def LDA_rep(window_data, n_topics=10, max_iters=20):
#data is in sliding window form
ui_matrix = window_data[["user", "artist", "sum"]]\
.pivot(index = "user", columns = "artist", values = "sum").fillna(0)
iu_matrix = window_data[["user", "artist", "sum"]]\
.pivot(index = "artist", columns = "user", values = "sum").fillna(0)
ui_lda = LatentDirichletAllocation(n_components=n_topics,
max_iter=max_iters,
learning_method='online',
learning_offset=5.).fit(ui_matrix)
iu_lda = LatentDirichletAllocation(n_components=n_topics,
max_iter=max_iters,
learning_method='online',
learning_offset=5.).fit(iu_matrix)
ti_matrix = ui_lda.components_
ut_matrix = ui_lda.fit_transform(ui_matrix)
tu_matrix = iu_lda.components_
it_matrix = iu_lda.fit_transform(iu_matrix)
res = {
"ui": ui_matrix,
"ti": ti_matrix,
"ut": ut_matrix,
"tu": tu_matrix,
"it": it_matrix,
"users": ui_matrix.index,
"artists": iu_matrix.index
}
res["user_sim"] = hf.LDASim(res["ut"])
res["artist_sim"] = hf.LDASim(res["it"])
return (res)
tf = iu_matrix / (iu_matrix.sum(axis=1)[:, np.newaxis])
idf = (iu_matrix > 0) / (iu_matrix > 0).sum(axis=0)
iu_matrix_tfidf = np.multiply(tf, idf)
#transform item_user matrix to item_topic and topic_user matrix
iu_lda = LatentDirichletAllocation(n_topics=n_topics,
max_iter=iters,
learning_method='online',
learning_offset=50.,
random_state=0).fit(iu_matrix_tfidf)
tu_matrix = iu_lda.components_
it_matrix = iu_lda.fit_transform(iu_matrix_tfidf)
user_sim = hf.LDASim(ut_matrix)
item_sim = hf.LDASim(it_matrix)
user_div = hf.LDADiversity(ut_matrix)
hf.LDAClustering(item_sim, ui_matrix)
hf.draw_graph(graph)
#Clustering bipartite graph
#forming bipartite graph W
m, n = ui_matrix.shape
axis_label = np.concatenate((ui_matrix.index, ui_matrix.columns))
W = np.block([[np.zeros((m, m)), ui_matrix], [ui_matrix.T, np.zeros((n, n))]])
D = np.diag(W.sum(axis=0))
W_scaled = np.linalg.inv(D)**(1 / 2) @ W @ np.linalg.inv(D)**(1 / 2)
