Here are 2 concerns I would have with the design (This discussion is mostly
to validate my own understanding)
1. if you have outliers "before" running k-means, aren't your centroids get
skewed? In other word, outliers by themselves may bias the cluster
evaluation, isn't it?
2. Typically microbatches are small, like 3 sec in your case. in this
window you may not have enough data to run any statistically sigficant
operation, can you?
My approach would have been: Run K-means on data without outliers (in batch
mode). Determine the model, ie centroids in case of kmeans. Then load the
model in your streaming app and just apply "outlier detection" function,
which takes the form of
def detectOutlier(model,data):
/// your code, like mean distance etc
return T or F
In response to your point about "alternet set of data", I would assume you
would accumulate the data you are receiving from streaming over few weeks
or months before running offline training.
Am I missing something?
On Sun, Nov 20, 2016 at 10:29 AM, Debasish Ghosh <ghosh.debas...@gmail.com>
wrote:
> Looking for alternative suggestions in case where we have 1 continuous
> stream of data. Offline training and online prediction can be one option if
> we can have an alternate set of data to train. But if it's one single
> stream you don't have separate sets for training or cross validation.
>
> So whatever data u get in each micro batch, train on them and u get the
> cluster centroids from the model. Then apply some heuristics like mean
> distance from centroid and detect outliers. So for every microbatch u get
> the outliers based on the model and u can control forgetfulness of the
> model through the decay factor that u specify for StramingKMeans.
>
> Suggestions ?
>
> regards.
>
> On Sun, 20 Nov 2016 at 3:51 AM, ayan guha <guha.a...@gmail.com> wrote:
>
>> Curious why do you want to train your models every 3 secs?
>> On 20 Nov 2016 06:25, "Debasish Ghosh" <ghosh.debas...@gmail.com> wrote:
>>
>> Thanks a lot for the response.
>>
>> Regarding the sampling part - yeah that's what I need to do if there's no
>> way of titrating the number of clusters online.
>>
>> I am using something like
>>
>> dstream.foreachRDD { rdd =>
>> if (rdd.count() > 0) { //.. logic
>> }
>> }
>>
>> Feels a little odd but if that's the idiom then I will stick to it.
>>
>> regards.
>>
>>
>>
>> On Sat, Nov 19, 2016 at 10:52 PM, Cody Koeninger <c...@koeninger.org>
>> wrote:
>>
>> So I haven't played around with streaming k means at all, but given
>> that no one responded to your message a couple of days ago, I'll say
>> what I can.
>>
>> 1. Can you not sample out some % of the stream for training?
>> 2. Can you run multiple streams at the same time with different values
>> for k and compare their performance?
>> 3. foreachRDD is fine in general, can't speak to the specifics.
>> 4. If you haven't done any transformations yet on a direct stream,
>> foreachRDD will give you a KafkaRDD. Checking if a KafkaRDD is empty
>> is very cheap, it's done on the driver only because the beginning and
>> ending offsets are known. So you should be able to skip empty
>> batches.
>>
>>
>>
>> On Sat, Nov 19, 2016 at 10:46 AM, debasishg <ghosh.debas...@gmail.com>
>> wrote:
>> > Hello -
>> >
>> > I am trying to implement an outlier detection application on streaming
>> data.
>> > I am a newbie to Spark and hence would like some advice on the
>> confusions
>> > that I have ..
>> >
>> > I am thinking of using StreamingKMeans - is this a good choice ? I have
>> one
>> > stream of data and I need an online algorithm. But here are some
>> questions
>> > that immediately come to my mind ..
>> >
>> > 1. I cannot do separate training, cross validation etc. Is this a good
>> idea
>> > to do training and prediction online ?
>> >
>> > 2. The data will be read from the stream coming from Kafka in
>> microbatches
>> > of (say) 3 seconds. I get a DStream on which I train and get the
>> clusters.
>> > How can I decide on the number of clusters ? Using StreamingKMeans is
>> there
>> > any way I can iterate on microbatches with different values of k to
>> find the
>> > optimal one ?
>> >
>> > 3. Even if I fix k, after training on every microbatch I get a DStream.
>> How
>> > can I compute things like clustering score on the DStream ?
>> > StreamingKMeansModel has a computeCost function but it takes an RDD. I
>> can
>> > use dstream.foreachRDD { // process RDD for the micro batch here } - is
>> this
>> > the idiomatic way ?
>> >
>> > 4. If I use dstream.foreachRDD { .. } and use functions like new
>> > StandardScaler().fit(rdd) to do feature normalization, then it works
>> when I
>> > have data in the stream. But when the microbatch is empty (say I don't
>> have
>> > data for some time), the fit method throws exception as it gets an empty
>> > collection. Things start working ok when data starts coming back to the
>> > stream. But is this the way to go ?
>> >
>> > any suggestion will be welcome ..
>> >
>> > regards.
>> >
>> >
>> >
>> > --
>> > View this message in context: http://apache-spark-user-list.
>> 1001560.n3.nabble.com/using-StreamingKMeans-tp28109.html
>> > Sent from the Apache Spark User List mailing list archive at Nabble.com.
>> >
>> > ---------------------------------------------------------------------
>> > To unsubscribe e-mail: user-unsubscr...@spark.apache.org
>> >
>>
>>
>>
>>
>> --
>> Debasish Ghosh
>> http://manning.com/ghosh2
>> http://manning.com/ghosh
>>
>> Twttr: @debasishg
>> Blog: http://debasishg.blogspot.com
>> Code: http://github.com/debasishg
>>
>> --
> Sent from my iPhone
>
--
Best Regards,
Ayan Guha
