Time warping algorithm is used often in similarity measure of time series with different lengths.
The question is can we use this method for similarity measure of 2 feature vectors with non equal length??
For example suppose that we have features of some people and want to classify them. But the feature vectors is not equal. For some patients we have some feature such as age, height, weight , ... but for some others we dont have some of those features.
Can dynamic time warping be used here??
Thanks all...
in this case no, because you cannot compare apple with orange, by doing DTW on different features, the values of different features will be compared yielding unpredictable or invalid results.
in your case, you may compare the mutual occurring features only, leaving other features without comparison.
for example if we have the 2 FV A and B:
A:height=80,length=70,weight=90,sinlgle=0
B:length=80,weight=60,sinlgle=0,education=1
the Manhattan distance will be |80-70|+|60-90|+|0-0|=40
you may also penalize the absent features, in this case the Manhattan distance will be
|80-0|+|80-70|+|60-90|+|0-0|+|1-0|=121
The heart of this question isn't whether or not DWT is feasible -- reading between the lines, you're actually asking how to impute values!
Search for methods to impute missing values, and you'll find answers for getting the missing cells of data when dealing with fixed-width feature vectors :D
You can try some robust classification algorithms that can be trained with and predict feature vectors with missing values.
Dynamic time warping requires having a metric to compare two vectors. From the moment you have a metric (Euclidian, Manhattan, or p-norm) you can certainly apply DTW.
The real question is whether you can define a metric that is robust against missing values.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series