Slope One Recommender on Hadoop YONG ZHENG Center for Web Intelligence DePaul University Nov 15, 2012
Overview • Introduction • Recommender Systems & Slope One Recommender • Distributed Slope One on Mahout and Hadoop • Experimental Setup and Analyses • Drive Mahout on Hadoop • Interesting Communities Center for Web Intelligence, DePaul University, USA
Introduction • About Me: a recommendation guy • My Research: data mining and recommender systems • Typical Experimental Research 1) Design or improve an algorithm; 2) Run algorithms and baseline algs on datasets; 3) Compare experimental results; 4) Try different parameters, find reasons and even re-design and improve algorithm itself; 5) Run algorithms and baseline algs on datasets; 6) Compare experimental results; 7) Try different parameters, find reasons and even re-design and improve algorithm itself; 8) And so on… Until it approaches expected results.
Introduction • Sometimes, data is large-scale. e.g. one algorithm may spend days to complete, how about experimental results are not as expected. Then improve algorithms and run it for days again, and again. How can we do previously? (for tasks not that complicated) 1). Paralleling but complicated synchronization and limited resources, such as CPU, memory, etc; 2). Take advantage of PC Labs, let’s do it with 10 PCs • Nearly all research will ultimately face the large-scale problems , especially in the domain of data mining. • But, we have Map-Reduce NOW!
Introduction • Do not need to distribute data and tasks manually. Instead we just simply generate configurations. • Do not need to care about more details, e.g. how data is distributed, when one specific task will be ran on which machine, or how they conduct tasks one by one. • Instead, we can pre-define working flow. We can take advantage of the functional contributions from mappers and reducers. • More benefits: replication, balancing, robustness, etc
Recommender Systems • Collaborative Filtering • Slope One and Simple Weighted Slope One • Slope One in Mahout • Distributed Slope One in Mahout • Mappers and Reducers Center for Web Intelligence, DePaul University, USA
Collaborative Filtering (CF) One of most popular recommendation algorithms. User-based: User-CF Item-based: Item-CF, Slope One User 5 Rating? 5 4 4 4 star 5 Example: User-based Collaborative Filtering
Slope One Recommender Reference: Daniel Lemire, Anna Maclachlan, Slope One Predictors for Online Rating-Based Collaborative Filtering, In SIAM Data Mining (SDM'05), April 21-23, 2005. http://lemire.me/fr/abstracts/SDM2005.html User Batman Spiderman U1 3 4 U2 2 4 U3 2 ? 1). How different two movies were rated? U1 rated Spiderman higher by (4-3) = 1 U2 rated Spiderman higher by (4-2) = 2 On average, Spiderman is rated (1+2)/2 = 1.5 higher 2). Rating difference can tell predictions If we know U3 gave Batman a 2-star, probably he will rated Spiderman by (2+1.5) = 3.5 star
Simple Weighted Slope One Usually user rated multiple items User HarryPotter Batman Spiderman U1 5 3 4 U2 ? 2 4 U3 4 2 ? 1). How different the two movies were rated? Diff(Batman, Spiderman) = [(4-3)+(4-2)]/2 = 1.5 Diff(HarryPotter, Spiderman) = (4-5)/1 = -1 “2” and “1” here we call them as “count”. 2). Weighted rating difference can tell predictions We use a simple weighted approach Refer to Batman only, rating = 2+1.5 = 3.5 Refer to HarryPotter only, rating = 4-1 = 3 Consider them all, predicted rating = (3.5*2 + 3*1])/ (2+1) = 3.33
Simple Weighted Slope One User HarryPotter Batman Spiderman u1 5 3 4 u2 ? 2 4 u3 4 2 ? Question: Online or Offline? To calculate the prediction ratings, we need 2 matrices: 1).Difference Matrix Movie1 Movie2 Movie3 Movie4 Movie1 Movie2 -1.5 Movie3 2 1 Movie4 -1 0.5 -2 2). Count Matrix Just number of users co-rated on two items
Slope One in Mahout org.apache.mahout.cf.taste.impl.recommender.slopeone.SlopeOneRecommender Pre-Processing Stage: (class MemoryDiffStorage with Map) for every item i for every other item j for every user u expressing preference for both i and j add the difference in u’s preference for i and j to an average Recommendation Stage: for every item i the user u expresses no preference for for every item j that user u expresses a preference for find the average preference difference between j and i add this diff to u’s preference value for j add this to a running average return the top items, ranked by these averages Simple weighting: as introduced previously StdDev weighting: item-item rating diffs with lower sd should be weighted highly
Distributed Slope One in Mahout Similar to our previous practice, e.g. the matrix factorization Process, what we need is the Difference Matrix. Suppose there are M users rated N items, the matrix requires N(N-1)/2 cells. Also, the density is another aspect – how user rated items. If there are several items and the rating matrix is dense, the computational costs will increase accordingly. Question again: Online or Offline? Depends on tasks & data. Large-scale data. Let’s do it offline!
Distributed Slope One in Mahout package org.apache.mahout.cf.taste.hadoop.slopeone; class SlopeOneAverageDiffsJob class SlopeOnePrefsToDiffsReducer class SlopeOneDiffsToAveragesReducer package org.apache.mahout.cf.taste.hadoop; class ToItemPrefsMapper org.apache.hadoop.mapreduce.Mapper Two Mapper-Reducer Stages: 1). Create DiffMatrix for each user 2). Collect AvgDiff info, counts, StdDev Let’s see how it works…
Mapper and Reducer - 1 User HarryPotter Batman Spiderman U1 5 3 4 U2 ? 2 4 U3 4 2 ? Mapper1 (ToItemPrefsMapper) <UserID, Pair<ItemID, Rating>> Reducer1 (PrefsToDiffsReducer) <Pair<Item1,Item2>, Diff> (for all three users) <U1> Potter Bat Spider <U2> Potter Bat Spider Potter Potter Bat -2 Bat NULL Spider -1 1 Spider NULL 2
Mapper and Reducer - 2 <U1> Potter Bat Spider <U2> Potter Bat Spider Potter Potter Bat -2 Bat NULL Spider -1 1 Spider NULL 2 Mapper2 (org.apache.hadoop.mapreduce.Mapper) Reducer2 (DiffsToAveragesReducer) Average Diffs, Count, StedDev <Aggregate> Potter Bat Spider Potter Bat -2, 1 Spider -1, 1 1.5, 2 Simply, <a,b> pair denotes a=averge diff, b=count Notice: we should use three matrices in practice, here I used 2.
Predictions User HarryPotter Batman Spiderman U1 5 3 4 U2 ? 2 4 U3 4 2 ? <Aggregate> Potter Bat Spider Potter Bat -2, 1 Spider -1, 1 1.5, 2 Simply, <a,b> pair denotes a=averge diff, b=count Notice: we should use three matrices in practice, here I used 2. Prediction(U3, Spiderman) = [(4-1)*1 + (2+1.5)*2] / (1+2) = 3.33333333333333333333
Experiments • Data • Hadoop Setup • Running Performances Center for Web Intelligence, DePaul University, USA
Experiment Setup Data: MovieLens-1M ratings # of users: 6,040 # of movies: 3,900 # of ratings: 1,000,209 Density of the ratings: each user has at least 20 ratings obviously, some users have many more ratings Rating format: UserID, ItemID, Rating (scale 1-5) Data Split: 80% training, 20% testing
Experiment Setup Hadoop Cluster Setup IBM SmartCloud 1 master node, 7 slave nodes Each node is as SUSE Linux Enterprise Server v11 SP1 Server Configuration: 64 bit (vCPU: 2, RAM: 4 GiB, Disk: 60 GiB) Hadoop v.0.20.205.0 Mahout distribution-0.6 The environment setup follows the typical workflow as: http://irecsys.blogspot.com/2012/11/configurate-map-reduce- environment-on.html Thanks Scott Young, neat writeup!!
Experimental Analyses Stage-1: SlopeOneAverageDiffsJob by Map-Reduce Goal: Build DiffStorage Output: DiffStorage txt file, 1.45GB Running Time: real 13m 34.228s user 0m 5.136s sys 0m 1.028s Item1 Item2 Diff Count StdDev 221 223 -1.02 197 0.5 Stage-2: Java evaluator to measure MAE on testing set Running Time: Load Testing Set (21K records), 299ms Load Training Set (79K records), 1,771ms Load DiffStorage, 176,352ms = 2.9m Prediction (21K records), 18,182ms = 0.3m MAE = 0.71330756
Experimental Experiences 1. Why not MovieLens 10M data? Map-Reduce on 10M data may cost several hrs; Running time depends on cluster and configuration; Also, DiffStorage file will be too large. 2. Java Evaluator Load full DiffStorage file is time-consuming. Also, incur Java heap space and GCOverlimit errors; Those errors can not be fixed by –Xmx or other solutions; Two solutions: 1). Just use simple weighting, discard StdDev weighting. 2). Simple Mapper and Reducer, run it on clusters. For MovieLens 1M, it is not that efficient compared with the live SlopeOne recommendation; 10M data may be better, will try MovieLens-10M data later; Slope One is simple but memory-expensive.
More … • Drive Mahout on Hadoop • Interesting Communities Center for Web Intelligence, DePaul University, USA
Mahout + Hadoop How to put more Mahout algorithms to Hadoop? 1. Pre-set Command in Mahout Let’s see bin/mahout – help, then it provides a list of available programs such as svd, fkmeans, etc. Some are basic functions, such as splitDataset Some can be executed as Hadoop tasks e.g. Run and evaluate Matrix Factorization on rating dataset bin/mahout parallelALS --input inputSource --output outputSource --tempDir tmpFolder --numFeatures 20 --numIterations 10 bin/mahout evaluateFactorization --input inputSource --output outputSource --userFeatures als/out/U/ --itemFeatures als/out/M/ --tempDir tmpFolder
Mahout + Hadoop 2. More Algorithms on Hadoop Mahout provides a way to run more Mahout algorithms. Simply, $HADOOP_HOME/bin/hadoop jar $MAHOUT_HOME/core/target/mahout-core- <version>.jar <Job Class> --recommenderClassName Class <OPTIONS> Which kinds of Jobs it supports? Mahout implemented some versions. Some popular ones: 1).org.apache.mahout.cf.taste.hadoop.pseudo.RecommenderJob --recommenderClassName ClassName 2).org.apache.mahout.cf.taste.hadoop.item.RecommenderJob 3).org.apache.mahout.cf.taste.hadoop.als.ParallelALSFactorizationJob 4).org.apache.mahout.cf.taste.hadoop.slopeone.SlopeOneAverageDiffsJob