aggregationslib 0.2.2__tar.gz → 0.2.3__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.
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/.gitignore +1 -0
- {aggregationslib-0.2.2/aggregationslib.egg-info → aggregationslib-0.2.3}/PKG-INFO +65 -1
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/README.md +58 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/README.rst +64 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/__init__.py +12 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/aggregation.py +175 -4
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/aggregations.py +16 -4
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/decimalaggregations.py +205 -10
- {aggregationslib-0.2.2 → aggregationslib-0.2.3/aggregationslib.egg-info}/PKG-INFO +65 -1
- aggregationslib-0.2.3/aggregationslib.egg-info/scm_version.json +8 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/fuzzy-integrals.md +71 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/tests/test_decimal_fuzzy_integrals.py +106 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/tests/test_fuzzy_integrals.py +86 -0
- aggregationslib-0.2.2/aggregationslib.egg-info/scm_version.json +0 -8
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/LICENSE +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/README.pl.md +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/conjuctive/__init__.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/conjuctive/purepython.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/disjunctive/__init__.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/example.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/feature_selection/__init__.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/feature_selection/binary_fuzzy_entropy.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/feature_selection/fuzzy_com_entropy.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/models/__init__.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/models/ensemble/__init__.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/models/ensemble/aggregation_regressor.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/models/single/__init__.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib/purepython.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib.egg-info/SOURCES.txt +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib.egg-info/dependency_links.txt +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib.egg-info/requires.txt +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib.egg-info/scm_file_list.json +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/aggregationslib.egg-info/top_level.txt +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/assets/icons/icon.ico +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/assets/icons/icon.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/assets/icons/icon.svg +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/build_docs.ps1 +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/commands.md +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/.nojekyll +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/aggregation-classes.md +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/aggregation-plots.md +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregation-surfaces-all.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregation-surfaces.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_amn_p=0.5.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_amx_p=0.5.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_ar.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_ex2_p=2,q=2.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_ex3_p=2,q=1.5.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_ex_r=1.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_gm.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_hm.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_lm_r=2.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_lo.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_md.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_mn.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_mx.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_ol.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_oln_p=1.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_pr.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_pw_r=2.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/aggregations/A_qd.png +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/assets/site.css +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/en/index.html +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/index.html +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/docs/pl/index.html +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/generate_aggregation_plots.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/pyproject.toml +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/setup.cfg +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/tests/__init__.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/tests/data/wolfram_integral_reference.json +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/tests/test_aggregations.py +0 -0
- {aggregationslib-0.2.2 → aggregationslib-0.2.3}/tests/test_decimal_exponential_mean.py +0 -0
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
Metadata-Version: 2.4
|
|
2
2
|
Name: aggregationslib
|
|
3
|
-
Version: 0.2.
|
|
3
|
+
Version: 0.2.3
|
|
4
4
|
Summary: Python implementation of Arithmetic, quasi arithmetic and other aggregating functions
|
|
5
5
|
Author: wgalka
|
|
6
6
|
Project-URL: Homepage, https://github.com/wgalka/Means
|
|
@@ -111,10 +111,44 @@ For an input vector :math:`(x_1,\dots,x_n)\in[0,1]^n`, let
|
|
|
111
111
|
|
|
112
112
|
.. math:: A_{ch}^{(\mu)}(x_1,\dots,x_n)=\sum_{i=1}^{n}(x_{\sigma(i)}-x_{\sigma(i-1)})\,\mu(A_{\sigma(i)})
|
|
113
113
|
|
|
114
|
+
Available fuzzy measures for the Choquet integral:
|
|
115
|
+
|
|
116
|
+
Each constructor returns a monotone fuzzy measure with
|
|
117
|
+
:math:`\mu(\emptyset)=0` and :math:`\mu(N)=1`.
|
|
118
|
+
|
|
119
|
+
- **Uniform/cardinality** (``uniform_measure(n)``):
|
|
120
|
+
:math:`\mu(A)=|A|/n`; every position has equal singleton weight.
|
|
121
|
+
- **Relative** (``relative_measure(n)``):
|
|
122
|
+
:math:`\mu(A)=\frac{\sum_{i\in A}i}{\sum_{j=1}^{n}j}`; singleton
|
|
123
|
+
weights are proportional to ``1, 2, ..., n``, so later positions
|
|
124
|
+
receive larger weights.
|
|
125
|
+
- **k-fuzzy/Sugeno** (``k_fuzzy_measure(weights)``): for ``k ≠ 0``,
|
|
126
|
+
:math:`\mu(A)=\frac{\prod_{i\in A}(1+k w_i)-1}{k}`; for ``k = 0``,
|
|
127
|
+
:math:`\mu(A)=\sum_{i\in A}w_i`. The parameter ``k`` is chosen so
|
|
128
|
+
:math:`\mu(N)=1`; incompatible singleton weights raise ``ValueError``.
|
|
129
|
+
- **Weighted mean** (``weighted_measure(weights)``):
|
|
130
|
+
:math:`\mu(A)=\sum_{i\in A}w_i`. The weights must be in ``[0, 1]`` and
|
|
131
|
+
sum to ``1``; with the Choquet integral this gives a weighted
|
|
132
|
+
arithmetic mean.
|
|
133
|
+
- **OWA** (``owa_measure(weights)``):
|
|
134
|
+
:math:`\mu(A)=\sum_{j=1}^{|A|}w_j`. The normalized weights are applied
|
|
135
|
+
by descending value rank, so the first belongs to the largest value;
|
|
136
|
+
with the Choquet integral this capacity gives the OWA aggregation.
|
|
137
|
+
- **Exponential cardinality**: :math:`\mu(A)=(|A|/n)^q` is not
|
|
138
|
+
implemented yet; choosing ``q`` is future work.
|
|
139
|
+
|
|
140
|
+
The constructors are in ``aggregationslib.aggregation``; Decimal
|
|
141
|
+
equivalents are available in ``aggregationslib.decimalaggregations``.
|
|
142
|
+
|
|
114
143
|
With the default cardinality measure :math:`\mu(A)=|A|/n`, this is the
|
|
115
144
|
arithmetic mean,
|
|
116
145
|
:math:`A_{ch}^{(\mu)}(x_1,\dots,x_n)=\frac{1}{n}\sum_{i=1}^{n}x_i`.
|
|
117
146
|
|
|
147
|
+
For the quasi-overlap generalization from Equation (9), pass a
|
|
148
|
+
two-argument function ``O(x, mu)`` as ``quasi_overlap`` to
|
|
149
|
+
``choquet_integral`` or ``A_ch``. The default remains the classical
|
|
150
|
+
product-based Choquet integral.
|
|
151
|
+
|
|
118
152
|
:math:`\mathbf{A_{sg}^{(\mu)}}` **- Sugeno integral [1]**
|
|
119
153
|
|
|
120
154
|
Tags: ``averaging`` ``aggregation interval: [0,1]``
|
|
@@ -127,6 +161,36 @@ For an input vector :math:`(x_1,\dots,x_n)\in[0,1]^n`, let
|
|
|
127
161
|
|
|
128
162
|
.. math:: A_{sg}^{(\mu)}(x_1,\dots,x_n)=\max_{1\leq i\leq n}\min(x_{\sigma(i)},\mu(A_{\sigma(i)}))
|
|
129
163
|
|
|
164
|
+
Available fuzzy measures for the Sugeno integral:
|
|
165
|
+
|
|
166
|
+
Each constructor returns a monotone fuzzy measure with
|
|
167
|
+
:math:`\mu(\emptyset)=0` and :math:`\mu(N)=1`. The same constructors are
|
|
168
|
+
available in ``aggregationslib.aggregation`` and
|
|
169
|
+
``aggregationslib.decimalaggregations``.
|
|
170
|
+
|
|
171
|
+
- **Uniform/cardinality** (``uniform_measure(n)``):
|
|
172
|
+
:math:`\mu(A)=|A|/n`; every position has equal singleton weight.
|
|
173
|
+
- **Relative** (``relative_measure(n)``):
|
|
174
|
+
:math:`\mu(A)=\frac{\sum_{i\in A}i}{\sum_{j=1}^{n}j}`; singleton
|
|
175
|
+
weights are proportional to ``1, 2, ..., n``, so later positions
|
|
176
|
+
receive larger weights.
|
|
177
|
+
- **k-fuzzy/Sugeno** (``k_fuzzy_measure(weights)``): for ``k ≠ 0``,
|
|
178
|
+
:math:`\mu(A)=\frac{\prod_{i\in A}(1+k w_i)-1}{k}`; for ``k = 0``,
|
|
179
|
+
:math:`\mu(A)=\sum_{i\in A}w_i`. The parameter ``k`` is chosen so
|
|
180
|
+
:math:`\mu(N)=1`; incompatible singleton weights raise ``ValueError``.
|
|
181
|
+
- **Weighted mean** (``weighted_measure(weights)``):
|
|
182
|
+
:math:`\mu(A)=\sum_{i\in A}w_i`. The weights must be in ``[0, 1]`` and
|
|
183
|
+
sum to ``1``. In the Sugeno integral this measure is used in the
|
|
184
|
+
max–min formula; it does not turn Sugeno into a weighted arithmetic
|
|
185
|
+
mean.
|
|
186
|
+
- **OWA** (``owa_measure(weights)``):
|
|
187
|
+
:math:`\mu(A)=\sum_{j=1}^{|A|}w_j`. The normalized weights are applied
|
|
188
|
+
by descending value rank, so the first belongs to the largest value.
|
|
189
|
+
In the Sugeno integral this capacity is used in the max–min formula;
|
|
190
|
+
the result is not the OWA aggregation.
|
|
191
|
+
- **Exponential cardinality**: :math:`\mu(A)=(|A|/n)^q` is not
|
|
192
|
+
implemented yet; choosing ``q`` is future work.
|
|
193
|
+
|
|
130
194
|
:math:`\mathbf{A_{ar}}` **- Arithmetic mean**
|
|
131
195
|
|
|
132
196
|
Tags: ``averaging`` ``symmetric``
|
|
@@ -75,9 +75,40 @@ Define $A_{\sigma(i)}=\{\sigma(i),\dots,\sigma(n)\}$ and $x_{\sigma(0)}=0$:
|
|
|
75
75
|
|
|
76
76
|
$$A_{ch}^{(\mu)}(x_1,\dots,x_n)=\sum_{i=1}^{n}(x_{\sigma(i)}-x_{\sigma(i-1)})\,\mu(A_{\sigma(i)})$$
|
|
77
77
|
|
|
78
|
+
Available fuzzy measures for the Choquet integral:
|
|
79
|
+
|
|
80
|
+
Each constructor returns a monotone fuzzy measure with
|
|
81
|
+
$\mu(\emptyset)=0$ and $\mu(N)=1$.
|
|
82
|
+
|
|
83
|
+
- **Uniform/cardinality** (`uniform_measure(n)`):
|
|
84
|
+
$\mu(A)=|A|/n$; every position has equal singleton weight.
|
|
85
|
+
- **Relative** (`relative_measure(n)`):
|
|
86
|
+
$\mu(A)=\frac{\sum_{i\in A}i}{\sum_{j=1}^{n}j}$; singleton weights are
|
|
87
|
+
proportional to `1, 2, ..., n`, so later positions receive larger weights.
|
|
88
|
+
- **k-fuzzy/Sugeno** (`k_fuzzy_measure(weights)`): for `k ≠ 0`,
|
|
89
|
+
$\mu(A)=\frac{\prod_{i\in A}(1+k w_i)-1}{k}$; for `k = 0`,
|
|
90
|
+
$\mu(A)=\sum_{i\in A}w_i$. The parameter `k` is chosen so
|
|
91
|
+
$\mu(N)=1$; incompatible singleton weights raise `ValueError`.
|
|
92
|
+
- **Weighted mean** (`weighted_measure(weights)`):
|
|
93
|
+
$\mu(A)=\sum_{i\in A}w_i$. The weights must be in `[0, 1]` and sum to
|
|
94
|
+
`1`; with the Choquet integral this gives a weighted arithmetic mean.
|
|
95
|
+
- **OWA** (`owa_measure(weights)`):
|
|
96
|
+
$\mu(A)=\sum_{j=1}^{|A|}w_j$. The normalized weights are applied by
|
|
97
|
+
descending value rank, so the first belongs to the largest value; with the
|
|
98
|
+
Choquet integral this capacity gives the OWA aggregation.
|
|
99
|
+
- **Exponential cardinality**: $\mu(A)=(|A|/n)^q$ is not implemented yet;
|
|
100
|
+
choosing `q` is future work.
|
|
101
|
+
|
|
102
|
+
The constructors are in `aggregationslib.aggregation`; Decimal equivalents
|
|
103
|
+
are available in `aggregationslib.decimalaggregations`.
|
|
104
|
+
|
|
78
105
|
With the default cardinality measure $\mu(A)=|A|/n$, this is the arithmetic
|
|
79
106
|
mean, $A_{ch}^{(\mu)}(x_1,\dots,x_n)=\frac{1}{n}\sum_{i=1}^{n}x_i$.
|
|
80
107
|
|
|
108
|
+
For the quasi-overlap generalization from Equation (9), pass a two-argument
|
|
109
|
+
function `O(x, mu)` as `quasi_overlap` to `choquet_integral` or `A_ch`. The
|
|
110
|
+
default remains the classical product-based Choquet integral.
|
|
111
|
+
|
|
81
112
|
$\mathbf{A_{sg}^{(\mu)}}$ **- Sugeno integral [1]**
|
|
82
113
|
|
|
83
114
|
Tags: `averaging` `aggregation interval: [0,1]` `symmetric if μ is symmetric`
|
|
@@ -88,6 +119,33 @@ $A_{\sigma(i)}=\{\sigma(i),\dots,\sigma(n)\}$:
|
|
|
88
119
|
|
|
89
120
|
$$A_{sg}^{(\mu)}(x_1,\dots,x_n)=\max_{1\leq i\leq n}\min(x_{\sigma(i)},\mu(A_{\sigma(i)}))$$
|
|
90
121
|
|
|
122
|
+
Available fuzzy measures for the Sugeno integral:
|
|
123
|
+
|
|
124
|
+
Each constructor returns a monotone fuzzy measure with
|
|
125
|
+
$\mu(\emptyset)=0$ and $\mu(N)=1$. The same constructors are available in
|
|
126
|
+
`aggregationslib.aggregation` and `aggregationslib.decimalaggregations`.
|
|
127
|
+
|
|
128
|
+
- **Uniform/cardinality** (`uniform_measure(n)`):
|
|
129
|
+
$\mu(A)=|A|/n$; every position has equal singleton weight.
|
|
130
|
+
- **Relative** (`relative_measure(n)`):
|
|
131
|
+
$\mu(A)=\frac{\sum_{i\in A}i}{\sum_{j=1}^{n}j}$; singleton weights are
|
|
132
|
+
proportional to `1, 2, ..., n`, so later positions receive larger weights.
|
|
133
|
+
- **k-fuzzy/Sugeno** (`k_fuzzy_measure(weights)`): for `k ≠ 0`,
|
|
134
|
+
$\mu(A)=\frac{\prod_{i\in A}(1+k w_i)-1}{k}$; for `k = 0`,
|
|
135
|
+
$\mu(A)=\sum_{i\in A}w_i$. The parameter `k` is chosen so
|
|
136
|
+
$\mu(N)=1$; incompatible singleton weights raise `ValueError`.
|
|
137
|
+
- **Weighted mean** (`weighted_measure(weights)`):
|
|
138
|
+
$\mu(A)=\sum_{i\in A}w_i$. The weights must be in `[0, 1]` and sum to
|
|
139
|
+
`1`. In the Sugeno integral this measure is used in the max–min formula; it
|
|
140
|
+
does not turn Sugeno into a weighted arithmetic mean.
|
|
141
|
+
- **OWA** (`owa_measure(weights)`):
|
|
142
|
+
$\mu(A)=\sum_{j=1}^{|A|}w_j$. The normalized weights are applied by
|
|
143
|
+
descending value rank, so the first belongs to the largest value. In the
|
|
144
|
+
Sugeno integral this capacity is used in the max–min formula; the result is
|
|
145
|
+
not the OWA aggregation.
|
|
146
|
+
- **Exponential cardinality**: $\mu(A)=(|A|/n)^q$ is not implemented yet;
|
|
147
|
+
choosing `q` is future work.
|
|
148
|
+
|
|
91
149
|
$\mathbf{A_{ar}}$ **- Arithmetic mean**
|
|
92
150
|
|
|
93
151
|
Tags: `averaging` `symmetric` `aggregation interval: any [a,b] ⊂ ℝ`
|
|
@@ -89,10 +89,44 @@ For an input vector :math:`(x_1,\dots,x_n)\in[0,1]^n`, let
|
|
|
89
89
|
|
|
90
90
|
.. math:: A_{ch}^{(\mu)}(x_1,\dots,x_n)=\sum_{i=1}^{n}(x_{\sigma(i)}-x_{\sigma(i-1)})\,\mu(A_{\sigma(i)})
|
|
91
91
|
|
|
92
|
+
Available fuzzy measures for the Choquet integral:
|
|
93
|
+
|
|
94
|
+
Each constructor returns a monotone fuzzy measure with
|
|
95
|
+
:math:`\mu(\emptyset)=0` and :math:`\mu(N)=1`.
|
|
96
|
+
|
|
97
|
+
- **Uniform/cardinality** (``uniform_measure(n)``):
|
|
98
|
+
:math:`\mu(A)=|A|/n`; every position has equal singleton weight.
|
|
99
|
+
- **Relative** (``relative_measure(n)``):
|
|
100
|
+
:math:`\mu(A)=\frac{\sum_{i\in A}i}{\sum_{j=1}^{n}j}`; singleton
|
|
101
|
+
weights are proportional to ``1, 2, ..., n``, so later positions
|
|
102
|
+
receive larger weights.
|
|
103
|
+
- **k-fuzzy/Sugeno** (``k_fuzzy_measure(weights)``): for ``k ≠ 0``,
|
|
104
|
+
:math:`\mu(A)=\frac{\prod_{i\in A}(1+k w_i)-1}{k}`; for ``k = 0``,
|
|
105
|
+
:math:`\mu(A)=\sum_{i\in A}w_i`. The parameter ``k`` is chosen so
|
|
106
|
+
:math:`\mu(N)=1`; incompatible singleton weights raise ``ValueError``.
|
|
107
|
+
- **Weighted mean** (``weighted_measure(weights)``):
|
|
108
|
+
:math:`\mu(A)=\sum_{i\in A}w_i`. The weights must be in ``[0, 1]`` and
|
|
109
|
+
sum to ``1``; with the Choquet integral this gives a weighted
|
|
110
|
+
arithmetic mean.
|
|
111
|
+
- **OWA** (``owa_measure(weights)``):
|
|
112
|
+
:math:`\mu(A)=\sum_{j=1}^{|A|}w_j`. The normalized weights are applied
|
|
113
|
+
by descending value rank, so the first belongs to the largest value;
|
|
114
|
+
with the Choquet integral this capacity gives the OWA aggregation.
|
|
115
|
+
- **Exponential cardinality**: :math:`\mu(A)=(|A|/n)^q` is not
|
|
116
|
+
implemented yet; choosing ``q`` is future work.
|
|
117
|
+
|
|
118
|
+
The constructors are in ``aggregationslib.aggregation``; Decimal
|
|
119
|
+
equivalents are available in ``aggregationslib.decimalaggregations``.
|
|
120
|
+
|
|
92
121
|
With the default cardinality measure :math:`\mu(A)=|A|/n`, this is the
|
|
93
122
|
arithmetic mean,
|
|
94
123
|
:math:`A_{ch}^{(\mu)}(x_1,\dots,x_n)=\frac{1}{n}\sum_{i=1}^{n}x_i`.
|
|
95
124
|
|
|
125
|
+
For the quasi-overlap generalization from Equation (9), pass a
|
|
126
|
+
two-argument function ``O(x, mu)`` as ``quasi_overlap`` to
|
|
127
|
+
``choquet_integral`` or ``A_ch``. The default remains the classical
|
|
128
|
+
product-based Choquet integral.
|
|
129
|
+
|
|
96
130
|
:math:`\mathbf{A_{sg}^{(\mu)}}` **- Sugeno integral [1]**
|
|
97
131
|
|
|
98
132
|
Tags: ``averaging`` ``aggregation interval: [0,1]``
|
|
@@ -105,6 +139,36 @@ For an input vector :math:`(x_1,\dots,x_n)\in[0,1]^n`, let
|
|
|
105
139
|
|
|
106
140
|
.. math:: A_{sg}^{(\mu)}(x_1,\dots,x_n)=\max_{1\leq i\leq n}\min(x_{\sigma(i)},\mu(A_{\sigma(i)}))
|
|
107
141
|
|
|
142
|
+
Available fuzzy measures for the Sugeno integral:
|
|
143
|
+
|
|
144
|
+
Each constructor returns a monotone fuzzy measure with
|
|
145
|
+
:math:`\mu(\emptyset)=0` and :math:`\mu(N)=1`. The same constructors are
|
|
146
|
+
available in ``aggregationslib.aggregation`` and
|
|
147
|
+
``aggregationslib.decimalaggregations``.
|
|
148
|
+
|
|
149
|
+
- **Uniform/cardinality** (``uniform_measure(n)``):
|
|
150
|
+
:math:`\mu(A)=|A|/n`; every position has equal singleton weight.
|
|
151
|
+
- **Relative** (``relative_measure(n)``):
|
|
152
|
+
:math:`\mu(A)=\frac{\sum_{i\in A}i}{\sum_{j=1}^{n}j}`; singleton
|
|
153
|
+
weights are proportional to ``1, 2, ..., n``, so later positions
|
|
154
|
+
receive larger weights.
|
|
155
|
+
- **k-fuzzy/Sugeno** (``k_fuzzy_measure(weights)``): for ``k ≠ 0``,
|
|
156
|
+
:math:`\mu(A)=\frac{\prod_{i\in A}(1+k w_i)-1}{k}`; for ``k = 0``,
|
|
157
|
+
:math:`\mu(A)=\sum_{i\in A}w_i`. The parameter ``k`` is chosen so
|
|
158
|
+
:math:`\mu(N)=1`; incompatible singleton weights raise ``ValueError``.
|
|
159
|
+
- **Weighted mean** (``weighted_measure(weights)``):
|
|
160
|
+
:math:`\mu(A)=\sum_{i\in A}w_i`. The weights must be in ``[0, 1]`` and
|
|
161
|
+
sum to ``1``. In the Sugeno integral this measure is used in the
|
|
162
|
+
max–min formula; it does not turn Sugeno into a weighted arithmetic
|
|
163
|
+
mean.
|
|
164
|
+
- **OWA** (``owa_measure(weights)``):
|
|
165
|
+
:math:`\mu(A)=\sum_{j=1}^{|A|}w_j`. The normalized weights are applied
|
|
166
|
+
by descending value rank, so the first belongs to the largest value.
|
|
167
|
+
In the Sugeno integral this capacity is used in the max–min formula;
|
|
168
|
+
the result is not the OWA aggregation.
|
|
169
|
+
- **Exponential cardinality**: :math:`\mu(A)=(|A|/n)^q` is not
|
|
170
|
+
implemented yet; choosing ``q`` is future work.
|
|
171
|
+
|
|
108
172
|
:math:`\mathbf{A_{ar}}` **- Arithmetic mean**
|
|
109
173
|
|
|
110
174
|
Tags: ``averaging`` ``symmetric``
|
|
@@ -1,3 +1,10 @@
|
|
|
1
|
+
from .aggregation import (
|
|
2
|
+
k_fuzzy_measure,
|
|
3
|
+
owa_measure,
|
|
4
|
+
relative_measure,
|
|
5
|
+
uniform_measure,
|
|
6
|
+
weighted_measure,
|
|
7
|
+
)
|
|
1
8
|
from .aggregations import (
|
|
2
9
|
A_amn,
|
|
3
10
|
A_amx,
|
|
@@ -23,6 +30,11 @@ from .aggregations import (
|
|
|
23
30
|
)
|
|
24
31
|
|
|
25
32
|
__all__ = [
|
|
33
|
+
"k_fuzzy_measure",
|
|
34
|
+
"owa_measure",
|
|
35
|
+
"relative_measure",
|
|
36
|
+
"uniform_measure",
|
|
37
|
+
"weighted_measure",
|
|
26
38
|
"A_amn",
|
|
27
39
|
"A_amx",
|
|
28
40
|
"A_ar",
|
|
@@ -307,6 +307,164 @@ def _as_unit_interval_vector(values: object) -> np.ndarray:
|
|
|
307
307
|
return vector
|
|
308
308
|
|
|
309
309
|
|
|
310
|
+
def _capacity_from_function(n: int, capacity_function) -> FuzzyMeasure:
|
|
311
|
+
values = {}
|
|
312
|
+
for mask in range(1 << n):
|
|
313
|
+
subset = frozenset(index for index in range(n) if mask & (1 << index))
|
|
314
|
+
values[subset] = capacity_function(subset)
|
|
315
|
+
return FuzzyMeasure(n, values)
|
|
316
|
+
|
|
317
|
+
|
|
318
|
+
def _measure_size(n: int) -> int:
|
|
319
|
+
if isinstance(n, bool) or not isinstance(n, Integral) or n < 1:
|
|
320
|
+
raise ValueError("n must be a positive integer")
|
|
321
|
+
return int(n)
|
|
322
|
+
|
|
323
|
+
|
|
324
|
+
def _normalized_weights(weights: object, name: str = "weights") -> np.ndarray:
|
|
325
|
+
vector = _as_unit_interval_vector(weights)
|
|
326
|
+
if not isclose(
|
|
327
|
+
math.fsum(float(weight) for weight in vector),
|
|
328
|
+
1.0,
|
|
329
|
+
abs_tol=_TOLERANCE,
|
|
330
|
+
rel_tol=0.0,
|
|
331
|
+
):
|
|
332
|
+
raise ValueError(f"{name} must sum to 1")
|
|
333
|
+
return vector
|
|
334
|
+
|
|
335
|
+
|
|
336
|
+
def uniform_measure(n: int) -> FuzzyMeasure:
|
|
337
|
+
"""Return the cardinality capacity ``mu(A) = |A| / n``."""
|
|
338
|
+
n = _measure_size(n)
|
|
339
|
+
return _capacity_from_function(n, lambda subset: len(subset) / n)
|
|
340
|
+
|
|
341
|
+
|
|
342
|
+
def relative_measure(n: int) -> FuzzyMeasure:
|
|
343
|
+
"""Return the relative capacity with singleton weights 1, ..., n."""
|
|
344
|
+
n = _measure_size(n)
|
|
345
|
+
denominator = n * (n + 1) / 2
|
|
346
|
+
return _capacity_from_function(
|
|
347
|
+
n,
|
|
348
|
+
lambda subset: sum(index + 1 for index in subset) / denominator,
|
|
349
|
+
)
|
|
350
|
+
|
|
351
|
+
|
|
352
|
+
def weighted_measure(weights: object) -> FuzzyMeasure:
|
|
353
|
+
"""Return the additive capacity induced by normalized positional weights."""
|
|
354
|
+
vector = _normalized_weights(weights)
|
|
355
|
+
n = vector.size
|
|
356
|
+
return _capacity_from_function(
|
|
357
|
+
n, lambda subset: sum(float(vector[index]) for index in subset)
|
|
358
|
+
)
|
|
359
|
+
|
|
360
|
+
|
|
361
|
+
def owa_measure(weights: object) -> FuzzyMeasure:
|
|
362
|
+
"""Return the symmetric capacity that yields OWA weights on sorted values.
|
|
363
|
+
|
|
364
|
+
``weights[0]`` is applied to the largest input, ``weights[1]`` to the
|
|
365
|
+
second largest, and so on. The weights must sum to one.
|
|
366
|
+
"""
|
|
367
|
+
vector = _normalized_weights(weights)
|
|
368
|
+
cumulative_weights = np.concatenate(([0.0], np.cumsum(vector)))
|
|
369
|
+
n = vector.size
|
|
370
|
+
return _capacity_from_function(
|
|
371
|
+
n, lambda subset: float(cumulative_weights[len(subset)])
|
|
372
|
+
)
|
|
373
|
+
|
|
374
|
+
|
|
375
|
+
def _sugeno_lambda(weights: np.ndarray) -> float:
|
|
376
|
+
total = math.fsum(float(weight) for weight in weights)
|
|
377
|
+
if total == 1.0:
|
|
378
|
+
return 0.0
|
|
379
|
+
|
|
380
|
+
positive_count = int(np.count_nonzero(weights))
|
|
381
|
+
if total < 1.0:
|
|
382
|
+
if positive_count < 2:
|
|
383
|
+
raise ValueError(
|
|
384
|
+
"singleton weights do not define a normalized k-fuzzy measure"
|
|
385
|
+
)
|
|
386
|
+
|
|
387
|
+
def log_gap(value):
|
|
388
|
+
return math.fsum(
|
|
389
|
+
math.log1p(value * float(weight)) for weight in weights
|
|
390
|
+
) - math.log1p(value)
|
|
391
|
+
|
|
392
|
+
lower, upper = 0.0, 1.0
|
|
393
|
+
while log_gap(upper) <= 0.0:
|
|
394
|
+
upper *= 2.0
|
|
395
|
+
if not isfinite(upper):
|
|
396
|
+
raise ValueError("could not find the k-fuzzy measure parameter")
|
|
397
|
+
for _ in range(100):
|
|
398
|
+
middle = (lower + upper) / 2.0
|
|
399
|
+
if log_gap(middle) < 0.0:
|
|
400
|
+
lower = middle
|
|
401
|
+
else:
|
|
402
|
+
upper = middle
|
|
403
|
+
return (lower + upper) / 2.0
|
|
404
|
+
|
|
405
|
+
if np.any(weights == 1.0):
|
|
406
|
+
return -1.0
|
|
407
|
+
|
|
408
|
+
def gap(value):
|
|
409
|
+
product = math.prod(1.0 + value * float(weight) for weight in weights)
|
|
410
|
+
return product - (1.0 + value)
|
|
411
|
+
|
|
412
|
+
lower, upper = -1.0, 0.0
|
|
413
|
+
for _ in range(100):
|
|
414
|
+
middle = (lower + upper) / 2.0
|
|
415
|
+
if gap(middle) > 0.0:
|
|
416
|
+
lower = middle
|
|
417
|
+
else:
|
|
418
|
+
upper = middle
|
|
419
|
+
return (lower + upper) / 2.0
|
|
420
|
+
|
|
421
|
+
|
|
422
|
+
def k_fuzzy_measure(singleton_weights: object) -> FuzzyMeasure:
|
|
423
|
+
"""Return the normalized Sugeno ``k``-fuzzy capacity.
|
|
424
|
+
|
|
425
|
+
The values are determined from the singleton weights by
|
|
426
|
+
``mu(A) = (prod(1 + k*w_i) - 1) / k``. The parameter ``k`` is selected
|
|
427
|
+
so that the measure of the full set is one; when ``k = 0``, this reduces
|
|
428
|
+
to the additive measure. Some singleton-weight vectors cannot define a
|
|
429
|
+
normalized ``k``-fuzzy measure and are rejected.
|
|
430
|
+
"""
|
|
431
|
+
vector = _as_unit_interval_vector(singleton_weights)
|
|
432
|
+
n = vector.size
|
|
433
|
+
parameter = _sugeno_lambda(vector)
|
|
434
|
+
|
|
435
|
+
def capacity(subset):
|
|
436
|
+
if not subset:
|
|
437
|
+
return 0.0
|
|
438
|
+
subset_weights = [float(vector[index]) for index in subset]
|
|
439
|
+
if parameter == 0.0:
|
|
440
|
+
return math.fsum(subset_weights)
|
|
441
|
+
if parameter == -1.0:
|
|
442
|
+
return 1.0 - math.prod(1.0 - weight for weight in subset_weights)
|
|
443
|
+
product = math.prod(1.0 + parameter * weight for weight in subset_weights)
|
|
444
|
+
return (product - 1.0) / parameter
|
|
445
|
+
|
|
446
|
+
values = {}
|
|
447
|
+
for mask in range(1 << n):
|
|
448
|
+
subset = frozenset(index for index in range(n) if mask & (1 << index))
|
|
449
|
+
values[subset] = capacity(subset)
|
|
450
|
+
values[frozenset()] = 0.0
|
|
451
|
+
values[frozenset(range(n))] = 1.0
|
|
452
|
+
return FuzzyMeasure(n, values)
|
|
453
|
+
|
|
454
|
+
|
|
455
|
+
def _overlap_value(quasi_overlap, value: float, measure_value: float) -> float:
|
|
456
|
+
result = quasi_overlap(value, measure_value)
|
|
457
|
+
if isinstance(result, (bool, np.bool_)):
|
|
458
|
+
raise TypeError("quasi_overlap must return a real number in [0, 1]")
|
|
459
|
+
try:
|
|
460
|
+
result = float(result)
|
|
461
|
+
except (TypeError, ValueError) as exc:
|
|
462
|
+
raise TypeError("quasi_overlap must return a real number in [0, 1]") from exc
|
|
463
|
+
if not isfinite(result) or not 0.0 <= result <= 1.0:
|
|
464
|
+
raise ValueError("quasi_overlap must return a finite value in [0, 1]")
|
|
465
|
+
return result
|
|
466
|
+
|
|
467
|
+
|
|
310
468
|
def _resolve_measure(measure: object, n: int):
|
|
311
469
|
if measure is None:
|
|
312
470
|
return lambda subset: len(subset) / n
|
|
@@ -348,8 +506,10 @@ def _resolve_measure(measure: object, n: int):
|
|
|
348
506
|
return lambda subset: float(cardinality_values[len(subset)])
|
|
349
507
|
|
|
350
508
|
|
|
351
|
-
def choquet_integral(
|
|
352
|
-
|
|
509
|
+
def choquet_integral(
|
|
510
|
+
values: object, measure: object = None, *, quasi_overlap=None
|
|
511
|
+
) -> float:
|
|
512
|
+
"""Return the discrete Choquet or quasi-overlap Choquet integral.
|
|
353
513
|
|
|
354
514
|
Values are sorted in ascending order. ``measure`` can be:
|
|
355
515
|
|
|
@@ -359,8 +519,13 @@ def choquet_integral(values: object, measure: object = None) -> float:
|
|
|
359
519
|
* a :class:`FuzzyMeasure` or a mapping that defines every subset.
|
|
360
520
|
|
|
361
521
|
Subset keys in mappings use zero-based vector positions. The result is a
|
|
362
|
-
Python ``float``.
|
|
522
|
+
Python ``float``. If ``quasi_overlap`` is supplied, it must be a callable
|
|
523
|
+
``O(x, mu)`` and the implementation uses the CQO formula
|
|
524
|
+
``sum(O(x_i, mu(A_i)) - O(x_(i-1), mu(A_i)))``. Omitting it uses the
|
|
525
|
+
classical product formula.
|
|
363
526
|
"""
|
|
527
|
+
if quasi_overlap is not None and not callable(quasi_overlap):
|
|
528
|
+
raise TypeError("quasi_overlap must be callable")
|
|
364
529
|
vector = _as_unit_interval_vector(values)
|
|
365
530
|
capacity = _resolve_measure(measure, vector.size)
|
|
366
531
|
order = np.argsort(vector, kind="stable")
|
|
@@ -370,7 +535,13 @@ def choquet_integral(values: object, measure: object = None) -> float:
|
|
|
370
535
|
previous_value = 0.0
|
|
371
536
|
for rank, value in enumerate(ordered_values):
|
|
372
537
|
upper_level_set = frozenset(int(index) for index in order[rank:])
|
|
373
|
-
|
|
538
|
+
measure_value = capacity(upper_level_set)
|
|
539
|
+
if quasi_overlap is None:
|
|
540
|
+
result += (float(value) - previous_value) * measure_value
|
|
541
|
+
else:
|
|
542
|
+
result += _overlap_value(
|
|
543
|
+
quasi_overlap, float(value), measure_value
|
|
544
|
+
) - _overlap_value(quasi_overlap, previous_value, measure_value)
|
|
374
545
|
previous_value = float(value)
|
|
375
546
|
return float(result)
|
|
376
547
|
|
|
@@ -85,16 +85,28 @@ class A_ch:
|
|
|
85
85
|
|
|
86
86
|
name = "ch"
|
|
87
87
|
|
|
88
|
-
def __init__(self, measure=None):
|
|
88
|
+
def __init__(self, measure=None, *, quasi_overlap=None):
|
|
89
89
|
self.__measure = measure
|
|
90
|
+
if quasi_overlap is not None and not callable(quasi_overlap):
|
|
91
|
+
raise TypeError("quasi_overlap must be callable")
|
|
92
|
+
self.__quasi_overlap = quasi_overlap
|
|
90
93
|
|
|
91
94
|
def __call__(self, array):
|
|
92
|
-
return choquet_integral(
|
|
95
|
+
return choquet_integral(
|
|
96
|
+
array,
|
|
97
|
+
measure=self.__measure,
|
|
98
|
+
quasi_overlap=self.__quasi_overlap,
|
|
99
|
+
)
|
|
93
100
|
|
|
94
101
|
def __repr__(self):
|
|
95
|
-
if self.__measure is None:
|
|
102
|
+
if self.__measure is None and self.__quasi_overlap is None:
|
|
96
103
|
return "A_ch()"
|
|
97
|
-
|
|
104
|
+
arguments = []
|
|
105
|
+
if self.__measure is not None:
|
|
106
|
+
arguments.append(f"measure={self.__measure!r}")
|
|
107
|
+
if self.__quasi_overlap is not None:
|
|
108
|
+
arguments.append(f"quasi_overlap={self.__quasi_overlap!r}")
|
|
109
|
+
return f"A_ch({', '.join(arguments)})"
|
|
98
110
|
|
|
99
111
|
def __str__(self):
|
|
100
112
|
return self.__repr__()
|