基于hadoop的推荐算法-mahout版

基于hadoop的推荐算法,讲其中mahout实现的基于项目的推荐算法

分为4步:

1.获得人-物 用户矩阵

    输入为所有人对物品的评价或关联

    map端输出key为人,value为物品+倾好度

    reeduce端输出key为人,vallue为多个物品+倾好度

 

2.获得物-物 项目矩阵

   输入为“用户矩阵”,讲每一行人-物数据中的物品做笛卡尔积,生产成物-物的关联

   map端输出为key为物,value为关联度

   reduce端输出key为物,value为多个物的关联度

(可以根据各种规则生成项目相似度矩阵表,此处算法带过)

修改:

求项目相似矩阵是基于项目的协同过滤算法的核心

公式有很多种,核心是物品i和物品j相关用户的交集与并集的商

mahout使用的公式是1.dot(i,j) = sum(Pi(u)*Pi(u))

                  2.norms(i) = sum(Pi(u)^2)

                  3.simi(i,j) = 1/(1+(norms(i)-2*dot(i,j)+noorm(i))^1/2)

 

mahout的实现方法是

第一个job,用物品-人的矩阵,求得norms,即物品的用户平方和,输出是物-norms

第二个job,Map:用人-物的矩阵,求Pi(u)*Pi(u),即相同用户的物品的评价的乘机,输出物-多个对端物品的Pi(u)*Pi(u)

           Reduce:用物-多个对端物品的Pi(u)*Pi(u)和物-norms,求得物品的相似矩阵(因为这个时候可以汇总所有和这个物品相关的物品的dot)

第三个job,补全物品的相似矩阵

 

 

3.获得用户-项目相似矩阵

输入为人-物 用户矩阵 和 物-物 项目矩阵

Map端输出key为物,value为类VectorOrPrefWritable,是包含物与人的倾好度,或是物与物的相似度

reduce端输出key为物,value为类VectorAndPrefWritable,是汇总当个物品到所有人的倾好度和到所有物品的相似度

 

4.获得用户推荐矩阵

输入为VectorAndPrefWritable

Map端输出为key:人,value:物+系数(map端根据单个物品贡献的系数生成推荐系数,也就是人到物品A的倾好度*物品A到其他物品的相似度)

reduce端输出为key:人,,value:推荐项目+系数(reduce端使用自定公式,汇总所有单物品贡献的四叔,求人到其他项目的倾好度,取topn作为当前用户的推荐项目)

 

再在这里贴几个mahout推荐算法分析的帖子:

http://eric-gcm.iteye.com/blog/1817822

http://eric-gcm.iteye.com/blog/1818033

http://eric-gcm.iteye.com/blog/1820060

 

以下是mahout代码:

 

 ItemSimilarityJob类是mahout使用hadoop做推荐引擎的主要实现类,下面开始分析。

run()函数是启动函数:

 

 

Java代码   收藏代码
  1. public final class RecommenderJob extends AbstractJob {  
  2.   
  3.   public static final String BOOLEAN_DATA = "booleanData";  
  4.   
  5.   private static final int DEFAULT_MAX_SIMILARITIES_PER_ITEM = 100;  
  6.   private static final int DEFAULT_MAX_PREFS_PER_USER = 1000;  
  7.   private static final int DEFAULT_MIN_PREFS_PER_USER = 1;  
  8.   
  9.   @Override  
  10.   public int run(String[] args) throws Exception {  
  11.     //这里原来有大一堆代码,都是用来载入配置项,不用管它  
  12.   
  13.     //第一步:准备矩阵,将原始数据转换为一个矩阵,在PreparePreferenceMatrixJob这个类中完成  
  14.     if (shouldRunNextPhase(parsedArgs, currentPhase)) {  
  15.       ToolRunner.run(getConf(), new PreparePreferenceMatrixJob(), new String[]{  
  16.               "--input", getInputPath().toString(),  
  17.               "--output", prepPath.toString(),  
  18.               "--maxPrefsPerUser", String.valueOf(maxPrefsPerUserInItemSimilarity),  
  19.               "--minPrefsPerUser", String.valueOf(minPrefsPerUser),  
  20.               "--booleanData", String.valueOf(booleanData),  
  21.               "--tempDir", getTempPath().toString()});  
  22.   
  23.       numberOfUsers = HadoopUtil.readInt(new Path(prepPath, PreparePreferenceMatrixJob.NUM_USERS), getConf());  
  24.     }  
  25.   
  26.     //第二步:计算协同矩阵  
  27.     if (shouldRunNextPhase(parsedArgs, currentPhase)) {  
  28.   
  29.       /* special behavior if phase 1 is skipped */  
  30.       if (numberOfUsers == -1) {  
  31.         numberOfUsers = (int) HadoopUtil.countRecords(new Path(prepPath, PreparePreferenceMatrixJob.USER_VECTORS),  
  32.                 PathType.LIST, null, getConf());  
  33.       }  
  34.   
  35.       /* Once DistributedRowMatrix uses the hadoop 0.20 API, we should refactor this call to something like 
  36.        * new DistributedRowMatrix(...).rowSimilarity(...) */  
  37.       //calculate the co-occurrence matrix  
  38.       ToolRunner.run(getConf(), new RowSimilarityJob(), new String[]{  
  39.               "--input", new Path(prepPath, PreparePreferenceMatrixJob.RATING_MATRIX).toString(),  
  40.               "--output", similarityMatrixPath.toString(),  
  41.               "--numberOfColumns", String.valueOf(numberOfUsers),  
  42.               "--similarityClassname", similarityClassname,  
  43.               "--maxSimilaritiesPerRow", String.valueOf(maxSimilaritiesPerItem),  
  44.               "--excludeSelfSimilarity", String.valueOf(Boolean.TRUE),  
  45.               "--threshold", String.valueOf(threshold),  
  46.               "--tempDir", getTempPath().toString()});  
  47.     }  
  48.   
  49.     //start the multiplication of the co-occurrence matrix by the user vectors  
  50.     if (shouldRunNextPhase(parsedArgs, currentPhase)) {  
  51.       Job prePartialMultiply1 = prepareJob(  
  52.               similarityMatrixPath, prePartialMultiplyPath1, SequenceFileInputFormat.class,  
  53.               SimilarityMatrixRowWrapperMapper.class, VarIntWritable.class, VectorOrPrefWritable.class,  
  54.               Reducer.class, VarIntWritable.class, VectorOrPrefWritable.class,  
  55.               SequenceFileOutputFormat.class);  
  56.       boolean succeeded = prePartialMultiply1.waitForCompletion(true);  
  57.       if (!succeeded)   
  58.         return -1;  
  59.       //continue the multiplication  
  60.       Job prePartialMultiply2 = prepareJob(new Path(prepPath, PreparePreferenceMatrixJob.USER_VECTORS),  
  61.               prePartialMultiplyPath2, SequenceFileInputFormat.class, UserVectorSplitterMapper.class, VarIntWritable.class,  
  62.               VectorOrPrefWritable.class, Reducer.class, VarIntWritable.class, VectorOrPrefWritable.class,  
  63.               SequenceFileOutputFormat.class);  
  64.       if (usersFile != null) {  
  65.         prePartialMultiply2.getConfiguration().set(UserVectorSplitterMapper.USERS_FILE, usersFile);  
  66.       }  
  67.       prePartialMultiply2.getConfiguration().setInt(UserVectorSplitterMapper.MAX_PREFS_PER_USER_CONSIDERED,  
  68.               maxPrefsPerUser);  
  69.       succeeded = prePartialMultiply2.waitForCompletion(true);  
  70.       if (!succeeded)   
  71.         return -1;  
  72.       //finish the job  
  73.       Job partialMultiply = prepareJob(  
  74.               new Path(prePartialMultiplyPath1 + "," + prePartialMultiplyPath2), partialMultiplyPath,  
  75.               SequenceFileInputFormat.class, Mapper.class, VarIntWritable.class, VectorOrPrefWritable.class,  
  76.               ToVectorAndPrefReducer.class, VarIntWritable.class, VectorAndPrefsWritable.class,  
  77.               SequenceFileOutputFormat.class);  
  78.       setS3SafeCombinedInputPath(partialMultiply, getTempPath(), prePartialMultiplyPath1, prePartialMultiplyPath2);  
  79.       succeeded = partialMultiply.waitForCompletion(true);  
  80.       if (!succeeded)   
  81.         return -1;  
  82.     }  
  83.   
  84.     if (shouldRunNextPhase(parsedArgs, currentPhase)) {  
  85.       //filter out any users we don't care about  
  86.       /* convert the user/item pairs to filter if a filterfile has been specified */  
  87.       if (filterFile != null) {  
  88.         Job itemFiltering = prepareJob(new Path(filterFile), explicitFilterPath, TextInputFormat.class,  
  89.                 ItemFilterMapper.class, VarLongWritable.class, VarLongWritable.class,  
  90.                 ItemFilterAsVectorAndPrefsReducer.class, VarIntWritable.class, VectorAndPrefsWritable.class,  
  91.                 SequenceFileOutputFormat.class);  
  92.         boolean succeeded = itemFiltering.waitForCompletion(true);  
  93.         if (!succeeded)   
  94.           return -1;  
  95.       }  
  96.   
  97.       String aggregateAndRecommendInput = partialMultiplyPath.toString();  
  98.       if (filterFile != null) {  
  99.         aggregateAndRecommendInput += "," + explicitFilterPath;  
  100.       }  
  101.       //extract out the recommendations  
  102.       Job aggregateAndRecommend = prepareJob(  
  103.               new Path(aggregateAndRecommendInput), outputPath, SequenceFileInputFormat.class,  
  104.               PartialMultiplyMapper.class, VarLongWritable.class, PrefAndSimilarityColumnWritable.class,  
  105.               AggregateAndRecommendReducer.class, VarLongWritable.class, RecommendedItemsWritable.class,  
  106.               TextOutputFormat.class);  
  107.       Configuration aggregateAndRecommendConf = aggregateAndRecommend.getConfiguration();  
  108.       if (itemsFile != null) {  
  109.         aggregateAndRecommendConf.set(AggregateAndRecommendReducer.ITEMS_FILE, itemsFile);  
  110.       }  
  111.   
  112.       if (filterFile != null) {  
  113.         setS3SafeCombinedInputPath(aggregateAndRecommend, getTempPath(), partialMultiplyPath, explicitFilterPath);  
  114.       }  
  115.       setIOSort(aggregateAndRecommend);  
  116.       aggregateAndRecommendConf.set(AggregateAndRecommendReducer.ITEMID_INDEX_PATH,  
  117.               new Path(prepPath, PreparePreferenceMatrixJob.ITEMID_INDEX).toString());  
  118.       aggregateAndRecommendConf.setInt(AggregateAndRecommendReducer.NUM_RECOMMENDATIONS, numRecommendations);  
  119.       aggregateAndRecommendConf.setBoolean(BOOLEAN_DATA, booleanData);  
  120.       boolean succeeded = aggregateAndRecommend.waitForCompletion(true);  
  121.       if (!succeeded)   
  122.         return -1;  
  123.     }  
  124.   
  125.     return 0;  
  126.   }  

 

 

 

       第二步,计算协同矩阵,主要在RowSimilarityJob 这个类中完成

 

  

Java代码   收藏代码
  1. ToolRunner.run(getConf(), new RowSimilarityJob(), new String[]{  
  2.               "--input", new Path(prepPath, PreparePreferenceMatrixJob.RATING_MATRIX).toString(),  
  3.               "--output", similarityMatrixPath.toString(),  
  4.               "--numberOfColumns", String.valueOf(numberOfUsers),  
  5.               "--similarityClassname", similarityClassname,  
  6.               "--maxSimilaritiesPerRow", String.valueOf(maxSimilaritiesPerItem),  
  7.               "--excludeSelfSimilarity", String.valueOf(Boolean.TRUE),  
  8.               "--threshold", String.valueOf(threshold),  
  9.               "--tempDir", getTempPath().toString()});  
  10.     }  

   可以看到这个job的输入路径就是上一篇中,PreparePreferenceMatrixJob中最后一个reducer的输出路径。

 

下边详细分析RowSimilarityJob类的实现:

 

Java代码   收藏代码
  1. public class RowSimilarityJob extends AbstractJob {  
  2.   
  3.   
  4.   @Override  
  5.   public int run(String[] args) throws Exception {  
  6.     //一大堆载入参数的代码,忽略  
  7.       
  8.     //第一个MapReduce  
  9.     if (shouldRunNextPhase(parsedArgs, currentPhase)) {  
  10.       Job normsAndTranspose = prepareJob(getInputPath(), weightsPath, VectorNormMapper.class, IntWritable.class,  
  11.           VectorWritable.class, MergeVectorsReducer.class, IntWritable.class, VectorWritable.class);  
  12.       normsAndTranspose.setCombinerClass(MergeVectorsCombiner.class);  
  13.       Configuration normsAndTransposeConf = normsAndTranspose.getConfiguration();  
  14.       normsAndTransposeConf.set(THRESHOLD, String.valueOf(threshold));  
  15.       normsAndTransposeConf.set(NORMS_PATH, normsPath.toString());  
  16.       normsAndTransposeConf.set(NUM_NON_ZERO_ENTRIES_PATH, numNonZeroEntriesPath.toString());  
  17.       normsAndTransposeConf.set(MAXVALUES_PATH, maxValuesPath.toString());  
  18.       normsAndTransposeConf.set(SIMILARITY_CLASSNAME, similarityClassname);  
  19.       boolean succeeded = normsAndTranspose.waitForCompletion(true);  
  20.       if (!succeeded) {  
  21.         return -1;  
  22.       }  
  23.     }  
  24.     //第二个MapReduce  
  25.     if (shouldRunNextPhase(parsedArgs, currentPhase)) {  
  26.       Job pairwiseSimilarity = prepareJob(weightsPath, pairwiseSimilarityPath, CooccurrencesMapper.class,  
  27.           IntWritable.class, VectorWritable.class, SimilarityReducer.class, IntWritable.class, VectorWritable.class);  
  28.       pairwiseSimilarity.setCombinerClass(VectorSumReducer.class);  
  29.       Configuration pairwiseConf = pairwiseSimilarity.getConfiguration();  
  30.       pairwiseConf.set(THRESHOLD, String.valueOf(threshold));  
  31.       pairwiseConf.set(NORMS_PATH, normsPath.toString());  
  32.       pairwiseConf.set(NUM_NON_ZERO_ENTRIES_PATH, numNonZeroEntriesPath.toString());  
  33.       pairwiseConf.set(MAXVALUES_PATH, maxValuesPath.toString());  
  34.       pairwiseConf.set(SIMILARITY_CLASSNAME, similarityClassname);  
  35.       pairwiseConf.setInt(NUMBER_OF_COLUMNS, numberOfColumns);  
  36.       pairwiseConf.setBoolean(EXCLUDE_SELF_SIMILARITY, excludeSelfSimilarity);  
  37.       boolean succeeded = pairwiseSimilarity.waitForCompletion(true);  
  38.       if (!succeeded) {  
  39.         return -1;  
  40.       }  
  41.     }  
  42.     //第三个MapReduce  
  43.     if (shouldRunNextPhase(parsedArgs, currentPhase)) {  
  44.       Job asMatrix = prepareJob(pairwiseSimilarityPath, getOutputPath(), UnsymmetrifyMapper.class,  
  45.           IntWritable.class, VectorWritable.class, MergeToTopKSimilaritiesReducer.class, IntWritable.class,  
  46.           VectorWritable.class);  
  47.       asMatrix.setCombinerClass(MergeToTopKSimilaritiesReducer.class);  
  48.       asMatrix.getConfiguration().setInt(MAX_SIMILARITIES_PER_ROW, maxSimilaritiesPerRow);  
  49.       boolean succeeded = asMatrix.waitForCompletion(true);  
  50.       if (!succeeded) {  
  51.         return -1;  
  52.       }  
  53.     }  
  54.   
  55.     return 0;  
  56.   }  

 

 

 可以看到RowSimilityJob也是分成三个MapReduce过程:

 

1、Mapper :VectorNormMapper类,输出 ( userid_index, <itemid_index, pref> )类型

Java代码   收藏代码
  1. public static class VectorNormMapper extends Mapper<IntWritable,VectorWritable,IntWritable,VectorWritable> {  
  2.   
  3.     @Override  
  4.     protected void map(IntWritable row, VectorWritable vectorWritable, Context ctx)  
  5.         throws IOException, InterruptedException {  
  6.   
  7.       Vector rowVector = similarity.normalize(vectorWritable.get());  
  8.   
  9.       int numNonZeroEntries = 0;  
  10.       double maxValue = Double.MIN_VALUE;  
  11.   
  12.       Iterator<Vector.Element> nonZeroElements = rowVector.iterateNonZero();  
  13.       while (nonZeroElements.hasNext()) {  
  14.         Vector.Element element = nonZeroElements.next();  
  15.         RandomAccessSparseVector partialColumnVector = new RandomAccessSparseVector(Integer.MAX_VALUE);  
  16.         partialColumnVector.setQuick(row.get(), element.get());  
  17.         //输出 ( userid_index, <itemid_index, pref> )类型  
  18.         ctx.write(new IntWritable(element.index()), new VectorWritable(partialColumnVector));  
  19.   
  20.         numNonZeroEntries++;  
  21.         if (maxValue < element.get()) {  
  22.           maxValue = element.get();  
  23.         }  
  24.       }  
  25.   
  26.       if (threshold != NO_THRESHOLD) {  
  27.         nonZeroEntries.setQuick(row.get(), numNonZeroEntries);  
  28.         maxValues.setQuick(row.get(), maxValue);  
  29.       }  
  30.       norms.setQuick(row.get(), similarity.norm(rowVector));  
  31.       //计算item的总数  
  32.       ctx.getCounter(Counters.ROWS).increment(1);  
  33.     }  
  34. }  

 

Reduer : MergeVectorsReducer类,输入的是(userid_index, <itemid_index, pref>),同一个userid_index在此进行合并,输出( userid_index, vector<itemid_index, pref> )

Java代码   收藏代码
  1.   public static class MergeVectorsReducer extends Reducer<IntWritable,VectorWritable,IntWritable,VectorWritable> {  
  2.   
  3.     @Override  
  4.     protected void reduce(IntWritable row, Iterable<VectorWritable> partialVectors, Context ctx)  
  5.         throws IOException, InterruptedException {  
  6.       Vector partialVector = Vectors.merge(partialVectors);  
  7.   
  8.       if (row.get() == NORM_VECTOR_MARKER) {  
  9.         Vectors.write(partialVector, normsPath, ctx.getConfiguration());  
  10.       } else if (row.get() == MAXVALUE_VECTOR_MARKER) {  
  11.         Vectors.write(partialVector, maxValuesPath, ctx.getConfiguration());  
  12.       } else if (row.get() == NUM_NON_ZERO_ENTRIES_VECTOR_MARKER) {  
  13.         Vectors.write(partialVector, numNonZeroEntriesPath, ctx.getConfiguration(), true);  
  14.       } else {  
  15.         ctx.write(row, new VectorWritable(partialVector));  
  16.       }  
  17.     }  
  18.   }  
  19. }  

 2、Mapper:CooccurrencesMapper类,对同一个userid_index下的vector<itemid_index ,pref>进行处理,

收集<item1, item2>对, 输出为( itemid_index, vector<itemid_index, value> )

Java代码   收藏代码
  1. public static class CooccurrencesMapper extends Mapper<IntWritable,VectorWritable,IntWritable,VectorWritable> {  
  2.   
  3.     @Override  
  4.     protected void map(IntWritable column, VectorWritable occurrenceVector, Context ctx)  
  5.         throws IOException, InterruptedException {  
  6.       Vector.Element[] occurrences = Vectors.toArray(occurrenceVector);  
  7.       Arrays.sort(occurrences, BY_INDEX);  
  8.   
  9.       int cooccurrences = 0;  
  10.       int prunedCooccurrences = 0;  
  11.       for (int n = 0; n < occurrences.length; n++) {  
  12.         Vector.Element occurrenceA = occurrences[n];  
  13.         Vector dots = new RandomAccessSparseVector(Integer.MAX_VALUE);  
  14.         for (int m = n; m < occurrences.length; m++) {  
  15.           Vector.Element occurrenceB = occurrences[m];  
  16.           if (threshold == NO_THRESHOLD || consider(occurrenceA, occurrenceB)) {  
  17.             dots.setQuick(occurrenceB.index(), similarity.aggregate(occurrenceA.get(), occurrenceB.get()));  
  18.             cooccurrences++;  
  19.           } else {  
  20.             prunedCooccurrences++;  
  21.           }  
  22.         }  
  23.         ctx.write(new IntWritable(occurrenceA.index()), new VectorWritable(dots));  
  24.       }  
  25.       ctx.getCounter(Counters.COOCCURRENCES).increment(cooccurrences);  
  26.       ctx.getCounter(Counters.PRUNED_COOCCURRENCES).increment(prunedCooccurrences);  
  27.     }  
  28.   }  

 

Reducer :SimilarityReducer类,生成协同矩阵

 

Java代码   收藏代码
  1. public static class SimilarityReducer  
  2.       extends Reducer<IntWritable,VectorWritable,IntWritable,VectorWritable> {  
  3.   
  4.     @Override  
  5.     protected void reduce(IntWritable row, Iterable<VectorWritable> partialDots, Context ctx)  
  6.         throws IOException, InterruptedException {  
  7.       Iterator<VectorWritable> partialDotsIterator = partialDots.iterator();  
  8.       //取一个vecotr作为该item的行向量  
  9.       Vector dots = partialDotsIterator.next().get();  
  10.       while (partialDotsIterator.hasNext()) {  
  11.         Vector toAdd = partialDotsIterator.next().get();  
  12.         Iterator<Vector.Element> nonZeroElements = toAdd.iterateNonZero();  
  13.         while (nonZeroElements.hasNext()) {  
  14.           Vector.Element nonZeroElement = nonZeroElements.next();  
  15.           //nonZeroElement.index()为itemid,将另一个vecotr中itemid的value加进去  
  16.           dots.setQuick(nonZeroElement.index(), dots.getQuick(nonZeroElement.index()) + nonZeroElement.get());  
  17.         }  
  18.       }  
  19.       //最后得到的dots是协同矩阵中行号为row的一行,行中元素是item对其他的item的相似度  
  20.       Vector similarities = dots.like();  
  21.       double normA = norms.getQuick(row.get());  
  22.       Iterator<Vector.Element> dotsWith = dots.iterateNonZero();  
  23.       while (dotsWith.hasNext()) {  
  24.         Vector.Element b = dotsWith.next();  
  25.         double similarityValue = similarity.similarity(b.get(), normA, norms.getQuick(b.index()), numberOfColumns);  
  26.         if (similarityValue >= treshold) {  
  27.           similarities.set(b.index(), similarityValue);  
  28.         }  
  29.       }  
  30.       if (excludeSelfSimilarity) {  
  31.         similarities.setQuick(row.get(), 0);  
  32.       }  
  33.       ctx.write(row, new VectorWritable(similarities));  
  34.     }  
  35.   }  

 

3、Mapper:UnsymmetrifyMapper类,用来生成对称矩阵的。上一步得到的是非对称矩阵,首先将矩阵偏转,得到偏转矩阵,用原矩阵加上偏转矩阵,可以得到对称矩阵

 

Java代码   收藏代码
  1. public static class UnsymmetrifyMapper extends Mapper<IntWritable,VectorWritable,IntWritable,VectorWritable>  {  
  2.   
  3.     private int maxSimilaritiesPerRow;  
  4.   
  5.     @Override  
  6.     protected void setup(Mapper.Context ctx) throws IOException, InterruptedException {  
  7.       maxSimilaritiesPerRow = ctx.getConfiguration().getInt(MAX_SIMILARITIES_PER_ROW, 0);  
  8.       Preconditions.checkArgument(maxSimilaritiesPerRow > 0, "Incorrect maximum number of similarities per row!");  
  9.     }  
  10.   
  11.     @Override  
  12.     protected void map(IntWritable row, VectorWritable similaritiesWritable, Context ctx)  
  13.         throws IOException, InterruptedException {  
  14.       Vector similarities = similaritiesWritable.get();  
  15.       // For performance reasons moved transposedPartial creation out of the while loop and reusing the same vector  
  16.       Vector transposedPartial = similarities.like();  
  17.       TopK<Vector.Element> topKQueue = new TopK<Vector.Element>(maxSimilaritiesPerRow, Vectors.BY_VALUE);  
  18.       Iterator<Vector.Element> nonZeroElements = similarities.iterateNonZero();  
  19.       //这个地方用来生成偏转矩阵的,非对称矩阵,用原矩阵加上偏转矩阵,可以得到对称矩阵  
  20.       while (nonZeroElements.hasNext()) {  
  21.         Vector.Element nonZeroElement = nonZeroElements.next();  
  22.         topKQueue.offer(new Vectors.TemporaryElement(nonZeroElement));  
  23.           
  24.         transposedPartial.setQuick(row.get(), nonZeroElement.get());  
  25.         //偏转矩阵中的每一个元素  
  26.         ctx.write(new IntWritable(nonZeroElement.index()), new VectorWritable(transposedPartial));  
  27.           
  28.         transposedPartial.setQuick(row.get(), 0.0);  
  29.       }  
  30.       Vector topKSimilarities = similarities.like();  
  31.       for (Vector.Element topKSimilarity : topKQueue.retrieve()) {  
  32.         topKSimilarities.setQuick(topKSimilarity.index(), topKSimilarity.get());  
  33.       }  
  34.       //这里只收集前maxSimilaritiesPerRow个得分最高的item,所以咱们最后的对称矩阵,实际上每行只有  
  35.       //maxSimilaritiesPerRow个是对称的,其他的位置也不管了  
  36.       ctx.write(row, new VectorWritable(topKSimilarities));  
  37.     }  
  38.   }  

 

 Reducer:MergeToTopKSimilaritiesReducer类,就是将上面Map偏转的元素都收集起来,也就是完成了偏转矩阵和(截取了得分前maxSimilaritiesPerRow个)的原矩阵相加的过程,得到了对称矩阵

Java代码   收藏代码
  1. public static class MergeToTopKSimilaritiesReducer  
  2.     extends Reducer<IntWritable,VectorWritable,IntWritable,VectorWritable> {  
  3.   
  4.   private int maxSimilaritiesPerRow;  
  5.   
  6.   @Override  
  7.   protected void setup(Context ctx) throws IOException, InterruptedException {  
  8.     maxSimilaritiesPerRow = ctx.getConfiguration().getInt(MAX_SIMILARITIES_PER_ROW, 0);  
  9.     Preconditions.checkArgument(maxSimilaritiesPerRow > 0, "Incorrect maximum number of similarities per row!");  
  10.   }  
  11.   
  12.   @Override  
  13.   protected void reduce(IntWritable row, Iterable<VectorWritable> partials, Context ctx)  
  14.       throws IOException, InterruptedException {  
  15.     Vector allSimilarities = Vectors.merge(partials);  
  16.     Vector topKSimilarities = Vectors.topKElements(maxSimilaritiesPerRow, allSimilarities);  
  17.     ctx.write(row, new VectorWritable(topKSimilarities));  
  18.   }  
  19. }  

 

至此,RowSimilarityJob类的全部工作就完成,最终生成的是一个对称矩阵,也就是协同矩阵

 

 

 

Java代码   收藏代码
  1. //协同矩阵与用户向量相乘  
  2.     //start the multiplication of the co-occurrence matrix by the user vectors  
  3.     if (shouldRunNextPhase(parsedArgs, currentPhase)) {  
  4.       //第一个MapReducer  
  5.       Job prePartialMultiply1 = prepareJob(  
  6.               similarityMatrixPath, prePartialMultiplyPath1, SequenceFileInputFormat.class,  
  7.               SimilarityMatrixRowWrapperMapper.class, VarIntWritable.class, VectorOrPrefWritable.class,  
  8.               Reducer.class, VarIntWritable.class, VectorOrPrefWritable.class,  
  9.               SequenceFileOutputFormat.class);  
  10.       boolean succeeded = prePartialMultiply1.waitForCompletion(true);  
  11.       if (!succeeded)   
  12.         return -1;  
  13.       //第二个MapReduce  
  14.       //continue the multiplication  
  15.       Job prePartialMultiply2 = prepareJob(new Path(prepPath, PreparePreferenceMatrixJob.USER_VECTORS),  
  16.               prePartialMultiplyPath2, SequenceFileInputFormat.class, UserVectorSplitterMapper.class, VarIntWritable.class,  
  17.               VectorOrPrefWritable.class, Reducer.class, VarIntWritable.class, VectorOrPrefWritable.class,  
  18.               SequenceFileOutputFormat.class);  
  19.       if (usersFile != null) {  
  20.         prePartialMultiply2.getConfiguration().set(UserVectorSplitterMapper.USERS_FILE, usersFile);  
  21.       }  
  22.       prePartialMultiply2.getConfiguration().setInt(UserVectorSplitterMapper.MAX_PREFS_PER_USER_CONSIDERED,  
  23.               maxPrefsPerUser);  
  24.       succeeded = prePartialMultiply2.waitForCompletion(true);  
  25.       if (!succeeded)   
  26.         return -1;  
  27.       //finish the job  
  28.       //第三个MapReduce  
  29.       Job partialMultiply = prepareJob(  
  30.               new Path(prePartialMultiplyPath1 + "," + prePartialMultiplyPath2), partialMultiplyPath,  
  31.               SequenceFileInputFormat.class, Mapper.class, VarIntWritable.class, VectorOrPrefWritable.class,  
  32.               ToVectorAndPrefReducer.class, VarIntWritable.class, VectorAndPrefsWritable.class,  
  33.               SequenceFileOutputFormat.class);  
  34.       setS3SafeCombinedInputPath(partialMultiply, getTempPath(), prePartialMultiplyPath1, prePartialMultiplyPath2);  
  35.       succeeded = partialMultiply.waitForCompletion(true);  
  36.       if (!succeeded)   
  37.         return -1;  
  38.     }  

 

 下边也是同样分析一下这个三个MapReduce的细节:

 

1、Mapper: SimilarityMatrixRowWrapperMapper 类,将协同矩阵的一行拿出来,通过包装,封装成VectorOrPrefWritable类,与那边的UserVectorSplitterMapper 的输出类型一致

Java代码   收藏代码
  1. public final class SimilarityMatrixRowWrapperMapper extends  
  2.     Mapper<IntWritable,VectorWritable,VarIntWritable,VectorOrPrefWritable> {  
  3.     
  4.   //将协同矩阵的一行拿出来,通过包装,封装成VectorOrPrefWritable类,与那边的UserVectorSplitterMapper  
  5.   //的输出类型一致  
  6.   @Override  
  7.   protected void map(IntWritable key,  
  8.                      VectorWritable value,  
  9.                      Context context) throws IOException, InterruptedException {  
  10.     Vector similarityMatrixRow = value.get();  
  11.     /* remove self similarity */  
  12.     similarityMatrixRow.set(key.get(), Double.NaN);  
  13.     context.write(new VarIntWritable(key.get()), new VectorOrPrefWritable(similarityMatrixRow));  
  14.   }  
  15.   
  16. }  

 

 

2、Mapper:UserVectorSplitterMapper类

Java代码   收藏代码
  1. //输入格式: theUserID:<itemid_index1,pref1>,<itemid_index2,pref2>........<itemid_indexN,prefN>  
  2.   //输出格式:  itemid1:<theUserID,pref1>  
  3.   //          itemid2:<theUserID,pref2>  
  4.   //          itemid3:<theUserID,pref3>  
  5.   //          ......  
  6.   //          itemidN:<theUserID,prefN>  
Java代码   收藏代码
  1. public final class UserVectorSplitterMapper extends  
  2.     Mapper<VarLongWritable,VectorWritable, VarIntWritable,VectorOrPrefWritable> {  
  3.     
  4.   @Override  
  5.   protected void map(VarLongWritable key,  
  6.                      VectorWritable value,  
  7.                      Context context) throws IOException, InterruptedException {  
  8.     long userID = key.get();  
  9.     if (usersToRecommendFor != null && !usersToRecommendFor.contains(userID)) {  
  10.       return;  
  11.     }  
  12.     Vector userVector = maybePruneUserVector(value.get());  
  13.     Iterator<Vector.Element> it = userVector.iterateNonZero();  
  14.     VarIntWritable itemIndexWritable = new VarIntWritable();  
  15.     VectorOrPrefWritable vectorOrPref = new VectorOrPrefWritable();  
  16.     while (it.hasNext()) {  
  17.       Vector.Element e = it.next();  
  18.       itemIndexWritable.set(e.index());  
  19.       vectorOrPref.set(userID, (float) e.get());  
  20.       context.write(itemIndexWritable, vectorOrPref);  
  21.     }  
  22.   }  

 

3、Reduce:ToVectorAndPrefReducer类,收集协同矩阵为itemid的一行,并且收集评价过该item的用户和评分,最后的输出是 itemid_index,VectorAndPrefsWritable(vector,List<userid>,List<pref>)

 

Java代码   收藏代码
  1. public final class ToVectorAndPrefReducer extends  
  2.     Reducer<VarIntWritable,VectorOrPrefWritable,VarIntWritable,VectorAndPrefsWritable> {  
  3.   
  4.   //收集所有key为itemid的  
  5.   @Override  
  6.   protected void reduce(VarIntWritable key,  
  7.                         Iterable<VectorOrPrefWritable> values,  
  8.                         Context context) throws IOException, InterruptedException {  
  9.   
  10.     List<Long> userIDs = Lists.newArrayList();  
  11.     List<Float> prefValues = Lists.newArrayList();  
  12.     Vector similarityMatrixColumn = null;  
  13.     for (VectorOrPrefWritable value : values) {  
  14.       if (value.getVector() == null) {  
  15.         // Then this is a user-pref value  
  16.         userIDs.add(value.getUserID());  
  17.         prefValues.add(value.getValue());  
  18.       } else {  
  19.         // Then this is the column vector  
  20.         //协同矩阵的一个行(行号为itemid的一行)  
  21.         if (similarityMatrixColumn != null) {  
  22.           throw new IllegalStateException("Found two similarity-matrix columns for item index " + key.get());  
  23.         }  
  24.         similarityMatrixColumn = value.getVector();  
  25.       }  
  26.     }  
  27.   
  28.     if (similarityMatrixColumn == null) {  
  29.       return;  
  30.     }  
  31.     //收集协同矩阵为itemid的一行,并且手机评价过该item的用户和评分  
  32.     VectorAndPrefsWritable vectorAndPrefs = new VectorAndPrefsWritable(similarityMatrixColumn, userIDs, prefValues);  
  33.     context.write(key, vectorAndPrefs);  
  34.   }  
  35.   
  36. }  

 

第四步,协同矩阵和用户向量相乘,得到推荐结果

 

Java代码   收藏代码
  1. //extract out the recommendations  
  2.      Job aggregateAndRecommend = prepareJob(  
  3.              new Path(aggregateAndRecommendInput), outputPath, SequenceFileInputFormat.class,  
  4.              PartialMultiplyMapper.class, VarLongWritable.class, PrefAndSimilarityColumnWritable.class,  
  5.              AggregateAndRecommendReducer.class, VarLongWritable.class, RecommendedItemsWritable.class,  
  6.              TextOutputFormat.class);  
  7.      Configuration aggregateAndRecommendConf = aggregateAndRecommend.getConfiguration();  

 

 

Mapper:PartialMultiplyMapper类

 

Java代码   收藏代码
  1. //输入类型:( itemid_index, <userid的数组,pref的数组,协同矩阵行号为itemid_index的行> )  
  2. //输出类型: userid,<该用户对itemid_index1的评分,协同矩阵行号为itemid_index1的行> )  
  3. //        userid,<该用户对itemid_index2的评分,协同矩阵行号为itemid_index2的行> )  
  4. //                       .....    
  5. //                       .....  
  6. //          userid,<该用户对itemid_indexN的评分,协同矩阵行号为itemid_indexN的行> )  

 

 

 

Java代码   收藏代码
  1. public final class PartialMultiplyMapper extends  
  2.     Mapper<VarIntWritable,VectorAndPrefsWritable,VarLongWritable,PrefAndSimilarityColumnWritable> {  
  3.   
  4.   @Override  
  5.   protected void map(VarIntWritable key,  
  6.                      VectorAndPrefsWritable vectorAndPrefsWritable,  
  7.                      Context context) throws IOException, InterruptedException {  
  8.   
  9.     Vector similarityMatrixColumn = vectorAndPrefsWritable.getVector();  
  10.     List<Long> userIDs = vectorAndPrefsWritable.getUserIDs();  
  11.     List<Float> prefValues = vectorAndPrefsWritable.getValues();  
  12.   
  13.     VarLongWritable userIDWritable = new VarLongWritable();  
  14.     PrefAndSimilarityColumnWritable prefAndSimilarityColumn = new PrefAndSimilarityColumnWritable();  
  15.   
  16.     for (int i = 0; i < userIDs.size(); i++) {  
  17.       long userID = userIDs.get(i);  
  18.       float prefValue = prefValues.get(i);  
  19.       if (!Float.isNaN(prefValue)) {  
  20.         prefAndSimilarityColumn.set(prefValue, similarityMatrixColumn);  
  21.         userIDWritable.set(userID);  
  22.         context.write(userIDWritable, prefAndSimilarityColumn);  
  23.       }  
  24.     }  
  25.   }  
  26.   
  27. }  

 

 Reducer:AggregateAndRecommendReducer类,Reducer中进行PartialMultiply,按乘积得到的推荐度的大小取出最大的几个item。对于非booleanData,是用pref和相似度矩阵的PartialMultiply得到推荐度的值来进行排序。
而booleanData的pref值都是1.0f,所以去计算矩阵相乘的过程没有意义,直接累加相似度的值即可。
用这个数据排序就可得到推荐结果

 

Java代码   收藏代码
  1. public final class AggregateAndRecommendReducer extends  
  2.     Reducer<VarLongWritable,PrefAndSimilarityColumnWritable,VarLongWritable,RecommendedItemsWritable> {  
  3.  @Override  
  4.   protected void reduce(VarLongWritable userID,  
  5.                         Iterable<PrefAndSimilarityColumnWritable> values,  
  6.                         Context context) throws IOException, InterruptedException {  
  7.     if (booleanData) {  
  8.       reduceBooleanData(userID, values, context);  
  9.     } else {  
  10.       reduceNonBooleanData(userID, values, context);  
  11.     }  
  12.   }  
  13.   
  14.   private void reduceBooleanData(VarLongWritable userID,  
  15.                                  Iterable<PrefAndSimilarityColumnWritable> values,  
  16.                                  Context context) throws IOException, InterruptedException {  
  17.     /* having boolean data, each estimated preference can only be 1, 
  18.      * however we can't use this to rank the recommended items, 
  19.      * so we use the sum of similarities for that. */  
  20.     Vector predictionVector = null;  
  21.     for (PrefAndSimilarityColumnWritable prefAndSimilarityColumn : values) {  
  22.       predictionVector = predictionVector == null  
  23.           ? prefAndSimilarityColumn.getSimilarityColumn()  
  24.           : predictionVector.plus(prefAndSimilarityColumn.getSimilarityColumn());  
  25.     }  
  26.     writeRecommendedItems(userID, predictionVector, context);  
  27.   }  
  28.   
  29.   private void reduceNonBooleanData(VarLongWritable userID,  
  30.                         Iterable<PrefAndSimilarityColumnWritable> values,  
  31.                         Context context) throws IOException, InterruptedException {  
  32.     /* each entry here is the sum in the numerator of the prediction formula */  
  33.     Vector numerators = null;  
  34.     /* each entry here is the sum in the denominator of the prediction formula */  
  35.     Vector denominators = null;  
  36.     /* each entry here is the number of similar items used in the prediction formula */  
  37.     Vector numberOfSimilarItemsUsed = new RandomAccessSparseVector(Integer.MAX_VALUE, 100);  
  38.   
  39.     for (PrefAndSimilarityColumnWritable prefAndSimilarityColumn : values) {  
  40.       Vector simColumn = prefAndSimilarityColumn.getSimilarityColumn();  
  41.       float prefValue = prefAndSimilarityColumn.getPrefValue();  
  42.       /* count the number of items used for each prediction */  
  43.       Iterator<Vector.Element> usedItemsIterator = simColumn.iterateNonZero();  
  44.       while (usedItemsIterator.hasNext()) {  
  45.         int itemIDIndex = usedItemsIterator.next().index();  
  46.         numberOfSimilarItemsUsed.setQuick(itemIDIndex, numberOfSimilarItemsUsed.getQuick(itemIDIndex) + 1);  
  47.       }  
  48.       //vector.times(float) 是向量乘于一个数,也就是向量的每一个值都乘以这个数  
  49.       //vector.plus(vector) 是两个向量相加,每一个位置上的值相加  
  50.         
  51.       //numerators是一个vecotr,每一个元素是这样的  
  52.       /* 
  53.                 例如index为item1的元素的值为: 
  54.        simility(item1, item_2)*pref(userid, item_2) 
  55.       + simility(item_1, item_3)*pref(userid, item_3) 
  56.       + simility(item1, item_4)*pref(userid, item_4) 
  57.       + ……  
  58.       + simility(item_1, item_2)*pref(userid, item_N) 
  59.       */  
  60.       // 注:其中simility(item1, item2)代表物品item1和物品item2的相似度 ,pref(userid, item)代表用于userid对item打分分值   
  61.        
  62.       numerators = numerators == null  
  63.           ? prefValue == BOOLEAN_PREF_VALUE ? simColumn.clone() : simColumn.times(prefValue)  
  64.           : numerators.plus(prefValue == BOOLEAN_PREF_VALUE ? simColumn : simColumn.times(prefValue));  
  65.         
  66.         
  67.         
  68.       simColumn.assign(ABSOLUTE_VALUES);  
  69.       //denominators是一个vecotr,每一个元素是这样的  
  70.       /* 
  71.                 例如index为item1的元素的值为: 
  72.        simility(item1, item_2)+ simility(item_1, item_3)+ …… + simility(item_1, item_2)*pref(userid, item_N) 
  73.       */  
  74.       // 注:其中simility(item1, item2)代表物品item1和物品item2的相似度  
  75.       denominators = denominators == null ? simColumn : denominators.plus(simColumn);  
  76.     }  
  77.   
  78.     if (numerators == null) {  
  79.       return;  
  80.     }  
  81.   
  82.     Vector recommendationVector = new RandomAccessSparseVector(Integer.MAX_VALUE, 100);  
  83.     Iterator<Vector.Element> iterator = numerators.iterateNonZero();  
  84.     while (iterator.hasNext()) {  
  85.       Vector.Element element = iterator.next();  
  86.       int itemIDIndex = element.index();  
  87.       /* preference estimations must be based on at least 2 datapoints */  
  88.       if (numberOfSimilarItemsUsed.getQuick(itemIDIndex) > 1) {  
  89.         /* compute normalized prediction */  
  90.         //计算归一化预测值  
  91.         double prediction = element.get() / denominators.getQuick(itemIDIndex);  
  92.         recommendationVector.setQuick(itemIDIndex, prediction);  
  93.       }  
  94.     }  
  95.     writeRecommendedItems(userID, recommendationVector, context);  
  96.   }  
  97. }  

 

 

 

你可能感兴趣的:(hadoop,Mahout,推荐算法)