def impute_conf_ax(means, stds, ns, ax=None):
models = []
xs = []
for mean, std, n in zip(means, stds, ns):
x = np.random.randn(n)*std + mean
model = copy.deepcopy(MODEL_SCHEMA)
model['col_suffstats'][0][0]['n'] = n
model['col_suffstats'][0][0]['sum_x'] = np.sum(x)
model['col_suffstats'][0][0]['sum_x_sq'] = np.sum(x**2)
models.append(model)
xs.append(x)
sns.distplot(x, hist=False, ax=ax, norm_hist=False, label='model')
sns.distplot(np.hstack(xs), hist=False, ax=ax, norm_hist=False,
kde_kws={'color': 'crimson', 'lw': 3}, label='combined')
x, conf = mu.impute(0, 0, ROW2IDX*len(models), models, (-20, 10))
ax.plot([x, x], [0, ax.get_ylim()[1]], lw=2, ls='--', c='black')
ax.set_title('Confidence = %f' % conf)
def impute(self, col, rows=None):
""" Infer the most likely value and reuturn a measure of the
confidence.
The imputed value is simply the most likely value.
Confidence is a meausure of the agreement between models; not an
interval. For example, the continuous confidence of row `r` in column
`c` is calculated like so:
* M is the set of all component models to which which `r` in `c` is
assigned.
* mu is the set of the means on M
* a -> min(mu), b -> max(mu)
* conf is one minus the integral of the probability of observing
a values between `a` and `b`.
The confidence is higher the narrower the imputation region and is 1 if
`x`'s component models are have the same mean.
Notes
-----
I'm not sure that categorical confidence behaves intuitively in the
event that the predictve PMF has multiple maximum-probability values.
For example if :code:`PMF = [.4, .4, .2]`.
Parameters
----------
col : column name
The column name to impute
rows : list(row name), optional
A list of the rows to impute. If None (default), all missing values
will be imputed.
Returns
-------
impdata : pandas.DataFrame
Row-indexed DataFrame with a columns for the imputed values and
the confidence (`conf`) in those values.
"""
if rows is None:
# Get row indices where col is null
rows = self._df[pd.isnull(self._df[col])].index
col_idx = self._converters['col2idx'][col]
row2idx = self._converters['row2idx_sf']
# FIXME: In the future we'll want a better way to determine
# optimization bounds for different dtypes. If statements are gross.
dtype = self._dtypes[col_idx]
if dtype == 'continuous':
lower = np.nanmin(self._data[:, col_idx])
upper = np.nanmax(self._data[:, col_idx])
bounds = (
lower,
upper,
)
elif dtype == 'categorical':
bounds = self._converters['valmaps'][col]['val2idx'].values()
else:
raise ValueError('Unsupported dtype: {}'.format(dtype))
impdata = []
for row in rows:
x, conf = mu.impute(row, col_idx, row2idx, self._models, bounds)
if dtype == 'categorical':
x = self._converters['valmaps'][col]['idx2val'][x]
impdata.append({col: x, 'conf': conf})
return pd.DataFrame(impdata, index=rows)
def impute(self, col, rows=None):
""" Infer the most likely value and reuturn a measure of the
confidence.
The imputed value is simply the most likely value.
Confidence is a meausure of the agreement between models; not an
interval. For example, the continuous confidence of row `r` in column
`c` is calculated like so:
* M is the set of all component models to which which `r` in `c` is
assigned.
* mu is the set of the means on M
* a -> min(mu), b -> max(mu)
* conf is one minus the integral of the probability of observing
a values between `a` and `b`.
The confidence is higher the narrower the imputation region and is 1 if
`x`'s component models are have the same mean.
Notes
-----
I'm not sure that categorical confidence behaves intuitively in the
event that the predictve PMF has multiple maximum-probability values.
For example if :code:`PMF = [.4, .4, .2]`.
Parameters
----------
col : column name
The column name to impute
rows : list(row name), optional
A list of the rows to impute. If None (default), all missing values
will be imputed.
Returns
-------
impdata : pandas.DataFrame
Row-indexed DataFrame with a columns for the imputed values and
the confidence (`conf`) in those values.
"""
if rows is None:
# Get row indices where col is null
rows = self._df[pd.isnull(self._df[col])].index
col_idx = self._converters['col2idx'][col]
row2idx = self._converters['row2idx_sf']
# FIXME: In the future we'll want a better way to determine
# optimization bounds for different dtypes. If statements are gross.
dtype = self._dtypes[col_idx]
if dtype == 'continuous':
lower = np.nanmin(self._data[:, col_idx])
upper = np.nanmax(self._data[:, col_idx])
bounds = (lower, upper,)
elif dtype == 'categorical':
bounds = self._converters['valmaps'][col]['val2idx'].values()
else:
raise ValueError('Unsupported dtype: {}'.format(dtype))
impdata = []
for row in rows:
x, conf = mu.impute(row, col_idx, row2idx, self._models, bounds)
if dtype == 'categorical':
x = self._converters['valmaps'][col]['idx2val'][x]
impdata.append({col: x, 'conf': conf})
return pd.DataFrame(impdata, index=rows)
