Speed of Time Series Classification Algorithms #6168
Replies: 2 comments 4 replies
-
|
@muriloasouza, that's an interesting question! We haven't done profiling or systematic experiments with datasets this large. I think this is an interesting development or research project, if you are interested in - just carrying out performance benchmarking for some of the more common algorithms. Further, if you have your own models - would you consider contributing them to |
Beta Was this translation helpful? Give feedback.
-
|
Hello @fkiraly, not really sure if this is the correct approach, but i tried this today: Got this as result: If this is the correct approach, maybe tomorrow i can try increasing the range in samples. Or another algorithm maybe. One question regarding Deep Learning models, that we specify the number of epochs to train. Do you have any idea how this comparation can be done? How many epochs one should set when comparing to other algorithms that does not use epochs to train? |
Beta Was this translation helpful? Give feedback.
-
|
Interesting! Btw, have you seen the The increase around 50.000 is probably when your computer or cluster runs out of memory, as you are using the memory-intensive Also, there are two other knn algorithms, from |
Beta Was this translation helpful? Give feedback.
-
|
FYI @srggrs, this may be of interest to you. |
Beta Was this translation helpful? Give feedback.
-
|
thx. At the moment I end up using the FB library FAISS, so far perform better and faster at searches and clustering using simple metrics (ie euclidean distance), though it does not have DTW and other complex distances |
Beta Was this translation helpful? Give feedback.
-
|
Interesting! Can we somehow leverage FAISS? |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Greetings,
I am working in a multiclass TSC problem (12 classes in total). My dataset contains around 4 millions of rows (4 millions of timeseries) and 48 columns (half-hourly measurements for a single day). Consider a train/test split of 0.7/0.3. And again, split that train part into train/validation with a 0.8/0.2 split.
How fast the TSC algorithms (Dictionary-based, Distance-based or Shapelet-based) would be to solve this problem? For example, currently i have developed a LSTM model from Keras library, that took around 1,5 hour to train/test.
Usually, are these TSC algorithms faster or slower than those traditional Deep Learning models we find in Keras or other DL libraries? If they are slower, any other TSC algorithm could be faster?
Beta Was this translation helpful? Give feedback.
All reactions