def find_best_weights(train, test, clf, thresh=0.5):
y_pred_cm = run_cm(train, test, '../data/ADR-lexicon.txt')
_, y_prob_tfidf = run_tfidf(train, test, grams='123', n_dim=40000, clf=clf)
_, y_prob_nblcr = run_nblcr(train, test, '../data/nblcr', grams='123', clf=clf)
_, y_prob_we = run_we(train, test, '../data/w2v_150.txt', 150, clf=clf)
alphas = np.float32(np.linspace(0, 1, 21))
max_f1 = 0
best_weights = [0, 0, 0]
for alpha1 in alphas:
for alpha2 in alphas:
for alpha3 in alphas:
if alpha1 + alpha2 + alpha3 > 1: continue
y_pred = []
for i in xrange(len(y_pred_cm)):
val = alpha1*y_pred_cm[i] + alpha2*y_prob_tfidf[i,1] + alpha3*y_prob_nblcr[i,1] + (1-alpha1-alpha2-alpha3)*y_prob_we[i,1]
if val >= thresh: y_pred.append(1)
else: y_pred.append(0)
f1 = f1_score(test['label'], y_pred)
if f1 > max_f1:
best_weights = [alpha1, alpha2, alpha3]
max_f1 = f1
return best_weights, max_f1
def find_best_weights(train, test, clf, thresh=0.5):
y_pred_cm = run_cm(train, test, '../data/ADR-lexicon.txt')
_, y_prob_tfidf = run_tfidf(train, test, grams='123', n_dim=40000, clf=clf)
_, y_prob_nblcr = run_nblcr(train,
test,
'../data/nblcr',
grams='123',
clf=clf)
_, y_prob_we = run_we(train, test, '../data/w2v_150.txt', 150, clf=clf)
alphas = np.float32(np.linspace(0, 1, 21))
max_f1 = 0
best_weights = [0, 0, 0]
for alpha1 in alphas:
for alpha2 in alphas:
for alpha3 in alphas:
if alpha1 + alpha2 + alpha3 > 1: continue
y_pred = []
for i in xrange(len(y_pred_cm)):
val = alpha1 * y_pred_cm[i] + alpha2 * y_prob_tfidf[
i, 1] + alpha3 * y_prob_nblcr[i, 1] + (
1 - alpha1 - alpha2 - alpha3) * y_prob_we[i, 1]
if val >= thresh: y_pred.append(1)
else: y_pred.append(0)
f1 = f1_score(test['label'], y_pred)
if f1 > max_f1:
best_weights = [alpha1, alpha2, alpha3]
max_f1 = f1
return best_weights, max_f1
def run_ensemble(train, test, weights, clf, thresh=0.5):
y_pred_cm = run_cm(train, test, '../data/ADR-lexicon.txt')
_, y_prob_tfidf = run_tfidf(train, test, grams='123', n_dim=40000, clf=clf)
_, y_prob_nblcr = run_nblcr(train, test, '../data/nblcr', grams='123', clf=clf)
_, y_prob_we = run_we(train, test, '../data/w2v_150.txt', 150, clf=clf)
y_pred = []
for i in xrange(len(y_pred_cm)):
val = weights[0]*y_pred_cm[i] + weights[1]*y_prob_tfidf[i,1] + weights[2]*y_prob_nblcr[i,1] + (1-weights[0]-weights[1]-weights[2])*y_prob_we[i,1]
if val >= thresh: y_pred.append(1)
else: y_pred.append(0)
return y_pred
def run_ensemble(train, test, weights, clf, thresh=0.5):
y_pred_cm = run_cm(train, test, '../data/ADR-lexicon.txt')
_, y_prob_tfidf = run_tfidf(train, test, grams='123', n_dim=40000, clf=clf)
_, y_prob_nblcr = run_nblcr(train,
test,
'../data/nblcr',
grams='123',
clf=clf)
_, y_prob_we = run_we(train, test, '../data/w2v_150.txt', 150, clf=clf)
y_pred = []
for i in xrange(len(y_pred_cm)):
val = weights[0] * y_pred_cm[i] + weights[1] * y_prob_tfidf[
i, 1] + weights[2] * y_prob_nblcr[i, 1] + (
1 - weights[0] - weights[1] - weights[2]) * y_prob_we[i, 1]
if val >= thresh: y_pred.append(1)
else: y_pred.append(0)
return y_pred
