Skip to content

Commit

Permalink
new commit due to failed github connexion
Browse files Browse the repository at this point in the history
  • Loading branch information
@paulfogeladvestis
paulfogeladvestis committed Apr 9, 2024
1 parent 381ece8 commit de8712d
Show file tree
Hide file tree
Showing 42 changed files with 35,509 additions and 200 deletions.
47 changes: 46 additions & 1 deletion README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -6,12 +6,57 @@ pip install adilsm

# Description

ILSM is Integrated Longitudinal Multi Source Model.
ILSM is Integrated Latent Multi Source Model.

# Usage

```python
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import adilsm.adilsm as ilsm

max_noise_level = 0.1
# Generate a random non-negative matrix with 100 rows and 10 columns
A = np.random.rand(100, 10)
# Swap the columns of the A and add some noise to generate B
B = np.random.permutation(A.T).T + np.random.uniform(low=0, high=max_noise_level, size=A.shape)
# Add noise to A
A += np.random.uniform(low=0, high=max_noise_level, size=A.shape)

# ISM is expected to recognize that A and B convey the same information up to some noise,
# albeit with the columns of B swapped around. Heatmaps of the loadings of A and B columns
# on ISM components show the effective permutation.

Xs = [A, B]
n_embedding, n_themes = [10,10]

ilsm_result = ilsm.ism(Xs, n_embedding, n_themes, norm_columns=False, update_h4_ism=True,
max_iter_mult=200, fast_mult_rules=True, sparsity_coeff=.8)
hv = ilsm_result['HV']
hv_sparse = ilsm_result['HV_SPARSE']
hhii = ilsm_result['HHII']
w_ism = ilsm_result['W']
h_ism = ilsm_result['H']
q_ism = ilsm_result['Q']
Xs_emb = ilsm_result['EMBEDDING']
Xs_norm = ilsm_result['NORMED_VIEWS']

fig, ax = plt.subplots(1, 2, figsize=(10, 5), constrained_layout=True)
ax[0].imshow(hv[0], cmap='viridis', aspect='auto')
# Add labels and title
ax[0].set_xlabel('Component')
ax[0].set_ylabel('Column')
ax[0].set_title('Loadings of A columns on ISM components')
ax[1].imshow(hv[1], cmap='viridis', aspect='auto')
# Add labels and title
ax[1].set_xlabel('Component')
ax[1].set_ylabel('Column')
ax[1].set_title('Loadings of B columns on ISM components')

# Show the plot
plt.show()

```


124 changes: 90 additions & 34 deletions adilsm/adilsm.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,8 @@
import pandas as pd
import numpy as np

print("coucou")

def format_loadings(h4, list_columns):
# Format loadings
df_h4 = pd.DataFrame(data=h4)
Expand Down Expand Up @@ -55,7 +57,7 @@ def integrate_scores(m0_nan_0, m0_weight, h4_sparse, w4_ism, h4_ism, q4_ism, n_s
h4_score = h4_sparse[i1:i2, :].copy()
m0_score = m0_nan_0[:, i1:i2]
m0_weight_score = m0_weight[:, i1:i2]
i1=i2
i1 = i2

# Apply multiplicative updates to preserve h sparsity
for _ in range(0, max_iter_mult):
Expand Down Expand Up @@ -100,23 +102,19 @@ def integrate_scores(m0_nan_0, m0_weight, h4_sparse, w4_ism, h4_ism, q4_ism, n_s

return h4_updated, h4_updated_sparse, hhii_updated, w4_ism, h4_ism, q4_ism, tensor_score

def ism(m0:np.array, n_embedding:int, n_themes:int, n_scores:int, n_items:list[int], norm_m0:bool = True, max_iter:int=200, tol:float=1.e-6, verbose:int=-1, random_state:int=0,
def ism(Xs:list[np.array], n_embedding:int, n_themes:int, norm_columns:bool = True, max_iter:int=200, tol:float=1.e-6, verbose:int=-1, random_state:int=0,
max_iter_integrate:int=20, max_iter_mult:int=200, fast_mult_rules:bool=True, update_h4_ism:bool=False, sparsity_coeff:float=.8):
"""Estimate ISM model
Parameters
----------
m0: NDArray
Matrix of views, concatenated horizontally.
Xs: List of NDArray
List of matrices of views.
n_embedding: integer
Dimension of the embedding space.
n_themes: integer
Dimension of the latent space.
n_scores: integer
Number of views.
n_items: integer
List of numbers of attributes (features) per view
norm_m0: boolean
norm_columns: boolean
Scale each column of the concatenated matrix
max_iter: integer, default: 200
Maximum number of iterations.
Expand Down Expand Up @@ -144,7 +142,15 @@ def ism(m0:np.array, n_embedding:int, n_themes:int, n_scores:int, n_items:list[i
Returns
-------
ISM decomposition W, H*, Q and view mapping matrix H
Dictionary
ilsm_result['Hv']: View-mapping
ilsm_result['Hv_sparse']: Sparse view-mapping
ilsm_result['HHII']: Number of non-negligable values by Hv component
ilsm_result['W']: ISM meta-scores
ilsm_result['H']: NTF loadings in latent space
ilsm_result['Q']: NTF view loadings
ilsm_result['EMBEDDING']: Embedded views (concatenated)
ilsm_result['NORMED_VIEWS']: Normed views (concatenated)
Example
-------
Expand All @@ -157,21 +163,38 @@ def ism(m0:np.array, n_embedding:int, n_themes:int, n_scores:int, n_items:list[i
----------
Fogel, P., Boldina, G., Augé, F., Geissler, C., & Luta, G. (2024).
ISM: A New Space-Learning Model for Heterogenous Multi-view Data Reduction, Visualization and Clustering.
Preprints. https://doi.org/10.20944/preprints202402.1001.v1
Preprints. https://doi.org/10.20944/preprints202402.1001.v2
"""


Xs_concat = Xs[0].copy()
for X in Xs[1:]:
Xs_concat = np.hstack((Xs_concat, X))

m0 = Xs_concat

n_items = [Xs[i].shape[1] for i in range(len(Xs))]
n_scores = len(n_items)

m0_nan_0 = m0.copy()

# create m0_weight with ones and zeros if not_missing/missing value
m0_weight = np.where(np.isnan(m0), 0, 1)
m0_nan_0[np.isnan(m0_nan_0)]=0

if norm_m0 is True:
if norm_columns is True:
#Scale each column of m0
max_values = np.max(m0_nan_0, axis=0)
# # Replace maximum values equal to 0 with 1
m0 = np.divide(m0, np.where(max_values == 0, 1, max_values))
m0_nan_0 = np.divide(m0_nan_0, np.where(max_values == 0, 1, max_values))
Xs_norm = []
i1 = 0
for i_score in range(n_scores):
i2 = i1+n_items[i_score]
Xs_norm.append(m0[:,i1:i2])
i1 = i2
else:
Xs_norm = Xs

# Initial Embedding
my_nmfmodel = NMF(n_components=n_embedding, leverage=None, max_iter=max_iter, tol=tol, verbose=verbose, random_state=random_state)
Expand All @@ -188,10 +211,10 @@ def ism(m0:np.array, n_embedding:int, n_themes:int, n_scores:int, n_items:list[i
# Embed using scores w4 found in preliminary NMF and initialize themes through NTF
h4_updated, h4_updated_sparse, hhii_updated, w4_ism, h4_ism, q4_ism, tensor_score = \
integrate_scores(m0_nan_0, m0_weight, h4_sparse, w4, None, None, n_scores, n_items, n_themes, update_h4_ism=True,
max_iter_mult=max_iter_mult, fast_mult_rules=True, sparsity_coeff=sparsity_coeff)
max_iter_mult=max_iter_mult, fast_mult_rules=fast_mult_rules, sparsity_coeff=sparsity_coeff)

error = np.linalg.norm(m0 - w4_ism @ h4_updated_sparse.T) / np.linalg.norm(m0)
# print('error ism before straightening: ',round(error, 2))
print('error ism before straightening: ',round(error, 2))

# Iterate embedding with themes subtensor until sparsity becomes stable
flag = 0
Expand All @@ -202,11 +225,11 @@ def ism(m0:np.array, n_embedding:int, n_themes:int, n_scores:int, n_items:list[i
if iter_integrate == 0:
h4_updated, h4_updated_sparse, hhii_updated, w4_ism, h4_ism, q4_ism, tensor_score = \
integrate_scores(m0_nan_0, m0_weight, h4_updated_sparse, w4_ism, np.identity(n_themes), q4_ism, n_scores, n_items, n_themes, update_h4_ism=update_h4_ism,
max_iter_mult=max_iter_mult, fast_mult_rules=True, sparsity_coeff=sparsity_coeff)
max_iter_mult=max_iter_mult, fast_mult_rules=fast_mult_rules, sparsity_coeff=sparsity_coeff)
else:
h4_updated, h4_updated_sparse, hhii_updated, w4_ism, h4_ism, q4_ism, tensor_score = \
integrate_scores(m0_nan_0, m0_weight, h4_updated_sparse, w4_ism, h4_ism, q4_ism, n_scores, n_items, n_themes, update_h4_ism=update_h4_ism,
max_iter_mult=max_iter_mult, fast_mult_rules=True, sparsity_coeff=sparsity_coeff)
max_iter_mult=max_iter_mult, fast_mult_rules=fast_mult_rules, sparsity_coeff=sparsity_coeff)

if (hhii_updated == hhii_updated_0).all():
flag+=1
Expand All @@ -216,20 +239,44 @@ def ism(m0:np.array, n_embedding:int, n_themes:int, n_scores:int, n_items:list[i
if flag==3:
break

error = np.linalg.norm(m0 - w4_ism @ h4_updated_sparse.T) / np.linalg.norm(m0)
# print('error ism after straightening: ',round(error, 2))

return h4_updated, h4_updated_sparse, w4_ism, h4_ism, q4_ism, tensor_score, m0
Xs_emb = []
i1 = 0
for i_score in range(n_scores):
i2 = i1+n_embedding
Xs_emb.append(tensor_score[:,i1:i2])
i1 = i2

hv = []
hv_sparse = []
i1 = 0
for i_score in range(n_scores):
i2 = i1+n_items[i_score]
hv.append(h4_updated[i1:i2,:])
hv_sparse.append(h4_updated_sparse[i1:i2,:])
i1 = i2

def ism_expand(m0, h4_sparse, h4_ism, q4_ism, n_themes, n_scores, n_items, max_iter=200, tol=1.e-6, verbose=-1, random_state=0,
max_iter_mult=200):
error = np.linalg.norm(m0 - w4_ism @ h4_updated_sparse.T) / np.linalg.norm(m0)
print('error ism after straightening: ',round(error, 2))
ilsm_result = {}
ilsm_result['HV'] = hv
ilsm_result['HV_SPARSE'] = hv_sparse
ilsm_result['HHII'] = hhii_updated
ilsm_result['W'] = w4_ism
ilsm_result['H'] = h4_ism
ilsm_result['Q'] = q4_ism
ilsm_result['EMBEDDING'] = Xs_emb
ilsm_result['NORMED_VIEWS'] = Xs_norm

return ilsm_result

def ism_expand(Xs:list[np.array], h4_sparse:float, h4_ism:float, q4_ism:float, n_themes:int, norm_features:bool = True, max_iter:int=200, tol:float=1.e-6, verbose:int=-1, random_state:int=0,
max_iter_mult:int=200):
"""Expand meta-scores to new observations
Parameters
----------
m0: float
Matrix of views, concatenated horizontally.
Xs: List of NDArray
List of matrices of views.
h4_sparse: float
View-mapping matrix H.
h4_ism: float
Expand All @@ -238,12 +285,10 @@ def ism_expand(m0, h4_sparse, h4_ism, q4_ism, n_themes, n_scores, n_items, max_i
View loading Q.
n_themes:
Dimension of the latent space.
n_scores: integer
Number of views.
n_items: integer
List of numbers of attributes (features) per view
leverage: None | 'standard' | 'robust', default 'standard'
Calculate leverage of W and H rows on each component.
norm_features: boolean
Scale each column of the concatenated matrix
max_iter: integer, default: 200
Maximum number of iterations.
tol: float, default: 1e-6
Expand Down Expand Up @@ -278,17 +323,28 @@ def ism_expand(m0, h4_sparse, h4_ism, q4_ism, n_themes, n_scores, n_items, max_i
Preprints. https://doi.org/10.20944/preprints202402.1001.v1
"""
EPSILON = np.finfo(np.float32).eps
#Scale each column of m0

Xs_concat = Xs[0].copy()
for X in Xs[1:]:
Xs_concat = np.hstack((Xs_concat, X))

m0 = Xs_concat

n_items = [Xs[i].shape[1] for i in range(len(Xs))]
n_scores = len(n_items)

m0_nan_0 = m0.copy()

# create m0_weight with ones and zeros if not_missing/missing value
m0_weight = np.where(np.isnan(m0), 0, 1)
m0_nan_0[np.isnan(m0_nan_0)]=0

max_values = np.max(m0_nan_0, axis=0)
# Replace maximum values equal to 0 with 1
m0 = np.divide(m0, np.where(max_values == 0, 1, max_values))
m0_nan_0 = np.divide(m0_nan_0, np.where(max_values == 0, 1, max_values))
if norm_features is True:
#Scale each column of m0
max_values = np.max(m0_nan_0, axis=0)
# Replace maximum values equal to 0 with 1
m0 = np.divide(m0, np.where(max_values == 0, 1, max_values))
m0_nan_0 = np.divide(m0_nan_0, np.where(max_values == 0, 1, max_values))

i1 = 0

Expand Down
Loading

0 comments on commit de8712d

Please sign in to comment.