mjboost 0.1.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- mjboost-0.1.0/LICENSE +21 -0
- mjboost-0.1.0/MANIFEST.in +5 -0
- mjboost-0.1.0/PKG-INFO +140 -0
- mjboost-0.1.0/README.md +114 -0
- mjboost-0.1.0/examples/quickstart.py +39 -0
- mjboost-0.1.0/pyproject.toml +35 -0
- mjboost-0.1.0/setup.cfg +4 -0
- mjboost-0.1.0/src/mjboost/__init__.py +13 -0
- mjboost-0.1.0/src/mjboost/_core.py +1395 -0
- mjboost-0.1.0/src/mjboost/_estimators.py +173 -0
- mjboost-0.1.0/src/mjboost/_explain.py +171 -0
- mjboost-0.1.0/src/mjboost/_fanova.py +363 -0
- mjboost-0.1.0/src/mjboost/_numba.py +749 -0
- mjboost-0.1.0/src/mjboost/_prep.py +170 -0
- mjboost-0.1.0/src/mjboost.egg-info/PKG-INFO +140 -0
- mjboost-0.1.0/src/mjboost.egg-info/SOURCES.txt +18 -0
- mjboost-0.1.0/src/mjboost.egg-info/dependency_links.txt +1 -0
- mjboost-0.1.0/src/mjboost.egg-info/requires.txt +12 -0
- mjboost-0.1.0/src/mjboost.egg-info/top_level.txt +1 -0
- mjboost-0.1.0/tests/test_mjboost.py +97 -0
mjboost-0.1.0/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 Minjong Cheon
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
mjboost-0.1.0/PKG-INFO
ADDED
|
@@ -0,0 +1,140 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: mjboost
|
|
3
|
+
Version: 0.1.0
|
|
4
|
+
Summary: Gradient boosting with Kolmogorov-Arnold units: interpretable regression and binary classification beyond pairwise interactions, with an exact functional ANOVA of the fitted model
|
|
5
|
+
Author: Minjong Cheon
|
|
6
|
+
License-Expression: MIT
|
|
7
|
+
Keywords: gradient boosting,interpretable machine learning,generalized additive model,Kolmogorov-Arnold,functional ANOVA,tabular data
|
|
8
|
+
Classifier: Programming Language :: Python :: 3
|
|
9
|
+
Classifier: Operating System :: OS Independent
|
|
10
|
+
Classifier: Intended Audience :: Science/Research
|
|
11
|
+
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
|
|
12
|
+
Requires-Python: >=3.9
|
|
13
|
+
Description-Content-Type: text/markdown
|
|
14
|
+
License-File: LICENSE
|
|
15
|
+
Requires-Dist: numpy>=1.22
|
|
16
|
+
Requires-Dist: scipy>=1.8
|
|
17
|
+
Requires-Dist: numba>=0.56
|
|
18
|
+
Requires-Dist: pandas>=1.4
|
|
19
|
+
Requires-Dist: scikit-learn>=1.1
|
|
20
|
+
Provides-Extra: plot
|
|
21
|
+
Requires-Dist: matplotlib>=3.5; extra == "plot"
|
|
22
|
+
Provides-Extra: test
|
|
23
|
+
Requires-Dist: pytest; extra == "test"
|
|
24
|
+
Requires-Dist: matplotlib>=3.5; extra == "test"
|
|
25
|
+
Dynamic: license-file
|
|
26
|
+
|
|
27
|
+
# MJBoost
|
|
28
|
+
|
|
29
|
+
Gradient boosting with Kolmogorov–Arnold units for interpretable regression and binary classification
|
|
30
|
+
beyond pairwise interactions.
|
|
31
|
+
|
|
32
|
+
Each boosting round adds one Kolmogorov–Arnold (KA) unit
|
|
33
|
+
|
|
34
|
+
h(x) = g( φ_1(x_1) + φ_2(x_2) + … )
|
|
35
|
+
|
|
36
|
+
whose inner functions φ and outer function g are penalized splines, fitted in closed form by alternating
|
|
37
|
+
penalized least squares from principal Hessian directions of the residual. An additive spline backbone
|
|
38
|
+
carries the main effects, and categorical and missing inputs enter the units directly. Because a unit depends
|
|
39
|
+
on each input only through one inner function, the fitted model has an **exact functional ANOVA
|
|
40
|
+
decomposition** into main effects, pair effects and a higher-order remainder, which one-dimensional
|
|
41
|
+
convolutions compute. The model reads as curves and surfaces where the data allow, and the decomposition
|
|
42
|
+
measures the higher-order part where they do not.
|
|
43
|
+
|
|
44
|
+
## Installation
|
|
45
|
+
|
|
46
|
+
```bash
|
|
47
|
+
pip install mjboost # once released on PyPI
|
|
48
|
+
pip install "mjboost[plot]" # with matplotlib for the figures
|
|
49
|
+
```
|
|
50
|
+
|
|
51
|
+
From a copy of this repository:
|
|
52
|
+
|
|
53
|
+
```bash
|
|
54
|
+
pip install .
|
|
55
|
+
```
|
|
56
|
+
|
|
57
|
+
Requirements: Python ≥ 3.9, numpy, scipy, numba, pandas, scikit-learn. The first fit in a new environment
|
|
58
|
+
compiles the numba kernels (tens of seconds); later sessions load them from the cache.
|
|
59
|
+
|
|
60
|
+
## Quick start
|
|
61
|
+
|
|
62
|
+
```python
|
|
63
|
+
import pandas as pd
|
|
64
|
+
from mjboost import MJBoostRegressor
|
|
65
|
+
|
|
66
|
+
model = MJBoostRegressor().fit(X_train, y_train) # pandas DataFrame or NumPy array
|
|
67
|
+
pred = model.predict(X_test)
|
|
68
|
+
```
|
|
69
|
+
|
|
70
|
+
DataFrame columns of type `category`, `object`, `bool` or `string` are categorical (each enters a unit as one
|
|
71
|
+
input); numeric columns may contain `NaN`. A level not seen in training gets the reference value of its
|
|
72
|
+
input. `cat_features` overrides the detection with a list of column names or positions.
|
|
73
|
+
|
|
74
|
+
Binary classification:
|
|
75
|
+
|
|
76
|
+
```python
|
|
77
|
+
from mjboost import MJBoostClassifier
|
|
78
|
+
|
|
79
|
+
clf = MJBoostClassifier().fit(X_train, y_train) # any two labels
|
|
80
|
+
proba = clf.predict_proba(X_test) # columns follow clf.classes_
|
|
81
|
+
```
|
|
82
|
+
|
|
83
|
+
Both estimators follow the scikit-learn API (`get_params`, `clone`, `Pipeline`, `cross_val_score`, pickling).
|
|
84
|
+
|
|
85
|
+
## Reading the fitted model
|
|
86
|
+
|
|
87
|
+
```python
|
|
88
|
+
ex = model.explain() # exact functional ANOVA (log-odds scale for a classifier)
|
|
89
|
+
|
|
90
|
+
ex.shares(X_test) # variance shares of main / pair / higher-order parts, R² of truncations
|
|
91
|
+
ex.importance(X_test) # spread of every main and pair effect, largest first
|
|
92
|
+
ex.main_effect("age") # DataFrame: value, effect (one row per level for a categorical input)
|
|
93
|
+
xa, xb, F = ex.pair_effect("lat", "long")
|
|
94
|
+
C = ex.components(X_test) # f0, main effects, pair effects, remainder, prediction per row
|
|
95
|
+
|
|
96
|
+
ex.plot_main("age"); ex.plot_pair("lat", "long"); ex.plot_shares(X_test)
|
|
97
|
+
```
|
|
98
|
+
|
|
99
|
+
The decomposition is taken with respect to the product of the training marginals:
|
|
100
|
+
f0 = E f, f_j = E[f | x_j] − f0, f_jk = E[f | x_j, x_k] − f_j − f_k − f0, and the remainder holds the
|
|
101
|
+
interactions of order three and higher. At every row the parts sum to the prediction.
|
|
102
|
+
|
|
103
|
+
## Model selection and options
|
|
104
|
+
|
|
105
|
+
By default (`selection="two_race"`, the setting of the paper) five-fold cross-validation inside the training
|
|
106
|
+
data chooses the additive backbone (on or off), the unit resolution (16 or 32 interior knots), the column
|
|
107
|
+
fraction (0.5 or 1) and support pruning (on or off). The candidates are raced on the first inner fold with
|
|
108
|
+
stops at 100 and 200 units, and three finalists continue on the other folds. Two races run on the same
|
|
109
|
+
folds, one over the eight candidates with support pruning and one over all sixteen; the winner of the second
|
|
110
|
+
replaces that of the first when its cross-validated loss is lower. The prediction is the average of the five
|
|
111
|
+
fold models. `model.selection_` reports the winning race, its candidate and the inner loss of each race.
|
|
112
|
+
|
|
113
|
+
| Parameter | Default | Meaning |
|
|
114
|
+
|---|---|---|
|
|
115
|
+
| `selection` | `"two_race"` | `"pruned"`: the first race only (about a third of the computation); `"all"`: the second only |
|
|
116
|
+
| `cat_features` | `"auto"` | categorical columns: `"auto"`, `None`, or a list of names / positions |
|
|
117
|
+
| `max_units` | 4000 | maximum number of units per fold model (early stopping decides) |
|
|
118
|
+
| `learning_rate` | 0.3 | shrinkage of every unit |
|
|
119
|
+
| `max_inputs_per_unit` | 12 | cap on the inputs a unit may use |
|
|
120
|
+
| `n_jobs` | 1 | processes for the inner cross-validation (the result does not depend on it) |
|
|
121
|
+
| `random_state` | 0 | seed of the inner folds and column draws |
|
|
122
|
+
| `params` | `None` | dict of further settings of the underlying `KAUnitBoost` (for experts) |
|
|
123
|
+
|
|
124
|
+
## Results
|
|
125
|
+
|
|
126
|
+
On the 35 regression tasks of OpenML-CTR23 (ten folds, one thread per fit), MJBoost has a lower test RMSE than
|
|
127
|
+
EBM on 21 tasks, a statistically equal one on 12 and a higher one on 2, with a 29.1% lower geometric-mean
|
|
128
|
+
RMSE; against tuned LightGBM the geometric-mean RMSE is 13.2% lower. On synthetic targets of interaction
|
|
129
|
+
order three to six its test error is up to 13 times lower than that of EBM. On 32 binary tasks of
|
|
130
|
+
OpenML-CC18 its test log-loss is on par with that of EBM. Over the CTR23 run MJBoost used 130 CPU-hours,
|
|
131
|
+
EBM 68 and tuned LightGBM 19.
|
|
132
|
+
|
|
133
|
+
## Citation
|
|
134
|
+
|
|
135
|
+
Minjong Cheon, Hanseon Joo and Eunji Lee. Gradient Boosting with Kolmogorov–Arnold Units for Interpretable
|
|
136
|
+
Regression beyond Pairwise Interactions. 2026.
|
|
137
|
+
|
|
138
|
+
## License
|
|
139
|
+
|
|
140
|
+
MIT
|
mjboost-0.1.0/README.md
ADDED
|
@@ -0,0 +1,114 @@
|
|
|
1
|
+
# MJBoost
|
|
2
|
+
|
|
3
|
+
Gradient boosting with Kolmogorov–Arnold units for interpretable regression and binary classification
|
|
4
|
+
beyond pairwise interactions.
|
|
5
|
+
|
|
6
|
+
Each boosting round adds one Kolmogorov–Arnold (KA) unit
|
|
7
|
+
|
|
8
|
+
h(x) = g( φ_1(x_1) + φ_2(x_2) + … )
|
|
9
|
+
|
|
10
|
+
whose inner functions φ and outer function g are penalized splines, fitted in closed form by alternating
|
|
11
|
+
penalized least squares from principal Hessian directions of the residual. An additive spline backbone
|
|
12
|
+
carries the main effects, and categorical and missing inputs enter the units directly. Because a unit depends
|
|
13
|
+
on each input only through one inner function, the fitted model has an **exact functional ANOVA
|
|
14
|
+
decomposition** into main effects, pair effects and a higher-order remainder, which one-dimensional
|
|
15
|
+
convolutions compute. The model reads as curves and surfaces where the data allow, and the decomposition
|
|
16
|
+
measures the higher-order part where they do not.
|
|
17
|
+
|
|
18
|
+
## Installation
|
|
19
|
+
|
|
20
|
+
```bash
|
|
21
|
+
pip install mjboost # once released on PyPI
|
|
22
|
+
pip install "mjboost[plot]" # with matplotlib for the figures
|
|
23
|
+
```
|
|
24
|
+
|
|
25
|
+
From a copy of this repository:
|
|
26
|
+
|
|
27
|
+
```bash
|
|
28
|
+
pip install .
|
|
29
|
+
```
|
|
30
|
+
|
|
31
|
+
Requirements: Python ≥ 3.9, numpy, scipy, numba, pandas, scikit-learn. The first fit in a new environment
|
|
32
|
+
compiles the numba kernels (tens of seconds); later sessions load them from the cache.
|
|
33
|
+
|
|
34
|
+
## Quick start
|
|
35
|
+
|
|
36
|
+
```python
|
|
37
|
+
import pandas as pd
|
|
38
|
+
from mjboost import MJBoostRegressor
|
|
39
|
+
|
|
40
|
+
model = MJBoostRegressor().fit(X_train, y_train) # pandas DataFrame or NumPy array
|
|
41
|
+
pred = model.predict(X_test)
|
|
42
|
+
```
|
|
43
|
+
|
|
44
|
+
DataFrame columns of type `category`, `object`, `bool` or `string` are categorical (each enters a unit as one
|
|
45
|
+
input); numeric columns may contain `NaN`. A level not seen in training gets the reference value of its
|
|
46
|
+
input. `cat_features` overrides the detection with a list of column names or positions.
|
|
47
|
+
|
|
48
|
+
Binary classification:
|
|
49
|
+
|
|
50
|
+
```python
|
|
51
|
+
from mjboost import MJBoostClassifier
|
|
52
|
+
|
|
53
|
+
clf = MJBoostClassifier().fit(X_train, y_train) # any two labels
|
|
54
|
+
proba = clf.predict_proba(X_test) # columns follow clf.classes_
|
|
55
|
+
```
|
|
56
|
+
|
|
57
|
+
Both estimators follow the scikit-learn API (`get_params`, `clone`, `Pipeline`, `cross_val_score`, pickling).
|
|
58
|
+
|
|
59
|
+
## Reading the fitted model
|
|
60
|
+
|
|
61
|
+
```python
|
|
62
|
+
ex = model.explain() # exact functional ANOVA (log-odds scale for a classifier)
|
|
63
|
+
|
|
64
|
+
ex.shares(X_test) # variance shares of main / pair / higher-order parts, R² of truncations
|
|
65
|
+
ex.importance(X_test) # spread of every main and pair effect, largest first
|
|
66
|
+
ex.main_effect("age") # DataFrame: value, effect (one row per level for a categorical input)
|
|
67
|
+
xa, xb, F = ex.pair_effect("lat", "long")
|
|
68
|
+
C = ex.components(X_test) # f0, main effects, pair effects, remainder, prediction per row
|
|
69
|
+
|
|
70
|
+
ex.plot_main("age"); ex.plot_pair("lat", "long"); ex.plot_shares(X_test)
|
|
71
|
+
```
|
|
72
|
+
|
|
73
|
+
The decomposition is taken with respect to the product of the training marginals:
|
|
74
|
+
f0 = E f, f_j = E[f | x_j] − f0, f_jk = E[f | x_j, x_k] − f_j − f_k − f0, and the remainder holds the
|
|
75
|
+
interactions of order three and higher. At every row the parts sum to the prediction.
|
|
76
|
+
|
|
77
|
+
## Model selection and options
|
|
78
|
+
|
|
79
|
+
By default (`selection="two_race"`, the setting of the paper) five-fold cross-validation inside the training
|
|
80
|
+
data chooses the additive backbone (on or off), the unit resolution (16 or 32 interior knots), the column
|
|
81
|
+
fraction (0.5 or 1) and support pruning (on or off). The candidates are raced on the first inner fold with
|
|
82
|
+
stops at 100 and 200 units, and three finalists continue on the other folds. Two races run on the same
|
|
83
|
+
folds, one over the eight candidates with support pruning and one over all sixteen; the winner of the second
|
|
84
|
+
replaces that of the first when its cross-validated loss is lower. The prediction is the average of the five
|
|
85
|
+
fold models. `model.selection_` reports the winning race, its candidate and the inner loss of each race.
|
|
86
|
+
|
|
87
|
+
| Parameter | Default | Meaning |
|
|
88
|
+
|---|---|---|
|
|
89
|
+
| `selection` | `"two_race"` | `"pruned"`: the first race only (about a third of the computation); `"all"`: the second only |
|
|
90
|
+
| `cat_features` | `"auto"` | categorical columns: `"auto"`, `None`, or a list of names / positions |
|
|
91
|
+
| `max_units` | 4000 | maximum number of units per fold model (early stopping decides) |
|
|
92
|
+
| `learning_rate` | 0.3 | shrinkage of every unit |
|
|
93
|
+
| `max_inputs_per_unit` | 12 | cap on the inputs a unit may use |
|
|
94
|
+
| `n_jobs` | 1 | processes for the inner cross-validation (the result does not depend on it) |
|
|
95
|
+
| `random_state` | 0 | seed of the inner folds and column draws |
|
|
96
|
+
| `params` | `None` | dict of further settings of the underlying `KAUnitBoost` (for experts) |
|
|
97
|
+
|
|
98
|
+
## Results
|
|
99
|
+
|
|
100
|
+
On the 35 regression tasks of OpenML-CTR23 (ten folds, one thread per fit), MJBoost has a lower test RMSE than
|
|
101
|
+
EBM on 21 tasks, a statistically equal one on 12 and a higher one on 2, with a 29.1% lower geometric-mean
|
|
102
|
+
RMSE; against tuned LightGBM the geometric-mean RMSE is 13.2% lower. On synthetic targets of interaction
|
|
103
|
+
order three to six its test error is up to 13 times lower than that of EBM. On 32 binary tasks of
|
|
104
|
+
OpenML-CC18 its test log-loss is on par with that of EBM. Over the CTR23 run MJBoost used 130 CPU-hours,
|
|
105
|
+
EBM 68 and tuned LightGBM 19.
|
|
106
|
+
|
|
107
|
+
## Citation
|
|
108
|
+
|
|
109
|
+
Minjong Cheon, Hanseon Joo and Eunji Lee. Gradient Boosting with Kolmogorov–Arnold Units for Interpretable
|
|
110
|
+
Regression beyond Pairwise Interactions. 2026.
|
|
111
|
+
|
|
112
|
+
## License
|
|
113
|
+
|
|
114
|
+
MIT
|
|
@@ -0,0 +1,39 @@
|
|
|
1
|
+
"""MJBoost on two data sets bundled with scikit-learn (no download)."""
|
|
2
|
+
import numpy as np
|
|
3
|
+
from sklearn.datasets import load_breast_cancer, load_diabetes
|
|
4
|
+
from sklearn.model_selection import train_test_split
|
|
5
|
+
|
|
6
|
+
from mjboost import MJBoostClassifier, MJBoostRegressor
|
|
7
|
+
|
|
8
|
+
# regression
|
|
9
|
+
X, y = load_diabetes(return_X_y=True, as_frame=True)
|
|
10
|
+
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.2, random_state=0)
|
|
11
|
+
reg = MJBoostRegressor().fit(Xtr, ytr)
|
|
12
|
+
print("diabetes test RMSE:", np.sqrt(np.mean((reg.predict(Xte) - yte) ** 2)))
|
|
13
|
+
print("selection:", reg.selection_["race"], reg.selection_["candidate"])
|
|
14
|
+
|
|
15
|
+
ex = reg.explain()
|
|
16
|
+
print(ex.shares(Xte, yte))
|
|
17
|
+
print(ex.importance(Xte).head(8))
|
|
18
|
+
print(ex.main_effect("bmi", n=9))
|
|
19
|
+
|
|
20
|
+
# binary classification
|
|
21
|
+
X, y = load_breast_cancer(return_X_y=True, as_frame=True)
|
|
22
|
+
Xtr, Xte, ytr, yte = train_test_split(X, y, test_size=0.2, random_state=0, stratify=y)
|
|
23
|
+
clf = MJBoostClassifier(selection="pruned").fit(Xtr, ytr)
|
|
24
|
+
print("breast cancer test accuracy:", np.mean(clf.predict(Xte) == yte))
|
|
25
|
+
|
|
26
|
+
try: # figures (pip install "mjboost[plot]")
|
|
27
|
+
import matplotlib
|
|
28
|
+
matplotlib.use("Agg")
|
|
29
|
+
import matplotlib.pyplot as plt
|
|
30
|
+
fig, axes = plt.subplots(1, 3, figsize=(10, 2.8))
|
|
31
|
+
ex.plot_main("bmi", ax=axes[0])
|
|
32
|
+
top = ex.importance(Xte).query("kind == 'pair'")["term"].iloc[0].split(" x ")
|
|
33
|
+
ex.plot_pair(top[0], top[1], ax=axes[1])
|
|
34
|
+
ex.plot_shares(Xte, ax=axes[2])
|
|
35
|
+
fig.tight_layout()
|
|
36
|
+
fig.savefig("mjboost_diabetes.png", dpi=150)
|
|
37
|
+
print("figure: mjboost_diabetes.png")
|
|
38
|
+
except ImportError:
|
|
39
|
+
pass
|
|
@@ -0,0 +1,35 @@
|
|
|
1
|
+
[build-system]
|
|
2
|
+
requires = ["setuptools>=77"]
|
|
3
|
+
build-backend = "setuptools.build_meta"
|
|
4
|
+
|
|
5
|
+
[project]
|
|
6
|
+
name = "mjboost"
|
|
7
|
+
version = "0.1.0"
|
|
8
|
+
description = "Gradient boosting with Kolmogorov-Arnold units: interpretable regression and binary classification beyond pairwise interactions, with an exact functional ANOVA of the fitted model"
|
|
9
|
+
readme = "README.md"
|
|
10
|
+
license = "MIT"
|
|
11
|
+
license-files = ["LICENSE"]
|
|
12
|
+
authors = [{name = "Minjong Cheon"}]
|
|
13
|
+
requires-python = ">=3.9"
|
|
14
|
+
dependencies = [
|
|
15
|
+
"numpy>=1.22",
|
|
16
|
+
"scipy>=1.8",
|
|
17
|
+
"numba>=0.56",
|
|
18
|
+
"pandas>=1.4",
|
|
19
|
+
"scikit-learn>=1.1",
|
|
20
|
+
]
|
|
21
|
+
keywords = ["gradient boosting", "interpretable machine learning", "generalized additive model",
|
|
22
|
+
"Kolmogorov-Arnold", "functional ANOVA", "tabular data"]
|
|
23
|
+
classifiers = [
|
|
24
|
+
"Programming Language :: Python :: 3",
|
|
25
|
+
"Operating System :: OS Independent",
|
|
26
|
+
"Intended Audience :: Science/Research",
|
|
27
|
+
"Topic :: Scientific/Engineering :: Artificial Intelligence",
|
|
28
|
+
]
|
|
29
|
+
|
|
30
|
+
[project.optional-dependencies]
|
|
31
|
+
plot = ["matplotlib>=3.5"]
|
|
32
|
+
test = ["pytest", "matplotlib>=3.5"]
|
|
33
|
+
|
|
34
|
+
[tool.setuptools.packages.find]
|
|
35
|
+
where = ["src"]
|
mjboost-0.1.0/setup.cfg
ADDED
|
@@ -0,0 +1,13 @@
|
|
|
1
|
+
"""MJBoost: gradient boosting with Kolmogorov-Arnold units for interpretable regression and binary
|
|
2
|
+
classification beyond pairwise interactions.
|
|
3
|
+
|
|
4
|
+
from mjboost import MJBoostRegressor
|
|
5
|
+
model = MJBoostRegressor().fit(X_train, y_train)
|
|
6
|
+
model.predict(X_test)
|
|
7
|
+
ex = model.explain() # exact functional ANOVA: main effects, pair effects, higher-order part
|
|
8
|
+
"""
|
|
9
|
+
from ._estimators import MJBoostClassifier, MJBoostRegressor
|
|
10
|
+
from ._explain import Explanation
|
|
11
|
+
|
|
12
|
+
__version__ = "0.1.0"
|
|
13
|
+
__all__ = ["MJBoostRegressor", "MJBoostClassifier", "Explanation", "__version__"]
|