def cv_warning(model, data, tournament='bernie', kfold=5, nsamples=100):
"Hold out a sample of eras not rows when doing cross validation."
data = data['train']
results_cve = pd.DataFrame()
results_cv = pd.DataFrame()
for i in range(nsamples):
# cv across eras
cve = nx.CVSplitter(data, kfold=kfold, seed=i)
prediction = nx.run(model, cve, tournament, verbosity=0)
df = prediction.performance(data, tournament)
results_cve = results_cve.append(df, ignore_index=True)
# cv ignoring eras but y balanced
cv = nx.IgnoreEraCVSplitter(data,
tournament=tournament,
kfold=kfold,
seed=i)
prediction = nx.run(model, cv, tournament, verbosity=0)
df = prediction.performance(data, tournament)
results_cv = results_cv.append(df, ignore_index=True)
# display results
rcve = results_cve.mean(axis=0)
rcv = results_cv.mean(axis=0)
rcve.name = 'cve'
rcv.name = 'cv'
r = pd.concat([rcve, rcv], axis=1)
print("\n{} runs".format(i + 1))
print(r)
def test_splitter_overlap():
"prediction data should not overlap"
d = nx.play_data()
splitters = [
nx.TournamentSplitter(d),
nx.ValidationSplitter(d),
nx.CheatSplitter(d),
nx.CVSplitter(d),
nx.IgnoreEraCVSplitter(d),
nx.SplitSplitter(d, fit_fraction=0.5)
]
for splitter in splitters:
predict_ids = []
for dfit, dpredict in splitter:
predict_ids.extend(dpredict.ids.tolist())
ok_(len(predict_ids) == len(set(predict_ids)), "ids overlap")
def test_splitter_reset():
"splitter reset should not change results"
d = nx.play_data()
splitters = [
nx.TournamentSplitter(d),
nx.ValidationSplitter(d),
nx.CheatSplitter(d),
nx.CVSplitter(d),
nx.IgnoreEraCVSplitter(d),
nx.SplitSplitter(d, fit_fraction=0.5)
]
for splitter in splitters:
ftups = [[], []]
ptups = [[], []]
for i in range(2):
for dfit, dpredict in splitter:
ftups[i].append(dfit)
ptups[i].append(dpredict)
splitter.reset()
ok_(ftups[0] == ftups[1], "splitter reset changed fit split")
ok_(ptups[0] == ptups[1], "splitter reset changed predict split")
def cv_warning(data, nsamples=100):
model = nx.logistic()
for i in range(nsamples):
report = nx.Report()
# cv across eras
cve = nx.CVSplitter(data, seed=i)
prediction = nx.run(model, cve, verbosity=0)
report.append_prediction(prediction, 'cve')
# cv ignoring eras but y balanced
cv = nx.IgnoreEraCVSplitter(data, seed=i)
prediction = nx.run(model, cv, verbosity=0)
report.append_prediction(prediction, 'cv')
# save performance results
df = report.performance_df(data)
cols = df.columns.tolist()
cols[-1] = 'cv_type'
df.columns = cols
if i == 0:
results = df
else:
results = results.append(df, ignore_index=True)
# display results
rcve = results[results.cv_type == 'cve'].mean(axis=0)
rcv = results[results.cv_type == 'cv'].mean(axis=0)
rcve.name = 'cve'
rcv.name = 'cv'
r = pd.concat([rcve, rcv], axis=1)
print("\n{} runs".format(i+1))
print(r)