def fit(self, df_original, topics):
#Create Dictionary
self.dictionary = self._create_dictionary(df_original)
#Create corpus
self.corpus = self._create_corpus(df_original)
#Train Model
hdp = HdpModel(self.corpus, id2word=self.dictionary, T=topics)
self.model = hdp.suggested_lda_model()
feature_vecs = []
for i in range(len(self.corpus)):
top_topics = self.model.get_document_topics(
self.corpus[i], minimum_probability=0.0)
topic_vec = [0] * topics
for j in top_topics:
index = j[0]
topic_vec[index] = j[1]
feature_vecs.append(topic_vec)
df_lda_reduced = pd.DataFrame(feature_vecs,
columns=list(range(len(
feature_vecs[0]))))
df_lda_reduced.insert(0, 'Name', list(df_original['Name'].values),
False)
df_lda_reduced = df_lda_reduced.sort_values(by=['Name'])
return df_lda_reduced
def test(rat, res, nt):
global origCorpus
corpus = origCorpus.copy()
id_topic_ratio = rat
resistance = res
done = False
numTops = nt
topicPath = "data\\topics_init" + str(nt) + "_rat"+str(rat)+"_res_"+str(res)+".csv"
relationPath = "data\\relations_init" + str(nt) + "_rat"+str(rat)+"_res_"+str(res)+".csv"
topicFile = open(topicPath, 'w')
topicOut = csv.writer(topicFile, lineterminator = '\n')
topicOut.writerow(["", "run", "topic", "terms", "p"])
relationFile = open(relationPath, 'w')
relationOut = csv.writer(relationFile, lineterminator = '\n')
relationOut.writerow(["run", "topic", "no IDs", "ID/strength"])
run = 1
totalTopics = 0
averageCoherence = 0
badIDs = docIDs
while not done:
print("Run #" + str(run))
doc2topic = dict()
topic2doc = dict()
oldIDs = badIDs.copy()
badIDs = list()
totalTopics += numTops
#perform LDA
hdp = HdpModel(corpus, dictionary, T=numTops)
lda_model = hdp.suggested_lda_model()
coherenceModel = CoherenceModel(model=lda_model, texts=data, dictionary=dictionary, coherence='c_v')
coherence = coherenceModel.get_coherence()
averageCoherence = ((totalTopics-numTops) * averageCoherence + numTops*coherence)/totalTopics
# tag documents
for ID in oldIDs:
doc = docTokens[ID]
vec = dictionary.doc2bow(doc)
store = lda_model[vec]
bestRel = 0
# build relations
for pair in store:
bestRel = max(bestRel, pair[1])
if pair[0] in topic2doc:
topic2doc[pair[0]] += [(ID, pair[1])]
else:
topic2doc[pair[0]] = [(ID, pair[1])]
# collect bad docs
if bestRel < resistance:
badIDs.append(ID)
#write terms
top_words_per_topic = []
for t in range(lda_model.num_topics):
top_words_per_topic.extend([(run, t, ) + x for x in lda_model.show_topic(t, topn = 10)])
terms = pd.DataFrame(top_words_per_topic, columns=['Run', 'Topic', 'Word', 'P']).to_csv(topicPath, mode='a', header=False)
# print relations
for topic in topic2doc:
relationOut.writerow([run, topic, len(topic2doc[topic])]+ sorted(topic2doc[topic], key=operator.itemgetter(1), reverse=True))
# done?
if len(badIDs) == 0:
done = True
print("Done!")
# if not, build new corpus
else:
print("Remaining: " + str(len(badIDs)))
corpus = [dictionary.doc2bow(docTokens[docID]) for docID in badIDs]
len(corpus)
numTops = math.ceil(len(badIDs) * id_topic_ratio)
run += 1
results.writerow([nt, rat, res, averageCoherence, totalTopics])
topicFile.close()
relationFile.close()
