def richrank(cities, names):
candls = ['-', '--', '.-']
mks = ['o', '^', '*', 'v', 's']
for idx in range(len(cities)):
lms = dict()
test = Dataset()
for pid in cities[idx]:
twtp = loadrows(GEOTWEET, ('place_id', 'text', 'created_at'),
('place_id=\'{0}\''.format(pid), ),
'sample_switch_place_cate',
'order by rand() limit 110')
lms[pid] = LanguageModel(twtp['text'][:100])
for cnt in range(100, 110):
test.append({
'label': twtp['place_id'][cnt],
'lm': LanguageModel([
twtp['text'][cnt],
])
})
lmranks = list()
for twtlm in test:
lmranks.append(ranke(lms, twtlm['lm']))
lmeval = batcheval(lmranks, test['label'])
plt.plot(lmeval['pos'],
lmeval['rate'],
ls=candls[idx % 2],
marker=mks[idx / 2],
label='{0}($s=100$)'.format(names[idx]))
plt.legend(loc='lower right')
plt.ylabel('Rate containing referece POI')
plt.xlabel('Top $p$ places')
plt.show()
def richrank(cities, names):
candls = ['-', '--']
mks = ['o', '^', '*']
for idx in range(len(cities)):
lms = dict()
test = Dataset()
for pid in cities[idx]:
twtp = loadrows(GEOTWEET, ('place_id', 'text', 'created_at'),
('place_id=\'{0}\''.format(pid),), 'sample',
'order by rand() limit 110')
lms[pid] = LanguageModel(twtp['text'][:100])
for cnt in range(100, 110):
test.append({'label': twtp['place_id'][cnt],
'lm': LanguageModel([twtp['text'][cnt],])})
lmranks = list()
randranks = list()
for twtlm in test:
lmranks.append(ranke(lms, twtlm['lm']))
randranks.append(randranke(cities[idx]))
lmeval = batcheval(lmranks, test['label'])
print names[idx], 'P@1', (lmeval['rate'][1] - 0.1)
plt.plot(lmeval['pos'], lmeval['rate'], ls=candls[idx%2], marker=mks[idx/2],
label='{0}($s=100$)'.format(names[idx]))
plt.plot(lmeval['pos'], [float(r) / max(lmeval['pos']) for r in lmeval['pos']],
ls='-.', marker='s',
label='Random Baseline')
plt.legend(loc='lower right')
plt.ylabel('Rate containing referece POI')
plt.xlabel('Top $p$ places')
plt.show()
def linearjoin(ranks, balance):
""" Join a set of ranks by the weights in balance
@arg ranks a list() of ranks that need to combine
@arg balance a list() of floats indicating combine weights
@return a combined rank
"""
for rak in ranks:
rak["score"] = unitscore(rak["score"])
sortedranks = [[item for item in r.sorted_items("label")] for r in ranks]
unitrank = list()
for idx in range(len(sortedranks[0])):
unitrank.append(
(
sortedranks[0][idx]["label"],
sum(
[
score * bl
for score, bl in zip([sortedranks[i][idx]["score"] for i in range(len(sortedranks))], balance)
]
),
)
)
unitrank = sorted(unitrank, key=itemgetter(1), reverse=False)
res = Dataset()
for item in unitrank:
res.append({"label": item[0], "score": item[1]})
return res
def f_tf(text):
"""get the term frequency feature
@arg text list() of str()
@return Dataset() of term vectors
"""
dset = Dataset()
for line in text:
dset.append(line2tf(line))
return dset
def randranke(cand):
""" Return a random sorted cand
"""
rak = list()
for lbl in cand:
rak.append((lbl, random.random()))
rak = sorted(rak, key=itemgetter(1))
res = Dataset()
for item in rak:
res.append({"label": item[0], "score": item[1]})
return res
def randranke(cand):
""" Return a random sorted cand
"""
rak = list()
for lbl in cand:
rak.append((lbl, random.random()))
rak = sorted(rak, key=itemgetter(1))
res = Dataset()
for item in rak:
res.append({'label': item[0], 'score': item[1]})
return res
def batcheval(ranks, expects, maxpos=-1):
""" The batch version of evaluate() which test a set of pos
@arg ranks list() of ranks returned by webguess()
@arg expects list() of expected labels
@arg maxpos the max position
@return a list of rates of goals
"""
if maxpos == -1:
maxpos = ranks[0].size()
goalrates = Dataset()
for pos in range(maxpos + 1):
goalrates.append({'pos': pos, 'rate': evaluate(ranks, expects, pos)})
return goalrates
def batcheval(ranks, expects, maxpos=-1):
""" The batch version of evaluate() which test a set of pos
@arg ranks list() of ranks returned by webguess()
@arg expects list() of expected labels
@arg maxpos the max position
@return a list of rates of goals
"""
if maxpos == -1:
maxpos = ranks[0].size()
goalrates = Dataset()
for pos in range(maxpos + 1):
goalrates.append({'pos': pos, 'rate': evaluate(ranks, expects, pos)})
return goalrates
def norm_v2(dset):
"""normalize values in vector wise, row normalization (L_2)
@arg dset Dataset() of vectors
@return Dataset() of vectors normalized
"""
ndset = Dataset()
for idx in range(dset.size()):
sqrval = math.sqrt(sum(dset[key][idx]**2 for key in dset))
item = DataItem()
for key in dset.iterkeys():
item[key] = dset[key][idx] / sqrval
ndset.append(item)
return ndset
def gen_crs_arff(self, dst, fold, key_lst=None, \
typemap=dict({'__CLASS__': 'DISC'}), \
default_type = 'NUMERIC'):
"""generate dataset for cross validation"""
clses = dict()
for i in range(len(self)):
if self[i]['__CLASS__'] not in clses:
clses[self[i]['__CLASS__']] = dict()
clses[self[i]['__CLASS__']]['list'] = list()
clses[self[i]['__CLASS__']]['list'].append(i)
for cls in clses:
random.shuffle(clses[cls]['list'])
clses[cls]['fold'] = list_split(clses[cls]['list'], fold)
for i in range(fold):
test = Dataset()
train = Dataset()
for cls in clses.iterkeys():
test.extend([self[f] for f in clses[cls]['fold'][i]])
for j in range(fold):
if j != i:
train.extend([self[f] for f in clses[cls]['fold'][j]])
gen_arff(test, '{0}.test.{1}.arff'.format(dst, i), key_lst, \
typemap, default_type)
gen_arff(train, '{0}.train.{1}.arff'.format(dst, i), key_lst, \
typemap, default_type)
def ranke(cand, ref, **kargs):
""" Rank the models in cand according to the scoring methods in the models
@arg cand dict() of (label, model)
@arg ref model for testing
@return an ordered Dataset() of {'label', 'score'}
"""
rak = list()
for lbl in cand.iterkeys():
rak.append((lbl, cand[lbl].score(ref, **kargs)))
rak = sorted(rak, key=itemgetter(1), reverse=not ref.isasc())
res = Dataset()
for item in rak:
res.append({"label": item[0], "score": item[1]})
return res
def ranke(cand, ref, **kargs):
""" Rank the models in cand according to the scoring methods in the models
@arg cand dict() of (label, model)
@arg ref model for testing
@return an ordered Dataset() of {'label', 'score'}
"""
rak = list()
for lbl in cand.iterkeys():
rak.append((lbl, cand[lbl].score(ref, **kargs)))
rak = sorted(rak, key=itemgetter(1), reverse=not ref.isasc())
res = Dataset()
for item in rak:
res.append({'label': item[0], 'score': item[1]})
return res
def cmpsparse(cities, numtwts, numtest):
""" Compare the model performance trained with different amount of tweets
"""
lmranks = [list() for i in range(len(numtwts))]
randranks = list()
lmtmranks = [list() for i in range(len(numtwts))]
test = Dataset()
for places in cities:
lms = [dict() for i in range(len(numtwts))]
tst = Dataset()
for pid in places:
twtp = loadrows(
GEOTWEET, ('place_id', 'text', 'created_at'),
('place_id=\'{0}\''.format(pid), ), 'sample',
'order by rand() limit {0}'.format(max(numtwts) + numtest))
for i in range(len(numtwts)):
lms[i][pid] = LanguageModel(twtp['text'][:numtwts[i]])
# test data
for i in range(max(numtwts), max(numtwts) + numtest):
tst.append({
'label': pid,
'lm': LanguageModel([
twtp['text'][i],
]),
})
test.extend(tst)
# rank
for item in tst:
for i in range(len(numtwts)):
lmranks[i].append(ranke(lms[i], item['lm']))
randranks.append(randranke(places))
# plot
candls = ['-', '--']
mks = ['o', '^', '*', 'v', 's']
for i, n in enumerate(numtwts):
lmeval = batcheval(lmranks[i], test['label'])
plt.plot(lmeval['pos'],
lmeval['rate'],
label='tweet(s={0})'.format(n),
marker=mks[i])
plt.plot(lmeval['pos'],
[float(r) / max(lmeval['pos']) for r in lmeval['pos']],
ls='-.',
marker='s',
label='Random Baseline')
plt.legend(loc='lower right')
plt.ylabel('Rate containing Reference POI')
plt.xlabel('Top $p$ places')
plt.show()
def log_parse(src):
"""parse predication output from WEKA"""
ins_lst = Dataset()
with open(src) as fsrc:
for line in fsrc:
line, dummy = _SYMBOL.subn(' ', line)
col = _SPACE.split(line)
ins = DataItem()
ins['ref'] = int((_CLSNO.split(col[2]))[0])
ins['refN'] = (_CLSNO.split(col[2]))[1]
ins['prd'] = int((_CLSNO.split(col[3]))[0])
ins['prdN'] = (_CLSNO.split(col[3]))[1]
ins['err'] = True if col[4] == '+' else False
ins['score'] = [float(col[i]) for i in range(4, len(col) - 2)]
ids, dummy = _PARATH.subn('', col[len(col) - 2])
ins['id'] = int(ids)
ins_lst.append(ins)
return ins_lst
def placetotalrank(ranks, expects):
""" Return the sum of position of reference POIs in ranks
@arg ranks a list() of ranks(dataset(){'label', 'score'})
@arg expects a list() of testitems(dataset(){'label', ...})
@return a sorted dataset(){'label', 'totalrank'}
"""
plcrak = dict()
for i in range(expects.size()):
if expects['label'][i] not in plcrak:
plcrak[expects['label'][i]] = 0
for j in range(ranks[i].size()):
if ranks[i]['label'][j] == expects['label'][i]:
plcrak[expects['label'][i]] += j
break
items = sorted(plcrak.iteritems(), key=itemgetter(1))
res = Dataset()
for item in items:
res.append({'label': item[0], 'totalrank': item[1]})
return res
def placetotalrank(ranks, expects):
""" Return the sum of position of reference POIs in ranks
@arg ranks a list() of ranks(dataset(){'label', 'score'})
@arg expects a list() of testitems(dataset(){'label', ...})
@return a sorted dataset(){'label', 'totalrank'}
"""
plcrak = dict()
for i in range(expects.size()):
if expects['label'][i] not in plcrak:
plcrak[expects['label'][i]] = 0
for j in range(ranks[i].size()):
if ranks[i]['label'][j] == expects['label'][i]:
plcrak[expects['label'][i]] += j
break
items = sorted(plcrak.iteritems(), key=itemgetter(1))
res = Dataset()
for item in items:
res.append({'label': item[0], 'totalrank': item[1]})
return res
def norm_e(dset):
"""normalize values in element wise, i.e., column normalization
@arg dset Dataset() of vectors
@return Dataset() of vectors normalized
"""
ndset = Dataset()
for key in dset.iterkeys():
maxval = max(dset[key])
ndset[key] = list()
for val in dset[key]:
ndset[key].append(val/maxval)
return ndset
def linearjoin(ranks, balance):
""" Join a set of ranks by the weights in balance
@arg ranks a list() of ranks that need to combine
@arg balance a list() of floats indicating combine weights
@return a combined rank
"""
for rak in ranks:
rak['score'] = unitscore(rak['score'])
sortedranks = [[item for item in r.sorted_items('label')] for r in ranks]
unitrank = list()
for idx in range(len(sortedranks[0])):
unitrank.append((sortedranks[0][idx]['label'],
sum([score * bl for score, bl in \
zip([sortedranks[i][idx]['score'] for i in range(len(sortedranks))],
balance)])))
unitrank = sorted(unitrank, key=itemgetter(1), reverse=False)
res = Dataset()
for item in unitrank:
res.append({'label': item[0], 'score': item[1]})
return res
def cmpsparse(cities, numtwts, numtest):
""" Compare the model performance trained with different amount of tweets
"""
lmranks = [list() for i in range(len(numtwts))]
randranks = list()
lmtmranks = [list() for i in range(len(numtwts))]
test = Dataset()
for places in cities:
lms = [dict() for i in range(len(numtwts))]
tst = Dataset()
for pid in places:
twtp = loadrows(GEOTWEET, ('place_id', 'text', 'created_at'),
('place_id=\'{0}\''.format(pid),), 'sample',
'order by rand() limit {0}'.format(max(numtwts) + numtest))
for i in range(len(numtwts)):
lms[i][pid] = LanguageModel(twtp['text'][:numtwts[i]])
# test data
for i in range(max(numtwts), max(numtwts) + numtest):
tst.append({'label': pid,
'lm': LanguageModel([twtp['text'][i],]),
})
test.extend(tst)
# rank
for item in tst:
for i in range(len(numtwts)):
lmranks[i].append(ranke(lms[i], item['lm']))
randranks.append(randranke(places))
# plot
candls = ['-', '--']
mks = ['o', '^', '*', 'v', 's']
for i, n in enumerate(numtwts):
lmeval = batcheval(lmranks[i], test['label'])
plt.plot(lmeval['pos'], lmeval['rate'],
label='tweet(s={0})'.format(n),
marker=mks[i])
plt.plot(lmeval['pos'], [float(r) / max(lmeval['pos']) for r in lmeval['pos']],
ls='-.', marker='s',
label='Random Baseline')
plt.legend(loc='lower right')
plt.ylabel('Rate containing Reference POI')
plt.xlabel('Top $p$ places')
plt.show()
def cmptimeweb(cities, numtwts, numtest):
""" compare the time model + web model to original pure text model
"""
lmranks = [list() for _ in numtwts]
tmranks = [list() for _ in numtwts]
wmranks = list()
randranks = list()
lmtmranks = [list() for _ in numtwts]
wmlmranks = [list() for _ in numtwts]
wmlmtmranks = [list() for _ in numtwts]
test = Dataset()
for city in cities:
lms = [dict() for _ in numtwts]
tms = [dict() for _ in numtwts]
wms = dict()
tst = Dataset()
for pid in city:
twtp = loadrows(
GEOTWEET, ('place_id', 'text', 'created_at'),
('place_id=\'{0}\''.format(pid), ), 'sample_switch_place_cate',
'order by rand() limit {0}'.format(max(numtwts) + numtest))
for i, n in enumerate(numtwts):
lms[i][pid] = LanguageModel(twtp['text'][:n])
tms[i][pid] = TimeModel(twtp['created_at'][:n])
web = loadrows(GEOTWEET, ('place_id', 'web'),
('place_id=\'{0}\''.format(pid), ), 'web',
'order by rand() limit 30')
try:
wms[pid] = LanguageModel(web['web'])
except KeyError:
wms[pid] = LanguageModel('')
# Prepare test data by the tail part of the data retrieved from db
test_pos = max(numtwts)
for i in range(test_pos, test_pos + numtest):
tst.append({
'label': pid,
'lm': LanguageModel([
twtp['text'][i],
]),
'tm': TimeModel([
twtp['created_at'][i],
])
})
test.extend(tst)
# rank
for item in tst:
for i, _ in enumerate(numtwts):
lmranks[i].append(ranke(lms[i], item['lm']))
tmranks[i].append(ranke(tms[i], item['tm']))
wmranks.append(ranke(wms, item['lm']))
randranks.append(randranke(city))
for i in range(len(numtwts)):
for ranklm, ranktm in zip(lmranks[i], tmranks[i]):
lmtmranks[i].append(linearjoin([ranklm, ranktm], [0.5, 0.5]))
for ranklm, rankwm in zip(lmranks[i], wmranks):
wmlmranks[i].append(linearjoin([ranklm, rankwm], [0.5, 0.5]))
for ranklm, ranktm, rankwm in zip(lmranks[i], tmranks[i], wmranks):
wmlmtmranks[i].append(\
linearjoin([ranklm, ranktm, rankwm], [0.33, 0.33, 0.33]))
# plot
candls = ['-', '--', '-.']
# mks = ['o', '^', '*']
#for i in range(len(numtwts)):
#lmeval = batcheval(lmranks[i], test['label'])
#plt.plot(lmeval['pos'], lmeval['rate'],
#label='tweet(s={0})'.format(numtwts[i]),
#ls=candls[i%2], marker=mks[i/2])
#for i in range(len(numtwts)):
#for plc in placetotalrank(lmranks[i], test)['label'][-10:]:
#print place_name(plc), plc
#print placetotalrank(lmranks[i], test)['totalrank'][-10:]
#print wilcoxontest(lmranks[i], lmranks[i-1], test)
#plt.legend(loc='lower right')
#---------------------------------------------------------------
for i in range(len(numtwts)):
lmeval = batcheval(lmranks[i], test['label'])
plt.plot(lmeval['pos'],
lmeval['rate'],
label='tweet(s={0})'.format(numtwts[i]),
ls=candls[i],
marker='o')
# wmlmeval = batcheval(wmlmranks[i], test['label'])
# plt.plot(wmlmeval['pos'], wmlmeval['rate'],
# label='tweet(s={0})+web'.format(numtwts[i]),
# ls=candls[i], marker='^')
# print 'Wilcoxon (lm vs wmlm):', wilcoxontest(lmranks[i], wmlmranks[i], test)
# print 'Place id -> name:'
# for plc in placetotalrank(wmlmranks[i], test)['label'][-10:]:
# print place_name(plc), plc
# print 'Place Total Rank:'
# print placetotalrank(wmlmranks[i], test)['totalrank'][-10:]
plt.plot(lmeval['pos'],
[float(r) / max(lmeval['pos']) for r in lmeval['pos']],
ls='-.',
marker='s',
label='Random Baseline')
# wmeval = batcheval(wmranks, test['label'])
# print 'Place id -> name:'
# for plc in placetotalrank(wmranks, test)['label'][-10:]:
# print place_name(plc), plc
# print 'Place Total Rank'
# print placetotalrank(wmranks, test)['totalrank'][-10:]
# plt.plot(wmeval['pos'], wmeval['rate'],
# label='web',
# ls=':')
#---------------------------------------------------------------
#for i in range(len(numtwts)):
#plt.subplot(121 + i)
#plt.title('$s={0}$'.format(numtwts[i]))
#lmeval = batcheval(lmranks[i], test['label'])
#plt.plot(lmeval['pos'], lmeval['rate'],
#label='tweet',
#ls=candls[i], marker='o')
#lmtmeval = batcheval(lmtmranks[i], test['label'])
#plt.plot(lmtmeval['pos'], lmtmeval['rate'],
#label='tweet+time',
#ls=candls[i], marker='^')
#wmlmtmeval = batcheval(wmlmtmranks[i], test['label'])
#plt.plot(wmlmtmeval['pos'], wmlmtmeval['rate'],
#label='tweet+time+web',
#ls=candls[i], marker='*')
#plt.legend(loc='lower right')
#plt.ylabel('Rate containing Reference POI')
#plt.xlabel('Top $p$ places')
#plt.show()
#---------------------------------------------------------------
#i=0
#plt.subplot(121 + i)
#plt.title('$s={0}$'.format(numtwts[i]))
#tmeval = batcheval(tmranks[i], test['label'])
#plt.plot(tmeval['pos'], tmeval['rate'],
#label='time',
#ls=candls[i], marker='o')
#lmeval = batcheval(lmranks[i], test['label'])
#plt.plot(lmeval['pos'], lmeval['rate'],
#label='tweet',
#ls=candls[i], marker='^')
#lmtmeval = batcheval(lmtmranks[i], test['label'])
#plt.plot(lmtmeval['pos'], lmtmeval['rate'],
#label='tweet+time',
#ls=candls[i], marker='*')
#for plc in placetotalrank(tmranks[i], test)['label'][-10:]:
#print place_name(plc), plc
#print placetotalrank(tmranks[i], test)['totalrank'][-10:]
#for plc in placetotalrank(lmtmranks[i], test)['label'][-10:]:
#print place_name(plc), plc
#print placetotalrank(lmtmranks[i], test)['totalrank'][-10:]
#print wilcoxontest(lmranks[i], lmtmranks[i], test)
#plt.legend(loc='lower right')
#plt.ylabel('Rate containing Reference POI')
#plt.xlabel('Top $p$ places')
#i=1
#plt.subplot(121 + i)
#plt.title('$s={0}$'.format(numtwts[i]))
#tmeval = batcheval(tmranks[i], test['label'])
#plt.plot(tmeval['pos'], tmeval['rate'],
#label='time',
#ls=candls[i], marker='o')
#wmlmeval = batcheval(wmlmranks[i], test['label'])
#plt.plot(wmlmeval['pos'], wmlmeval['rate'],
#label='tweet + web',
#ls=candls[i], marker='^')
#wmlmtmeval = batcheval(wmlmtmranks[i], test['label'])
#plt.plot(wmlmtmeval['pos'], wmlmtmeval['rate'],
#label='tweet+time+web',
#ls=candls[i], marker='*')
#for plc in placetotalrank(wmlmranks[i], test)['label'][-10:]:
#print place_name(plc), plc
#print placetotalrank(wmlmranks[i], test)['totalrank'][-10:]
#for plc in placetotalrank(wmlmtmranks[i], test)['label'][-10:]:
#print place_name(plc), plc
#print placetotalrank(wmlmtmranks[i], test)['totalrank'][-10:]
#print wilcoxontest(wmlmranks[i], wmlmtmranks[i], test)
#plt.legend(loc='lower right')
#plt.ylabel('Rate containing Reference POI')
#plt.xlabel('Top $p$ places')
plt.legend(loc='lower right')
plt.tight_layout()
plt.show()
def cmpsparsecombine(cities, numtwts, numtest):
""" the combined model performance under the influence of sparseness
"""
lmranks = [list() for i in range(len(numtwts))]
tmranks = [list() for i in range(len(numtwts))]
wmranks = list()
randranks = list()
lmtmranks = [list() for i in range(len(numtwts))]
wmlmranks = [list() for i in range(len(numtwts))]
wmlmtmranks = [list() for i in range(len(numtwts))]
test = Dataset()
for places in cities:
lms = [dict() for i in range(len(numtwts))]
tms = [dict() for i in range(len(numtwts))]
wms = dict()
tst = Dataset()
for pid in places:
twtp = loadrows(
GEOTWEET, ('place_id', 'text', 'created_at'),
('place_id=\'{0}\''.format(pid), ), 'sample',
'order by rand() limit {0}'.format(max(numtwts) + numtest))
for i in range(len(numtwts)):
lms[i][pid] = LanguageModel(twtp['text'][:numtwts[i]])
tms[i][pid] = TimeModel(twtp['created_at'][:numtwts[i]])
web = loadrows(GEOTWEET, ('place_id', 'web'),
('place_id=\'{0}\''.format(pid), ), 'web',
'order by rand() limit 30')
wms[pid] = LanguageModel(web['web'])
# test data
for i in range(max(numtwts), max(numtwts) + numtest):
tst.append({
'label': pid,
'lm': LanguageModel([
twtp['text'][i],
]),
'tm': TimeModel([
twtp['created_at'][i],
])
})
test.extend(tst)
# rank
for item in tst:
for i in range(len(numtwts)):
lmranks[i].append(ranke(lms[i], item['lm']))
tmranks[i].append(ranke(tms[i], item['tm']))
wmranks.append(ranke(wms, item['lm']))
randranks.append(randranke(places))
for i in range(len(numtwts)):
for ranklm, ranktm in zip(lmranks[i], tmranks[i]):
lmtmranks[i].append(linearjoin([ranklm, ranktm], [0.5, 0.5]))
for ranklm, rankwm in zip(lmranks[i], wmranks):
wmlmranks[i].append(linearjoin([ranklm, rankwm], [0.5, 0.5]))
for ranklm, ranktm, rankwm in zip(lmranks[i], tmranks[i], wmranks):
wmlmtmranks[i].append(\
linearjoin([ranklm, ranktm, rankwm], [0.33, 0.33, 0.33]))
# plot
candls = ['-', '--']
mks = ['o', '^', '*', 'v', 's']
i = 0
plt.subplot(121 + i)
plt.title('$s={0}$'.format(numtwts[i]))
tmeval = batcheval(tmranks[i], test['label'])
plt.plot(tmeval['pos'],
tmeval['rate'],
label='time',
ls=candls[i],
marker='o')
lmeval = batcheval(lmranks[i], test['label'])
plt.plot(lmeval['pos'],
lmeval['rate'],
label='tweet',
ls=candls[i],
marker='^')
lmtmeval = batcheval(lmtmranks[i], test['label'])
plt.plot(lmtmeval['pos'],
lmtmeval['rate'],
label='tweet+time',
ls=candls[i],
marker='*')
for plc in placetotalrank(tmranks[i], test)['label'][-10:]:
print place_name(plc), plc
print placetotalrank(tmranks[i], test)['totalrank'][-10:]
for plc in placetotalrank(lmtmranks[i], test)['label'][-10:]:
print place_name(plc), plc
print placetotalrank(lmtmranks[i], test)['totalrank'][-10:]
print wilcoxontest(lmranks[i], lmtmranks[i], test)
plt.plot(lmeval['pos'],
[float(r) / max(lmeval['pos']) for r in lmeval['pos']],
ls='-.',
marker='s',
label='Random Baseline')
plt.legend(loc='lower right')
plt.ylabel('Rate containing Reference POI')
plt.xlabel('Top $p$ places')
i = 1
plt.subplot(121 + i)
plt.title('$s={0}$'.format(numtwts[i]))
tmeval = batcheval(tmranks[i], test['label'])
plt.plot(tmeval['pos'],
tmeval['rate'],
label='time',
ls=candls[i],
marker='o')
wmlmeval = batcheval(wmlmranks[i], test['label'])
plt.plot(wmlmeval['pos'],
wmlmeval['rate'],
label='tweet + web',
ls=candls[i],
marker='^')
wmlmtmeval = batcheval(wmlmtmranks[i], test['label'])
plt.plot(wmlmtmeval['pos'],
wmlmtmeval['rate'],
label='tweet+time+web',
ls=candls[i],
marker='*')
for plc in placetotalrank(wmlmranks[i], test)['label'][-10:]:
print place_name(plc), plc
print placetotalrank(wmlmranks[i], test)['totalrank'][-10:]
for plc in placetotalrank(wmlmtmranks[i], test)['label'][-10:]:
print place_name(plc), plc
print placetotalrank(wmlmtmranks[i], test)['totalrank'][-10:]
print wilcoxontest(wmlmranks[i], wmlmtmranks[i], test)
plt.plot(lmeval['pos'],
[float(r) / max(lmeval['pos']) for r in lmeval['pos']],
ls='-.',
marker='s',
label='Random Baseline')
plt.legend(loc='lower right')
plt.ylabel('Rate containing Reference POI')
plt.xlabel('Top $p$ places')
plt.show()
def cmptimeweb(cities, numtwts, numtest):
""" compare the time model + web model to original pure text model
"""
lmranks = [list() for i in range(len(numtwts))]
tmranks = [list() for i in range(len(numtwts))]
wmranks = list()
randranks = list()
lmtmranks = [list() for i in range(len(numtwts))]
wmlmranks = [list() for i in range(len(numtwts))]
wmlmtmranks = [list() for i in range(len(numtwts))]
test = Dataset()
for places in cities:
lms = [dict() for i in range(len(numtwts))]
tms = [dict() for i in range(len(numtwts))]
wms = dict()
tst = Dataset()
for pid in places:
twtp = loadrows(GEOTWEET, ('place_id', 'text', 'created_at'),
('place_id=\'{0}\''.format(pid),), 'sample',
'order by rand() limit {0}'.format(max(numtwts) + numtest))
for i in range(len(numtwts)):
lms[i][pid] = LanguageModel(twtp['text'][:numtwts[i]])
tms[i][pid] = TimeModel(twtp['created_at'][:numtwts[i]])
web = loadrows(GEOTWEET, ('place_id', 'web'),
('place_id=\'{0}\''.format(pid),), 'web',
'order by rand() limit 30')
wms[pid] = LanguageModel(web['web'])
# test data
for i in range(max(numtwts), max(numtwts) + numtest):
tst.append({'label': pid,
'lm': LanguageModel([twtp['text'][i],]),
'tm': TimeModel([twtp['created_at'][i],])})
test.extend(tst)
# rank
for item in tst:
for i in range(len(numtwts)):
lmranks[i].append(ranke(lms[i], item['lm']))
tmranks[i].append(ranke(tms[i], item['tm']))
wmranks.append(ranke(wms, item['lm']))
randranks.append(randranke(places))
for i in range(len(numtwts)):
for ranklm, ranktm in zip(lmranks[i], tmranks[i]):
lmtmranks[i].append(linearjoin([ranklm, ranktm], [0.5, 0.5]))
for ranklm, rankwm in zip(lmranks[i], wmranks):
wmlmranks[i].append(linearjoin([ranklm, rankwm], [0.5, 0.5]))
for ranklm, ranktm, rankwm in zip(lmranks[i], tmranks[i], wmranks):
wmlmtmranks[i].append(\
linearjoin([ranklm, ranktm, rankwm], [0.33, 0.33, 0.33]))
# plot
candls = ['-', '--']
mks = ['o', '^', '*', 'v', 's']
#for i in range(len(numtwts)):
#lmeval = batcheval(lmranks[i], test['label'])
#plt.plot(lmeval['pos'], lmeval['rate'],
#label='tweet(s={0})'.format(numtwts[i]),
#ls=candls[i%2], marker=mks[i/2])
#for i in range(len(numtwts)):
#for plc in placetotalrank(lmranks[i], test)['label'][-10:]:
#print place_name(plc), plc
#print placetotalrank(lmranks[i], test)['totalrank'][-10:]
#print wilcoxontest(lmranks[i], lmranks[i-1], test)
#plt.legend(loc='lower right')
#---------------------------------------------------------------
for i in range(len(numtwts)):
lmeval = batcheval(lmranks[i], test['label'])
plt.plot(lmeval['pos'], lmeval['rate'],
label='tweet(s={0})'.format(numtwts[i]),
ls=candls[i], marker='o')
wmlmeval = batcheval(wmlmranks[i], test['label'])
plt.plot(wmlmeval['pos'], wmlmeval['rate'],
label='tweet(s={0})+web'.format(numtwts[i]),
ls=candls[i], marker='^')
print wilcoxontest(lmranks[i], wmlmranks[i], test)
for plc in placetotalrank(wmlmranks[i], test)['label'][-10:]:
print place_name(plc), plc
print placetotalrank(wmlmranks[i], test)['totalrank'][-10:]
wmeval = batcheval(wmranks, test['label'])
for plc in placetotalrank(wmranks, test)['label'][-10:]:
print place_name(plc), plc
print placetotalrank(wmranks, test)['totalrank'][-10:]
plt.plot(wmeval['pos'], wmeval['rate'],
label='web',
ls=':')
plt.plot(lmeval['pos'], [float(r) / max(lmeval['pos']) for r in lmeval['pos']],
ls='-.', marker='s',
label='Random Baseline')
#---------------------------------------------------------------
#for i in range(len(numtwts)):
#plt.subplot(121 + i)
#plt.title('$s={0}$'.format(numtwts[i]))
#lmeval = batcheval(lmranks[i], test['label'])
#plt.plot(lmeval['pos'], lmeval['rate'],
#label='tweet',
#ls=candls[i], marker='o')
#lmtmeval = batcheval(lmtmranks[i], test['label'])
#plt.plot(lmtmeval['pos'], lmtmeval['rate'],
#label='tweet+time',
#ls=candls[i], marker='^')
#wmlmtmeval = batcheval(wmlmtmranks[i], test['label'])
#plt.plot(wmlmtmeval['pos'], wmlmtmeval['rate'],
#label='tweet+time+web',
#ls=candls[i], marker='*')
#plt.legend(loc='lower right')
#plt.ylabel('Rate containing Reference POI')
#plt.xlabel('Top $p$ places')
#plt.show()
#---------------------------------------------------------------
#i=0
#plt.subplot(121 + i)
#plt.title('$s={0}$'.format(numtwts[i]))
#tmeval = batcheval(tmranks[i], test['label'])
#plt.plot(tmeval['pos'], tmeval['rate'],
#label='time',
#ls=candls[i], marker='o')
#lmeval = batcheval(lmranks[i], test['label'])
#plt.plot(lmeval['pos'], lmeval['rate'],
#label='tweet',
#ls=candls[i], marker='^')
#lmtmeval = batcheval(lmtmranks[i], test['label'])
#plt.plot(lmtmeval['pos'], lmtmeval['rate'],
#label='tweet+time',
#ls=candls[i], marker='*')
#for plc in placetotalrank(tmranks[i], test)['label'][-10:]:
#print place_name(plc), plc
#print placetotalrank(tmranks[i], test)['totalrank'][-10:]
#for plc in placetotalrank(lmtmranks[i], test)['label'][-10:]:
#print place_name(plc), plc
#print placetotalrank(lmtmranks[i], test)['totalrank'][-10:]
#print wilcoxontest(lmranks[i], lmtmranks[i], test)
#plt.legend(loc='lower right')
#plt.ylabel('Rate containing Reference POI')
#plt.xlabel('Top $p$ places')
#i=1
#plt.subplot(121 + i)
#plt.title('$s={0}$'.format(numtwts[i]))
#tmeval = batcheval(tmranks[i], test['label'])
#plt.plot(tmeval['pos'], tmeval['rate'],
#label='time',
#ls=candls[i], marker='o')
#wmlmeval = batcheval(wmlmranks[i], test['label'])
#plt.plot(wmlmeval['pos'], wmlmeval['rate'],
#label='tweet + web',
#ls=candls[i], marker='^')
#wmlmtmeval = batcheval(wmlmtmranks[i], test['label'])
#plt.plot(wmlmtmeval['pos'], wmlmtmeval['rate'],
#label='tweet+time+web',
#ls=candls[i], marker='*')
#for plc in placetotalrank(wmlmranks[i], test)['label'][-10:]:
#print place_name(plc), plc
#print placetotalrank(wmlmranks[i], test)['totalrank'][-10:]
#for plc in placetotalrank(wmlmtmranks[i], test)['label'][-10:]:
#print place_name(plc), plc
#print placetotalrank(wmlmtmranks[i], test)['totalrank'][-10:]
#print wilcoxontest(wmlmranks[i], wmlmtmranks[i], test)
plt.legend(loc='lower right')
plt.ylabel('Rate containing Reference POI')
plt.xlabel('Top $p$ places')
plt.show()