Introduction to Latent Dirichlet Allocation (LDA)
In LDA model, first you need to create a vocabulary on probabilistic term distribution over each topic using a set of training documents.
In a simple scenario, assume there are 2 documents in the training set and their content has following unique, important terms. (Important terms is extracted using TF vectors as I have mentioned later)
Document 1: "car", "hybrid", "Toyota"
Document 2: "birds", "parrot", "Sri Lanka"
Using the above terms, LDA creates a vocabulary on probabilistic term distribution over each topic as given below: We define that we need to form 2 topics from this training content.
Topic 1: car: 0.7, hybrid: 0.1, Toyota: 0.1, birds: 0.02, parrot: 0.03, Sri Lanka: 0.05
 |
| Topic 1: Term-Topic distribution |
Topic 2: car: 0.05, hybrid: 0.03, Toyota: 0.02, birds: 0.4, parrot: 0.5, Sri Lanka: 0.1
 |
| Topic 2: Term-Topic distribution |
The topic model is created based on above training data which will be later used for inference.
For a new document, you need to infer the probabilistic topic distribution over document. Assume the document content is as follows:
Document 3: "Toyota", "Prius", "Hybrid", "For sale", "2003"
For the above document, probabilistic topic distribution over document will (roughly!) be a value like this:
Topic 1: 0.99, Topic 2: 0.01
 |
| Topic distribution over the new document |
So, we can use the terms in the topics with high probability (E.g., car, hybrid) as metadata for the document which can be used in different applications such as search indexing, document clustering, business analytic etc.
Pre-processing
Preparing input TF vectors
To bring out the important words within a document, we normally use TF-IDF vectors. However, in LDA, TF vectors are used instead of TF-IDF words to recognize the co-occurrence or correlation between words.
(In vector space model [VSM] it is assumed that occurrences of the words are independent of each other, but this assumption is wrong in many cases! n-gram generation is a solution for this problem)
Convert input documents to SequenceFile format
sequence file is a flat file consisting of binary key value pairs. This is used as input/ output file format for map-reduce jobs in Hadoop, which is the underlying framework which Mahout is running on.
Configuration conf = new Configuration();
HadoopUtil.delete(conf, new Path(infoDirectory));
SequenceFilesFromDirectory sfd = new SequenceFilesFromDirectory();
// input: directory contains number of text documents
// output: the directory where the sequence files will be created
String[] para = { "-i", targetInputDirectoryPath, "-o", sequenceFileDirectoryPath };
sfd.run(para);
Convert sequence files to TF vectors
Configuration conf = new Configuration();
Tokenization and Analyzing
During the tokenization, document content will be split in to set of terms/tokens. Different analyzers may use different tokenizers. Stemming and removing stop words can be done and customized in this stage. Please note that both stemming and stop words are language dependent.
You can specify your own analyzer if you want, specifying on how you
want the terms to be extracted. That has to be extended by the Lucene
Analyzer class.
StandardAnalyzer analyzer = new StandardAnalyzer(Version.LUCENE_43);
DocumentProcessor.tokenizeDocuments(new Path(sequenceFileinputDirectoryPath + "/" + "part-m-00000"), analyzer.getClass().asSubclass(Analyzer.class),
new Path(infoDirectory + "/" + DocumentProcessor.TOKENIZED_DOCUMENT_OUTPUT_FOLDER), conf);
analyzer.close();
There are couple of important parameters for generating TF vectors.
In mahout, DictionaryVectorizer class is used for TF weighting and n-gram collocation.
// Minimum frequency of the term in the entire collection to be considered as part of the dictionary file. Terms with lesser frequencies are ignored.
int minSupport = 5;
// Maximum size of n-grams to be selected. For more information, visit:
ngram collocation in Mahout
int maxNGramSize = 2;
// Minimum log likelihood ratio (This is related to ngram collocation. Read more
here.)
// This work only when maxNGramSize > 1 (Less significant ngrams have lower score here)
float minLLRValue = 50;
// Parameters for Hadoop map reduce operations
int reduceTasks = 1;
int chunkSize = 200;
boolean sequentialAccessOutput = true;
DictionaryVectorizer.createTermFrequencyVectors(new Path(infoDirectory + DocumentProcessor.TOKENIZED_DOCUMENT_OUTPUT_FOLDER),
new Path(infoDirectory), DictionaryVectorizer.DOCUMENT_VECTOR_OUTPUT_FOLDER, conf, minSupport, maxNGramSize, minLLRValue,
-1.0f, false, reduceTasks, chunkSize, sequentialAccessOutput, true);
Once the TF vectors are generated for each training document, the model can be created.
Training
Generate term distribution for each topic and generate topic distribution for each training document
(Read about the CVB algorithm in mahout here.)
CVB0Driver cvbDriver = new CVB0Driver();
I will explain the parameters and how you need to assign them values. Before that you need to read the training dictionary in to memory as given below:
Configuration conf = new Configuration();
FileSystem fs = FileSystem.get(conf);
SequenceFile.Reader reader = new SequenceFile.Reader(fs, new Path(
dictionaryFilePath), conf);
Text key = new Text();
IntWritable val = new IntWritable();
ArrayList
dictLst = new ArrayList();
while (reader.next(key,val)) {
System.out.println(key.toString()+" "+val.toString());
dictLst.add(key.toString());
}
String[] dictionary = new String[dictLst.size()];
dictionary = dictLst.toArray(dictionary);
Then, you have to convert vector representation of documents to a matrix, like this.
RowIdJob rowidjob = new RowIdJob();
String[] para = { "-i", inputVectorPath, "-o",
TRAINING_DOCS_OUTPUTMATRIX_PATH };
rowidjob.run(para);
Now, I will explain each parameters and factors you should consider on deciding values.
// Input path to the above created matrix using TF vectors
Path inputPath = new Path(TRAINING_DOCS_OUTPUTMATRIX_PATH + "/matrix");
// Path to save the model (Note: You may need this during inferring new documents)
Path topicModelOutputPath = new Path(TRAINING_MODEL_PATH);
// Numbe of topics (#important!) Lower value results in broader topics and higher value may result in niche topics. Optimal value for this parameter can vary depending on the given use case. Large number of topics may cause the system to slowdown.
int numTopics = 2;
// Number of terms in the training dictionary. Here's the method to read that:
private static int getNumTerms(Configuration conf, Path dictionaryPath) throws IOException {
FileSystem fs = dictionaryPath.getFileSystem(conf);
Text key = new Text();
IntWritable value = new IntWritable();
int maxTermId = -1;
for (FileStatus stat : fs.globStatus(dictionaryPath)) {
SequenceFile.Reader reader = new SequenceFile.Reader(fs, stat.getPath(), conf);
while (reader.next(key, value)) {
maxTermId = Math.max(maxTermId, value.get());
}
reader.close();
}
return maxTermId + 1;
}
int numTerms = getNumTerms(conf, new Path(TRAINING_DOCS_ROOT_PATH + "dictionary.file-0"));
// Smoothing parameters for p(topic|document) prior: This value can control how term topic likelihood is calculated for each document
double alpha = 0.0001;
double eta = 0.0001;
int maxIterations = 10;
int iterationBlockSize = 10;
double convergenceDelta = 0;
Path dictionaryPath = new Path(TRAINING_DOCS_ROOT_PATH + "dictionary.file-0");
// Final output path for probabilistic topic distribution training documents
Path docTopicOutputPath = new Path(TRAINING_DOCS_TOPIC_OUTPUT_PATH);
// Temporary output path for saving models in each iteration
Path topicModelStateTempPath = new Path(TRAINING_MODEL_TEMP_PATH);
long randomSeed = 1;
// This is a measurement of how well a probability distribution or probability model predicts a sample. LDA is a generative model, you start with a known model and try to explain the data by refining parameters to fit the model of the data. These values can be taken to evaluate the performance.
boolean backfillPerplexity = false;
int numReduceTasks = 1;
int maxItersPerDoc = 10;
int numUpdateThreads = 1;
int numTrainThreads = 4;
float testFraction = 0;
cvbDriver.run(conf, inputPath, topicModelOutputPath,
numTopics, numTerms, alpha, eta, maxIterations, iterationBlockSize, convergenceDelta, dictionaryPath, docTopicOutputPath, topicModelStateTempPath, randomSeed, testFraction, numTrainThreads, numUpdateThreads, maxItersPerDoc, numReduceTasks, backfillPerplexity) ;
Once this step is completed the training phase of topic modeling is over. Now, lets see how to infer new documents using the trained model.
Topic Inference for new document
To infer topic distribution for new document, you need to follow the same steps for the new document which I have mentioned earlier.
- Pre-processing - stop word removal
- Convert the document to sequence file format
- Convert the content in the sequence file to TF vectors
There is an important step here, (Even I missed this step at the first time and got wrong results as the outcome :( )
We need to map the new document's dictionary with the training documents' dictionary and identify the common terms, that appears in both. Then, a TF vector needs to be created for the new document with the cardinality of training documents' dictionary. This is how you should do that.
//Get the model dictionary file
HashMap
modelDictionary = new HashMap<>();
SequenceFile.Reader reader = new SequenceFile.Reader(fs, new Path("reuters-dir/dictionary.file-0"), conf);
Text keyModelDict = new Text();
IntWritable valModelDict = new IntWritable();
int cardinality = 0;
while(reader.next(keyModelDict, valModelDict)){
cardinality++;
modelDictionary.put(keyModelDict.toString(), Integer.parseInt(valModelDict.toString()));
}
RandomAccessSparseVector newDocVector = new RandomAccessSparseVector(cardinality);
reader.close();
//Get the new document dictionary file
ArrayList newDocDictionaryWords = new ArrayList<>();
reader = new SequenceFile.Reader(fs, new Path("reuters-test-dir/dictionary.file-0"), conf);
Text keyNewDict = new Text();
IntWritable newVal = new IntWritable();
while(reader.next(keyNewDict,newVal)){
System.out.println("Key: "+keyNewDict.toString()+" Val: "+newVal);
newDocDictionaryWords.add(keyNewDict.toString());
}
//Get the document frequency count of the new vector
HashMap newDocTermFreq = new HashMap<>();
reader = new SequenceFile.Reader(fs, new Path("reuters-test-dir/wordcount/ngrams/part-r-00000"), conf);
Text keyTFNew = new Text();
DoubleWritable valTFNew = new DoubleWritable();
while(reader.next(keyTFNew, valTFNew)){
newDocTermFreq.put(keyTFNew.toString(), Double.parseDouble(valTFNew.toString()));
}
//perform the process of term frequency vector creation
for (String string : newDocDictionaryWords) {
if(modelDictionary.containsKey(string)){
int index = modelDictionary.get(string);
double tf = newDocTermFreq.get(string);
newDocVector.set(index, tf);
}
}
System.out.println(newDocVector.asFormatString());
- Read the model (Term distribution for each topic)
// Dictionary is the training dictionary
double alpha = 0.0001; // default: doc-topic smoothing
double eta = 0.0001; // default: term-topic smoothing
double modelWeight = 1f;
TopicModel model = new TopicModel(conf, eta, alpha, dictionary, 1, modelWeight, TRAINING_MODEL_PATH));
- Infer topic distribution for the new document
The final result, which is probabilistic topic distribution over new document will be stored in this vector
If you have a prior guess as to what the topic distribution should be, you can start with it here, instead of the uniform prior
Vector docTopics = new DenseVector(new double[model.getNumTopics()]).assign(1.0/model.getNumTopics());
Empty matrix holding intermediate data - Term-Topic likelihoods for each term in the new document will be stored here.
Matrix docTopicModel = new SparseRowMatrix(model.getNumTopics(), newDocVector.size());
int maxIters = 100;
for(int i = 0; i < maxIters; i++) {
model.trainDocTopicModel(newDocVector, docTopics, docTopicModel);
}
model.stop();
To be continued...
References: Mahout In Action, Wikipedia
