Custome metrics in Mahout Recommendation
I am implementing recommendation techniques in Mahout. However, I have a requirement for a custom evaluation metrics other than predefined or built-in ones. So, Q: Can someone please point me to sample custom Evaluator or metric implementation in Mahout? Thanks Vikas
FPGrowth and Recommendations
Hi, New to mahout and fp growth. I havefollowed this example:https://chimpler.wordpress.com/2013/05/02/finding-association-rules-with-mahout-frequent-pattern-mining/ I generated nice output informationlike this (as an example): [abc,def,ghi] => klm,confidence:0.597, support:0.01, lift: 57.415, conviction: 2.453…... Now I am not clear on how to model“recommendations” where given items [qrs, tuv] recommend wxy basedon confidence level. Am I to make lookups based on the above results or use one of the several recommender and similarity classesin mahout? A bit lost on where to start. Thanks
Re: FPGrowth and Recommendations
Hi Jeff, as I recall the map-reduce-based fp-growth solution was problematic, and it's been either deprecated or removed. There are better solutions under the "recommendations" tab at http://mahout.apache.org And I would encourage your updating your version of Mahout to 0.9 or to the master branch at https://github.com/apache/mahout, since 0.7 from that blog is outdated. Best Andrew On Mon, Mar 2, 2015 at 3:26 PM, Jeff Isenhart wrote: > Hi, > New to mahout and fp growth. I havefollowed this example: > https://chimpler.wordpress.com/2013/05/02/finding-association-rules-with-mahout-frequent-pattern-mining/ > I generated nice output informationlike this (as an example): > [abc,def,ghi] => klm,confidence:0.597, support:0.01, lift: 57.415, > conviction: 2.453…... > > Now I am not clear on how to model“recommendations” where given items > [qrs, tuv] recommend wxy basedon confidence level. Am I to make lookups > based on the above results or use one of the several recommender and > similarity classesin mahout? A bit lost on where to start. > Thanks >
Re: Custome metrics in Mahout Recommendation
If you check out the latest master branch from https://github.com/apache/mahout you'll find classes like this in the map-reduce legacy package: /home/akm/mahout/mrlegacy/src/main/java/org/apache/mahout/common/distance/CosineDistanceMeasure.java I'm not sure if/where new ones are being written.. On Mon, Mar 2, 2015 at 3:12 PM, Vikas Kumar wrote: > I am implementing recommendation techniques in Mahout. However, I have a > requirement for a custom evaluation metrics other than predefined or > built-in ones. So, > > Q: Can someone please point me to sample custom Evaluator or metric > implementation in Mahout? > > > Thanks > Vikas >
Re: Custome metrics in Mahout Recommendation
Evaluation metric? You mean like the old recommender evaluator? I’d use MAP mean average precision, but none are implemented in the new Spark recommender code. On Mar 2, 2015, at 3:12 PM, Vikas Kumar wrote: I am implementing recommendation techniques in Mahout. However, I have a requirement for a custom evaluation metrics other than predefined or built-in ones. So, Q: Can someone please point me to sample custom Evaluator or metric implementation in Mahout? Thanks Vikas
Re: Custome metrics in Mahout Recommendation
Sorry for the confusion. Yes, I meant the recommender evaluation metric such as RMSE, Precision, Recall etc which are inbuilt. But, I am planning (or reusing - let me know if already done) to write the metrics such as nDCG, Popularity, Avg. Rating, diversity etc. Thanks Vikas On Mon, Mar 2, 2015 at 6:46 PM, Pat Ferrel wrote: > Evaluation metric? You mean like the old recommender evaluator? I’d use > MAP mean average precision, but none are implemented in the new Spark > recommender code. > > On Mar 2, 2015, at 3:12 PM, Vikas Kumar wrote: > > I am implementing recommendation techniques in Mahout. However, I have a > requirement for a custom evaluation metrics other than predefined or > built-in ones. So, > > Q: Can someone please point me to sample custom Evaluator or metric > implementation in Mahout? > > > Thanks > Vikas > >
Re: FPGrowth and Recommendations
Jeff, are you trying to build a general recommender? Or a shopping cart recommender? FP was used to find things often bought together, which means recommendations based on some partial group of items (watchlist, wishlist, shopping cart). FPG has been deprecated in favor of newer methods. There are references at the top of this page http://mahout.apache.org/users/recommender/intro-cooccurrence-spark.html that may help for a general recommender but things bought together would require a slightly different approach. Think of the shopping-cart-id as if it were a user-id. Create a cooccurrence matrix using spark-itemsimilarity as described in the above references. The shopping cart ids will disappear leaving an item-id followed by a list of similar items. Index this in a search engine and use the current partial shopping cart as the query. You’ll get back an ordered list of the items most commonly bought with items in the current shopping cart. On Mar 2, 2015, at 3:37 PM, Andrew Musselman wrote: Hi Jeff, as I recall the map-reduce-based fp-growth solution was problematic, and it's been either deprecated or removed. There are better solutions under the "recommendations" tab at http://mahout.apache.org And I would encourage your updating your version of Mahout to 0.9 or to the master branch at https://github.com/apache/mahout, since 0.7 from that blog is outdated. Best Andrew On Mon, Mar 2, 2015 at 3:26 PM, Jeff Isenhart wrote: > Hi, > New to mahout and fp growth. I havefollowed this example: > https://chimpler.wordpress.com/2013/05/02/finding-association-rules-with-mahout-frequent-pattern-mining/ > I generated nice output informationlike this (as an example): > [abc,def,ghi] => klm,confidence:0.597, support:0.01, lift: 57.415, > conviction: 2.453…... > > Now I am not clear on how to model“recommendations” where given items > [qrs, tuv] recommend wxy basedon confidence level. Am I to make lookups > based on the above results or use one of the several recommender and > similarity classesin mahout? A bit lost on where to start. > Thanks >
Re: Custome metrics in Mahout Recommendation
MAP is what I use, not precision. Mean average precision accounts for ranking better and ranking is usually what you want to optimize. As I said none are in the current iteration of the recommender code including RMSE, etc. http://mahout.apache.org/users/recommender/intro-cooccurrence-spark.html What use did you have for offline metrics? No offline metric is as good as A/B testing. Often things built in to recommenders will lower one metric or another but produce better results in the form of user engagement or sales. Metrics must be used with great caution and should never be relied on in a real world situation where user testing is available. On Mar 2, 2015, at 5:04 PM, Vikas Kumar wrote: Sorry for the confusion. Yes, I meant the recommender evaluation metric such as RMSE, Precision, Recall etc which are inbuilt. But, I am planning (or reusing - let me know if already done) to write the metrics such as nDCG, Popularity, Avg. Rating, diversity etc. Thanks Vikas On Mon, Mar 2, 2015 at 6:46 PM, Pat Ferrel wrote: > Evaluation metric? You mean like the old recommender evaluator? I’d use > MAP mean average precision, but none are implemented in the new Spark > recommender code. > > On Mar 2, 2015, at 3:12 PM, Vikas Kumar wrote: > > I am implementing recommendation techniques in Mahout. However, I have a > requirement for a custom evaluation metrics other than predefined or > built-in ones. So, > > Q: Can someone please point me to sample custom Evaluator or metric > implementation in Mahout? > > > Thanks > Vikas > >
spark-itemsimilarity cross-indicator-matrix and indicator-matrix
Hi Pat, Thanks for the spark-itemsimilarity. I just ran it against my csv file which looks like below. My two actions are Liked and AddToCart. AddToCart is the main action. usr000d3ca6655-c132-11e4-ac0a-0cc47a03334d Liked prdb5bc44b-fca8-4462-ba19-b9b2b823beb6 usr000e615c7ff-c132-11e4-ac0a-0cc47a03334d Liked prd0002da21ddf-6004-4611-9ef5-4fe96bbb969f usr000d4821d99-c132-11e4-ac0a-0cc47a03334d AddToCart prd75a5710-4890-4ec7-8faa-d9ffb0e94e6d usr000cfd955f3-c132-11e4-ac0a-0cc47a03334d Liked prdb5bc44b-fca8-4462-ba19-b9b2b823beb6 usrdcd0443-c133-11e4-ac0a-0cc47a03334d Liked prd000e8149493-f3a1-491c-abe1-e7886e0286b9 usr000dfc1e861-c132-11e4-ac0a-0cc47a03334d Liked prd0009d6ff4b1-b139-46de-8d4c-425dbee5b4ca usr000d2db7271-c132-11e4-ac0a-0cc47a03334d Liked prd000db1268fa-7fcb-4c84-8877-3b75f2748fda usr000d274251f-c132-11e4-ac0a-0cc47a03334d Liked prd00078f2d7a1-9d6b-4175-96b5-d50ee35ab6db usr000fb647da9-c132-11e4-ac0a-0cc47a03334d AddToCart prd000613c24dd-b838-433f-b1bc-e8e4dddae4d7 usr000d34d4713-c132-11e4-ac0a-0cc47a03334dLiked prd0007ca32be4-2443-4bba-83b4-f01382968c07 I can see there are two output files. One is under cross-indicator-matrix and the other under indicator-matrix. I use Solr. I created a field called recommendation. It's indexed multivalues string field. I loaded indicator-matrix/part-0 into this filed. My question is what is cross-indicator-matrix for? What's the difference between indicator-matrix and cross-indicator-matrix? Do I need to create another Solr field for this cross-indicator-matrix? When I should use it? Is indicator-matrix for AddToCart and cross-indicator-matrix for Liked? Thanks a lot, Kevin
