I have been studying time series prediction recently. As a newbie of HTM, I have run the demo with my own dataset. HTM is very suitable for data stream, but it requires data to be continuous in time. however, in most cases, there are train dataset and test dataset. They are discontinuous in time. For example, train data came from January and test data came from October. And I want to train HTM network with train data, and then use HTM network(I have trained) to predict test data. Can HTM solve this problem at present?
Appreciate for your help!
Would the October data contain the same temporal patterns that existed in the January data? I’m guessing so since train data is typically similar to test data. If it is, then you shouldn’t have any problem running predictions on October data after training on January data. In this case, I believe you are describing a single break in the data stream (which HTM should recover from in the initial 2 or 3 timesteps).
Thanks for your reply!
January and October are just examples. The real situation is that I have a dataset with 30000 data points. The dimension of the dataset is 23. The latter half of the dataset is the train data and the first half is the test data. And I need to use train data to train HTM network and predict test data. But the problem is that the dimension of train dataset is 24(23-dimension dataset plus 1-dimension prediction target) and the dimension of the test dataset is 23(there is no test target, because the test target is what I need to predict). So when I run HTM with test dataset, we lack the ‘prediction field’. How do I solve this problem?
Probably could just chop off the prediction target from the train dataset (HTM doesn’t do error propagation like other AI technologies).
If I remove the prediction target from the dataset, how can I predict it ? HTM is similar to autoregression (y(t) = f(y(t-1), y(t-2), …)), but can HTM solve the problem of regression (y = w1 * x1 + w2 * x2)? Just like in my case, I want to use 23-dimension dataset to predict target (target is not included in 23-dimension dataset)
Is this a classification problem then? (not a temporal pattern recognition problem). HTM may not be suited for this particular use case.
It is not a classification problem.
Sorry, then others on the forum will need to help. I’m not clear on the use case. I’m more of a visually-oriented person than a math guy
Your reply still gives me a lot of help. Thank you very much.
You will not be successful unless you take 23 single scalar streams and feed them into 23 HTM models, each one running SP/TM and anomaly likelihood.
OR you can do some data science and find out where the valuable semantics are in the data so you can reduce your temporal data feed.