is one of the key differences in working with image data than say text data for ML: high dimensionality.
Second... this is a duplicate, see How to approach machine learning problems with high dimensional input space?
I'm not a DSP expert, but I understand that there are two ways that I can apply a discrete time-domain filter to a discrete time-domain waveform. The first is to convolve them in the time domain, and the second is to take the FFT of both, multiply both complex spectrums, and take IFFT of the result. One key difference in these methods is the second approach is subject to circular convolution.
As an example, if the filter and waveforms are both N points long, the first approach (i.e. convolution) produces a result that is N+N-1 points long, where the first half of this response is the filter filling up and the 2nd half is the filter emptying. To get a steady-state response, the filter needs to have fewer points than the waveform to be filtered.
Continuing this example with the second approach, and assuming the discrete time-domain waveform data is all real (not complex), the FFT of the filter and the waveform both produce FFTs of N points long. Multiplying both spectrums IFFT'ing the result produces a time-domain result also N points long. Here the response where the filter fills up and empties overlap each other in the time domain, and there's no steady state response. This is the effect of circular convolution. To avoid this, typically the filter size would be smaller than the waveform size and both would be zero-padded to allow space for the frequency convolution to expand in time after IFFT of the product of the two spectrums.
My question is, I often see work in the literature from well-established experts/companies where they have a discrete (real) time-domain waveform (N points), they FFT it, multiply it by some filter (also N points), and IFFT the result for subsequent processing. My naive thinking is this result should contain no steady-state response and thus should contain artifacts from the filter filling/emptying that would lead to errors in interpreting the resulting data, but I must be missing something. Under what circumstances can this be a valid approach?
Any insight would be greatly appreciated
The basic problem is not about zero padding vs the assumed periodicity, but that Fourier analysis decomposes the signal into sine waves which, at the most basic level, are assumed to be infinite in extent. Both approaches are correct in that the IFFT using the full FFT will return the exact input waveform, and both approaches are incorrect in that using less than the full spectrum can lead to effects at the edges (that usually extend a few wavelengths). The only difference is in the details of what you assume fills in the rest of infinity, not in whether you are making an assumption.
Back to your first paragraph: Usually, in DSP, the biggest problem I run into with FFTs is that they are non-causal, and for this reason I often prefer to stay in the time domain, using, for example, FIR and IIR filters.
Update:
In the question statement, the OP correctly points out some of the problems that can arise when using FFTs to filter signals, for example, edge effects, that can be particularly problematic when doing a convolution that is comparable in the length (in the time domain) to the sampled waveform. It's important to note though that not all filtering is done using FFTs, and in the paper cited by the OP, they are not using FFT filters, and the problems that would arise with an FFT filter implementation do not arise using their approach.
Consider, for example, a filter that implements a simple average over 128 sample points, using two different implementations.
FFT: In the FFT/convolution approach one would have a sample of, say, 256, points and convolve this with a wfm that is constant for the first half and goes to zero in the second half. The question here is (even after this system has run a few cycles), what determines the value of the first point of the result? The FFT assumes that the wfm is circular (i.e. infinitely periodic) so either: the first point of the result is determined by the last 127 (i.e. future) samples of the wfm (skipping over the middle of the wfm), or by 127 zeros if you zero-pad. Neither is correct.
FIR: Another approach is to implement the average with an FIR filter. For example, here one could use the average of the values in a 128 register FIFO queue. That is, as each sample point comes in, 1) put it in the queue, 2) dequeue the oldest item, 3) average all of the 128 items remaining in the queue; and this is your result for this sample point. This approach runs continuously, handling one point at a time, and returning the filtered result after each sample, and has none of the problems that occur from the FFT as it's applied to finite sample chunks. Each result is just the average of the current sample and the 127 samples that came before it.
The paper that OP cites takes an approach much more similar to the FIR filter than to the FFT filter (note though that the filter in the paper is more complicated, and the whole paper is basically an analysis of this filter.) See, for example, this free book which describes how to analyze and apply different filters, and note also that the Laplace approach to analysis of the FIR and IIR filters is quite similar what what's found in the cited paper.
Here's an example of convolution without zero padding for the DFT (circular convolution) vs linear convolution. This is the convolution of a length M=32 sequence with a length L=128 sequence (using Numpy/Matplotlib):
f = rand(32); g = rand(128)
h1 = convolve(f, g)
h2 = real(ifft(fft(f, 128)*fft(g)))
plot(h1); plot(h2,'r')
grid()
The first M-1 points are different, and it's short by M-1 points since it wasn't zero padded. These differences are a problem if you're doing block convolution, but techniques such as overlap and save or overlap and add are used to overcome this problem. Otherwise if you're just computing a one-off filtering operation, the valid result will start at index M-1 and end at index L-1, with a length of L-M+1.
As to the paper cited, I looked at their MATLAB code in appendix A. I think they made a mistake in applying the Hfinal transfer function to the negative frequencies without first conjugating it. Otherwise, you can see in their graphs that the clock jitter is a periodic signal, so using circular convolution is fine for a steady-state analysis.
Edit: Regarding conjugating the transfer function, the PLLs have a real-valued impulse response, and every real-valued signal has a conjugate symmetric spectrum. In the code you can see that they're just using Hfinal[N-i] to get the negative frequencies without taking the conjugate. I've plotted their transfer function from -50 MHz to 50 MHz:
N = 150000 # number of samples. Need >50k to get a good spectrum.
res = 100e6/N # resolution of single freq point
f = res * arange(-N/2, N/2) # set the frequency sweep [-50MHz,50MHz), N points
s = 2j*pi*f # set the xfer function to complex radians
f1 = 22e6 # define 3dB corner frequency for H1
zeta1 = 0.54 # define peaking for H1
f2 = 7e6 # define 3dB corner frequency for H2
zeta2 = 0.54 # define peaking for H2
f3 = 1.0e6 # define 3dB corner frequency for H3
# w1 = natural frequency
w1 = 2*pi*f1/((1 + 2*zeta1**2 + ((1 + 2*zeta1**2)**2 + 1)**0.5)**0.5)
# H1 transfer function
H1 = ((2*zeta1*w1*s + w1**2)/(s**2 + 2*zeta1*w1*s + w1**2))
# w2 = natural frequency
w2 = 2*pi*f2/((1 + 2*zeta2**2 + ((1 + 2*zeta2**2)**2 + 1)**0.5)**0.5)
# H2 transfer function
H2 = ((2*zeta2*w2*s + w2**2)/(s**2 + 2*zeta2*w2*s + w2**2))
w3 = 2*pi*f3 # w3 = 3dB point for a single pole high pass function.
H3 = s/(s+w3) # the H3 xfer function is a high pass
Ht = 2*(H1-H2)*H3 # Final transfer based on the difference functions
subplot(311); plot(f, abs(Ht)); ylabel("abs")
subplot(312); plot(f, real(Ht)); ylabel("real")
subplot(313); plot(f, imag(Ht)); ylabel("imag")
As you can see, the real component has even symmetry and the imaginary component has odd symmetry. In their code they only calculated the positive frequencies for a loglog plot (reasonable enough). However, for calculating the inverse transform they used the values for the positive frequencies for the negative frequencies by indexing Hfinal[N-i] but forgot to conjugate it.
I can shed some light to the reason why "windowing" is applied before FFT is applied.
As already pointed out the FFT assumes that we have a infinite signal. When we take a sample over a finite time T this is mathematically the equivalent of multiplying the signal with a rectangular function.
Multiplying in the time domain becomes convolution in the frequency domain. The frequency response of a rectangle is the sync function i.e. sin(x)/x. The x in the numerator is the kicker, because it dies down O(1/N).
If you have frequency components which are exactly multiples of 1/T this does not matter as the sync function is zero in all points except that frequency where it is 1.
However if you have a sine which fall between 2 points you will see the sync function sampled on the frequency point. It lloks like a magnified version of the sync function and the 'ghost' signals caused by the convolution die down with 1/N or 6dB/octave. If you have a signal 60db above the noise floor, you will not see the noise for 1000 frequencies left and right from your main signal, it will be swamped by the "skirts" of the sync function.
If you use a different time window you get a different frequency response, a cosine for example dies down with 1/x^2, there are specialized windows for different measurements. The Hanning window is often used as a general purpose window.
The point is that the rectangular window used when not applying any "windowing function" creates far worse artefacts than a well chosen window. i.e by "distorting" the time samples we get a much better picture in the frequency domain which closer resembles "reality", or rather the "reality" we expect and want to see.
Although there will be artifacts from assuming that a rectangular window of data is periodic at the FFT aperture width, which is one interpretation of what circular convolution does without sufficient zero padding, the differences may or may not be large enough to swamp the data analysis in question.
I have a scenario where I have several thousand instances of data. The data itself is represented as a single integer value. I want to be able to detect when an instance is an extreme outlier.
For example, with the following example data:
a = 10
b = 14
c = 25
d = 467
e = 12
d is clearly an anomaly, and I would want to perform a specific action based on this.
I was tempted to just try an use my knowledge of the particular domain to detect anomalies. For instance, figure out a distance from the mean value that is useful, and check for that, based on heuristics. However, I think it's probably better if I investigate more general, robust anomaly detection techniques, which have some theory behind them.
Since my working knowledge of mathematics is limited, I'm hoping to find a technique which is simple, such as using standard deviation. Hopefully the single-dimensioned nature of the data will make this quite a common problem, but if more information for the scenario is required please leave a comment and I will give more info.
Edit: thought I'd add more information about the data and what I've tried in case it makes one answer more correct than another.
The values are all positive and non-zero. I expect that the values will form a normal distribution. This expectation is based on an intuition of the domain rather than through analysis, if this is not a bad thing to assume, please let me know. In terms of clustering, unless there's also standard algorithms to choose a k-value, I would find it hard to provide this value to a k-Means algorithm.
The action I want to take for an outlier/anomaly is to present it to the user, and recommend that the data point is basically removed from the data set (I won't get in to how they would do that, but it makes sense for my domain), thus it will not be used as input to another function.
So far I have tried three-sigma, and the IQR outlier test on my limited data set. IQR flags values which are not extreme enough, three-sigma points out instances which better fit with my intuition of the domain.
Information on algorithms, techniques or links to resources to learn about this specific scenario are valid and welcome answers.
What is a recommended anomaly detection technique for simple, one-dimensional data?
Check out the three-sigma rule:
mu = mean of the data
std = standard deviation of the data
IF abs(x-mu) > 3*std THEN x is outlier
An alternative method is the IQR outlier test:
Q25 = 25th_percentile
Q75 = 75th_percentile
IQR = Q75 - Q25 // inter-quartile range
IF (x < Q25 - 1.5*IQR) OR (Q75 + 1.5*IQR < x) THEN x is a mild outlier
IF (x < Q25 - 3.0*IQR) OR (Q75 + 3.0*IQR < x) THEN x is an extreme outlier
this test is usually employed by Box plots (indicated by the whiskers):
EDIT:
For your case (simple 1D univariate data), I think my first answer is well suited.
That however isn't applicable to multivariate data.
#smaclell suggested using K-means to find the outliers. Beside the fact that it is mainly a clustering algorithm (not really an outlier detection technique), the problem with k-means is that it requires knowing in advance a good value for the number of clusters K.
A better suited technique is the DBSCAN: a density-based clustering algorithm. Basically it grows regions with sufficiently high density into clusters which will be maximal set of density-connected points.
DBSCAN requires two parameters: epsilon and minPoints. It starts with an arbitrary point that has not been visited. It then finds all the neighbor points within distance epsilon of the starting point.
If the number of neighbors is greater than or equal to minPoints, a cluster is formed. The starting point and its neighbors are added to this cluster and the starting point is marked as visited. The algorithm then repeats the evaluation process for all the neighbors recursively.
If the number of neighbors is less than minPoints, the point is marked as noise.
If a cluster is fully expanded (all points within reach are visited) then the algorithm proceeds to iterate through the remaining unvisited points until they are depleted.
Finally the set of all points marked as noise are considered outliers.
There are a variety of clustering techniques you could use to try to identify central tendencies within your data. One such algorithm we used heavily in my pattern recognition course was K-Means. This would allow you to identify whether there are more than one related sets of data, such as a bimodal distribution. This does require you having some knowledge of how many clusters to expect but is fairly efficient and easy to implement.
After you have the means you could then try to find out if any point is far from any of the means. You can define 'far' however you want but I would recommend the suggestions by #Amro as a good starting point.
For a more in-depth discussion of clustering algorithms refer to the wikipedia entry on clustering.
This is an old topic but still it lacks some information.
Evidently, this can be seen as a case of univariate outlier detection. The approaches presented above have several pros and cons. Here are some weak spots:
Detection of outliers with the mean and sigma has the obvious disadvantage of dependence of mean and sigma on the outliers themselves.
The case of the small sample limit (see question for example) is not adequately covered by, 3 sigma, K-Means, IQR etc.
And I could go on... However the statistical literature offers a simple metric: the median absolute deviation. (Medians are insensitive to outliers)
Details can be found here: https://www.sciencedirect.com/book/9780128047330/introduction-to-robust-estimation-and-hypothesis-testing
I think this problem can be solved in a few lines of python code like this:
import numpy as np
import scipy.stats as sts
x = np.array([10, 14, 25, 467, 12]) # your values
np.abs(x - np.median(x))/(sts.median_abs_deviation(x)/0.6745) #MAD criterion
Subsequently you reject values above a certain threshold (97.5 percentile of the distribution of data), in case of an assumed normal distribution the threshold is 2.24. Here it translates to:
array([ 0.6745 , 0. , 1.854875, 76.387125, 0.33725 ])
or the 467 entry being rejected.
Of course, one could argue, that the MAD (as presented) also assumes a normal dist. Therefore, why is it that argument 2 above (small sample) does not apply here? The answer is that MAD has a very high breakdown point. It is easy to choose different threshold points from different distributions and come to the same conclusion: 467 is the outlier.
Both three-sigma rule and IQR test are often used, and there are a couple of simple algorithms to detect anomalies.
The three-sigma rule is correct
mu = mean of the data
std = standard deviation of the data
IF abs(x-mu) > 3*std THEN x is outlier
The IQR test should be:
Q25 = 25th_percentile
Q75 = 75th_percentile
IQR = Q75 - Q25 // inter-quartile range
If x > Q75 + 1.5 * IQR or x < Q25 - 1.5 * IQR THEN x is a mild outlier
If x > Q75 + 3.0 * IQR or x < Q25 – 3.0 * IQR THEN x is a extreme outlier