** I have a "row-vector" of fixed length (29) where each value represents the direction of a "vector" between times 'ti' and 'ti+1' where i ranges from 1 to 29.
eg: a= [10 20 30 40 50];
Here 'a' is a row-vector of length 5 (just for the example, in my data length is 29) where each values 10,20,30,40,50 are values of angular direction of the vector at times 1-2,2-3,3-4,4-5,5-6 seconds respectively.
** Similarly, there are other such row-vectors, all of the same length (29).
** Now, my problem is: I want to group the row vectors with similar angular pattern over the whole length of time.
** Grouping should take into consideration the overall trend of the pattern (should avoid similarity measures like euclidean distances which misses the trend in the vector).
** If I am not wrong, this can be assumed to be a statistical time series analysis problem, which involves some kind of time series clustering techniques.
It would of great help if someone could suggest me an appropriate time series clustering technique or any other technique that would enable me to solve this problem.
Hi,
I suggest you to use HMM. It seems a powerful tool for your problem.
Thank you. I will look into it. But, is it possible to cluster the vectors based on the overall trend of the orientation data using HMM?
Hi,
Please can you explain what do you mean by:
(should avoid similarity measures like euclidean distances which misses the trend in the vector).
Hi Fares,
Here, I want to group the vectors having similar orientation pattern for a span of 1 secs to 30 secs.
So, If I opt for a similarity measure, for eg: the Euclidean distance and if two vectors are similar for more than half the length of time (for eg:) and there is some change in pattern (beyond the allowed threshold) between the two vectors for the rest of time, will the euclidean distance between the two vectors be distinguishable enough so that i can cluster the two vectors into different groups??
Basically, i want to group the "similar trend" in the orientation patterns, say, for a threshold of + or - 30degrees.
Thanks.
Hi,
theorically, the euclidean distance will take into account the difference between your vectors, since it's a root of the sum of squared differences.
It is true that each vector is seen as a time series, but if you want a comparison between several vectors, you don't need time series.
So, I suggest you to use an adaptation of the SVM method, in order to take into account your threshold.
Best regards.
- Badges
- Science topic
- Similar topics
- Vectorization