Use "log lift" to display top topic terms in `topicexplorer launch`

[JaimieMurdock]

lift = p(w|k) / p(w)
log lift = log p(w/k) - log p(w)

Should give a better indicator of what makes a particular topic unique.

[JaimieMurdock]

An early implementation...

from corpus import *
from collections import Counter
import numpy as np

v = lda_v[20]

counts = Counter(c.corpus)
total_len = len(c.corpus)

counts_a = np.array([counts[i] / total_len for i in range(v.phi.T.shape[1])])

def get_lift(k, term_i):
    return np.log(v.phi.T[k, term_i]) - np.log(counts[term_i] / total_len)

for i in range(v.phi.T.shape[0]):
    lift_sort = np.log(v.phi.T[i]) - np.log(counts_a / total_len)
    #lift_sort[lift_sort.argsort()]
    print(i, ', '.join(c.words[lift_sort.argsort()][-15:]), sep='\t')