Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

ENH: Add non-negative reconciliation heuristic for MinTraceSparse #284

Open
wants to merge 4 commits into
base: main
Choose a base branch
from
Open

ENH: Add non-negative reconciliation heuristic for MinTraceSparse #284

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
8cd41c0
ENH: :sparkles: Add sparse non-negative reconciliation heuristic.
christophertitchen Sep 3, 2024
59a2b6b
BUG: :bug: Remove legacy absolute tolerance for BiCGSTAB convergence.
christophertitchen Sep 3, 2024
73355ce
TST: :white_check_mark: Add tests for sparse MinT.
christophertitchen Sep 3, 2024
a037eb9
Merge branch 'main' into enh/nn-heuristic
elephaint Sep 12, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions hierarchicalforecast/_modidx.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -102,6 +102,8 @@
'hierarchicalforecast/methods.py'),
'hierarchicalforecast.methods.MinTraceSparse._get_PW_matrices': ( 'methods.html#mintracesparse._get_pw_matrices',
'hierarchicalforecast/methods.py'),
'hierarchicalforecast.methods.MinTraceSparse.fit': ( 'methods.html#mintracesparse.fit',
'hierarchicalforecast/methods.py'),
'hierarchicalforecast.methods.OptimalCombination': ( 'methods.html#optimalcombination',
'hierarchicalforecast/methods.py'),
'hierarchicalforecast.methods.OptimalCombination.__init__': ( 'methods.html#optimalcombination.__init__',
Expand Down
78 changes: 69 additions & 9 deletions hierarchicalforecast/methods.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1123,11 +1123,6 @@ def _get_PW_matrices(
"Only the methods with diagonal W are supported as sparse operations"
)

if self.nonnegative:
raise NotImplementedError(
"Non-negative MinT is currently not implemented as sparse"
)

S = sparse.csr_matrix(S)

if self.method in res_methods and y_insample is None and y_hat_insample is None:
Expand Down Expand Up @@ -1186,7 +1181,7 @@ def get_P_action(y):
(b.size, b.size), matvec=lambda v: R @ (S @ v)
)

x_tilde, exit_code = sparse.linalg.bicgstab(A, b, atol="legacy")
x_tilde, exit_code = sparse.linalg.bicgstab(A, b)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do you remove this as SciPy no longer supports it? (I don't see it being mentioned as a valid value on bicgstab page)

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There was an issue on Github with this too, I think this argument is indeed deprecated and needs to removed.


return x_tilde

Expand All @@ -1197,7 +1192,72 @@ def get_P_action(y):

return P, W

# %% ../nbs/methods.ipynb 55
def fit(self,
S: sparse.csr_matrix,
y_hat: np.ndarray,
y_insample: Optional[np.ndarray] = None,
y_hat_insample: Optional[np.ndarray] = None,
sigmah: Optional[np.ndarray] = None,
intervals_method: Optional[str] = None,
num_samples: Optional[int] = None,
seed: Optional[int] = None,
tags: Dict[str, np.ndarray] = None,
idx_bottom: Optional[np.ndarray] = None):
# Clip the base forecasts if required to align them with their use in practice.
if self.nonnegative:
self.y_hat = np.clip(y_hat, 0, None)
else:
self.y_hat = y_hat
# Get the reconciliation matrices.
self.P, self.W = self._get_PW_matrices(
S=S,
y_hat=self.y_hat,
y_insample=y_insample,
y_hat_insample=y_hat_insample,
idx_bottom=idx_bottom,
)

if self.nonnegative:
# Get the number of leaf nodes.
_, n_bottom = S.shape
# Although it is now sufficient to ensure that all of the entries in P are
# positive, as it is implemented as a linear operator for the iterative
# method to solve the sparse linear system, we need to reconcile to find
# if any of the coherent bottom level point forecasts are negative.
y_tilde = self._reconcile(
S=S, P=self.P, y_hat=self.y_hat, level=None, sampler=None
)["mean"][-n_bottom:]
# Find if any of the forecasts are negative.
if np.any(y_tilde < 0):
# Clip the negative forecasts.
y_tilde = np.clip(y_tilde, 0, None)
# Force non-negative coherence by overwriting the base forecasts with
# the aggregated, clipped bottom level forecasts.
self.y_hat = S @ y_tilde
# Overwrite the attributes for the P and W matrices with those for
# bottom-up reconciliation to force projection onto the non-negative
# coherent subspace.
self.P, self.W = BottomUpSparse()._get_PW_matrices(S=S, idx_bottom=None)

# Get the sampler for probabilistic reconciliation.
self.sampler = self._get_sampler(
S=S,
P=self.P,
W=self.W,
y_hat=self.y_hat,
y_insample=y_insample,
y_hat_insample=y_hat_insample,
sigmah=sigmah,
intervals_method=intervals_method,
num_samples=num_samples,
seed=seed,
tags=tags,
)
# Set the instance as fitted.
self.fitted = True
return self

# %% ../nbs/methods.ipynb 56
class OptimalCombination(MinTrace):
"""Optimal Combination Reconciliation Class.

Expand Down Expand Up @@ -1231,7 +1291,7 @@ def __init__(self,
super().__init__(method=method, nonnegative=nonnegative, num_threads=num_threads)
self.insample = False

# %% ../nbs/methods.ipynb 64
# %% ../nbs/methods.ipynb 65
@njit
def lasso(X: np.ndarray, y: np.ndarray,
lambda_reg: float, max_iters: int = 1_000,
Expand Down Expand Up @@ -1263,7 +1323,7 @@ def lasso(X: np.ndarray, y: np.ndarray,
#print(it)
return beta

# %% ../nbs/methods.ipynb 65
# %% ../nbs/methods.ipynb 66
class ERM(HReconciler):
"""Optimal Combination Reconciliation Class.

Expand Down
96 changes: 89 additions & 7 deletions nbs/methods.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1735,11 +1735,6 @@
" \"Only the methods with diagonal W are supported as sparse operations\"\n",
" )\n",
"\n",
" if self.nonnegative:\n",
" raise NotImplementedError(\n",
" \"Non-negative MinT is currently not implemented as sparse\"\n",
" )\n",
"\n",
" S = sparse.csr_matrix(S)\n",
"\n",
" if self.method in res_methods and y_insample is None and y_hat_insample is None:\n",
Expand Down Expand Up @@ -1798,7 +1793,7 @@
" (b.size, b.size), matvec=lambda v: R @ (S @ v)\n",
" )\n",
"\n",
" x_tilde, exit_code = sparse.linalg.bicgstab(A, b, atol=\"legacy\")\n",
" x_tilde, exit_code = sparse.linalg.bicgstab(A, b)\n",
"\n",
" return x_tilde\n",
"\n",
Expand All @@ -1807,7 +1802,72 @@
" )\n",
" W = sparse.spdiags(W_diag, 0, W_diag.size, W_diag.size)\n",
"\n",
" return P, W"
" return P, W\n",
"\n",
" def fit(self,\n",
" S: sparse.csr_matrix,\n",
" y_hat: np.ndarray,\n",
" y_insample: Optional[np.ndarray] = None,\n",
" y_hat_insample: Optional[np.ndarray] = None,\n",
" sigmah: Optional[np.ndarray] = None,\n",
" intervals_method: Optional[str] = None,\n",
" num_samples: Optional[int] = None,\n",
" seed: Optional[int] = None, \n",
" tags: Dict[str, np.ndarray] = None,\n",
" idx_bottom: Optional[np.ndarray] = None):\n",
" # Clip the base forecasts if required to align them with their use in practice.\n",
" if self.nonnegative:\n",
" self.y_hat = np.clip(y_hat, 0, None)\n",
" else:\n",
" self.y_hat = y_hat\n",
" # Get the reconciliation matrices.\n",
" self.P, self.W = self._get_PW_matrices(\n",
" S=S, \n",
" y_hat=self.y_hat, \n",
" y_insample=y_insample, \n",
" y_hat_insample=y_hat_insample, \n",
" idx_bottom=idx_bottom,\n",
" )\n",
"\n",
" if self.nonnegative:\n",
" # Get the number of leaf nodes.\n",
" _, n_bottom = S.shape\n",
" # Although it is now sufficient to ensure that all of the entries in P are \n",
" # positive, as it is implemented as a linear operator for the iterative \n",
" # method to solve the sparse linear system, we need to reconcile to find \n",
" # if any of the coherent bottom level point forecasts are negative.\n",
" y_tilde = self._reconcile(\n",
" S=S, P=self.P, y_hat=self.y_hat, level=None, sampler=None\n",
" )[\"mean\"][-n_bottom:]\n",
" # Find if any of the forecasts are negative.\n",
" if np.any(y_tilde < 0):\n",
" # Clip the negative forecasts.\n",
" y_tilde = np.clip(y_tilde, 0, None)\n",
" # Force non-negative coherence by overwriting the base forecasts with \n",
" # the aggregated, clipped bottom level forecasts.\n",
" self.y_hat = S @ y_tilde\n",
" # Overwrite the attributes for the P and W matrices with those for \n",
" # bottom-up reconciliation to force projection onto the non-negative \n",
" # coherent subspace.\n",
" self.P, self.W = BottomUpSparse()._get_PW_matrices(S=S, idx_bottom=None) \n",
"\n",
" # Get the sampler for probabilistic reconciliation.\n",
" self.sampler = self._get_sampler(\n",
" S=S,\n",
" P=self.P,\n",
" W=self.W,\n",
" y_hat=self.y_hat,\n",
" y_insample=y_insample,\n",
" y_hat_insample=y_hat_insample,\n",
" sigmah=sigmah,\n",
" intervals_method=intervals_method,\n",
" num_samples=num_samples,\n",
" seed=seed,\n",
" tags=tags,\n",
" )\n",
" # Set the instance as fitted.\n",
" self.fitted = True\n",
" return self"
]
},
{
Expand Down Expand Up @@ -1950,6 +2010,28 @@
" )"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#| hide\n",
"for method in [\"ols\", \"wls_struct\"]:\n",
" for nonnegative in [False, True]:\n",
" cls_min_trace = MinTraceSparse(method=method, nonnegative=nonnegative)\n",
" test_close(\n",
" cls_min_trace(\n",
" S=S,\n",
" y_hat=S @ y_hat_bottom,\n",
" y_insample=S @ y_bottom,\n",
" y_hat_insample=S @ y_hat_bottom_insample,\n",
" idx_bottom=idx_bottom if nonnegative else None,\n",
" )[\"mean\"],\n",
" S @ y_hat_bottom,\n",
" )"
]
},
{
"cell_type": "markdown",
"metadata": {},
Expand Down
Loading