Zipeng365's picture
Add ICML 2026 TSDecompose benchmark release
17b7ba4 verified
Raw
History Blame Contribute Delete
7.35 kB
import numpy as np
from time import perf_counter
from typing import Dict, Any, List, Optional
from .._native import invoke_native
from ..backends import finalize_result, resolve_backend, result_from_native_payload, split_runtime_params
from ..core import DecompResult
from ..registry import MethodRegistry
from .utils import dominant_frequency
def _dominant_frequency(x: np.ndarray, fs: float = 1.0) -> float:
return dominant_frequency(x, fs)
def _diagonal_averaging(matrix: np.ndarray) -> np.ndarray:
L, K = matrix.shape
T = L + K - 1
recon = np.zeros(T)
counts = np.zeros(T)
for i in range(L):
for j in range(K):
recon[i + j] += matrix[i, j]
counts[i + j] += 1.0
counts[counts == 0.0] = 1.0
return recon / counts
def _basic_ssa(y: np.ndarray, window: int, rank: int) -> List[np.ndarray]:
y_arr = np.asarray(y, dtype=float)
T = y_arr.shape[0]
L = int(window)
if L < 2 or L > T - 1:
raise ValueError(f"Invalid SSA window length L={L} for T={T}")
K = T - L + 1
X = np.empty((L, K), dtype=float)
for i in range(K):
X[:, i] = y_arr[i : i + L]
U, s, Vt = np.linalg.svd(X, full_matrices=False)
d = min(rank, U.shape[1])
rc_list: List[np.ndarray] = []
for idx in range(d):
Xi = np.outer(U[:, idx], s[idx] * Vt[idx, :])
rc = _diagonal_averaging(Xi)[:T]
rc_list.append(rc)
return rc_list
def _diagonal_averaging_fast(matrix: np.ndarray) -> np.ndarray:
L, K = matrix.shape
T = L + K - 1
anti_diag_ids = np.add.outer(np.arange(L), np.arange(K)).ravel()
flat = np.asarray(matrix, dtype=float).ravel()
sums = np.bincount(anti_diag_ids, weights=flat, minlength=T)
counts = np.bincount(anti_diag_ids, minlength=T).astype(float)
counts[counts == 0.0] = 1.0
return sums / counts
def _basic_ssa_fast(y: np.ndarray, window: int, rank: int) -> List[np.ndarray]:
y_arr = np.asarray(y, dtype=float).reshape(-1)
T = y_arr.shape[0]
L = int(window)
if L < 2 or L > T - 1:
raise ValueError(f"Invalid SSA window length L={L} for T={T}")
X = np.lib.stride_tricks.sliding_window_view(y_arr, L).T.copy()
U, s, Vt = np.linalg.svd(X, full_matrices=False)
d = min(rank, U.shape[1])
rc_list: List[np.ndarray] = []
for idx in range(d):
Xi = np.outer(U[:, idx], s[idx] * Vt[idx, :])
rc = _diagonal_averaging_fast(Xi)[:T]
rc_list.append(rc)
return rc_list
def _sum_components(components: List[np.ndarray], indices: List[int], length: int) -> np.ndarray:
if not indices:
return np.zeros(length)
out = np.zeros(length)
for idx in indices:
if 0 <= idx < len(components):
out += components[idx]
return out
@MethodRegistry.register("SSA")
def ssa_decompose(
y: np.ndarray,
params: Dict[str, Any],
) -> DecompResult:
started_at = perf_counter()
cfg, runtime = split_runtime_params(params)
y_arr = np.asarray(y, dtype=float)
T = y_arr.shape[0]
window = int(cfg.get("window", max(4, T // 4)))
window = min(max(2, window), T - 1)
rank = int(cfg.get("rank", 10))
rank = max(1, min(rank, window, T - window + 1))
fs = float(cfg.get("fs", 1.0))
primary_period = cfg.get("primary_period")
primary_period = float(primary_period) if primary_period not in (None, 0) else None
backend = resolve_backend("SSA", runtime, native_methods=("ssa_decompose",))
if backend == "native":
payload = invoke_native(
"ssa_decompose",
y_arr,
window=window,
rank=rank,
fs=fs,
primary_period=primary_period,
trend_components=list(cfg.get("trend_components", [])),
season_components=list(cfg.get("season_components", [])),
season_freq_tol_ratio=float(cfg.get("season_freq_tol_ratio", 0.25)),
trend_freq_threshold=cfg.get("trend_freq_threshold"),
)
return finalize_result(
result_from_native_payload(payload, method="SSA"),
method="SSA",
runtime=runtime,
backend_used="native",
started_at=started_at,
)
if runtime.speed_mode == "fast":
rc_list = _basic_ssa_fast(y_arr, window=window, rank=rank)
else:
rc_list = _basic_ssa(y_arr, window=window, rank=rank)
num_rc = len(rc_list)
if num_rc == 0:
zeros = np.zeros_like(y_arr)
return finalize_result(
DecompResult(
trend=zeros,
season=zeros,
residual=y_arr.copy(),
components={"rc_list": np.zeros((0, T), dtype=float)},
meta={"method": "SSA", "window": window, "rank": rank, "rc_list_shape": [0, T]}
),
method="SSA",
runtime=runtime,
backend_used="python",
started_at=started_at,
)
trend_components = list(cfg.get("trend_components", []))
season_components = list(cfg.get("season_components", []))
# Auto-grouping logic
if not trend_components and not season_components:
if primary_period is not None and primary_period > 0:
dom_freqs = [_dominant_frequency(rc, fs=fs) for rc in rc_list]
f0 = 1.0 / primary_period
tol = float(cfg.get("season_freq_tol_ratio", 0.25)) * f0
low_freq_threshold = float(cfg.get("trend_freq_threshold", f0 / 4.0 if f0 else 0.05))
for idx, f_dom in enumerate(dom_freqs):
if f_dom <= max(low_freq_threshold, 1e-8):
trend_components.append(idx)
elif f0 > 0 and abs(f_dom - f0) <= max(tol, 1e-8):
season_components.append(idx)
# Fallback if empty
if not trend_components and num_rc >= 1:
trend_components.append(0)
if not season_components:
for idx in range(num_rc):
if idx not in trend_components:
season_components.append(idx)
break
else:
# Default heuristic
if num_rc >= 1: trend_components.append(0)
if num_rc >= 2: trend_components.append(1)
if num_rc >= 4: season_components.extend([2, 3])
elif num_rc >= 3: season_components.append(2)
trend = _sum_components(rc_list, trend_components, T)
season = _sum_components(rc_list, season_components, T)
residual = y_arr - trend - season
return finalize_result(
DecompResult(
trend=trend,
season=season,
residual=residual,
components={"rc_list": np.stack(rc_list, axis=0)},
meta={
"method": "SSA",
"window": window,
"rank": rank,
"trend_components": trend_components,
"season_components": season_components,
"rc_list_shape": [len(rc_list), T],
}
),
method="SSA",
runtime=runtime,
backend_used="python",
started_at=started_at,
)