I have a dataset which contains a timestamp and a response time. I want to recognize current patterns in the dataset with labelled learning or something like that. I've been looking at multi-class classification but that only takes 1 input/row and outputs a class which it things it belongs to.

The algorithm should look at the surrounding previous data to check the correct pattern the newest input value is in.


A single input could be 45ms but you wouldn't know which pattern it belongs to unless you check all recent inputs. When it gives the class prediction a score of 0.75 or higher I can check the previous down time of the same class and give a prediction as to when the service is going down. Would be nice if 1 type of machine learning could do this all but I'm currently looking for a pattern classification algorithm.

Any help is greatly appreciated.

  • $\begingroup$ I would suggest you read about bidirectional-recurrent/recurrent networks. The idea is that they operate on time-series type data, so they take the context of the sequence into account. $\endgroup$
    – Araymer
    Commented Feb 14, 2017 at 16:48
  • $\begingroup$ Do you know good algorithms/github projects so I can try it out? $\endgroup$
    – Appernicus
    Commented Feb 15, 2017 at 7:50
  • $\begingroup$ Tensorflow has support for recurrent networks, if you're using python. There are implementations available if you google and examples in their official repo. If you're using JAVA, Deeplearning4J also has support for RNNs (but not bidirectional until the next release - current at time of writing: 0.7.2) All of this should be easily searchable. $\endgroup$
    – Araymer
    Commented Feb 15, 2017 at 16:35

1 Answer 1


In your dataset, the input is a sequence(time series), and Recurrent Neural Networks are the best choice for these kinds of the classification problem. Here is a good blog post on implementation of Recurrent Neural Networks using keras library(high level API for tensorflow and theano):



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.