Open pierretoussing opened 4 years ago
Combining: In this paper, they tried to make time series classification using several similarity measures. They fuse the different similarity measures like this:
- For all the pairs of time series in the train data, we calculate the similarity values using all the considered elementary similarity measures.
- In some of the above pairs, both time series belong to the same class, in others they belong to different classes. We define the indicator I(t1, t2) of a pair of time series (t1, t2), as follows: I(t1, t2) = 0 if t1 and t2 belong to the same class, I(t1, t2) = 1 otherwise. 2 Note that we do not assume the elementary similarity measures to fulfill specific properties (such as triangular inequality). Fusion of Similarity Measures for Time Series Classification 5
- We train a regression model M. We use the similarity values (see first step) as training data along with the corresponding indicators as labels.
- We propose to use the output of M as the fused similarity measure. For a pair of time series (t0 , t), where either or both of them can be unlabeled (test) time series, we calculate the similarity values using all the considered elementary similarity measures. Then use M to predict (based on these similarity values) the likelihood that t and t0 belong to the different classes. Finally, we use this prediction as the distance of t and t0.
Combining: In this paper, they use the following approach:
Compute the similarity scores for the different similarity measures
Log-transform the values
Combine the transformed values using a linear regression classifier from the WEKA toolkit (Hall etal., 2009)
Combining: In this paper, they train a classifier using the weighted sum of the similarity measures values. The weights are randomly initiallized and then adapted over time in order to increase accuracy.
Since we are not basing on any approach, it would be interesting to find other approaches for the same problems. Look for works handling one of the following issues.
[ ] Making similarity measures comparable
[x] Combining similarity measures
[ ] Define agreeableness between similarity measures