我正在尝试构建一个NaiveBayes分类器,将数据从数据库加载为包含(标签,文本)的DataFrame。 这是数据样本(多项标签):
label| feature|
+-----+--------------------+
| 1|combusting prepar...|
| 1|adhesives for ind...|
| 1| |
| 1| salt for preserving|
| 1|auxiliary fluids ...|
我使用了以下转换来进行标记化,停用词,n-gram和hashTF:
val selectedData = df.select("label", "feature")
// Tokenize RDD
val tokenizer = new Tokenizer().setInputCol("feature").setOutputCol("words")
val regexTokenizer = new RegexTokenizer().setInputCol("feature").setOutputCol("words").setPattern("\\W")
val tokenized = tokenizer.transform(selectedData)
tokenized.select("words", "label").take(3).foreach(println)
// Removing stop words
val remover = new StopWordsRemover().setInputCol("words").setOutputCol("filtered")
val parsedData = remover.transform(tokenized)
// N-gram
val ngram = new NGram().setInputCol("filtered").setOutputCol("ngrams")
val ngramDataFrame = ngram.transform(parsedData)
ngramDataFrame.take(3).map(_.getAs[Stream[String]]("ngrams").toList).foreach(println)
//hashing function
val hashingTF = new HashingTF().setInputCol("ngrams").setOutputCol("hash").setNumFeatures(1000)
val featurizedData = hashingTF.transform(ngramDataFrame)
转型的输出:
+-----+--------------------+--------------------+--------------------+------ --------------+--------------------+
|label| feature| words| filtered| ngrams| hash|
+-----+--------------------+--------------------+--------------------+------ --------------+--------------------+
| 1|combusting prepar...|[combusting, prep...|[combusting, prep...| [combusting prepa...|(1000,[124,161,69...|
| 1|adhesives for ind...|[adhesives, for, ...|[adhesives, indus...| [adhesives indust...|(1000,[451,604],[...|
| 1| | []| []| []| (1000,[],[])|
| 1| salt for preserving|[salt, for, prese...| [salt, preserving]| [salt preserving]| (1000,[675],[1.0])|
| 1|auxiliary fluids ...|[auxiliary, fluid...|[auxiliary, fluid...|[auxiliary fluids...|(1000,[661,696,89...|
要构建朴素贝叶斯模型,我需要将标签和要素转换为LabelPoint
。以下方法我尝试将数据帧转换为RDD并创建labelpoint:
val rddData = featurizedData.select("label","hash").rdd
val trainData = rddData.map { line =>
val parts = line.split(',')
LabeledPoint(parts(0), parts(1))
}
val rddData = featurizedData.select("label","hash").rdd.map(r => (Try(r(0).asInstanceOf[Integer]).get.toDouble, Try(r(1).asInstanceOf[org.apache.spark.mllib.linalg.SparseVector]).get))
val trainData = rddData.map { line =>
val parts = line.split(',')
LabeledPoint(parts(0).toDouble, Vectors.dense(parts(1).split(',').map(_.toDouble)))
}
我收到以下错误:
scala> val trainData = rddData.map { line =>
| val parts = line.split(',')
| LabeledPoint(parts(0).toDouble, Vectors.dense(parts(1).split(',').map(_.toDouble)))
| }
<console>:67: error: value split is not a member of (Double, org.apache.spark.mllib.linalg.SparseVector)
val parts = line.split(',')
^
<console>:68: error: not found: value Vectors
LabeledPoint(parts(0).toDouble, Vectors.dense(parts(1).split(',').map(_.toDouble)))
编辑1:
根据以下建议,我创建了LabelPoint并训练模型。
val trainData = featurizedData.select("label","features")
val trainLabel = trainData.map(line => LabeledPoint(Try(line(0).asInstanceOf[Integer]).get.toDouble,Try(line(1).asInsta nceOf[org.apache.spark.mllib.linalg.SparseVector]).get))
val splits = trainLabel.randomSplit(Array(0.8, 0.2), seed = 11L)
val training = splits(0)
val test = splits(1)
val model = NaiveBayes.train(training, lambda = 1.0, modelType = "multinomial")
val predictionAndLabels = test.map { point =>
val score = model.predict(point.features)
(score, point.label)}
使用N-gram和没有N-gram以及不同的哈希特征数,我的准确率降低了大约40%。我的数据集包含5000行和45个mutlinomial标签。有没有办法改善模型性能?提前致谢
答案 0 :(得分:1)
您无需将featurizedData
转换为RDD
,因为Apache Spark
有两个库ML
和MLLib
,第一个可以使用DataFrame
s,而MLLib
使用RDD
s。因此,您可以使用ML
,因为您已经拥有DataFrame
。
为了实现这一目标,您只需将列重命名为(label
,features
),并使其适合您的模型,如NaiveBayes中所示,示例如下所示。< / p>
df = sqlContext.createDataFrame([
Row(label=0.0, features=Vectors.dense([0.0, 0.0])),
Row(label=0.0, features=Vectors.dense([0.0, 1.0])),
Row(label=1.0, features=Vectors.dense([1.0, 0.0]))])
nb = NaiveBayes(smoothing=1.0, modelType="multinomial")
model = nb.fit(df)
关于您获得的错误,原因是您已经拥有SparseVector
,并且该类没有split
方法。因此,考虑更多相关信息,RDD
几乎具有您实际需要的结构,但您必须将Tuple
转换为LabeledPoint
。
有一些技术可以提高性能,我想到的第一个技术是删除停用词(例如,a,an,to,尽管等等),第二个是计算数字您的文本中的不同单词,然后手动构造向量,即这是因为如果散列数低,则不同的单词可能具有相同的散列,因此性能较差。