def plot_margin_histograms(self):
for (lbl, group) in zip(['buyer_wins', 'seller_wins'],
[self.buyer_wins, self.seller_wins]):
margins = []
for ex in group:
winner = utils.get_winner(ex)
if winner is None:
continue
margin = utils.get_margin(ex)
if 0 <= margin <= MAX_MARGIN:
margins.append(margin)
b = np.linspace(0, MAX_MARGIN, num=int(MAX_MARGIN / 0.2) + 2)
print b
hist, bins = np.histogram(margins, bins=b)
width = np.diff(bins)
center = (bins[:-1] + bins[1:]) / 2
fig, ax = plt.subplots(figsize=(8, 3))
ax.bar(center, hist, align='center', width=width)
ax.set_xticks(bins)
save_path = os.path.join(
self.stats_path, '{:s}_wins_margins_histogram.png'.format(lbl))
plt.savefig(save_path)
def plot_length_vs_margin(self, out_name='turns_vs_margin.png'):
labels = ['buyer wins', 'seller wins']
plt.figure(figsize=(10, 6))
for (chats, lbl) in zip([self.buyer_wins, self.seller_wins], labels):
margins = defaultdict(list)
for ex in chats:
turns = utils.get_turns_per_agent(ex)
total_turns = turns[0] + turns[1]
margin = utils.get_margin(ex)
if margin > MAX_MARGIN or margin < 0.:
continue
margins[total_turns].append(margin)
sorted_keys = list(sorted(margins.keys()))
turns = []
means = []
errors = []
for k in sorted_keys:
if len(margins[k]) >= THRESHOLD:
turns.append(k)
means.append(np.mean(margins[k]))
errors.append(stats.sem(margins[k]))
plt.errorbar(turns, means, yerr=errors, label=lbl, fmt='--o')
plt.legend()
plt.xlabel('# of turns in dialogue')
plt.ylabel('Margin of victory')
save_path = os.path.join(self.stats_path, out_name)
plt.savefig(save_path)
def plot_speech_acts_old(self):
labels = ['buyer_wins', 'seller_wins']
for (group, lbl) in zip([self.buyer_wins, self.seller_wins], labels):
plt.figure(figsize=(10, 6))
speech_act_counts = dict(
(act, defaultdict(list)) for act in SpeechActs.ACTS)
for chat in group:
winner = utils.get_winner(chat)
margin = utils.get_margin(chat)
if margin > MAX_MARGIN or margin < 0.:
continue
if winner is None:
continue
margin = round_partial(
margin
) # round the margin to the nearest 0.1 to reduce noise
if winner == -1 or winner == 0:
speech_acts = self.get_speech_acts(chat, agent=0)
# print "Chat {:s}\tWinner: {:d}".format(chat['uuid'], winner)
# print speech_acts
for act in SpeechActs.ACTS:
frac = float(speech_acts.count(act)) / float(
len(speech_acts))
speech_act_counts[act][margin].append(frac)
if winner == -1 or winner == 1:
speech_acts = self.get_speech_acts(chat, agent=1)
# print "Chat {:s}\tWinner: {:d}".format(chat['uuid'], winner)
# print speech_acts
for act in SpeechActs.ACTS:
frac = float(speech_acts.count(act)) / float(
len(speech_acts))
speech_act_counts[act][margin].append(frac)
for act in SpeechActs.ACTS:
counts = speech_act_counts[act]
margins = []
fracs = []
errors = []
bin_totals = 0.
for m in sorted(counts.keys()):
if len(counts[m]) > THRESHOLD:
bin_totals += len(counts[m])
margins.append(m)
fracs.append(np.mean(counts[m]))
errors.append(stats.sem(counts[m]))
print bin_totals / float(len(margins))
plt.errorbar(margins, fracs, yerr=errors, label=act, fmt='--o')
plt.xlabel('Margin of victory')
plt.ylabel('Fraction of speech act occurences')
plt.title('Speech act frequency vs. margin of victory')
plt.legend()
save_path = os.path.join(self.stats_path,
'{:s}_speech_acts.png'.format(lbl))
plt.savefig(save_path)
def plot_price_trends(self, top_n=10):
labels = ['buyer_wins', 'seller_wins']
for (group, lbl) in zip([self.buyer_wins, self.seller_wins], labels):
plt.figure(figsize=(10, 6))
trends = []
for chat in group:
winner = utils.get_winner(chat)
margin = utils.get_margin(chat)
if margin > 1.0 or margin < 0.:
continue
if winner is None:
continue
# print "Winner: Agent {:d}\tWin margin: {:.2f}".format(winner, margin)
if winner == -1 or winner == 0:
trend = self.get_price_trend(self.price_tracker,
chat,
agent=0)
if len(trend) > 1:
trends.append((margin, chat, trend))
if winner == -1 or winner == 1:
trend = self.get_price_trend(self.price_tracker,
chat,
agent=1)
if len(trend) > 1:
trends.append((margin, chat, trend))
# print ""
sorted_trends = sorted(trends, key=lambda x: x[0], reverse=True)
for (idx, (margin, chat,
trend)) in enumerate(sorted_trends[:top_n]):
print '{:s}: Chat {:s}\tMargin: {:.2f}'.format(
lbl, chat['uuid'], margin)
print 'Trend: ', trend
print chat['scenario']['kbs']
print ""
plt.plot(trend, label='Margin={:.2f}'.format(margin))
plt.legend()
plt.xlabel('N-th price mentioned in chat')
plt.ylabel('Value of mentioned price')
out_path = os.path.join(self.stats_path,
'{:s}_trend.png'.format(lbl))
plt.savefig(out_path)
def plot_speech_acts_by_role(self):
labels = utils.ROLES
for lbl in labels:
plt.figure(figsize=(10, 6))
speech_act_counts = dict(
(act, defaultdict(list)) for act in SpeechActs.ACTS)
for chat in self.dataset:
if utils.get_winner(chat) is None:
# skip chats with no outcomes
continue
speech_acts = self.get_speech_acts(chat, role=lbl)
agent = 1 if chat['scenario']['kbs'][1]['personal'][
'Role'] == lbl else 0
margin = utils.get_margin(chat, agent=agent)
if margin > MAX_MARGIN:
continue
margin = round_partial(margin)
for act in SpeechActs.ACTS:
frac = float(speech_acts.count(act)) / float(
len(speech_acts))
speech_act_counts[act][margin].append(frac)
for act in SpeechActs.ACTS:
counts = speech_act_counts[act]
margins = []
fracs = []
errors = []
for m in sorted(counts.keys()):
if len(counts[m]) > THRESHOLD:
margins.append(m)
fracs.append(np.mean(counts[m]))
errors.append(stats.sem(counts[m]))
plt.errorbar(margins, fracs, yerr=errors, label=act, fmt='--o')
plt.xlabel('Margin of victory')
plt.ylabel('Fraction of speech act occurences')
plt.title('Speech act frequency vs. margin of victory')
plt.legend()
save_path = os.path.join(self.stats_path,
'{:s}_speech_acts.png'.format(lbl))
plt.savefig(save_path)
def __init__(self, transcripts, featurizer, split=0.8):
self.featurizer = featurizer
X = [featurizer.create_feature_vector(t) for t in transcripts]
margins = [utils.get_margin(t, role=utils.BUYER) for t in transcripts]
new_X = []
y = []
for (idx, margin) in enumerate(margins):
if margin >= self.POSITIVE_THRESHOLD or margin < self.NEGATIVE_THRESHOLD:
new_X.append(X[idx])
y.append(margin)
print "# of chats with buyer margin >= {:.1f}: {:d}".format(self.POSITIVE_THRESHOLD, y.count(1))
print "# of chats with buyer margin < {:.1f}: {:d}".format(self.NEGATIVE_THRESHOLD, y.count(0))
X, y = self.shuffle_data(X, y)
self.train_X = None
self.train_y = None
self.test_X = None
self.test_y = None
self.split(X, y, split=split)
print "# of training examples: {:d}".format(self.train_X.shape[0])
print "# of test examples: {:d}".format(self.test_X.shape[0])
self.model = LogisticRegression(C=1.5)
