Re: Why each time the classification model trained by using TrainNewsGroup are not the same?
Hi, Ted Dunning I comment the line Collection.shuffle(files); in TrainNewsGroups.java, let the model trained with same order of examples each time. After recompile the code and redo the experiment, the model are still not the same each time. :( And I make sure the vectors before input to train are the same value and order each time. How do I fixed the order of examples?? Thanks and regards 於 西元2012年07月03日 23:19, Ted Dunning 提到: Because the order of the examples is randomized. On Tue, Jul 3, 2012 at 8:13 AM, Caspar Hsieh caspar.hs...@9x9.tv wrote: I use Mahout classification example TrainNewsGroup to train the model with leak type 3, then use TestNewsGroups to test the model, then I re-trained the model and test again, the test result are not the same before. why each time the model trained are not the same? Thanks.
Re: Approaches for combining multiple types of item data for user-user similarity
The best default answer is to put them all in one model. The math doesn't care what the things are. Unless you have a strong reason to weight one data set I wouldn't. If you do, then two models is best. It is hard to weight a subset of the data within most similarity functions. I don't think it would in Pearson for instance but could work in Tanimoto. On Wed, Jul 4, 2012 at 1:20 AM, Ken Krugler kkrugler_li...@transpac.com wrote: Hi all, I'm curious what approaches are recommended for generating user-user similarity, when I've got two (or more) distinct types of item data, both of which are fairly large. E.g. let's say I had a set of users where I knew both (a) what books they had bought on Amazon, and (b) what YouTube videos they had watched. For each user, I want to find the 10 most similar other users. - I could create two separate models, find the nearest 30 users for each user, and combine (maybe with weighting) the results. - I could toss all of the data into one model - and I could use a value of 1.0 for whichever type of preference is less important. Any other suggestions? Input on the above two approaches? Thanks! -- Ken -- Ken Krugler http://www.scaleunlimited.com custom big data solutions training Hadoop, Cascading, Mahout Solr
Re: Generating similarity file(s) for item recommender?
Hi Sean, Myrrix does look interesting! I'll keep an eye on it. What I'd like to do is recommend items to users yes. I looked at the IdRescorer and it did the job perfectly (pre filtering). I was a little misleading in regard to the size of the data. The raw data files are around 1GB. But after the interesting data is extracted -- session-id, item-id and type-of-event (product image clicked, product description viewed etc.), the data file comes out to about 10MB. Not so bad. Btw, just bought the Mahout in Action book! - Matt On Tue, Jul 3, 2012 at 10:40 AM, Sean Owen sro...@gmail.com wrote: I'm not sure if Mridul's suggestion does what you want. Do you want to recommend items to users? then no, you do not start with item IDs and recommend to them. It sounds like your question is how to compute similarity data. The first answer is that you do not use Hadoop unless you must use Hadoop. You don't compute it yourself, you let the framework do it with LogLikelihoodSimilarity. It just happens automatically. You can use caching, you can use precomputation, but that comes after you decide that you have too much data to do it all in real-time. 1GB of input data suggests you have a lot of data. Is that tens of millions of user-item associations? then yes you are not in simple non-Hadoop land anymore and you need to look at RecommenderJob / Hadoop. This doesn't have anything to do with FileDataModel or the non-distributed bits. To your second point -- this is really what Rescorer does for you, lets you filter or boost certain results at query time. But this is part of the non-distributed code. You could try stitching together some offline similarities from the Hadoop job, and loading them selectively in memory as part of the real-time Recommender, but it's going to be a bit dicey to get it to work fast. I don't mind mentioning that this is exactly the kind of problem I'm working on in Myrrix (myrrix.com). It does the offline model building on Hadoop and still lets you do real-time recommendations, with Rescorer objects if you want. The whole point is to fix up this dicey hard part mentioned above. Might take a look. On Tue, Jul 3, 2012 at 3:15 PM, Matt Mitchell goodie...@gmail.com wrote: Thanks Mridul, I'll try this out. Does getItemIDs return every item id from the file in your example? This kind of leads me to another, related question... I want to have my recommender engine recommend items to a user, but the items should be from a known set of item ids. For example, if a user is doing a search for gaming system, I only want recommendations for gaming system items. I was thinking I could feed the recommendation engine a set of item IDs that are known to be gaming systems as a candidate set *when executing that actual recommendation*. Does this make sense? If so, do you know how I can do this? I basically want to constrain the recommendations to a set of known item IDs at recommendation time. Thanks again! - Matt On Tue, Jul 3, 2012 at 8:01 AM, Mridul Kapoor mridulkap...@gmail.com wrote: I'm thinking the session ID (in the cookie) would be used as the user ID. The events are tied to product IDs, so these would be used in generating the preferences. I guess if you consider product-preference on a per session-basis (i.e. only items for which a user expresses preference for, in a single session, are similar to each other, in some way or the other). This way, you would be considering the session-ids as dummy user-ids, which I think should be good. I'd like to eventually run this on Hadoop, but it'd also be nice to know if there is a way to do this locally, while developing the app, maybe using a smaller dataset? Yes just writing a small offline recommender (made to run on a local machine) should do; you could take a subset of the data, use a FileDataModel, then do something like LongPrimitiveIterator itemIDs = dataModel.getItemIDs(); and iterate over these; getting _n_ recommended items for each, storing them somewhere (and maybe use this evaluating the recommender somehow) Best, Mridul
recommendations for new users
Hi, Slowly prototyping a recommendation here. The system does not have user accounts. Since the users on the system don't have accounts, I'm struggling a bit with completely new users, and what to recommend them. I do have information about the user, like what referring site they came from (1 of n partner sites), the city they want to shop in, and rating value of the product. I wonder if I could use this information, to find the most similar existing user. Then use that most similar user to generate recommendations? Anyone have tips for dealing this this? I'm not sure if Mahout supports finding most similar user based on user attributes, if not, this should be a simple sql/where-like select from the database. - Matt
Re: Generating similarity file(s) for item recommender?
If your input is 10MB then the good news is you are not near the scale where you need Hadoop. A simple non-distributed Mahout recommender works well, and includes the Rescorer capability you need. That's a fine place to start. The book ought to give a pretty good tour of how that works in chapter 2-5 yes. Separately we can talk offline about Myrrix as needed. Sean On Wed, Jul 4, 2012 at 4:50 PM, Matt Mitchell goodie...@gmail.com wrote: Hi Sean, Myrrix does look interesting! I'll keep an eye on it. What I'd like to do is recommend items to users yes. I looked at the IdRescorer and it did the job perfectly (pre filtering). I was a little misleading in regard to the size of the data. The raw data files are around 1GB. But after the interesting data is extracted -- session-id, item-id and type-of-event (product image clicked, product description viewed etc.), the data file comes out to about 10MB. Not so bad. Btw, just bought the Mahout in Action book! - Matt
Re: recommendations for new users
Have a look at the PlusAnonymousUserDataModel, which is a bit of a hack but a decent sort of solution for this case. It lets you temporarily add a user to the system and then everything else works as normal, so you can make recommendations to these new / temp users. There isn't a way to inject anything but rating/pref information directly, no. You can use this info in a Rescorer to influence recommendations; this is not specific to the case of a new user. You can also decide to make recommendations by a completely different means for new users -- for example, some canned list of top-10 recs that is appropriate for their city or referring site. That's legitimate too in practice. Yes you can also find most-similar users based on this info. You'd have to write the similarity metric yourself. I assume this is also not the metric you use in your real recommender. So maybe you could use it to find the nearest 1 real user and sub in those recommendations? or a neighborhood. You would have to rewrite a bit of what a recommender does to go this way but it's not so hard. No there is no content-based similarity metric; this is so domain specific. On Wed, Jul 4, 2012 at 4:54 PM, Matt Mitchell goodie...@gmail.com wrote: Hi, Slowly prototyping a recommendation here. The system does not have user accounts. Since the users on the system don't have accounts, I'm struggling a bit with completely new users, and what to recommend them. I do have information about the user, like what referring site they came from (1 of n partner sites), the city they want to shop in, and rating value of the product. I wonder if I could use this information, to find the most similar existing user. Then use that most similar user to generate recommendations? Anyone have tips for dealing this this? I'm not sure if Mahout supports finding most similar user based on user attributes, if not, this should be a simple sql/where-like select from the database. - Matt
Re: Generating similarity file(s) for item recommender?
Thanks Sean! Nice to know I can stay simple for now. - Matt On Wed, Jul 4, 2012 at 9:59 AM, Sean Owen sro...@gmail.com wrote: If your input is 10MB then the good news is you are not near the scale where you need Hadoop. A simple non-distributed Mahout recommender works well, and includes the Rescorer capability you need. That's a fine place to start. The book ought to give a pretty good tour of how that works in chapter 2-5 yes. Separately we can talk offline about Myrrix as needed. Sean On Wed, Jul 4, 2012 at 4:50 PM, Matt Mitchell goodie...@gmail.com wrote: Hi Sean, Myrrix does look interesting! I'll keep an eye on it. What I'd like to do is recommend items to users yes. I looked at the IdRescorer and it did the job perfectly (pre filtering). I was a little misleading in regard to the size of the data. The raw data files are around 1GB. But after the interesting data is extracted -- session-id, item-id and type-of-event (product image clicked, product description viewed etc.), the data file comes out to about 10MB. Not so bad. Btw, just bought the Mahout in Action book! - Matt
A bunch of SVD questions...
Hi, I'm exploring Mahout SVD parallel implementation over Hadoop (ALS), and I would like to clarify a few things : 1. How do you recommend top K items with this job? Does the job factorize the ranking matrix, than compute a predicted ranking for each cell in the matrix, so when you need a recommendation you only need to retrieve the top K items according to prediction value for the user? Or is it factorize the matrix and require some online logic when the recommendation is being asked? 2. From my knowledge, applying a SVD technique require first to fill in all empty cells in the ranking matrix (with average ranking for example). Is it something done during the ALS job (and if so, what is the way it's being filled), or should it be done as a preprocessing step? 3. From my understanding SVD recommenders are used to predict user implicit preference. By doing so you can recommend top K items (top K items over descending orders according to the prediction). I wonder, could it be applied on a binary dataset (explicit), where my ranking matrix contain only 1\0? 4. From doing some readings I found that the timeSVD++ developed by Yehuda Koren is considered as the superior SVD implementation for SVD recommenders. I wondered if there is any kind of a parallel implementation of it on top of Hadoop? I found this proposal: https://issues.apache.org/jira/browse/MAHOUT-371 I wonder, what is the status of it? Was it being checked already? Is it stable? Did anyone experienced with it? Thanks, Oren - Intel Electronics Ltd. This e-mail and any attachments may contain confidential material for the sole use of the intended recipient(s). Any review or distribution by others is strictly prohibited. If you are not the intended recipient, please contact the sender and delete all copies.
Re: Why each time the classification model trained by using TrainNewsGroup are not the same?
On Wed, Jul 4, 2012 at 12:09 AM, Caspar Hsieh caspar.hs...@9x9.tv wrote: Hi, Ted Dunning I comment the line Collection.shuffle(files); in TrainNewsGroups.java, let the model trained with same order of examples each time. This will prevent effective learning. You must shuffle the data at least once. After recompile the code and redo the experiment, the model are still not the same each time. :( There is also non-determinism in the AdaptiveLogisticRegression in the CrossFoldLearner. And I make sure the vectors before input to train are the same value and order each time. How do I fixed the order of examples?? One thing that might help is to tell RandomUtils that you are running a test. This will make the random number generators that Mahout controls become deterministic. You can add a random number generator argument to the shuffle call to get determinism there as well.
custom file data model?
Hi, I'd like to store additional information in my user preference data files. Is it possible, to add more columns to the file that FileDataModel uses? For example, an additional ID that maps to my applications database ID for item-ids, a simple 3 char code for possible use in custom user-user similarity etc.. Is this possible? Thanks! - Matt
Re: custom file data model?
Sure. It will ignore columns beyond the fourth, which is an optional timestamp. If you just want it to read some common input file but ignore the unused columns, that's easy. You can copy and modify FileDataModel to do whatever you like, if you want it to use this data. You'd have to change other code to use your new data somehow. On Wed, Jul 4, 2012 at 6:52 PM, Matt Mitchell goodie...@gmail.com wrote: Hi, I'd like to store additional information in my user preference data files. Is it possible, to add more columns to the file that FileDataModel uses? For example, an additional ID that maps to my applications database ID for item-ids, a simple 3 char code for possible use in custom user-user similarity etc.. Is this possible? Thanks! - Matt
Re: Extracting document/topic inference with the new lda cvb algorithm
Hi Caroline, Jake Mannix and I wrote the LDA CVB implementation. Apologies for the light documentation. When you invoked Mahout, did you supply the --doc_topic_output path parameter? If this is present, after training a model the driver app will apply the model to the input term-vectors, storing inference results in the specified path. If the parameter isn't specified, this final inference run is skipped: https://github.com/apache/mahout/blob/trunk/core/src/main/java/org/apache/mahout/clustering/lda/cvb/CVB0Driver.java#L74 https://github.com/apache/mahout/blob/trunk/core/src/main/java/org/apache/mahout/clustering/lda/cvb/CVB0Driver.java#L331 So, assuming you did generate inference output, I should note that both the model and inference output have the *same* format: Both the topic-term matrix and doc-topic inference output are stored as SequenceFileIntWritable, VectorWritable data. If you point the vectordump util at either data set and supply a dictionary, it'll happily map term ids or topic ids into term strings using that dictionary... Quite confusing. Just make sure that when you run vectordump against the doc-topic data that you don't supply the dictionary-- This way, you'll see the raw topic ids (zero-based indices) in output, instead of whatever terms those indices might correspond to in your dictionary. Best, Andy @sagemintblue On Wed, Jul 4, 2012 at 2:30 AM, Caroline Meyer caromeye...@gmail.comwrote: Hey Guys, I have been able to successfully execute the new lda algorithm as well as extract the topic/term inference with vectordump. What I was not able to do was get the document/topic inference. When I run the same vectordump command I get the same kinds of vectors (term:probability) as before. Should the vectors not be (topic:probability)? The command I run is: vectordump -s temp/lda-cvb-doc/part-m-0 -d temp/vectors/dictionary.file-* -dt sequencefile -o temp/lda-cvb-topics.txt I have not been able to find any documentation except what's in the code. Thanks for the help. Cheers, Caroline
Re: Extracting document/topic inference with the new lda cvb algorithm
I haven't looked into the vector dumper code in detail, but I remember having successfully run some version of it without an input dictionary. Perhaps you've stumbled into a legitimate bug with the utility? For the time being you might also try the sequence file dumper util which is somewhat more generic but may suit your purpose here. Andy On Wed, Jul 4, 2012 at 9:42 AM, Caroline Meyer caromeye...@gmail.comwrote: Hi Andy If I only use the -s and -o options I get this null pointer exception: Exception in thread main java.lang.NullPointerException at org.apache.mahout.utils.vectors.VectorHelper$1.apply(VectorHelper.java:118) at org.apache.mahout.utils.vectors.VectorHelper$1.apply(VectorHelper.java:115) at com.google.common.collect.Iterators$8.next(Iterators.java:765) at java.util.AbstractCollection.toArray(AbstractCollection.java:124) at java.util.ArrayList.init(ArrayList.java:131) at com.google.common.collect.Lists.newArrayList(Lists.java:119) at org.apache.mahout.utils.vectors.VectorHelper.toWeightedTerms(VectorHelper.java:114) at org.apache.mahout.utils.vectors.VectorHelper.vectorToJson(VectorHelper.java:124) at org.apache.mahout.utils.vectors.VectorDumper.main(VectorDumper.java:241) In the code it looks like it is looking for a dictionary that is not specified. Is there another option i am missing? Cheers, Caroline On Wed, Jul 4, 2012 at 6:10 PM, Andy Schlaikjer andrew.schlaik...@gmail.com wrote: Hi Caroline, Jake Mannix and I wrote the LDA CVB implementation. Apologies for the light documentation. When you invoked Mahout, did you supply the --doc_topic_output path parameter? If this is present, after training a model the driver app will apply the model to the input term-vectors, storing inference results in the specified path. If the parameter isn't specified, this final inference run is skipped: https://github.com/apache/mahout/blob/trunk/core/src/main/java/org/apache/mahout/clustering/lda/cvb/CVB0Driver.java#L74 https://github.com/apache/mahout/blob/trunk/core/src/main/java/org/apache/mahout/clustering/lda/cvb/CVB0Driver.java#L331 So, assuming you did generate inference output, I should note that both the model and inference output have the *same* format: Both the topic-term matrix and doc-topic inference output are stored as SequenceFileIntWritable, VectorWritable data. If you point the vectordump util at either data set and supply a dictionary, it'll happily map term ids or topic ids into term strings using that dictionary... Quite confusing. Just make sure that when you run vectordump against the doc-topic data that you don't supply the dictionary-- This way, you'll see the raw topic ids (zero-based indices) in output, instead of whatever terms those indices might correspond to in your dictionary. Best, Andy @sagemintblue On Wed, Jul 4, 2012 at 2:30 AM, Caroline Meyer caromeye...@gmail.com wrote: Hey Guys, I have been able to successfully execute the new lda algorithm as well as extract the topic/term inference with vectordump. What I was not able to do was get the document/topic inference. When I run the same vectordump command I get the same kinds of vectors (term:probability) as before. Should the vectors not be (topic:probability)? The command I run is: vectordump -s temp/lda-cvb-doc/part-m-0 -d temp/vectors/dictionary.file-* -dt sequencefile -o temp/lda-cvb-topics.txt I have not been able to find any documentation except what's in the code. Thanks for the help. Cheers, Caroline
Re: custom file data model?
Thanks Sean. I'll have a look at creating a custom model! A somewhat related question here... I've also thought about using a separate database for user prefs, either riak or amazons dynamo db. Any tips on how to create a custom data source? - Matt On Jul 4, 2012, at 11:55 AM, Sean Owen sro...@gmail.com wrote: Sure. It will ignore columns beyond the fourth, which is an optional timestamp. If you just want it to read some common input file but ignore the unused columns, that's easy. You can copy and modify FileDataModel to do whatever you like, if you want it to use this data. You'd have to change other code to use your new data somehow. On Wed, Jul 4, 2012 at 6:52 PM, Matt Mitchell goodie...@gmail.com wrote: Hi, I'd like to store additional information in my user preference data files. Is it possible, to add more columns to the file that FileDataModel uses? For example, an additional ID that maps to my applications database ID for item-ids, a simple 3 char code for possible use in custom user-user similarity etc.. Is this possible? Thanks! - Matt
Re: custom file data model?
Look at the example DataModels in integration. The pattern is the same: load it all into memory! it's too slow for real-time otherwise. So there is no point in say moving your data from a DB to Dynamo for scalability if you're using non-distributed code. If you're using Hadoop, DataModel is not relevant. On Wed, Jul 4, 2012 at 9:09 PM, Matt Mitchell goodie...@gmail.com wrote: Thanks Sean. I'll have a look at creating a custom model! A somewhat related question here... I've also thought about using a separate database for user prefs, either riak or amazons dynamo db. Any tips on how to create a custom data source?
Difference when we don't use partial implementation
Hi All, I am running Decision forest in Mahout, below are the commands that I have used to implement the algo: Info file: mahout org.apache.mahout.df.tools.Describe -p /user/an32665/KDD/KDDTrain+.arff -f /user/an32665/KDD/KDDTrain+.info -d N 3 C 2 N C 4 N C 8 N 2 C 19 N L Building Forest: mahout org.apache.mahout.df.mapreduce.BuildForest -Dmapred.max.split.size=1874231 -oob -d /user/an32665/KDD/KDDTrain+.arff -ds /user/an32665/KDD/KDDTrain+.info -sl 5 -p -t 100 -o nsl-forest Testing Forest: mahout org.apache.mahout.df.mapreduce.TestForest -i /user/an32665/KDD/KDDTest+.arff -ds /user/an32665/KDD/KDDTrain+.info -m nsl-forest -a -mr -o predictions So while building the forest we use -P for implementing partial implementation. I just wanted to know the difference in algorithm when we use -p and when we don't use -p. Regards, Akshay Nowal
Re: Difference when we don't use partial implementation
Hi Akshay, when you don't use the -p parameter, the builder loads the whole dataset in memory in every computing node, so every tree grown is trained on the whole dataset (of course using bagging to select a subset of it). When using -p, every computing node loads a part of the dataset (thus the name partial) so the trees are trained on parts of the dataset. The training algorithm is the same in both implementations, and the partial implementation is used when the dataset is too big to fit in memory. On Thu, Jul 5, 2012 at 4:38 AM, Nowal, Akshay akshay_no...@syntelinc.comwrote: Hi All, I am running Decision forest in Mahout, below are the commands that I have used to implement the algo: Info file: mahout org.apache.mahout.df.tools.Describe -p /user/an32665/KDD/KDDTrain+.arff -f /user/an32665/KDD/KDDTrain+.info -d N 3 C 2 N C 4 N C 8 N 2 C 19 N L Building Forest: mahout org.apache.mahout.df.mapreduce.BuildForest -Dmapred.max.split.size=1874231 -oob -d /user/an32665/KDD/KDDTrain+.arff -ds /user/an32665/KDD/KDDTrain+.info -sl 5 -p -t 100 -o nsl-forest Testing Forest: mahout org.apache.mahout.df.mapreduce.TestForest -i /user/an32665/KDD/KDDTest+.arff -ds /user/an32665/KDD/KDDTrain+.info -m nsl-forest -a -mr -o predictions So while building the forest we use -P for implementing partial implementation. I just wanted to know the difference in algorithm when we use -p and when we don't use -p. Regards, Akshay Nowal
