Each training example should have recent historical predictor variables and the future state.
I would generate training data relatively often. One way to do that is to take all points where you see instability and all minutes where you don't see near future instability. There will be a fair bit of repetition which you could decrease by looking only at a sampling of negative examples. Millions of data points would be excellent. On Mon, Jun 6, 2011 at 6:46 AM, Svetlomir Dimitrov Kasabov < [email protected]> wrote: > Thanks for the useful replies, I really appreciate that! > > @Ted and Hector: My initial parameters (predictors) are blood pressures, > heart rates, etc: they come every minute from a patient's monitor. > In my implementation, I plan refering to this Paper : > http://www.multi-science.co.uk/acce-free.pdf on page 7 (Table 1) you can > see the parameters used. On page 17, figure 4 you can see vizualization of > the prediction using time series: > > I think I still plan using the logistic regression implementation (since I > am already worked into it), but I am confuzed how to implement time series > with Mahout. Should I create periodically (for example every 15 minutes) a > new logistic regression model, in order to predict the probability of > instability? Then the amount of training data depends on the 'time window > for the past' that I will be using. For example, for data only two hours > from the past, I will have only circa 60 * 2 = 120 examples for creating a > temporal model (I assume that I will need one compound data vector pro > minute, including blood pressures, heart rates, etc...) > > Or should I implement the time series so, that I train the model only once > with old data of many patients and the training algorithm will be so, that > it checks what is the patient's hemodynamic stability in two hours (since > this information is also known during the training)? In this case, I will > potentually have many more examples (one million or more...) > > > Many thanks, best regards and sorry for the long post. > > Svetlomir. > > > > > Zitat von Ted Dunning <[email protected]>: > > What Hector said. >> >> You will need to extract features from your time history. >> >> The question also comes up about how large is your data set. If it is >> less >> than 100,000 training examples or so, then you will probably be better off >> using a system like R which handles that much data easily and has >> essentially every kind of classifier available for you to try. >> >> If you have 1 million training examples or more, then Mahout begins to >> dominate alternatives. Even there, Mahout is currently optimized for >> sparse >> data which is not what you have. My guess is that using the >> OnlineLogisticRegression or some of Hector's recent patches is the way to >> go. The AdaptiveLogisticRegression is heavily oriented around per term >> annealing and magic knob tuning in the context of sparse data. >> >> Can you post your data? >> >> On Sun, Jun 5, 2011 at 10:04 AM, Hector Yee <[email protected]> wrote: >> >> You can also try HMMs: >>> >>> >>> >>> https://builds.apache.org/job/Mahout-Quality/javadoc/org/apache/mahout/classifier/sequencelearning/hmm/package-tree.html >>> >>> If you want to do it with a classifier you can window your time series >>> and >>> make a training set >>> >>> e.g. >>> >>> label, feature >>> stable, (last X seconds of time series) >>> unstable, (last X seconds of time series) >>> >>> On Sun, Jun 5, 2011 at 8:08 AM, Svetlomir Dimitrov Kasabov < >>> [email protected]> wrote: >>> >>> > Hello, >>> > >>> > I plan using Apache Mahout's Logistic Regression (LR) implementation in >>> my >>> > Master-Thesis. We plan using time series in order to predict, whether a >>> > particular patient will have an instable blood flow soon or not. >>> Thats's >>> why >>> > I want to ask you if it is possible to use Mahout in connection with >>> time >>> > series ? Do you see any potential problems / risks ? >>> > >>> > Many thanks and best regards! >>> > >>> > Svetlomir Kasabov. >>> > >>> > >>> > >>> > -- >>> > Svetlomir Dimitrov Kasabov >>> > >>> > ---------------------------------------------------------------- >>> > This message was sent using IMP, the Internet Messaging Program. >>> > >>> > >>> >>> >>> -- >>> Yee Yang Li Hector >>> http://hectorgon.blogspot.com/ (tech + travel) >>> http://hectorgon.com (book reviews) >>> >>> >> > > > -- > Svetlomir Dimitrov Kasabov > > ---------------------------------------------------------------- > This message was sent using IMP, the Internet Messaging Program. > >
