Hi,
I am up for building a item based recommender using Mahout. I have
humongous amount of data in a Mongodb collection. But I am not sure that
the MongoDBDataModel provided with Mahout will be able to handle my case. I
see that in the buildModel() function, it creates a

> FastByIDMap<Collection<Preference>> userIDPrefMap = new
> FastByIDMap<Collection<Preference>>();
>
[line 556]
Does the subsequent code refer to creating an in-memory-model of the data
from the mongodb collection(which I think it does); if yes - is there any
current immediate alternative to that ?

Thanks
Mridul

Yep it's all in memory -- it would be too slow to access it out of Mongo.
The purpose is just making it easy to read and re-read data into Mongo, and
facilitate updates.

If the data is too big to fit in memory you should look first at pruning
your data -- can sampling 10% of it still give you good results?
If not, you are in Hadoop territory then and would want to look at a
distributed algorithm here.

On Sun, Mar 18, 2012 at 8:12 PM, Mridul Kapoor <[EMAIL PROTECTED]>wrote:

> Hi,
> I am up for building a item based recommender using Mahout. I have
> humongous amount of data in a Mongodb collection. But I am not sure that
> the MongoDBDataModel provided with Mahout will be able to handle my case. I
> see that in the buildModel() function, it creates a
>
> > FastByIDMap<Collection<Preference>> userIDPrefMap = new
> > FastByIDMap<Collection<Preference>>();
> >
> [line 556]
> Does the subsequent code refer to creating an in-memory-model of the data
> from the mongodb collection(which I think it does); if yes - is there any
> current immediate alternative to that ?
>
> Thanks
> Mridul
>

Mridul,

What is the humongous amount of data in Mongo?  Is it really item->item
links?  Or is it session information?

With a recommender, it is unusual to have more than a few hundred links to
other items for any given item.  This means that even for 10 million items,
you only have about a billion links in total and that can usually fit in
memory on a single machine pretty easily.  Recommenders with 10 million
items are pretty rare and can often be factored down by some content
characteristic.

So, are you sure your data is too large for memory?

On Sun, Mar 18, 2012 at 1:12 PM, Mridul Kapoor <[EMAIL PROTECTED]>wrote:

> Hi,
> I am up for building a item based recommender using Mahout. I have
> humongous amount of data in a Mongodb collection. But I am not sure that
> the MongoDBDataModel provided with Mahout will be able to handle my case. I
> see that in the buildModel() function, it creates a
>
> > FastByIDMap<Collection<Preference>> userIDPrefMap = new
> > FastByIDMap<Collection<Preference>>();
> >
> [line 556]
> Does the subsequent code refer to creating an in-memory-model of the data
> from the mongodb collection(which I think it does); if yes - is there any
> current immediate alternative to that ?
>
> Thanks
> Mridul
>

On 19 March 2012 02:24, Ted Dunning <[EMAIL PROTECTED]> wrote:

> Mridul,
>
> What is the humongous amount of data in Mongo?  Is it really item->item
> links?  Or is it session information?
>

It is session information. So this should anyway exceed, I suppose. And in
my case, I need to pre-compute the item-item similarities. These would be
refreshed periodically, maybe three times a week. I then need a recommender
to run as a web-service, which would read from these pre-computed
similarities(should be stored in some persistent place, right?). Could you
suggest the way forward here, then.

Thanks
Mridul

I've created a guide for scaling out a recommender system, maybe it is
useful for you:
http://ssc.io/deploying-a-massively-scalable-recommender-system-with-apache-mahout/
Am 19.03.2012 06:20 schrieb "Mridul Kapoor" <[EMAIL PROTECTED]>:

> On 19 March 2012 02:24, Ted Dunning <[EMAIL PROTECTED]> wrote:
>
> > Mridul,
> >
> > What is the humongous amount of data in Mongo?  Is it really item->item
> > links?  Or is it session information?
> >
>
> It is session information. So this should anyway exceed, I suppose. And in
> my case, I need to pre-compute the item-item similarities. These would be
> refreshed periodically, maybe three times a week. I then need a recommender
> to run as a web-service, which would read from these pre-computed
> similarities(should be stored in some persistent place, right?). Could you
> suggest the way forward here, then.
>
> Thanks
> Mridul
>

Session data never needs to be in memory.  It can be processed sequentially or using map reduce.

The item item data is all you need in memory.

Sent from my iPhone

On Mar 18, 2012, at 10:19 PM, Mridul Kapoor <[EMAIL PROTECTED]> wrote:

> On 19 March 2012 02:24, Ted Dunning <[EMAIL PROTECTED]> wrote:
>
>> Mridul,
>>
>> What is the humongous amount of data in Mongo?  Is it really item->item
>> links?  Or is it session information?
>>
>
> It is session information. So this should anyway exceed, I suppose. And in
> my case, I need to pre-compute the item-item similarities. These would be
> refreshed periodically, maybe three times a week. I then need a recommender
> to run as a web-service, which would read from these pre-computed
> similarities(should be stored in some persistent place, right?). Could you
> suggest the way forward here, then.
>
> Thanks
> Mridul

Thanks Sebastian,
That guide turned out useful for me. In accordance I have changed my
approach now. I have written a script to put the virtual sessions data in a
preferences.csv file, from mongodb.
Now, I need to pre-compute ItemSimilarities. I probably wont be using a
hadoop cluster. Is there a way that I run the ItemSimilarityJob on a single
machine ?

Thanks a lot!
Mridul

On Mon, Mar 19, 2012 at 10:06 PM, Mridul Kapoor <[EMAIL PROTECTED]>wrote:

> .... Is there a way that I run the ItemSimilarityJob on a single
> machine ?
>
>
Yes.  There is a sequential invocation as well.

On 20 March 2012 11:24, Ted Dunning <[EMAIL PROTECTED]> wrote:

> Yes.  There is a sequential invocation as well.
>

And pardon me for this, but I couldn't really get that, couldn't find it
through Mahout in Action as well. Could someone point me to the job/class I
should use to pre-compute item similarities -- in a way like *
org.apache.mahout.cf.taste.hadoop.similarity.item.ItemSimilarityJob*
but I really do not want to leverage hadoop right now -- and run it on a
single machine

Thanks
Mridul
 **

If you don't need Hadoop then this is pretty simple. You can just write a
nested loop that computes all pairs off an ItemSimilarity implementation.

If I recall rightly GenericItemSimilarity will do that for you off an
existing ItemSimilarity and then has the results in memory as a new
ItemSimilarity.

On Tue, Mar 20, 2012 at 6:25 AM, Mridul Kapoor <[EMAIL PROTECTED]>wrote:

> On 20 March 2012 11:24, Ted Dunning <[EMAIL PROTECTED]> wrote:
>
> > Yes.  There is a sequential invocation as well.
> >
>
> And pardon me for this, but I couldn't really get that, couldn't find it
> through Mahout in Action as well. Could someone point me to the job/class I
> should use to pre-compute item similarities -- in a way like *
> org.apache.mahout.cf.taste.hadoop.similarity.item.ItemSimilarityJob*
> but I really do not want to leverage hadoop right now -- and run it on a
> single machine
>
> Thanks
> Mridul
>  **
>

There's no magic involved in precomputing item similarities, just setup
a recommender, ask it for similar items and store them somewhere:

DataModel dataModel = ...
ItemSimilarity similarity = new CachingItemSimilarity(...);
ItemBasedRecommender recommender = new
    GenericItemBasedRecommender(dataModel, similarity);

LongPrimitiveIterator itemIDs = dataModel.getItemIDs();
while (itemIDs.hasNext()) {
  long itemID = itemIDs.nextLong();
  for (RecommendedItem similarItem :
      recommender.mostSimilarItems(itemID, ...)) {
    // save similar item to a file or a database
  }
}

With a little bit of engineering, this code can also run multithreaded.

Best,
Sebastian
On 20.03.2012 07:25, Mridul Kapoor wrote:
> On 20 March 2012 11:24, Ted Dunning <[EMAIL PROTECTED]> wrote:
>
>> Yes.  There is a sequential invocation as well.
>>
>
> And pardon me for this, but I couldn't really get that, couldn't find it
> through Mahout in Action as well. Could someone point me to the job/class I
> should use to pre-compute item similarities -- in a way like *
> org.apache.mahout.cf.taste.hadoop.similarity.item.ItemSimilarityJob*
> but I really do not want to leverage hadoop right now -- and run it on a
> single machine
>
> Thanks
> Mridul
>  **
>

On 20 March 2012 15:21, Sebastian Schelter <[EMAIL PROTECTED]> wrote:

> There's no magic involved in precomputing item similarities, just setup
> a recommender, ask it for similar items and store them somewhere:
>
> DataModel dataModel = ...
> ItemSimilarity similarity = new CachingItemSimilarity(...);
> ItemBasedRecommender recommender = new
>    GenericItemBasedRecommender(dataModel, similarity);
>
> LongPrimitiveIterator itemIDs = dataModel.getItemIDs();
> while (itemIDs.hasNext()) {
>  long itemID = itemIDs.nextLong();
>  for (RecommendedItem similarItem :
>      recommender.mostSimilarItems(itemID, ...)) {
>    // save similar item to a file or a database
>  }
> }
>
> With a little bit of engineering, this code can also run multithreaded.
>
> Best,
> Sebastian
Did a bit of exploring. Could I just do this instead. I mean this would
obviously just save to a file instead of a database which might instead be
better for larger data. But is the following ok ?

$ bin/mahout itemsimilarity --input ~/path/to/preferences.csv --output
~/path/to/output --similarityClassname SIMILARITY_LOGLIKELIHOOD
--maxSimilaritiesPerItem 10 --tempDir ~/path/to/temp

Thanks
Mridul

You can do this, but it will be awfully slow on a single box...

On 20.03.2012 10:56, Mridul Kapoor wrote:
> On 20 March 2012 15:21, Sebastian Schelter <[EMAIL PROTECTED]> wrote:
>
>> There's no magic involved in precomputing item similarities, just setup
>> a recommender, ask it for similar items and store them somewhere:
>>
>> DataModel dataModel = ...
>> ItemSimilarity similarity = new CachingItemSimilarity(...);
>> ItemBasedRecommender recommender = new
>>    GenericItemBasedRecommender(dataModel, similarity);
>>
>> LongPrimitiveIterator itemIDs = dataModel.getItemIDs();
>> while (itemIDs.hasNext()) {
>>  long itemID = itemIDs.nextLong();
>>  for (RecommendedItem similarItem :
>>      recommender.mostSimilarItems(itemID, ...)) {
>>    // save similar item to a file or a database
>>  }
>> }
>>
>> With a little bit of engineering, this code can also run multithreaded.
>>
>> Best,
>> Sebastian
>
>
> Did a bit of exploring. Could I just do this instead. I mean this would
> obviously just save to a file instead of a database which might instead be
> better for larger data. But is the following ok ?
>
> $ bin/mahout itemsimilarity --input ~/path/to/preferences.csv --output
> ~/path/to/output --similarityClassname SIMILARITY_LOGLIKELIHOOD
> --maxSimilaritiesPerItem 10 --tempDir ~/path/to/temp
>
> Thanks
> Mridul
>

Thanks a ton people!
Mridul

Hi
Thanks a lot Sebastian, Sean and Ted for your continuing help! Following
your examples, I tried to create my own implementation. I ran it over a
part of the dataset.

DataModel dataModel = new FileDataModel(new File("preferences.csv"));
>
>     LogLikelihoodSimilarity llls = new LogLikelihoodSimilarity(dataModel);
>
>     ItemSimilarity similarity = new CachingItemSimilarity(llls, dataModel);
>
>     ItemBasedRecommender recommender = new
>
>       GenericBooleanPrefItemBasedRecommender(dataModel, similarity);
>
>
>>     BufferedWriter f = new BufferedWriter(new
>> FileWriter("output-similarities.txt",true));
>
>
>>     LongPrimitiveIterator itemIDs = dataModel.getItemIDs();
>
>     while (itemIDs.hasNext()) {
>
>       long itemID = itemIDs.nextLong();
>
>       for (RecommendedItem similarItem :
>
>              recommender.mostSimilarItems(itemID, 10)) {
>
>         f.write(itemID + "\t" + similarItem.getID() + "\t" +
>> similarItem.getValue());
>
>       }// inner for loop ends
>
>     }// outer while loop ends
>
>     f.close();
>
>
My preferences.csv is of the form

> user, item

user, item
So I successfully ran this over a small part of my dataset.
>From whatever I have read about in the mahout api doc, in the FileDataModel
documentation :
[Class FileDataModel]

> This class will also look for update "delta" files in the same directory,
> with file names that start the same way (up to the first period). These
> files have the same format, and provide updated data that supersedes what
> is in the main data file.
Please correct me if I got this code/concept wrong here : in this case I
want to update output_similarities.txt with similarities calculated from
new update data. Can I have more and more datafiles with names like
preferences.csv.001, preferences.csv.002 and so on.. and then just run this
code again to update the output_similarity ? Will it totally recalculate
everything, or just use these update files to modify and add the
item-item-similarities.

Thanks
Mridul

Correction :
On 21 March 2012 15:57, Mridul Kapoor <[EMAIL PROTECTED]> wrote:

> Hi
> Thanks a lot Sebastian, Sean and Ted for your continuing help! Following
> your examples, I tried to create my own implementation.
>
> DataModel dataModel = new FileDataModel(new File("preferences.csv"));
>>
>>     LogLikelihoodSimilarity llls = new LogLikelihoodSimilarity(dataModel);
>>
>>     ItemSimilarity similarity = new CachingItemSimilarity(llls,
>>> dataModel);
>>
>>     ItemBasedRecommender recommender = new
>>
>>       GenericBooleanPrefItemBasedRecommender(dataModel, similarity);
>>
>>
>>>     BufferedWriter f = new BufferedWriter(new
>>> FileWriter("output-similarities.txt",true));
>>
>>
>>>     LongPrimitiveIterator itemIDs = dataModel.getItemIDs();
>>
>>     while (itemIDs.hasNext()) {
>>
>>       long itemID = itemIDs.nextLong();
>>
>>       for (RecommendedItem similarItem :
>>
>>              recommender.mostSimilarItems(itemID, 10)) {
>>
>>         f.write(itemID + "\t" + similarItem.getID() + "\t" +
>>> similarItem.getValue());
>>
>>       }// inner for loop ends
>>
>>     }// outer while loop ends
>>
>>     f.close();
>>
>>
> My preferences.csv is of the form
>
>> user, item
>
> user, item
>
>
> *Correction :*

> So I successfully ran this over a small part of my dataset.
>

I ran this using
$ bin/mahout itemsimilarity --
and now want to run in the way Sebastian suggested

>From whatever I have read about in the mahout api doc, in the FileDataModel
> documentation :
> [Class FileDataModel]
>
>> This class will also look for update "delta" files in the same directory,
>> with file names that start the same way (up to the first period). These
>> files have the same format, and provide updated data that supersedes what
>> is in the main data file.
>
>
> Please correct me if I got this code/concept wrong here : in this case I
> want to update output_similarities.txt with similarities calculated from
> new update data. Can I have more and more datafiles with names like
> preferences.csv.001, preferences.csv.002 and so on.. and then just run this
> code again to update the output_similarity ? Will it totally recalculate
> everything, or just use these update files to modify and add the
> item-item-similarities.
>
> Thanks
> Mridul
>

On 21 March 2012 16:01, Mridul Kapoor <[EMAIL PROTECTED]> wrote:

> Correction :
>
>
> On 21 March 2012 15:57, Mridul Kapoor <[EMAIL PROTECTED]> wrote:
>
>> Hi
>> Thanks a lot Sebastian, Sean and Ted for your continuing help! Following
>> your examples, I tried to create my own implementation.
>>
>> DataModel dataModel = new FileDataModel(new File("preferences.csv"));
>>>
>>>     LogLikelihoodSimilarity llls = new
>>>> LogLikelihoodSimilarity(dataModel);
>>>
>>>     ItemSimilarity similarity = new CachingItemSimilarity(llls,
>>>> dataModel);
>>>
>>>     ItemBasedRecommender recommender = new
>>>
>>>       GenericBooleanPrefItemBasedRecommender(dataModel, similarity);
>>>
>>>
>>>>     BufferedWriter f = new BufferedWriter(new
>>>> FileWriter("output-similarities.txt",true));
>>>
>>>
>>>>     LongPrimitiveIterator itemIDs = dataModel.getItemIDs();
>>>
>>>     while (itemIDs.hasNext()) {
>>>
>>>       long itemID = itemIDs.nextLong();
>>>
>>>       for (RecommendedItem similarItem :
>>>
>>>              recommender.mostSimilarItems(itemID, 10)) {
>>>
>>>         f.write(itemID + "\t" + similarItem.getID() + "\t" +
>>>> similarItem.getValue());
>>>
>>>       }// inner for loop ends
>>>
>>>     }// outer while loop ends
>>>
>>>     f.close();
>>>
>>>
>> My preferences.csv is of the form
>>
>>> user, item
>>
>> user, item
>>
>>
>> *Correction :*
>
>>  So I successfully ran this over a small part of my dataset.
>>
>
> I ran this using
> $ bin/mahout itemsimilarity --
> and now want to run in the way Sebastian suggested
>
> From whatever I have read about in the mahout api doc, in the
>> FileDataModel documentation :
>> [Class FileDataModel]
>>
>>> This class will also look for update "delta" files in the same
>>> directory, with file names that start the same way (up to the first
>>> period). These files have the same format, and provide updated data that
>>> supersedes what is in the main data file.
>>
>>
>> Please correct me if I got this code/concept wrong here : in this case I
>> want to update output_similarities.txt with similarities calculated from
>> new update data. Can I have more and more datafiles with names like
>> preferences.csv.001, preferences.csv.002 and so on.. and then just run this
>> code again to update the output_similarity ? Will it totally recalculate
>> everything, or just use these update files to modify and add the
>> item-item-similarities.
>>
>> Thanks
>> Mridul
>>
>
I think I found the answer to my questions. Mahout in Action to the rescue
! Really well written.
I guess I'll name the files preferences.01.csv preferences.02.csv and so
on. And whenever I am want to use these updates, I should run
recommender.refresh(null). That should pick up the data from these update
files. Correct me if I am wrong here.

Mridul