Last week in #135, I talked about mlsauce’s v0.13.0, and LSBoost in particular. When using LSBoost, it’s now possible to:

  • Obtain prediction intervals for regression, notably by employing Split Conformal Prediction.
  • Take into account an a priori heterogeneity in explanatory variables through clustering.

In v0.17.0, I added a 2 new features to LSBoost:

  • The possibility to add polynomial interaction functions of explanatory variables to the mix (see sklearn.preprocessing.PolynomialFeatures for more details). This is done by setting the degree parameter of LSBoostRegressor or LSBoostClassifier to a positive integer value.
  • The possibility to use Elastic Net as a solver (a base learner), in addition to ridge, and lasso. /!\ enet (for Elastic Net) will become the default value for the solver parameter next week, as it’s fast (uses coordinate descent) and gracefully combines both ridge regression and the lasso. ridge, and lasso will remain available to avoid breaking existing pipelines, but you’d have to specify them explicitly as solvers. For enet, reg_lambda is still used as a regularization parameter, and an alpha (in [0, 1]) parameter defines a compromise between lasso (alpha = 1) and ridge (alpha = 0) penalties. The default value for alpha is 0.5.

This week, I’ll show how to use the new features, to gain an intuition of how they work. Keep in mind however: these examples only show that it’s possible to overfit the training set (hence reducing the loss function’s magnitude) by adding some clusters. The whole model’s hyperparameters need to be ‘fine-tuned’, the learning_rate and n_iterations in particular, for example by using  GPopt. Next week, I’ll update the documentation and notably this working paper in a more comprehensive way.

The best way (feel free to answer this question on stackoverflow) to install the package is still to use the development version (tested in colab…):

pip install git+https://github.com/Techtonique/mlsauce.git --verbose

You can reproduce the results with this notebook.

# 0 - Install and import data

!pip uninstall mlsauce --yes

!pip install git+https://github.com/Techtonique/mlsauce.git --verbose

import mlsauce as ms
from sklearn.datasets import fetch_california_housing
from sklearn.model_selection import train_test_split
from time import time
import matplotlib.pyplot as plt


dataset = fetch_california_housing()
X = dataset.data
y = dataset.target
# split data into training test and test set
X_train, X_test, y_train, y_test = train_test_split(X, y,
                                                    test_size=0.2)

1 - solver = 'ridge' and polynomial degree >= 2

obj1 = ms.LSBoostRegressor()
print(obj1.get_params())
start = time()
obj1.fit(X_train, y_train)
print(f"\n Elapsed: {time()-start}")
print(f"loss: {obj1.obj['loss']}")

obj2 = ms.LSBoostRegressor(n_clusters=2, learning_rate=0.2)
print(obj2.get_params())
start = time()
obj2.fit(X_train, y_train)
print(f"\n Elapsed: {time()-start}")
print(f"loss: {obj2.obj['loss']}")

obj3 = ms.LSBoostRegressor(n_clusters=2, degree=2)
print(obj3.get_params())
start = time()
obj3.fit(X_train, y_train)
print(f"\n Elapsed: {time()-start}")
print(f"loss: {obj3.obj['loss']}")

obj4 = ms.LSBoostRegressor(n_clusters=3, degree=3)
print(obj4.get_params())
start = time()
obj4.fit(X_train, y_train)
print(f"\n Elapsed: {time()-start}")
print(f"loss: {obj4.obj['loss']}")


{'activation': 'relu', 'alpha': 0.5, 'backend': 'cpu', 'cluster_scaling': 'standard', 'clustering_method': 'kmeans', 'col_sample': 1, 'degree': 0, 'direct_link': 1, 'dropout': 0, 'kernel': None, 'learning_rate': 0.1, 'n_clusters': 0, 'n_estimators': 100, 'n_hidden_features': 5, 'reg_lambda': 0.1, 'replications': None, 'row_sample': 1, 'seed': 123, 'solver': 'ridge', 'tolerance': 0.0001, 'type_pi': None, 'verbose': 1}


100%|██████████| 100/100 [00:01<00:00, 79.75it/s]



 Elapsed: 1.2682864665985107
loss: [139.32111847614541, 131.70308253447635, 125.19499783631174, 119.59820193020641, 114.93497551784296, 110.8527083582196, 107.54354587416081, 104.7134674297217, 102.39708217776042, 100.42933365689483, 98.83144460196893, 97.52844685361345, 96.30852369223383, 95.39003032176558, 94.54352885226272, 93.96512308838865, 93.47502447209713, 93.03977520683647, 92.65829051516376, 92.2802164369876, 92.01046242081102, 91.74049787360917, 91.52039293368806, 91.35783846316978, 91.18787261074638, 91.04354476856568, 90.94852971198672, 90.86008655659633, 90.78674222651686, 90.74000692736819, 90.63215837544449, 90.52950586401757, 90.48233169529428, 90.37423492871945, 90.32472893384154, 90.21890835467052, 90.21111046873041, 90.13909134552976, 90.0139953107389, 89.97612995083936, 89.94348148438813, 89.93304401122035, 89.89959616857168, 89.85807584682216, 89.8392966602071, 89.80667250002107, 89.77924928861444, 89.77367456574875, 89.75559610481965, 89.71397020176407, 89.59175806869736, 89.56664373311831, 89.54091445963215, 89.46168703639418, 89.45189787914167, 89.42845994767242, 89.36573344897069, 89.34909990107334, 89.31251207605847, 89.30960378231673, 89.2900193168802, 89.15571629926559, 89.1332794039042, 89.06143615145245, 89.04323756985643, 89.00592023420243, 89.0020187793033, 88.96391864010155, 88.94535574532689, 88.93949958765987, 88.91891525045274, 88.90760445979019, 88.89798522168671, 88.88338871701343, 88.85385469298946, 88.82386465259158, 88.81106654969633, 88.7874323259479, 88.7748363821829, 88.74795322218621, 88.74667050075695, 88.68556071633893, 88.67176501250366, 88.66029100611858, 88.65336168694326, 88.62276144125545, 88.6201402067312, 88.60335631240642, 88.57280352448245, 88.5594483820922, 88.49023189096201, 88.44785594877571, 88.43544524381494, 88.40251935329863, 88.39618944976945, 88.37910622847218, 88.37611957899422, 88.3640605723657, 88.33872924544201, 88.33664108588486]
{'activation': 'relu', 'alpha': 0.5, 'backend': 'cpu', 'cluster_scaling': 'standard', 'clustering_method': 'kmeans', 'col_sample': 1, 'degree': 0, 'direct_link': 1, 'dropout': 0, 'kernel': None, 'learning_rate': 0.2, 'n_clusters': 2, 'n_estimators': 100, 'n_hidden_features': 5, 'reg_lambda': 0.1, 'replications': None, 'row_sample': 1, 'seed': 123, 'solver': 'ridge', 'tolerance': 0.0001, 'type_pi': None, 'verbose': 1}


100%|██████████| 100/100 [00:01<00:00, 73.32it/s]



 Elapsed: 1.5329296588897705
loss: [130.83367893567655, 118.36978853470147, 109.70263745563274, 103.66906516909664, 99.41491268844976, 96.62374388556682, 94.83298574697442, 93.68222947764735, 92.9153846160194, 92.31803473535508, 91.97115696869, 91.71083711206805, 91.51983855278394, 91.38012699363394, 91.28552640926058, 91.23390055534463, 91.18292155812665, 91.13266090009023, 91.01351872357111, 90.94152103319904, 90.88256741236287, 90.76917649363118, 90.68916515483345, 90.6241463204776, 90.54149853760747, 90.50983879162368, 90.43465235971378, 90.33848797211253, 90.2680297351062, 90.24095813688729, 90.18312455033704, 90.05454272754294, 90.03731488611776, 89.9471623200098, 89.89669185915396, 89.75343778455495, 89.70676861370553, 89.56565026994494, 89.45908888711088, 89.36680063879811, 89.34911086865165, 89.31088815546408, 89.17808729324963, 89.1125761522851, 89.09901562204598, 89.01167429798109, 88.95076279488383, 88.92192476032442, 88.89143552678257, 88.8399353076311, 88.7733300141213, 88.76804706309517, 88.69814287249048, 88.5788453268133, 88.57819734317886, 88.48007430990472, 88.38166860953304, 88.35506600185093, 88.33691014634925, 88.32996570082982, 88.22934992687436, 88.10191428809424, 88.08363243633644, 88.05564920162551, 88.03926225732519, 88.01212148638815, 87.97514703192572, 87.93122870673841, 87.92194926097288, 87.88218854141395, 87.86070378212733, 87.85385147379762, 87.75543031113773, 87.73143263493998, 87.69174842759524, 87.67965321647605, 87.673454211299, 87.63804462078319, 87.5854261384591, 87.54857396987046, 87.5293009106566, 87.5269864514613, 87.52241658670458, 87.46538794704216, 87.3685761090272, 87.24895344995836, 87.23889062286989, 87.22965005021842, 87.19078151706046, 87.17004581387963, 87.08081926597161, 87.04897060083277, 87.02453881475769, 87.00386156116492, 86.99266101056178, 86.98340206274179, 86.93831520947467, 86.93420075031149, 86.86839781973956, 86.78001242587659]
{'activation': 'relu', 'alpha': 0.5, 'backend': 'cpu', 'cluster_scaling': 'standard', 'clustering_method': 'kmeans', 'col_sample': 1, 'degree': 2, 'direct_link': 1, 'dropout': 0, 'kernel': None, 'learning_rate': 0.1, 'n_clusters': 2, 'n_estimators': 100, 'n_hidden_features': 5, 'reg_lambda': 0.1, 'replications': None, 'row_sample': 1, 'seed': 123, 'solver': 'ridge', 'tolerance': 0.0001, 'type_pi': None, 'verbose': 1}


100%|██████████| 100/100 [00:03<00:00, 31.79it/s]



 Elapsed: 3.2886624336242676
loss: [138.46406679289797, 130.02911687784407, 122.74598097793128, 116.47657742727863, 111.08524444228846, 106.55680329178956, 102.72017473071023, 99.47042863602837, 96.75628668834405, 94.50553551404269, 92.63949144013449, 91.0691722107126, 89.80317774899414, 88.74124411409235, 87.87069719885474, 87.16695784585367, 86.55517523465173, 86.02400455678165, 85.62603500112755, 85.27070382572722, 85.00266337877103, 84.76565021927969, 84.58713376109829, 84.40933853463417, 84.28716539837815, 84.18976818357947, 84.10173059703584, 84.03046303014399, 83.97274729370876, 83.92075520312329, 83.88192039529922, 83.84111176784998, 83.79511863502773, 83.76136287330272, 83.73337968833519, 83.71127142629713, 83.6918744941698, 83.64818431151006, 83.63210752232445, 83.62237186924207, 83.59768445289845, 83.57998394117516, 83.56990256435788, 83.56076546063005, 83.54908605112888, 83.54167871651399, 83.52967231659126, 83.51665242112935, 83.49544601625972, 83.47893395060991, 83.45981431662462, 83.45316456129692, 83.42405210925749, 83.41671254058762, 83.41456973986233, 83.40860977678491, 83.40464082684298, 83.40020895250508, 83.39602524105483, 83.3853736635742, 83.37780008790719, 83.37597304768038, 83.3687469839559, 83.36297069936599, 83.36029035126958, 83.35203272637273, 83.35059500671038, 83.34967485591902, 83.34000115609149, 83.32722314668666, 83.32358467659589, 83.31136126470331, 83.29259945057044, 83.28914212751293, 83.28521039760028, 83.28481746246688, 83.27460743074492, 83.26962034466729, 83.2628548274941, 83.26036082589354, 83.25827409034957, 83.25321243448998, 83.24859393673994, 83.24568366451265, 83.23549679513344, 83.23430138376206, 83.22909756587205, 83.22599323519538, 83.22550682810837, 83.21019176234417, 83.20724081337342, 83.20574882859364, 83.2042111859721, 83.19948926198819, 83.193568246941, 83.15961915958252, 83.1577172169777, 83.15182952964035, 83.15132244242335, 83.14764227408152]
{'activation': 'relu', 'alpha': 0.5, 'backend': 'cpu', 'cluster_scaling': 'standard', 'clustering_method': 'kmeans', 'col_sample': 1, 'degree': 3, 'direct_link': 1, 'dropout': 0, 'kernel': None, 'learning_rate': 0.1, 'n_clusters': 3, 'n_estimators': 100, 'n_hidden_features': 5, 'reg_lambda': 0.1, 'replications': None, 'row_sample': 1, 'seed': 123, 'solver': 'ridge', 'tolerance': 0.0001, 'type_pi': None, 'verbose': 1}


100%|██████████| 100/100 [00:10<00:00,  9.89it/s]


 Elapsed: 10.495529413223267
loss: [138.14931908588397, 129.4027435243869, 121.88694206439065, 115.41145625256686, 109.87726578132731, 105.14963912878896, 101.17914759601891, 97.83191553199156, 95.0289111825631, 92.67406781644411, 90.73041397940855, 89.11116067394822, 87.70840560127974, 86.60244317194072, 85.6780514819482, 84.92440650331801, 84.29736111426118, 83.79278115563103, 83.35892332229895, 83.00162692480497, 82.72002881320502, 82.47126331888968, 82.2781242828601, 82.12232589978221, 81.98856994702976, 81.87415215829667, 81.77693471717099, 81.6961133540323, 81.6090635806128, 81.54414873216453, 81.49038650109277, 81.43905501208901, 81.39981895004144, 81.37411574944969, 81.33550631900393, 81.30587634281169, 81.28408474336716, 81.25579561110607, 81.23994744567902, 81.21942411130279, 81.20594539477105, 81.19324314448927, 81.17555560987032, 81.16151415480925, 81.1482104356931, 81.12852873652301, 81.11831172643035, 81.07918029614552, 81.06147632347029, 81.0524168083218, 81.04378841205214, 81.02537802368047, 80.9925138598936, 80.97930632272424, 80.96601996808367, 80.95349353716256, 80.94038307845716, 80.91914687808031, 80.88390169235261, 80.86194574294117, 80.85448718971496, 80.8423225915586, 80.83661148464881, 80.82743562581062, 80.8187318876223, 80.8122028241299, 80.80500996384802, 80.79800245234749, 80.79156632394738, 80.78025350191442, 80.77210849733146, 80.76065721716013, 80.75626236456611, 80.74383580649274, 80.7337232134181, 80.72810744836396, 80.72038672045795, 80.71322853377548, 80.70518351193886, 80.69680451558467, 80.69294035424154, 80.68319307292629, 80.67887178809434, 80.66527265856205, 80.65506840664501, 80.6477872646086, 80.63140309323022, 80.6203078711991, 80.60896984379703, 80.60379161057065, 80.59523925524817, 80.57821467893916, 80.57061121781464, 80.56348346600964, 80.55992990139632, 80.55021895814008, 80.5440861559823, 80.5409277611609, 80.53768386936636, 80.53084993480212]

# Plotting the lines with labels
plt.plot(obj1.obj['loss'], label='Loss - learning_rate=0.1')
plt.plot(obj2.obj['loss'], label='Loss - n_clusters=2, learning_rate=0.2')
plt.plot(obj3.obj['loss'], label='Loss - n_clusters=2, degree=2')
plt.plot(obj4.obj['loss'], label='Loss - n_clusters=3, degree=3')

# Displaying the legend
plt.legend()

# Show the plot
plt.show()

pres-image

2 - solver = 'elasticnet'

obj1 = ms.LSBoostRegressor(solver="enet", n_estimators=25)
print(obj1.get_params())
start = time()
obj1.fit(X_train, y_train)
print(f"\n Elapsed: {time()-start}")
print(f"loss: {obj1.obj['loss']}")

obj2 = ms.LSBoostRegressor(n_clusters=2, learning_rate=0.2, solver="enet", 
                           n_estimators=25)
print(obj2.get_params())
start = time()
obj2.fit(X_train, y_train)
print(f"\n Elapsed: {time()-start}")
print(f"loss: {obj2.obj['loss']}")

obj3 = ms.LSBoostRegressor(n_clusters=2, learning_rate=0.2, solver="enet", 
                           n_estimators=25, alpha=0)
print(obj3.get_params())
start = time()
obj3.fit(X_train, y_train)
print(f"\n Elapsed: {time()-start}")
print(f"loss: {obj3.obj['loss']}")

obj4 = ms.LSBoostRegressor(n_clusters=2, learning_rate=0.2, solver="enet", 
                           n_estimators=25, alpha=1)
print(obj4.get_params())
start = time()
obj4.fit(X_train, y_train)
print(f"\n Elapsed: {time()-start}")
print(f"loss: {obj4.obj['loss']}")

obj5 = ms.LSBoostRegressor(n_clusters=2, learning_rate=0.1, solver="enet", 
                           n_estimators=25, alpha=1, degree=2)
print(obj5.get_params())
start = time()
obj5.fit(X_train, y_train)
print(f"\n Elapsed: {time()-start}")
print(f"loss: {obj5.obj['loss']}")

obj6 = ms.LSBoostRegressor(n_clusters=2, learning_rate=0.06, solver="enet", 
                           n_estimators=25, alpha=1, degree=3)
print(obj6.get_params())
start = time()
obj6.fit(X_train, y_train)
print(f"\n Elapsed: {time()-start}")
print(f"loss: {obj6.obj['loss']}")

# Plotting the lines with labels
plt.plot(obj1.obj['loss'], label='Loss - learning_rate=0.1, alpha=0.5 (L1+L2)')
plt.plot(obj2.obj['loss'], label='Loss - n_clusters=2, learning_rate=0.2, alpha=0.5 (L1+L2)')
plt.plot(obj3.obj['loss'], label='Loss - learning_rate=0.1, alpha=0 (L2 pen.)')
plt.plot(obj4.obj['loss'], label='Loss - n_clusters=2, learning_rate=0.2, alpha=1 (L1 pen.)')
plt.plot(obj5.obj['loss'], label='Loss - n_clusters=2, learning_rate=0.1, alpha=1, degree=2')
plt.plot(obj6.obj['loss'], label='Loss - n_clusters=2, learning_rate=0.06, alpha=1, degree=3')

# Displaying the legend
plt.legend()

# Show the plot
plt.show()

pres-image

Comments powered by Talkyard.