Improvements on benchmarks.knn

pyFTS/benchmarks/benchmarks.py

@@ -643,7 +643,7 @@ def run_point2(fts_method, order, partitioner_method, partitions, transformation
     import time
     from pyFTS.models import yu, chen, hofts, pwfts,ismailefendi,sadaei, song, cheng, hwang
     from pyFTS.partitioners import Grid, Entropy, FCM
-    from pyFTS.benchmarks import Measures, naive, arima, quantreg
+    from pyFTS.benchmarks import Measures, naive, arima, quantreg, benchmarks
     from pyFTS.common import Transformations
 
     tmp = [song.ConventionalFTS, chen.ConventionalFTS, yu.WeightedFTS, ismailefendi.ImprovedWeightedFTS,
@@ -664,7 +664,7 @@ def run_point2(fts_method, order, partitioner_method, partitions, transformation
     method = kwargs.get('method', None)
     parameters = kwargs.get('parameters', {})
 
-    mfts, pttr = __build_model(fts_method, order, parameters, partitioner_method, partitions, train_data,
+    mfts, pttr = benchmarks.__build_model(fts_method, order, parameters, partitioner_method, partitions, train_data,
                                transformation)
 
     _start = time.time()
@@ -691,7 +691,7 @@ def run_interval2(fts_method, order, partitioner_method, partitions, transformat
     import time
     from pyFTS.models import hofts,ifts,pwfts
     from pyFTS.partitioners import Grid, Entropy, FCM
-    from pyFTS.benchmarks import Measures, arima, quantreg, BSTS
+    from pyFTS.benchmarks import Measures, arima, quantreg, BSTS, benchmarks
 
     tmp = [hofts.HighOrderFTS, ifts.IntervalFTS, ifts.WeightedIntervalFTS,  pwfts.ProbabilisticWeightedFTS]
 
@@ -705,7 +705,7 @@ def run_interval2(fts_method, order, partitioner_method, partitions, transformat
     method = kwargs.get('method', None)
     parameters = kwargs.get('parameters',{})
 
-    mfts, pttr = __build_model(fts_method, order, parameters, partitioner_method, partitions, train_data,
+    mfts, pttr = benchmarks.__build_model(fts_method, order, parameters, partitioner_method, partitions, train_data,
                                transformation)
     _start = time.time()
     mfts.fit(train_data, **kwargs)
@@ -735,7 +735,7 @@ def run_probabilistic2(fts_method, order, partitioner_method, partitions, transf
     from pyFTS.models import hofts, ifts, pwfts
     from pyFTS.models.ensemble import ensemble
     from pyFTS.partitioners import Grid, Entropy, FCM
-    from pyFTS.benchmarks import Measures, arima, quantreg, knn
+    from pyFTS.benchmarks import Measures, arima, quantreg, knn, benchmarks
     from pyFTS.models.seasonal import SeasonalIndexer
 
     tmp = [hofts.HighOrderFTS, ifts.IntervalFTS, pwfts.ProbabilisticWeightedFTS, arima.ARIMA,
@@ -751,9 +751,8 @@ def run_probabilistic2(fts_method, order, partitioner_method, partitions, transf
     method = kwargs.get('method', None)
     parameters = kwargs.get('parameters', {})
 
-    mfts, pttr = __build_model(fts_method, order, parameters, partitioner_method, partitions, train_data,
+    mfts, pttr = benchmarks.__build_model(fts_method, order, parameters, partitioner_method, partitions, train_data,
                                transformation)
-
     if mfts.has_seasonality:
         mfts.indexer = indexer
 

pyFTS/benchmarks/knn.py

@@ -5,17 +5,20 @@ import numpy as np
 from statsmodels.tsa.tsatools import lagmat
 from pyFTS.common import fts
 from pyFTS.probabilistic import ProbabilityDistribution
-
+from sklearn.neighbors import KDTree
+from itertools import product
+from pyFTS.models.ensemble.ensemble import sampler
 
 class KNearestNeighbors(fts.FTS):
     """
-    K-Nearest Neighbors
+    A façade for sklearn.neighbors
     """
     def __init__(self, **kwargs):
         super(KNearestNeighbors, self).__init__(**kwargs)
         self.name = "kNN"
         self.shortname = "kNN"
         self.detail = "K-Nearest Neighbors"
+        self.uod_clip = False
         self.is_high_order = True
         self.has_point_forecasting = True
         self.has_interval_forecasting = True
@@ -26,30 +29,113 @@ class KNearestNeighbors(fts.FTS):
         self.lag = None
         self.k = kwargs.get("k", 30)
         self.uod = None
+        self.kdtree = None
+        self.values = None
+
+    def _prepare_x(self, data):
+        l = len(data)
+        X = []
+
+        if l == self.order:
+            l += 1
+
+        for t in np.arange(self.order, l):
+            X.append([data[t - k - 1] for k in np.arange(self.order)])
+
+        return X
+
+    def _prepare_xy(self, data):
+        l = len(data)
+        X = []
+        Y = []
+
+        for t in np.arange(self.order, l):
+            X.append([data[t - k - 1] for k in np.arange(self.order)])
+            Y.append(data[t])
+
+        return (X,Y)
 
     def train(self, data, **kwargs):
-        self.data = np.array(data)
+        X,Y = self._prepare_xy(data)
 
+        self.kdtree = KDTree(X)
+        self.values = Y
 
     def knn(self, sample):
+        X = self._prepare_x(sample)
+        _, ix = self.kdtree.query(X, self.k)
 
-        if self.order == 1:
+        return [self.values[k] for k in ix.flatten() ]
-            dist = np.apply_along_axis(lambda x: (x - sample) ** 2, 0, self.data)
-            ix = np.argsort(dist) + 1
-        else:
-            dist = []
-            for k in np.arange(self.order, len(self.data)):
-                dist.append(sum([ (self.data[k - kk] - sample[kk])**2 for kk in range(self.order)]))
-            ix = np.argsort(np.array(dist)) + self.order + 1
 
-        ix2 = np.clip(ix[:self.k], 0, len(self.data)-1)
+    def forecast(self, data, **kwargs):
-        return self.data[ix2]
+        ret = []
+        for k in np.arange(self.order, len(data)):
+
+            sample = data[k-self.order : k]
+
+            forecasts = self.knn(sample)
+
+            ret.append(np.nanmean(forecasts))
+
+        return ret
+
+    def forecast_ahead(self, data, steps, **kwargs):
+        start = kwargs.get('start', self.order)
+
+        sample = [k for k in data[start - self.order: start]]
+
+        for k in np.arange(self.order, steps + self.order):
+            tmp = self.forecast(sample[k-self.order:k])
+            sample.append(tmp)
+
+        return sample[-steps]
+
+    def forecast_interval(self, data, **kwargs):
+
+        alpha = kwargs.get('alpha',self.alpha)
+
+        ret = []
+        for k in np.arange(self.order, len(data)):
+
+            sample = data[k-self.order : k]
+
+            forecasts = self.knn(sample)
+
+            i = np.percentile(forecasts, [alpha*100, (1-alpha)*100]).tolist()
+            ret.append(i)
+
+        return ret
+
+    def forecast_ahead_interval(self, data, steps, **kwargs):
+        alpha = kwargs.get('alpha', self.alpha)
+
+        ret = []
+
+        start = kwargs.get('start', self.order)
+
+        sample = [[k] for k in data[start - self.order: start]]
+
+        for k in np.arange(self.order, steps + self.order):
+            forecasts = []
+
+            lags = [sample[k - i - 1] for i in np.arange(0, self.order)]
+
+            # Trace the possible paths
+            for path in product(*lags):
+                forecasts.extend(self.knn(path))
+
+            sample.append(sampler(forecasts, np.arange(.1, 1, 0.1), bounds=True))
+
+            interval = np.percentile(forecasts, [alpha*100, (1-alpha)*100]).tolist()
+
+            ret.append(interval)
+
+        return ret
 
     def forecast_distribution(self, data, **kwargs):
         ret = []
 
-        smooth = kwargs.get("smooth", "KDE")
+        smooth = kwargs.get("smooth", "histogram")
-        alpha = kwargs.get("alpha", None)
 
         uod = self.get_UoD()
 
@@ -65,4 +151,32 @@ class KNearestNeighbors(fts.FTS):
 
         return ret
 
+    def forecast_ahead_distribution(self, data, steps, **kwargs):
+        smooth = kwargs.get("smooth", "histogram")
+
+        ret = []
+
+        start = kwargs.get('start', self.order)
+
+        uod = self.get_UoD()
+
+        sample = [[k] for k in data[start - self.order: start]]
+
+        for k in np.arange(self.order, steps + self.order):
+            forecasts = []
+
+            lags = [sample[k - i - 1] for i in np.arange(0, self.order)]
+
+            # Trace the possible paths
+            for path in product(*lags):
+                forecasts.extend(self.knn(path))
+
+            dist = ProbabilityDistribution.ProbabilityDistribution(smooth, uod=uod, data=forecasts,
+                                                                   name="", **kwargs)
+            ret.append(dist)
+
+            sample.append(sampler(forecasts, np.arange(.1, 1, 0.1), bounds=True))
+
+        return ret
+
 

pyFTS/common/fts.py

@@ -519,8 +519,10 @@ class FTS(object):
             return data
 
     def get_UoD(self):
-        #return [self.original_min, self.original_max]
+        if self.partitioner is not None:
             return [self.partitioner.min, self.partitioner.max]
+        else:
+            return [self.original_min, self.original_max]
 
     def __str__(self):
         """String representation of the model"""

pyFTS/models/ensemble/ensemble.py

@@ -167,8 +167,7 @@ class EnsembleFTS(fts.FTS):
         if "method" in kwargs:
             self.interval_method = kwargs.get('method','quantile')
 
-        if 'alpha' in kwargs:
+        self.alpha = kwargs.get('alpha', self.alpha)
-            self.alpha = kwargs.get('alpha',0.05)
 
         l = len(data)
 
@@ -189,15 +188,12 @@ class EnsembleFTS(fts.FTS):
         if 'method' in kwargs:
             self.interval_method = kwargs.get('method','quantile')
 
-        if 'alpha' in kwargs:
         self.alpha = kwargs.get('alpha', self.alpha)
 
         ret = []
 
         start = kwargs.get('start', self.order)
 
-        uod = self.get_UoD()
-
         sample = [[k] for k in data[start - self.order: start]]
 
         for k in np.arange(self.order, steps + self.order):

pyFTS/tests/general.py

@@ -13,29 +13,51 @@ from pyFTS.partitioners import Grid, Entropy, Util as pUtil, Simple
 from pyFTS.benchmarks import benchmarks as bchmk, Measures
 from pyFTS.models import chen, yu, cheng, ismailefendi, hofts, pwfts, tsaur, song, sadaei, ifts
 from pyFTS.models.ensemble import ensemble
-from pyFTS.common import Transformations, Membership
+from pyFTS.common import Transformations, Membership, Util
-from pyFTS.benchmarks import arima, quantreg, BSTS, gaussianproc
+from pyFTS.benchmarks import arima, quantreg, BSTS, gaussianproc, knn
 from pyFTS.fcm import fts, common, GA
 
 from pyFTS.data import TAIEX, NASDAQ, SP500
 
+train = TAIEX.get_data()[:800]
+test = TAIEX.get_data()[800:1000]
+
+order = 2
+model = knn.KNearestNeighbors(order=order)
+model.fit(train)
+
+horizon=7
+
+intervals05 = model.predict(test[:10], type='interval', alpha=.05, steps_ahead=horizon)
+
+print(test[:10])
+print(intervals05)
+
+intervals25 = model.predict(test[:10], type='interval', alpha=.25, steps_ahead=horizon)
+distributions = model.predict(test[:10], type='distribution', steps_ahead=horizon, smoothing=0.01, num_bins=100)
+
+fig, ax = plt.subplots(nrows=1, ncols=1,figsize=[15,5])
+ax.plot(test[:10], label='Original',color='black')
+Util.plot_interval2(intervals05, test[:10], start_at=model.order, ax=ax, color='green', label='alpha=.05'.format(model.order))
+Util.plot_interval2(intervals25, test[:10], start_at=model.order, ax=ax, color='green', label='alpha=.25'.format(model.order))
+Util.plot_distribution2(distributions, test[:10], start_at=model.order, ax=ax, cmap="Blues")
+
+print("")
+'''
+
 datasets = {}
 
 datasets['TAIEX'] = TAIEX.get_data()[:5000]
 datasets['NASDAQ'] = NASDAQ.get_data()[:5000]
 datasets['SP500'] = SP500.get_data()[10000:15000]
 
-methods = [ensemble.SimpleEnsembleFTS]*8
+methods = [ensemble.SimpleEnsembleFTS]*4
 
 methods_parameters = [
-    {'name': 'EnsembleFTS-HOFTS-10-.05', 'fts_method': hofts.HighOrderFTS, 'partitions': np.arange(20,50,10), 'alpha': .05},
+    {'name': 'EnsembleFTS-HOFTS-10', 'fts_method': hofts.HighOrderFTS, 'partitions': np.arange(20,50,10)},
-    {'name': 'EnsembleFTS-HOFTS-5-.05', 'fts_method': hofts.HighOrderFTS, 'partitions': np.arange(20,50,5), 'alpha': .05},
+    {'name': 'EnsembleFTS-HOFTS-5', 'fts_method': hofts.HighOrderFTS, 'partitions': np.arange(20,50,5)},
-    {'name': 'EnsembleFTS-HOFTS-10-.25', 'fts_method': hofts.HighOrderFTS, 'partitions': np.arange(20,50,10), 'alpha': .25},
+    {'name': 'EnsembleFTS-WHOFTS-10', 'fts_method': hofts.WeightedHighOrderFTS, 'partitions': np.arange(20,50,10)},
-    {'name': 'EnsembleFTS-HOFTS-5-.25', 'fts_method': hofts.HighOrderFTS, 'partitions': np.arange(20,50,5), 'alpha': .25},
+    {'name': 'EnsembleFTS-WHOFTS-5', 'fts_method': hofts.WeightedHighOrderFTS, 'partitions': np.arange(20,50,5)}
-    {'name': 'EnsembleFTS-WHOFTS-10-.05', 'fts_method': hofts.WeightedHighOrderFTS, 'partitions': np.arange(20,50,10), 'alpha': .05},
-    {'name': 'EnsembleFTS-WHOFTS-5-.05', 'fts_method': hofts.WeightedHighOrderFTS, 'partitions': np.arange(20,50,5), 'alpha': .05},
-    {'name': 'EnsembleFTS-WHOFTS-10-.25', 'fts_method': hofts.WeightedHighOrderFTS, 'partitions': np.arange(20,50,10), 'alpha': .25},
-    {'name': 'EnsembleFTS-WHOFTS-5-.25', 'fts_method': hofts.WeightedHighOrderFTS, 'partitions': np.arange(20,50,5), 'alpha': .25},
 ]
 
 for dataset_name, dataset in datasets.items():
@@ -46,6 +68,7 @@ for dataset_name, dataset in datasets.items():
                                     transformations=[None],
                                     orders=[3],
                                     partitions=[None],
-                                    type='interval',
+                                    type='distribution',
-                                    #distributed=True, nodes=['192.168.0.110', '192.168.0.107','192.168.0.106'],
+                                    distributed=True, nodes=['192.168.0.110', '192.168.0.107','192.168.0.106'],
-                                    file="tmp.db", dataset=dataset_name, tag="gridsearch")
+                                    file="experiments.db", dataset=dataset_name, tag="gridsearch")
+'''