i have a problem with clustering time series in R.
I googled a lot and found nothing that fits my problem.
I have made a STL-Decomposition of Timeseries.
The trend component is in a matrix with 64 columns, one for every series.
Now i want to cluster these series in simular groups, involve the curve shapes and the timely shift. I found some functions that imply one of these aspects but not both.
First i tried to calculte a distance matrix with the dtw-distance so i
found clusters based on the values and inply the time shift but not on the shape of the timeseries. After this i tried some correlation based clustering, but then the timely shift
we're not recognized and the result dont satisfy my claims.
Is there a function that could cover my problem or have i to build up something
on my own. Im thankful for every kind of help, after two days of tutorials and examples i totaly uninspired. I hope i could explain the problem well enough to you.
I attached a picture. Here you can see some example time series.
There you could see the problem. The two series in the middle are set to one cluster,
although the upper and the one on the bottom have the same shape as one of the middle.
Have you tried the R package dtwclust
https://cran.r-project.org/web/packages/dtwclust/index.html
(I'm just starting to explore this package, but it seems like it covers a lot of aspects of time series clustering and it has lots of good references.)
you can use the kml package. It is used specifically to longitudinal data. You can consult its help. It has the next example:
### Generation of some data
cld1 <- generateArtificialLongData(25)
### We suspect 3, 4 or 6 clusters, we want 3 redrawing.
### We want to "see" what happen (so printCal and printTraj are TRUE)
kml(cld1,c(3,4,6),3,toPlot='both')
### 4 seems to be the best. We want 7 more redrawing.
### We don't want to see again, we want to get the result as fast as possible.
kml(cld1,4,10)
Example cluster
Related
I'm working with a set of co-ordinates, and want to dynamically (I have many sets that need to go through this process) understand how many distinct groups there are within the data. My approach was to apply k-means to investigate whether it would find the centroids and I could go from there.
When plotting some data with 6 distinct clusters (visually) the k-means algorithm continues to ignore two significant clusters while putting many centroids into another.
See image below:
Red are the co-ordinate data points and blue are centroids that k-means has provided. In this specific case I've gone for 15 (arbitrary), but it still doesn't recognise those patches of data on the right hand side, rather putting a mid point between them while putting in 8 in the cluster in the top right.
Admittedly there are slightly more data points in the top right, but not by much.
I'm using the standard k-means algorithm in R and just feeding in x and y co-ordinates. I've tried standardising the data, but this doesn't make any difference.
Any thoughts on why this is, or other potential methodologies that could be applied to try and dynamically understand the number of distinct clusters there are in the data?
You could try with Self-organizing map:
this is a clustering algorithm based on Neural Networks which create a discretized representation of the input space of the training samples, called a map, and is, therefore, a method to do dimensionality reduction (SOM).
This algorithm is very good for clustering also because does not require a priori selection of the number of clusters (in k-mean you need to choose k, here no). In your case, it hopefully finds automatically the optimal number of cluster, and you can actually visualize it.
You can find a very nice python package called somoclu which has got this algorithm implemented and an easy way to visualize the result. Else you can go with R. Here you can find a blog post with a tutorial, and Cran package manual for SOM.
K-means is a randomized algorithm and it will get stuck in local minima.
Because of these problems, it is common to run k-means several times, and keep the result with least squares, I.e., the best of the local minima found.
I have a pretty big data table (about 100.000 observations) that I'd like to use for clustering. Since some of the data is categorical, I've tried using "gower distance" and then hclust() with the "ward" method.
The data itself is very heterogeneous, which is why I'd like to sort of "pre-cluster" the data and then do the actual cluster analysis. Have any of you done this before and can point me in the right direction? I'm at a loss at the moment :(
With the mentioned methods, I don't really get useful clusters.
Thanks guys, I really appreciate every tip I can get.
Edit: I think that I didn't really explain my problem right, so here's another attempt: let's say, that I have a dataset containing brands of cars and some of their features. Before clustering them by features I would like to precluster them by brand. So all BMW e.g. are in the same cluster and so on.. and only after that I would like to cluster by features, so I should get a cluster with fast cars etc.
does anybody know, how to do this in R?
this does not describe my dataset, but maybe the question I'm having is clearer now.
You should start with a sample first.
Once you get good results on the sample, try to reproduce it on a different sample. Once the results are stable, you can either try to scale the algorithm to the entire data set (maybe try doubling first), or you can train a classifier and predict the clusters of the remaining data. With most clustering algorithms, a 1 nearest neighbor classifier will be very good.
At work when I want to understand a dataset (I work with portfolio data in life insurance), I would normally use pivot tables in Excel to look at e.g. the development of variables over time or dependencies between variables.
I remembered from university the nice R-function where you can plot every column of a dataframe against every other column like in:
For the dependency between issue.age and duration this plot is actually interesting because you can clearly see that high issue ages come with shorter policy durations (because there is a maximum age for each policy). However the plots involving the issue year iss.year are much less "visual". In fact you cant see anything from them. I would like to see with once glance if the distribution of issue ages has changed over the different issue.years, something like
where you could see immediately that the average age of newly issue policies has been increasing from 2014 to 2016.
I don't want to write code that needs to be customized for every dataset that I put in because then I can also do it faster manually in Excel.
So my question is, is there an easy way to plot each column of a matrix against every other column with more flexible chart types than with the standard plot(data.frame)?
The ggpairs() function from the GGally library. It has a lot of capability for visualizing columns of all different types, and provides a lot of control over what to visualize.
For example, here is a snippet from the vignette linked to above:
data(tips, package = "reshape")
ggpairs(tips)
I have the following time series data. It has 60 data points shown below. Please see a simple plot of this data below. I am using R for plotting this. I think that if I draw a moving average curve on the points in the graph, then we can better understand the patterns in the data. I don't know how to do it in R. Could some one help me to do that. Additionally, I am not sure whether this is a good way to identify patterns or not. Please also suggest me if there is any better way. Thank you.
x <- c(18,21,18,14,8,14,10,14,14,12,12,14,10,10,12,6,10,8,
14,10,10,6,6,4,6,2,8,6,2,6,4,4,2,8,6,6,8,12,8,8,6,6,2,2,4,
4,4,8,14,8,6,6,2,6,6,4,4,8,6,6)
To answer your question about moving averages, you could accomplish it with the help of rollmean which is in package zoo.
From Joshua's comment: You could also look into TTR package that depends on xts that depends on zoo. Also, there are other moving averages in the package TTR: check ?MA.
require(TTR)
# assuming your vector is loaded in dat
# sliding window / moving average of size 5
dat.k5 <- rollmean(dat, k=5)
One reasonable possibility:
d <- data.frame(x=scan("tmp.dat"))
qplot(x=seq(nrow(d)),x,data=d)+geom_smooth(method="loess")
edit: moved from comment to answer, based on https://meta.stackexchange.com/questions/164783/why-was-a-seemingly-relevant-non-offensive-comment-removed
With regard to "is this a good way to identify patterns" (which is a little off-topic for StackOverflow, but whatever); I think rolling means are perfectly respectable, although more sophisticated methods (such as the locally-weighted regression [loess/lowess] shown here) do exist. However, it doesn't look to me as though there is much of a complicated pattern to detect here: the data seem to initially decline with time, then level off. Rolling means and more sophisticated approaches may look prettier, but I don't think they will identify any deeper patterns in this data set ...
If you want to do this sort of thing for multiple data sets at once (as indicated in your comment), you may like ggplot's capabilities for automatically producing multi-line or faceted versions of the same plot.
I am using R software (R commander) to cluster my data. I have a smaller subset of my data containing 200 rows and about 800 columns. I am getting the following error when trying kmeans cluster and plot on a graph.
"'princomp' can only be used with more units than variables"
I then created a test doc of 10 row and 10 columns whch plots fine but when I add an extra column I get te error again.
Why is this? I need to be able to plot my cluster. When I view my data set after performing kmeans on it I can see the extra results column which shows which clusters they belong to.
IS there anything I am doing wrong, can I ger rid of this error and plot my larger sample???
Please help, been wrecking my head for a week now.
Thanks guys.
The problem is that you have more variables than sample points and the principal component analysis that is being done is failing.
In the help file for princomp it explains (read ?princomp):
‘princomp’ only handles so-called R-mode PCA, that is feature
extraction of variables. If a data matrix is supplied (possibly
via a formula) it is required that there are at least as many
units as variables. For Q-mode PCA use ‘prcomp’.
Principal component analysis is underspecified if you have fewer samples than data point.
Every data point will be it's own principal component. For PCA to work, the number of instances should be significantly larger than the number of dimensions.
Simply speaking you can look at the problems like this:
If you have n dimensions, you can encode up to n+1 instances using vectors that are all 0 or that have at most one 1. And this is optimal, so PCA will do this! But it is not very helpful.
you can use prcomp instead of princomp