OK, I still wonder whether it's not better to make one big model. The
usual assumption is that the user's identity isn't predictive per se.
If every customer in your shop is truly unlike the others, most
predictive analytics goes out the window. It's factors like our
location, income, etc that are predictive and there aren't a million
of those.
But let's say it's so and you really need 1M RDDs. I think I'd just
repeatedly filter the source RDD. That really won't be the slow step.
I think the right way to do it is to create a list of all user IDs on
the driver, turn it into a parallel collection (and override the # of
threads it uses on the driver to something reasonable) and map each
one to the result of filtering and modeling that user subset.
The problem is just the overhead of scheduling millions and millions
of tiny modeling jobs. It will still probably take a long time. Could
be fine if you have still millions of data points per user. It's even
appropriate. But then the challenge here is that you're processing
trillions of data points! that will be fun.
I think any distributed system is overkill and not designed for the
case where data fits into memory. You can always take a local
collection and call parallelize to make it into an RDD, so in that
sense Spark can handle a tiny data set if you really want.
I'm still not sure I've seen a case where you want to partition by
user but trust you really need that.
On Tue, Jan 13, 2015 at 1:30 AM, Josh Buffum <jbuf...@gmail.com> wrote:
> You are right... my code example doesn't work :)
>
> I actually do want a decision tree per user. So, for 1 million users, I want
> 1 million trees. We're training against time series data, so there are still
> quite a few data points per users. My previous message where I mentioned
> RDDs with no length was, I think, a result of the way the random
> partitioning worked (I was partitioning into N groups where N was the number
> of users... total).
>
> Given this, I'm thinking the mlllib is not designed for this particular
> case? It appears optimized for training across large datasets. I was just
> hoping to leverage it since creating my feature sets for the users was
> already in Spark.
>
>
> On Mon, Jan 12, 2015 at 5:05 PM, Sean Owen <so...@cloudera.com> wrote:
>>
>> A model partitioned by users?
>>
>> I mean that if you have a million users surely you don't mean to build a
>> million models. There would be little data per user right? Sounds like you
>> have 0 sometimes.
>>
>> You would typically be generalizing across users not examining them in
>> isolation. Models are built on thousands or millions of data points.
>>
>> I assumed you were subsetting for cross validation in which case we are
>> talking about making more like say 10 models. You usually take random
>> subsets. But it might be as fine to subset as a function of a user ID if you
>> like. Or maybe you do have some reason for segregating users and modeling
>> them differently (e.g. different geographies or something).
>>
>> Your code doesn't work as is since you are using RDDs inside RDDs. But I
>> am also not sure you should do what it looks like you are trying to do.
>>
>> On Jan 13, 2015 12:32 AM, "Josh Buffum" <jbuf...@gmail.com> wrote:
>>>
>>> Sean,
>>>
>>> Thanks for the response. Is there some subtle difference between one
>>> model partitioned by N users or N models per each 1 user? I think I'm
>>> missing something with your question.
>>>
>>> Looping through the RDD filtering one user at a time would certainly give
>>> me the response that I am hoping for (i.e a map of user => decisiontree),
>>> however, that seems like it would yield poor performance? The userIDs are
>>> not integers, so I either need to iterator through some in-memory array of
>>> them (could be quite large) or have some distributed lookup table. Neither
>>> seem great.
>>>
>>> I tried the random split thing. I wonder if I did something wrong there,
>>> but some of the splits got RDDs with 0 tuples and some got RDDs with > 1
>>> tuple. I guess that's to be expected with some random distribution? However,
>>> that won't work for me since it breaks the "one tree per user" thing. I
>>> guess I could randomly distribute user IDs and then do the "scan everything
>>> and filter" step...
>>>
>>> How bad of an idea is it to do:
>>>
>>> data.groupByKey.map( kvp => {
>>> val (key, data) = kvp
>>> val tree = DecisionTree.train( sc.makeRDD(data), ... )
>>> (key, tree)
>>> })
>>>
>>> Is there a way I could tell spark not to distribute the RDD created by
>>> sc.makeRDD(data) but just to deal with it on whatever spark worker is
>>> handling kvp? Does that question make sense?
>>>
>>> Thanks!
>>>
>>> Josh
>>>
>>> On Sun, Jan 11, 2015 at 4:12 AM, Sean Owen <so...@cloudera.com> wrote:
>>>>
>>>> You just mean you want to divide the data set into N subsets, and do
>>>> that dividing by user, not make one model per user right?
>>>>
>>>> I suppose you could filter the source RDD N times, and build a model
>>>> for each resulting subset. This can be parallelized on the driver. For
>>>> example let's say you divide into N subsets depending on the value of
>>>> the user ID modulo N:
>>>>
>>>> val N = ...
>>>> (0 until N).par.map(d => DecisionTree.train(data.filter(_.userID % N
>>>> == d), ...))
>>>>
>>>> data should be cache()-ed here of course.
>>>>
>>>> However it may be faster and more principled to take random subsets
>>>> directly:
>>>>
>>>> data.randomSplit(Array.fill(N)(1.0 / N)).par.map(subset =>
>>>> DecisionTree.train(subset, ...))
>>>>
>>>> On Sun, Jan 11, 2015 at 1:53 AM, Josh Buffum <jbuf...@gmail.com> wrote:
>>>> > I've got a data set of activity by user. For each user, I'd like to
>>>> > train a
>>>> > decision tree model. I currently have the feature creation step
>>>> > implemented
>>>> > in Spark and would naturally like to use mllib's decision tree model.
>>>> > However, it looks like the decision tree model expects the whole RDD
>>>> > and
>>>> > will train a single tree.
>>>> >
>>>> > Can I split the RDD by user (i.e. groupByKey) and then call the
>>>> > DecisionTree.trainClassifer in a reduce() or aggregate function to
>>>> > create a
>>>> > RDD[DecisionTreeModels]? Maybe train the model with an in-memory
>>>> > dataset
>>>> > instead of an RDD? Call sc.parallelize on the Iterable values in a
>>>> > groupBy
>>>> > to create a mini-RDD?
>>>> >
>>>> > Has anyone else tried something like this with success?
>>>> >
>>>> > Thanks!
>>>
>>>
>
---------------------------------------------------------------------
To unsubscribe, e-mail: user-unsubscr...@spark.apache.org
For additional commands, e-mail: user-h...@spark.apache.org
