how can I decouple the time cforest/ctree takes to construct a tree from the number of columns in the data?
I thought the option mtry could be used to do just that, i.e. the help says
number of input variables randomly sampled as candidates at each node for random forest like algorithms.
But while that does randomize the output trees it doesn't decouple the CPU time from the number of columns, e.g.
p<-proc.time()
ctree(gs.Fit~.,
data=Aspekte.Fit[,1:60],
controls=ctree_control(mincriterion=0,
maxdepth=2,
mtry=1))
proc.time()-p
takes twice as long as the same with Aspekte.Fit[,1:30] (btw. all variables are boolean). Why? Where does it scale with the number of columns?
As I see it the algorithm should:
At each node randomly select two columns.
Use them to split the response. (no scaling because of mincriterion=0)
Proceed to the next node (for a total of 3 due to maxdepth=2)
without being influenced by the column total.
Thx for pointing out the error of my ways
Related
My question is somewhat similar to this one. But I want to ask whether the column order matters or not. I have some time series data. For each cycle I computed some features (let's call them var1, var2,.... I now train the model using the following column order which of course will be consistent for the test set.
X_train=data['var1','var2','var3',var4']
After watching this video I've concluded that the order in which the columns appear is significant i.e. swapping var 1 and var 3 as:
X_train=data['var3','var2','var1',var4']
I would get a different loss function.
If the above is true, then how does one figure out the correct feature order to minimize the loss function, especially when the number of features could be in dozens.
In order to to ensure a good population representation I have created custom validation sets from my training data. However, I am not sure how I interface this in PCR in R
I have tried to add a list in the segments argument with each index similar to what you do in python predefined splits cv iterator, which runs but takes forever. So I feel I must be making an error somewhere
pcr(y~X,scale=FALSE,data=tdata,validation="CV",segments=test_fold)
where test fold is a list containing the validation set which belongs in the index
For example if the training data is composed on 9 samples and I want to use the first three as the first validation set on son
test_fold<-c(1,1,1,2,2,2,3,3,3)
This runs but it is very slow where if I do regular "CV" it runs in minutes. So far the results look okay but I have a over a thousand runs I need to do and it took 1 hr to get through one. So if anybody knows how I can speed this up I would be grateful.
So the segments parameters needs to be a list of multiple vectors. So going again with 9 samples if I want the first three to be in the first validation set, the next three in the second validation set and so on it should be
test_vec<-list(c(1,2,3),c(4,5,6),c(7,8,9))
Lets say I have a very big dataset (billions of records), one that doesnt fit on a single machine and I want to have multiple unknown queries (its a service where a user can choose a certain subset of the dataset and I need to return the max of that subset).
For the computation itself I was thinking about Spark or something similar, problem is Im going to have a lot of IO/network activity since Spark is going to have to keep re-reading the data set from the disk and distributing it to the workers, instead of, for instance, having Spark divide the data among the workers when the cluster goes up and then just ask from each worker to do the work on certain records (by their number, for example).
So, to the big data people here, what do you usually do? Just have Spark redo the read and distribution for every request?
If I want to do what I said above I have no choice but to write something of my own?
If the queries are known but the subsets unknown, you could precalculate the max (or whatever the operator) for many smaller windows / slices of the data. This gives you a small and easily queried index of sorts, which might allow you to calculate the max for an arbitrary subset. In case a subset does not start and end neatly where your slices do, you just need to process the ‘outermost’ partial slices to get the result.
If the queries are unknown, you might want to consider storing the data in a MPP database or use OLAP cubes (Kylin, Druid?) depending on the specifics; or you could store the data in a columnar format such as Parquet for efficient querying.
Here's a precalculating solution based on the problem description in the OP's comment to my other answer:
Million entries, each has 3k name->number pairs. Given a subset of the million entries and a subset of the names, you want the average for each name for all the entries in the subset. So each possible subset (of each possible size) of a million entries is too much to calculate and keep.
Precalculation
First, we split the data into smaller 'windows' (shards, pages, partitions).
Let's say each window contains around 10k rows with roughly 20k distinct names and 3k (name,value) pairs in each row (choosing the window size can affect performance, and you might be better off with smaller windows).
Assuming ~24 bytes per name and 2 bytes for the value, each window contains 10k*3k*(24+2 bytes) = 780 MB of data plus some overhead that we can ignore.
For each window, we precalculate the number of occurrences of each name, as well as the sum of the values for that name. With those two values we can calculate the average for a name over any set of windows as:
Average for name N = (sum of sums for N)/(sum of counts for N)
Here's a small example with much less data:
Window 1
{'aaa':20,'abcd':25,'bb':10,'caca':25,'ddddd':50,'bada':30}
{'aaa':12,'abcd':31,'bb':15,'caca':24,'ddddd':48,'bada':43}
Window 2
{'abcd':34,'bb':8,'caca':22,'ddddd':67,'bada':9,'rara':36}
{'aaa':21,'bb':11,'caca':25,'ddddd':56,'bada':17,'rara':22}
Window 3
{'caca':20,'ddddd':66,'bada':23,'rara':29,'tutu':4}
{'aaa':10,'abcd':30,'bb':8,'caca':42,'ddddd':38,'bada':19,'tutu':6}
The precalculated Window 1 'index' with sums and counts:
{'aaa':[32,2],'abcd':[56,2],'bb':[25,2],'caca':[49,2],'ddddd':[98,2],'bada':[73,2]}
This 'index' will contain around 20k distinct names and two values for each name, or 20k*(24+2+2 bytes) = 560 KB of data. That's one thousand times less than the data itself.
Querying
Now let's put this in action: given an input spanning 1 million rows, you'll need to load (1M/10k)=100 indices or 56 MB, which fits easily in memory on a single machine (heck, it would fit in memory on your smartphone).
But since you are aggregating the results, you can do even better; you don't even need to load all of the indices at once, you can load them one at a time, filter and sum the values, and discard the index before loading the next. That way you could do it with just a few megabytes of memory.
More importantly, the calculation should take no more than a few seconds for any set of windows and names. If the names are sorted alphabetically (another worthwhile pre-optimization) you get the best performance, but even with unsorted lists it should run more than fast enough.
Corner cases
The only thing left to do is handle the case where the input span doesn't line up exactly with the precalculated windows. This requires a little bit of logic for the two 'ends' of the input span, but it can be easily built into your code.
Say each window contains exactly one week of data, from Monday through Sunday, but your input specifies a period starting on a Wednesday. In that case you would have to load the actual raw data from Wednesday through Sunday of the first week (a few hundred megabytes as we noted above) to calculate the (count,sum) tuples for each name first, and then use the indices for the rest of the input span.
This does add some processing time to the calculation, but with an upper bound of 2*780MB it still fits very comfortably on a single machine.
At least that's how I would do it.
I'm trying to analyse multiple sequences with TraMineR at once. I've had a look at seqdef but I'm struggling to understand how I'd create a TraMineR dataset when I'm dealing with multiple variables. I guess I'm working with something similar to the dataset used by Aassve et al. (as mentioned in the tutorial), whereby each wave has information about several states (e.g. children, marriage, employment). All my variables are binary. Here's an example of a dataset with three waves (D,W2,W3) and three variables.
D<-data.frame(ID=c(1:4),A1=c(1,1,1,0),B1=c(0,1,0,1),C1=c(0,0,0,1))
W2<-data.frame(A2=c(0,1,1,0),B2=c(1,1,0,1),C2=c(0,1,0,1))
W3<-data.frame(A3=c(0,1,1,0),B3=c(1,1,0,1),C3=c(0,1,0,1))
L<-data.frame(D,W2,W3)
I may be wrong but the material I found deals with the data management and analysis of one variable at a time only (e.g. employment status across several waves). My dataset is much larger than the above so I can't really impute these manually as shown on page 48 of the tutorial. Has anyone dealt with this type of data using TraMineR (or similar package)?
1) How would you feed the data above to TraMineR?
2) How would you compute the substitution costs and then cluster them?
Many thanks
When using sequence analysis, we are interested in the evolution of one variable (for instance, a sequence of one variable across several waves). You have then multiple possibilities to analyze several variables:
Create on sequences per variable and then analyze the links between the cluster of sequences. In my opinion, this is the best way to go, if your variables measure different concepts (for instance, family and employment).
Create a new variable for each wave that is the interaction of the different variables of one wave using the interaction function. For instance, for wave one, use L$IntVar1 <- interaction(L$A1, L$B1, L$C1, drop=T) (use drop=T to remove unused combination of answers). And then analyze the sequence of this newly created variable. In my opinion, this is the prefered way if your variables are different dimensions of the same concept. For instance, marriage, children and union are all related to familly life.
Create one sequence object per variable and then use seqdistmc to compute the distance (multi-channel sequence analysis). This is equivalent to the previous method depending on how you will set substitution costs (see below).
If you use the second strategy, you could use the following substitution costs. You can count the differences between the original variable to set the substition costs. For instance, between states "Married, Child" and "Not married and Child", you could set the substitution to "1" because there is only a difference on the "marriage" variable. Similarly, you would set the substition cost between states "Married, Child" and "Not married and No Child" to "2" because all of your variables are different. Finally, you set the indel cost to half the maximum substitution cost. This is the strategy used by seqdistmc.
Hope this helps.
In Biemann and Datta (2013) they talk about multi dimensional analysis. That means creating multiple sequences for the same "individuals".
I used the following approach to do so:
1) define 3 dimensional sequences
comp.seq <- seqdef(comp,NULL,states=comp.scodes,labels=comp.labels, alphabet=comp.alphabet,missing="Z")
titles.seq <- seqdef(titles,NULL,states=titles.scodes,labels=titles.labels, alphabet=titles.alphabet,missing="Z")
member.seq <- seqdef(member,NULL,states=member.scodes,labels=member.labels, alphabet=member.alphabet,missing="Z")
2) Compute the multi channel (multi dimension) distance
mcdist <- seqdistmc(channels=list(comp.seq,member.seq,titles.seq),method="OM",sm=list("TRATE","TRATE","TRATE"),with.missing=TRUE)
3) cluster it with ward's method:
library(cluster)
clusterward<- agnes(mcdist,diss=TRUE,method="ward")
plot(clusterward,which.plots=2)
Nevermind the parameters like "missing" or "left" and etc. but i hope the brief code sample helps.
I am using randomForest package in R platform to build a binary classifier. There are about 30,000 rows with 14,000 being in positive class and 16,000 in negative class. I have 15 variables that have been known to be important for classification.
I have some additional variables (about 5) which have missing information. These variables have values 1 or 0. 1 means presence of something but 0 means that it is not known whether it is present or absent. It is widely known that these variables would be the most important variable for classification (increase reliability of classification and its more likely that the sample lies in positive class) if there is 1 but useless if there is 0. And, only 5% of the rows have value 1. So, one variable is useful for only 5% of the cases. The 5 variables are independent of each other, so I expect that these will be highly useful for 15-25% of the data I have.
Is there a way to make use of available data but neglect the missing/unknown data present in a single column? Your ideas and suggestions would be appreciated. The implementation does not have to be specific to random forest and R-platform. If this is possible using other machine learning techniques or in other platforms, they are also most welcome.
Thank you for your time.
Regards
I can see at least the following approaches. Personally, I prefer the third option.
1) Discard the extra columns
You can choose to discard those 5 extra columns. Obviously this is not optimal, but it is good to know the performance of this option, to compare with the following.
2) Use the data as it is
In this case, those 5 extra columns are left as they are. The definite presence (1) or unknown presence/absence (0) in each of those 5 columns is used as information. This is the same as saying "if I'm not sure whether something is present or absent, I'll treat it as absent". I know this is obvious, but if you haven't tried this, you should, to compare it to option 1.
3) Use separate classifiers
If around 95% of each of those 5 columns has zeroes, and they are roughly independent of each other, that's 0.95^5 = 77.38% of data (roughly 23200 rows) which has zeroes in ALL of those columns. You can train a classifier on those 23200 rows, removing the 5 columns which are all zeroes (since those columns are equal for all points, they have zero predictive utility anyway). You can then train a separate classifier for the remaining points, which will have at least one of those columns set to 1. For these points, you leave the data as it is.
Then, for your test point, if all those columns are zeroes you use the first classifier, otherwise you use the second.
Other tips
If the 15 "normal" variables are not binary, make sure you use a classifier which can handle variables with different normalizations. If you're not sure, normalize the 15 "normal" variables to lie in the interval [0,1] -- you probably won't lose anything by doing this.
I'd like to add a further suggestion to Herr Kapput's: if you use a probabilistic approach, you can treat "missing" as a value which you have a certain probability of observing, either globally or within each class (not sure which makes more sense). If it's missing, it has probability of occurring p(missing), and if it's present it has probability p(not missing) * p(val | not missing). This allows you to gracefully handle the case where the values have arbitrary range when they are present.