Text classification - how to approach

问题:

I'll try do describe what I have in mind.

There is a text content stored in MS SQL database. Content comes daily as a stream. Some people go through the content every day and, if the content fits certain criteria, mark it as validated. There is only one category. It's either "valid" or not.

What I want is to create a model based on already validated content, save it and using this model to "pre-validate" or mark new incoming content. Also once in a while to update the model based on a newly validated content. Hopefully I explained myself clearly.

I am thinking to use Spark streaming for data classification based on model created. And Naive Bayes algorithm. But how would you approach creating, updating and storing the model? There is ~200K+ validated resultants (texts) of various length. Do I need so many for the model? And how to use this model in Spark Streaming.

Thanks in advance.

回答1:

Wow this question is very broad, and more related to Machine Learning than to Apache Spark, however I will try to give you some hints or steps to follow (I won't do the work for you).

1. Import all the libraries you need

   from pyspark.mllib.classification import LogisticRegressionWithSGD, LogisticRegressionModel
   from pyspark.mllib.linalg import SparseVector
   from pyspark.mllib.regression import LabeledPoint
   import re
   

2. Load your data into an RDD

   msgs = [("I love Star Wars but I can't watch it today", 1.0),
           ("I don't love Star Wars and people want to watch it today", 0.0),
           ("I dislike not being able to watch Star Wars", 1.0),
           ("People who love Star Wars are my friends", 1.0),
           ("I preffer to watch Star Wars on Netflix", 0.0),
           ("George Lucas shouldn't have sold the franchise", 1.0),
           ("Disney makes better movies than everyone else", 0.0)]
   
   rdd = sc.parallelize(msgs)
   

3. Tokenize your data (if you use ML it might be easier) and

   rdd = rdd.map(lambda (text, label): ([w.lower() for w in re.split(" +", text)], label))
   

4. Remove all unnecessary words (widely known as stop-words), and symbols e.g. ,.&

   commons = ["and", "but", "to"]
   rdd = rdd.map(lambda (tokens, label): (filter(lambda token: token not in commons, tokens), label))
   

5. Create a dictionary with all distinct words in all your dataset, it sounds huge but they are not so many as you would expect, and I bet they will fit in your master node (however there are other ways to approach this, but for simplicity I will keep this way).

   # finds different words
   words = rdd.flatMap(lambda (tokens, label): tokens).distinct().collect()
   diffwords = len(words)
   

6. Convert your features into a DenseVector or SparseVector, I would obviously recommend the second way because normally a SparseVector requires less space to be represented, however it depends on the data. Note, there are better alternatives like hashing, but I am trying to keep loyal to my verbose approach. After that transform the tuple into a LabeledPoint

   def sparsify(length, tokens):
       indices = [words.index(t) for t in set(tokens)]
       quantities = [tokens.count(words[i]) for i in indices]
   
       return SparseVector(length, [(indices[i], quantities[i]) for i in xrange(len(indices))])
   
   rdd = rdd.map(lambda (tokens, label): LabeledPoint(label, sparsify(diffwords, tokens)))
   

7. Fit your favorite model, in this case I used LogisticRegressionWithSGD due ulterior motives.

   lrm = LogisticRegressionWithSGD.train(rdd)
   

8. Save your model.

   lrm.save(sc, "mylovelymodel.model")
   

9. Load your LogisticRegressionModel in another application.

   lrm = LogisticRegressionModel.load(sc, "mylovelymodel.model")
   

10. Predict the categories.

    lrm.predict(SparseVector(37,[2,4,5,13,15,19,23,26,27,29],[1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0]))
    # outputs 0
    

Note that I didn't evaluate the accuracy of the model, however it looks very doesn't it?