Kratos Kratos - 20 days ago 15
Python Question

PrefixSpan sequence extraction misunderstanding

I have a set of tuples of size three in a list that represent windowed sequences.
What I need is using pyspask to be able to get (given the two first parts of the tuple) the third one.

So I need it to create sequences of three elements based on their frequency.

This is what I am doing:

data = [[['a','b','c'],['b','c','d'],['c','d','e'],['d','e','f'],['e','f','g'],['f','g','h'],['a','b','c'],['d','e','f'],['a','b','c'],['b','c','d'],['f','g','h'],['d','e','f'],['b','c','d']]]
rdd = spark.sparkContext.parallelize(data,2)
rdd.cache()
model = PrefixSpan.train( rdd, 0.2, 3)

print(sorted(model.freqSequences().take(100)))


Although, I would expect to see the sequences and the frequencies o them to follow the alphabet, they don't.

And I am getting sequences like:

FreqSequence(sequence=[[u'c'], [u'd'], [u'b']], freq=1)
FreqSequence(sequence=[[u'g'], [u'c'], [u'c']], freq=1)


which are not appearing in the defined ones. Obviously there is a problem in the way I have structure my features or I am missing something in the purpose and functionality of this algorithm..

Thank you!

Answer

First let's look at your input:

rdd.count()
1

As you can see you created a dataset with only one sequence. It can be described as:

<(abc)(bcd)(cde)(def)(efg)(fgh)(abc)(def)(abc)(bcd)(fgh)(def)(bcd)>

So patterns you get are indeed correct given the input. For example

FreqSequence(sequence=[[u'c'], [u'd'], [u'b']], freq=1)

corresponds to:

...(abc)(def)(abc)...

If each element of the dataset represents individual sequence data could have the following shape:

rdd = sc.parallelize([
    [['a'], ['b'], ['c']], [['b'], ['c'], ['d']], [['c'], ['d'], ['e']],
    [['d'], ['e'], ['f']], [['e'], ['f'], ['g']], [['f'], ['g'], ['h']],
    [['a'], ['b'], ['c']], [['d'], ['e'], ['f']], [['a'], ['b'], ['c']],
    [['b'], ['c'], ['d']], [['f'], ['g'], ['h']], [['d'], ['e'], ['f']],
    [['b'], ['c'], ['d']]
])

rdd.count()
13
rdd.first()
[['a'], ['b'], ['c']]

where:

  • Each element is a list of lists.
  • Each internal list represents possible alternatives at the given position.

With data structured like this:

model = PrefixSpan.train(rdd, 0.2, 3)
model.freqSequences().top(5, key=lambda x: len(x.sequence))
[FreqSequence(sequence=[['d'], ['e'], ['f']], freq=3),
 FreqSequence(sequence=[['b'], ['c'], ['d']], freq=3),
 FreqSequence(sequence=[['a'], ['b'], ['c']], freq=3),
 FreqSequence(sequence=[['f'], ['g']], freq=3),
 FreqSequence(sequence=[['d'], ['f']], freq=3)]
model.freqSequences().top(5, key=lambda x: x.freq)
[FreqSequence(sequence=[['d']], freq=7),
 FreqSequence(sequence=[['c']], freq=7),
 FreqSequence(sequence=[['f']], freq=6),
 FreqSequence(sequence=[['b']], freq=6),
 FreqSequence(sequence=[['b'], ['c']], freq=6)]
Comments