I want to classify some time series events. Due to the occurrence of an event, particular parameter continuously changes w.r.t time. This changing trend of the parameter is different for different events. But this change is very small. How can I classify these events? Is this a time series classification problem? Does transforming the parameter data to higher dimension is of any use? If yes, how to do this and if not then what are the other appropriate techniques to do this.
Please provide relevant papers or documents regarding this.
Thanks in advance.
One technique I believe is through an in-depth, interpretative qualitative analysis.
There are many options to classify time series events. If you have labeled training data, you can treat the problem as a classification problem. To be more specific, you can label time points with respect to event they are in, then train a classification model to learn what is related with each event. This process involves in deciding which features to use in the classification model. Observed time series values can be used as input to the learning algorithm. You can also extract certain features (such as slopes, variances over time series intervals).
The other option you can consider is to do some sort event detection using anomaly detection techniques. Under the assumption that each event will show a different distribution over time, you can try to model the probability distribution of event over time and each event potentially has different distribution parameters. However this is harder compared to first option as it is unsupervised.
As mentioned earlier, there are many possibilities for time series classification and some comprehensive reviews are:
http://www.sciencedirect.com/science/article/pii/S0031320305001305
http://research.ijcaonline.org/volume52/number15/pxc3881278.pdf
As the other response indicate, the answer is "it depends". Mostly it depends on the data and the events you are actually looking at. Nonetheless, there are generic tools (again, most of the earlier responders give good suggestions). For the particular question in your title though - there is a well established method to convert a low-dimensional (scalar) time series into a higher dimensional one. Google "Takens Theorem" - the underlying premise is that you are looking at a deterministic and stationary dynamical system and the theorem provides techniques to reconstruct that.
Again, depending on what you are actually trying to do, machine learning or statistical approaches could be good too...
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series