Problem Constraints
Size of the data set, but not the data itself, is known.
Data set grows by one data point at a time.
Trend line is graphed one data point at a time (using a spline/Bezier curve).
Graphs
The collage below shows data sets with reasonably accurate trend lines:
The graphs are:
Upper-left. By hour, with ~24 data points.
Upper-right. By day for one year, with ~365 data points.
Lower-left. By week for one year, with ~52 data points.
Lower-right. By month for one year, with ~12 data points.
User Inputs
The user can select:
the type of time series (hourly, daily, monthly, quarterly, annual); and
the start and end dates for the time series.
For example, the user could select a daily report for 30 days in June.
Trend Weight
To calculate the window size (i.e., the number of data points to average when calculating the trend line), the following expression is used:
data points / trend weight
Where data points is derived from user inputs and trend weight is 6.4. Even though a trend weight of 6.4 produces good fits, it is rather arbitrary, and might not be appropriate for different user inputs.
Question
How should trend weight be calculated given the constraints of this problem?
Based on the looks of the graphs I would say you have too many points for your 12 point graph (it is just a spline of the points given... which is visually pleasing, but actually does more harm than good when trying to understand the trend) and too few points for your 365 point graph. Perhaps try doing something a little exponential like:
(Data points)^1.2/14.1
I do realize this is even more arbitrary than what you already have, but arbitrary isn't the worst thing in the world.
(I got 14.1 by trying to keep the 52 point graph fixed, since that one looks nice, by taking (52^(1.2)/52)*6.4=14.1. You using this technique you could try other powers besides 1.2 to see what you visually get.
Dan
I voted this up for the quality of your results and the clarity of your write-up. I wish I could offer an answer that could improve on your already excellent work.
I fear that it might be a matter of trial and error with the trend weight until you see an improved fit.
It could be that you could make this an input from users as well: allow them to fiddle with the value, given realistic constraints, until they get satisfactory values.
I also wondered if the weight would be different for each graph, since the number of points in each is different. Are you trying to get a single weighting that works for all graphs?
Excellent work; a nice question. Well done. I wish I was more helpful. Perhaps someone else will have more wisdom to impart than I do.
It might look like the trend lines are accurate in those 4 graphs but its really quite off. (This is best seen in the begging of the lower left one and the beginning of the upper right. I would think that you would want to use no less than half of your points when finding the trend line (though really you should use much more than half). I would suggest a Trend Weight of 2 at a maximum. Though really you ought to stick closer to the 1-1.5 range. Since it is arbitrary i would suggest you give your user an "accuracy of trend line" slider that they can use where the most accurate setting uses a trend weight of 1 and the least accurate uses a weight of #of data points +1. This would use 0 points (amusing you always round down) and, i would assume, though your statistics software might be different, will generate a strait horizontal line.
Related
Given a set of X/Y co-ordinates ([(x,y)] with increasing X(representing a timestamp) and Y representing a value/measurement at that timestamp.
This set can possibly be huge and i would like to avoid returning every single point in the set for display but rather find a smaller subset that would represent the overall trend of the measurement(some level of accuracy loss in the line graph will be acceptable).
So far, i tried the simple uniform sampling of measurement skipping points at uniform interval, then adding the max/min measurement value to the subset. While this is simple, It doesn't really account well for local peaks or valleys if the measurement fluctuates often.
I'm wondering if there are any standard algorithms that deal with solving this type of problems on server side?
Appreciate if anyone has solved it or know of any util/common libraries solving such problems. I'm on Java, but if there is any reference to standard algorithms i might try to implement one in Java.
It's hard to give a general answer to this question. It all depends on how your datapoints are stored, what properties your chart has, how it is rendered etc.
But as #dmuir suggested, you should check out the Douglas-Peucker algorithm. Another approach I just thought up could be to split the input data into chunks of some size (maybe corresponding to a single horizontal pixel) and then using some statistic (min, max, or average) for rendering chunk. If you use running statistics when adding data points to a chunk, this should be O(n), so it's not more expensive than the reading on of your data points.
What I seek is to turn a grid into a somewhat "random" plane of tiles.
I tried just multiplying Math.random() individually with the width and height of the plane (in this case its 800 / 600). The circles you see there are points that intersect each other and have been removed from the scene.
As you can see, it looks very far from an "evenly distributed" field of points. There are large holes and just as bad, clusters of points can be seen.
What I am looking for is a way to distribute these points better to have a minimum amount of clusters and holes. Ideally, to have a value that is the minimum distance between any two points, while having the maximum number of points that can fit in the area. I am fine with approximations of all kinds, I just don't want to attempt to do a greedy distribution.
Whatever ecma solution you give its fine, I can convert it to Actionscript.
I have found a visual example. The left side is what I got and the right is what I aim for.
You can try Loyds algorithm, i.e. centroidal weighted voronoi diagrams. Compute the vd and then the center of gravity of each cell. Replace the old points and rinse and repeat: http://www-cs-students.stanford.edu/~amitp/game-programming/polygon-map-generation/.
In general, it is a non-trivial problem, and there are many different approaches.
One that I have liked, since it is fast and produces decent results, is the quasi-random number generator from this article: "The Unreasonable Effectiveness of Quasirandom Sequences"
Other approaches are generally iterative, where the more iterations you do, the better results. You could look up "Mitchell's Best Candidate", for one. Another is "Poisson Disc Sampling".
There are innumerable variations on the different algorithms depending on what you want — some applications demand certain frequencies of noise, for instance. But if you just want something that "looks okay", I think the quasirandom one is a good starting point.
Another cheap and easy one is a "jittered grid", where you evenly space the points on your plane, then randomly adjust each one a small amount.
I have a data set of time series data I would like to display on a line graph. The data is currently stored in an oracle table and the data is sampled at 1 point / second. The question is how do I plot the data over a 6 month period of time? Is there a way to down sample the data once it has been returned from oracle (this can be done in various charts, but I don't want to move the data over the network)? For example, if a query returns 10K points, how can I down sample this to 1K points and still have the line graph and keep the visual characteristics (peaks/valley)of the 10K points?
I looked at apache commons but without know exactly what the statistical name for this is I'm a bit at a loss.
The data I am sampling is indeed time series data such as page hits.
It sounds like what you want is to segment the 10K data points into 1K buckets -- the value of each one of these buckets may be any statistic computation that makes sense for your data (sorry, without actual context it's hard to say) For example, if you want to spot the trend of the data, you might want to use Median Percentile to summarize the 10 points in each bucket. Apache Commons Math have helper functions for that. Then, with the 1K downsampled datapoints, you can plot the chart.
For example, if I have 10K data points of page load times, I might map that to 1K data points by doing a median on every 10 points -- that will tell me the most common load time within the range -- and point that. Or, maybe I can use Max to find the maximum load time in the period.
There are two options: you can do as #Adrian Pang suggests and use time bins, which means you have bins and hard boundaries between them. This is perfectly fine, and it's called downsampling if you're working with a time series.
You can also use a smooth bin definition by applying a sliding window average/function convolution to points. This will give you a time series at the same sampling rate as your original, but much smoother. Prominent examples are the sliding window average (mean/median of all points in the window, equally weighted average) and Gaussian convolution (weighted average where the weights come from a Gaussian density curve).
My advice is to average the values over shorter time intervals. Make the length of the shorter interval dependent on the overall time range. If the overall time range is short enough, just display the raw data. E.g.:
overall = 1 year: let subinterval = 1 day
overall = 1 month: let subinterval = 1 hour
overall = 1 day: let subinterval = 1 minute
overall = 1 hour: no averaging, just use raw data
You will have to make some choices about where to shift from one subinterval to another, e.g., for overall = 5 months, is subinterval = 1 day or 1 hour?
My advice is to make a simple scheme so that it is easy for others to comprehend. Remember that the purpose of the plot is to help someone else (not you) understand the data. A simple averaging scheme will help get you to that goal.
If all you need is reduce the points of your visuallization without losing any visuall information, I suggest to use the code here. The tricky part of this approach is to find the correct threshold. Where threshold is the amount of data point you target to have after the downsampling. The less the threshold the more visual information you lose. However from 10K to 1K, is feasible, since I have tried it with a similar amount of data.
As a side note you should have in mind
The quality of your visualization depends one the amount of points and the size (in pixels) of your charts. Meaning that for bigger charts you need more data.
Any further analysis many not return the corrected results if it is applied at the downsampled data. Or at least I haven't seen anyone prooving the opposite.
I wrote an object tracker that will try to detect and follow a moving object in a recorded video. In order to maximize the detection rate, my algorithm is using a bunch of detection & tracking algorithms (cascade, foreground & particle tracker). Each tracking algorithm will return me some point of interest that might be part of the object that I'm trying to track. Let's assume (for the simplicity of this example) that my object is a rectangle and that the three tracking algorithms returned the points 1, 2 and 3:
Based on the relation / distance of these three points it is possible to calculate the center of gravity (blue X in above image) of the tracked object. So for each frame I might be able to come up with some good estimate of the center of gravity. However, the object might move from one frame to the next:
In this example I merely rotated the original object. My algorithm will give me three new points of interest: 1',2' and 3'. I could again calculate the center of gravity based on these three new points, but I would throw away important information that I've acquired from the previous frame: based on points 1, 2 and 3 I already do know something about the relationship of these points and thus by combining the information from 1, 2 and 3 and 1',2' and 3' I should be able to come up with a better estimate of the center of gravity.
Furthermore, the next frame might yield a forth data point:
This is what I would like to do (but I don't know how):
based on the individual points (and their relationship to each other) that are returned from the different tracking algorithms, I want to build up a localization map of the tracked object. Intuitively I feel like I need to come up with A) an identification function that will identify individual points across frames and B) some cost function that will determine how similar tracked points (and the relationship / distance between them) are from frame to frame, but I can't get my head around on how to implement this. Alternatively, maybe some kind of map buildup based on the points will work. But again, I don't know how to approach this.
Any advice (and example code) is highly appreciated!
EDIT1
a simple particle filter might probably work too, but I again don't know how to define the cost function. A particle filter for tracking a certain color is easy to program: for each pixel you calculate the difference between target color and pixel color. But how would I do the same for estimating the relationship between tracked points?
EDIT2 intuitively I feel like Kalman filters could also help with the prediction step. See slides 24 - 32 of this pdf. Or am I misled?
What I think you're trying to do is essentially build up a state space of features, which can be applied to a filtering process, such as an Extended Kalman Filter. This is a useful framework when you have multiple observations in every frame, and you're trying to estimate or measure something indicated by these observations.
To determine the similarity of the tracked points, you can perform simple template matching from frame to frame for small regions around the points. One way of doing this is to extract an NxN (say, 7x7) region around point a in frame n and point a' in frame n+1, followed by normalised cross correlation between the extracted regions. This will give you a reasonable measure of how similar the patches are. If the patches are not similar, then you've probably lost track of that point.
There is an enormous literature on this and related problems starting in the 80's. Try searching for "optical flow" algorithms". The input for such algorithms is two successive frames of the same scene. The output is a vector field, one vector per pixel in the second image, which shows what the direction and speed of movement of the feature in that field. This presentation is a pretty nice summary.
A nice thing about optical flow is that many algorithms for it parallelize nicely and map onto your favorite video card GPU, so they can run in real time. Think ESPN overlays.
According to me, in order to identify who is who in each frame, you will have to use a greater dimension. For example if you want to know which point is where between two frame (considering your extracted point are same), you will have to build vectors or simplex and then deduce an organisation between your points (like angles values).
The main problem is that combinations increase with point number. If your camera is a fixed point then, you could use background as a reference in order to deduce object rotations and translations, i mean build vectors between background interest points and object points in order to clearly identify them.
hope that help go forward.
I would recommend looking in to the divided difference filter (DDF), which is similar to the extended Kalman filter (EKF), but does not require an approximate model of the dynamics of your system (which you may not have). Basically the DDF approximates the derivatives used in the EKF using a difference equation. There are plenty of papers online about this, but I do not know whether you have access to them so I have not linked them here. If you are working from a university or a company that has access to online journals (like IEEE Explore), then just Google "divided difference filter" and check out some of the papers.
I am implementing a project which needs to cluster geographical points. OPTICS algorithm seems to be a very nice solution. It needs just 2 parameters as input(MinPts and Epsilon), which are, respectively, the minimum number of points needed to consider them as a cluster, and the distance value used to compare if two points are in can be placed in same cluster.
My problem is that, due to the extreme variety of the points, I can't set a fixed epsilon.
Just look at the image below.
The same points structure but in a different scale would result very different. Suppose to set MinPts=2 and epsilon = 1Km.
On the left, the algorithm would create 2 clusters(red and blue), but on the right it would create one single cluster containing all of the points(red), but I would like to obtain 2 clusters even on the right.
So my question is: is there any kind of way to calculate dynamically the epsilon value to get this result?
EDIT 05 June 2012 3.15pm:
I thought I was using the OPTICS algorithm implementation from the javaml library, but it seems it is actually a DBSCAN algorithm implementation.
So the question now is: does anybody know a java based implementation of OPTICS algorithm?
Thank you very much and excuse my for my poor english.
Marco
The epsilon value in OPTICS is solely to limit the runtime complexity when using index structures. If you do not have an index for acceleration, you can set it to infinity.
To quote Wikipedia on OPTICS
The parameter \varepsilon is strictly speaking not necessary. It can be set to a maximum value. When a spatial index is available, it does however play a practical role when it comes to complexity.
What you seem to have looks much more like DBSCAN than OPTICS. In OPTICS, you should not need to choose epsilon (it should have been called max-epsilon by the authors!), but your cluster extraction method will take care of that. Are you using the Xi extraction proposed in the OPTICS paper?
minPts is much more important. You should try a value of at least 5 or 10, not 2. With 2, you are essentially performing single-linkage clustering!
The example you gave above should work fine once you increase minPts!
Re: edit: As you can even see in the Wikipedia article, ELKI has a proper OPTICS implementation and it's in Java.
You'd can try to scale epsilon by the total size of the enclosing rectangle. For example, your left data is about 4km x 6km (using my Mark I eyeball to measure) and the right is about 2km x 2km. So, epsilon on the right should be about 2.5 times smaller.
Of course, this doesn't work reliably. If, on your right hand data, there were an additional single point 4km to the right and 2km down, that would make the enclosing rectangle for the right the same as on the left, and you'd get similar (wrong) results.
You can try a minimum spanning tree and then remove the longest edge. The remaining spanning tree and the center of them is the best center for OPTICS and you can count the numbers of points around it.
In your explanation above, it is the change in scale which creates the uncertainty. When your scale gets bigger, your epsilon should change accordingly. Because they are at two very different scales, the two images you've presented are NOT the same set of points. They will not respond identically to your OPTICS algorithm without changing the parameters.
In short, no. there is no way to dynamically calculate epsilon to get this result. Clustering like this is already NP-Hard, and these clustering algorithims (optics, k-means, veroni) can only approximate the optimal solution.