Active Learning

This module implements many active learning algorithms in an objected-oriented fashion, similar to sklearn and modAL. The usage is similar for all algorithms, and we give some high-level examples of how to use this module with each of the provided acquisition functions.

The common workflow, however, is as follows:

import graphlearning as gl

# define ssl model and acquisition function
model = gl.ssl.laplace(W)                   # graph-based ssl classifier with a given graph
acq_func = gl.active_learning.unc_sampling  # acquisition function for prioritizing which points to query


# instantiate active learner object
AL = gl.active_learner(
     model=model,                                    
     acq_function=acq_func,   
     labeled_ind=..,                        # indices of initially labeled nodes 
     labels=,                               # (integer) labels for initially labeled nodes
     policy='max',                         # active learning policy (i.e., 'max', or 'prop')
     **kwargs=...                           # other keyword arguments for the specified acq_function
    )

# select next query points
query_point = AL.select_queries(
                batch_size=1               # number of query points to select at this iteration
               ) 

# acquire label for query points
query_labels = y[query_point] 


# update the labeled data of active_learner object (including the graph-based ssl ``model`` outputs)
AL.update(query_points, query_labels) 

Some clarification of terms: - acquisition function: a function that quantifies "how useful" it would be to label a currently unlabeled node. Oftentimes, this is reflected in the "uncertainty" of the current classifier's output for each node. - NOTE: users can provide their own acquisition functions that inherit from the acquisition_function class, being sure to implement it so that larger values of the acquisition function correspond to more desirable nodes to be labeled. - policy: the active learning policy determines which node(s) will be selected as query points, given the set of acquisition function values evaluated on the unlabeled nodes. - The default value max indicates that query points will be the maximizers of the acquisition function on the unlabeled nodes. The policy prop selects the query points proportional to the ''softmax'' of the acquisition function values; namely, $$\mathbb{P}(X = x) \propto e^{\gamma \mathcal{A}(x)}$$

"""
Active Learning
========================

This module implements many active learning algorithms in an objected-oriented
fashion, similar to [sklearn](https://scikit-learn.org/stable/) and [modAL](https://modal-python.readthedocs.io/en/latest/index.html). The usage is similar for all algorithms, and we give some high-level examples of how to use this module with each of the provided acquisition functions.

The common workflow, however, is as follows:
```py
import graphlearning as gl

# define ssl model and acquisition function
model = gl.ssl.laplace(W)                   # graph-based ssl classifier with a given graph
acq_func = gl.active_learning.unc_sampling  # acquisition function for prioritizing which points to query


# instantiate active learner object
AL = gl.active_learner(
     model=model,                                    
     acq_function=acq_func,   
     labeled_ind=..,                        # indices of initially labeled nodes 
     labels=,                               # (integer) labels for initially labeled nodes
     policy='max',                         # active learning policy (i.e., 'max', or 'prop')
     **kwargs=...                           # other keyword arguments for the specified acq_function
    )

# select next query points
query_point = AL.select_queries(
                batch_size=1               # number of query points to select at this iteration
               ) 

# acquire label for query points
query_labels = y[query_point] 


# update the labeled data of active_learner object (including the graph-based ssl ``model`` outputs)
AL.update(query_points, query_labels) 
```

Some clarification of terms:
- ``acquisition function``: a function that quantifies "how useful" it would be to label a currently unlabeled node. Oftentimes, this is reflected in the "uncertainty" of the current classifier's output for each node. 
- __NOTE:__ users can provide their own acquisition functions that inherit from the ``acquisition_function`` class, being sure to implement it so that __larger values__ of the acquisition function correspond to __more desirable__ nodes to be labeled.
- ``policy``: the active learning policy determines which node(s) will be selected as query points, given the set of acquisition function values evaluated on the unlabeled nodes. 
- The default value ``max`` indicates that query points will be the maximizers of the acquisition function on the unlabeled nodes. The policy ``prop`` selects the query points proportional to the ''softmax'' of the acquisition function values; namely, 
\\[\\mathbb{P}(X = x) \\propto e^{\\gamma \\mathcal{A}(x)}\\]
"""

import numpy as np
from scipy.special import softmax
from abc import ABCMeta, abstractmethod
import matplotlib.pyplot as plt

from . import graph


class active_learner:
    def __init__(self, model, acq_function, labeled_ind, labels, policy='max', **kwargs):
        self.model = model
        self.labeled_ind = labeled_ind.copy()
        self.labels = labels.copy()
        self.acq_function = acq_function(**kwargs)
        self.acq_function.update(labeled_ind, labels)
        self.policy = policy
        self.u = self.model.fit(self.labeled_ind, self.labels) # initialize the ssl model on the initially labeled data
        self.n = self.model.graph.num_nodes
        self.all_inds = np.arange(self.n)
        self.unlabeled_ind = np.setdiff1d(self.all_inds, self.labeled_ind)
        self.printed_warning = False
        

    def select_queries(self, batch_size=1, policy=None, candidate_ind='full', rand_frac=0.1, return_acq_vals=False, prop_gamma=1.0, 
                       allow_repeat=False):
        if policy is None:
            policy = self.policy
        
        if isinstance(candidate_ind, np.ndarray):
            if (candidate_ind.min() < 0) or (candidate_ind.max() > self.n):
                raise ValueError(f"candidate_ind must have integer values between 0 and {self.n}")
        elif candidate_ind == 'full':
            if allow_repeat:
                candidate_ind = np.arange(self.all_inds)
            else:
                candidate_ind = np.setdiff1d(self.all_inds, self.labeled_ind)
        elif (candidate_ind == 'rand') and (rand_frac>0 and rand_frac<1):
            if allow_repeat:
                candidate_ind = np.random.choice(self.all_inds, size=int(rand_frac * self.n), replace=False)
            else:
                candidate_ind = np.random.choice(self.unlabeled_ind, size=int(rand_frac * len(self.unlabeled_ind)), replace=False)
        else:
            raise ValueError("Invalid input for candidate_ind")
        
        acq_vals = self.acq_function.compute(self.u, candidate_ind)
        
        if policy == 'max':
            query_ind = candidate_ind[(-acq_vals).argsort()[:batch_size]]
        elif policy == 'prop':
            probs = np.exp(prop_gamma*(acq_vals - acq_vals.max()))
            probs /= probs.sum()
            query_ind = np.random.choice(candidate_ind, batch_size, p=probs)
        else:
            query_ind = policy(candidate_ind, acq_vals, batch_size) # user-defined policy

        if return_acq_vals:
            return query_ind, acq_vals
        return query_ind
    

    def update(self, query_ind, query_labels):
        if np.intersect1d(query_ind, self.labeled_ind).size > 0 and not self.printed_warning:
            print("WARNING: Having multiple observations at a single node detected")
            self.printed_warning = True
        self.labeled_ind = np.append(self.labeled_ind, query_ind)
        self.labels = np.append(self.labels, query_labels)
        self.u = self.model.fit(self.labeled_ind, self.labels)
        self.unlabeled_ind = np.setdiff1d(self.all_inds, self.labeled_ind)
        self.acq_function.update(query_ind, query_labels)
        return

class acquisition_function:  
    """Acquisition Function
    ======
    
    Object that computes a measure of ''utility'' for labeling nodes that are currently unlabeled. Users can define their own acqusition functions to inherit from this class as follows:
    
    ```py
    import graphlearning as gl
    
    class new_acq_func(gl.acquisition_function):
        def __init__(self, arg1=None):
            self.arg1 = arg1           # any arguments that are passed into this acquisition function are given 
                                       # as kwargs in active_learner instantiation

        def compute(self, u, candidate_ind):
            vals = ...                 # compute the acquisition function so that larger values are more desired
            return vals  
    
    ```
    
    
    """
    @abstractmethod
    def compute(self, u, candidate_ind):
        """Internal Compute Acquisition Function Values Function
        ======

        Internal function that any acquisition function object must override. 

        Parameters
        ----------
        u : numpy array
            score matrix from GSSL classifier. 
        candidate_ind : numpy array (or list)
            (sub)set of indices on which to compute the acquisition function

        Returns
        -------
        acquisition_values : numpy array, float
            acquisition function values
        """
        raise NotImplementedError("Must override compute")

    @abstractmethod
    def update(self, query_ind, query_labels):
        return 



class unc_sampling(acquisition_function):
    """Uncertainty Sampling
    ===================

    Active learning algorithm that selects points that the classifier is most uncertain of.

    Examples
    --------
    ```py
    import graphlearning.active_learning as al
    import graphlearning as gl
    import numpy as np
    import matplotlib.pyplot as plt
    import sklearn.datasets as datasets

    X,labels = datasets.make_moons(n_samples=500,noise=0.1)
    W = gl.weightmatrix.knn(X,10)
    train_ind = gl.trainsets.generate(labels, rate=5)
    plt.scatter(X[:,0],X[:,1], c=labels)
    plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
    plt.show()

    
    model = gl.ssl.laplace(W)
    AL = gl.active_learning.active_learner(model, gl.active_learning.unc_sampling, train_ind, y[train_ind])

    for i in range(10):
        query_points = AL.select_queries() # return this iteration's newly chosen points
        query_labels = y[query_points] # simulate the human in the loop process
        AL.update(query_points, query_labels) # update the active_learning object's labeled set

        # plot
        plt.scatter(X[:,0],X[:,1], c=y)
        plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
        plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
        plt.show()
        print(AL.labeled_ind)
        print(AL.labels)
    ```

    Reference
    ---------
    [1] Settles, B., [Active Learning], vol. 6, Morgan & Claypool Publishers LLC (June 2012).
    """
    def __init__(self, unc_method='smallest_margin'):
        self.unc_method = unc_method

    def compute(self, u, candidate_ind):
        if self.unc_method == "norm":
            u_probs = softmax(u[candidate_ind], axis=1)
            one_hot_predicted_labels = np.eye(u.shape[1])[np.argmax(u[candidate_ind], axis=1)]
            unc_terms = np.linalg.norm((u_probs - one_hot_predicted_labels), axis=1)
        elif self.unc_method == "entropy":
            u_probs = softmax(u[candidate_ind], axis=1)
            unc_terms = np.max(u_probs, axis=1) - np.sum(u_probs*np.log(u_probs +.00001), axis=1)
        elif self.unc_method == "least_confidence":
            unc_terms = np.ones((u[candidate_ind].shape[0],)) - np.max(u[candidate_ind], axis=1)
        elif self.unc_method == "smallest_margin":
            u_sort = np.sort(u[candidate_ind])
            unc_terms = 1.-(u_sort[:,-1] - u_sort[:,-2])
        elif self.unc_method == "largest_margin":
            u_sort = np.sort(u[candidate_ind])
            unc_terms = 1.-(u_sort[:,-1] - u_sort[:,0])
        elif self.unc_method == "unc_2norm":
            unc_terms = 1. - np.linalg.norm(u[candidate_ind], axis=1)
        return unc_terms



class var_opt(acquisition_function):
    """Variance Optimization
    ===================

    Active learning algorithm that selects points that minimizes the variance of the distribution of unlabeled nodes.

    Examples
    --------
    ```py
    import graphlearning.active_learning as al
    import graphlearning as gl
    import numpy as np
    import matplotlib.pyplot as plt
    import sklearn.datasets as datasets

    X,labels = datasets.make_moons(n_samples=500,noise=0.1)
    W = gl.weightmatrix.knn(X,10)
    train_ind = gl.trainsets.generate(labels, rate=5)
    plt.scatter(X[:,0],X[:,1], c=labels)
    plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
    plt.show()

    # compute initial, low-rank (spectral truncation) covariance matrix 
    evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
    C = np.diag(1. / (evals + 1e-11))
    AL = gl.active_learning.active_learner(model, gl.active_learning.var_opt, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

    for i in range(10):
        query_points = AL.select_queries() # return this iteration's newly chosen points
        query_labels = y[query_points] # simulate the human in the loop process
        AL.update(query_points, query_labels) # update the active_learning object's labeled set

        # plot
        plt.scatter(X[:,0],X[:,1], c=y)
        plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
        plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
        plt.show()
        print(AL.labeled_ind)
        print(AL.labels)
    ```

    Reference
    ---------
    [1] Ji, M. and Han, J., “A variance minimization criterion to active learning on graphs,” in [Artificial Intelligence
    and Statistics ], 556–564 (Mar. 2012).
    """

    ## this only handles the FULL C computation or the spectral truncation
    def __init__(self, C, V=None, gamma2=0.1**2.):
        assert (C.shape[0] == C.shape[1]) or (V is not None)
        self.C = C.copy()
        self.V = V
        self.gamma2 = gamma2
        if self.V is None:
            self.storage = 'full'
        else:
            self.storage = 'trunc'
        
        
    def compute(self, u, candidate_ind):
        if self.storage == 'full':
            col_norms = np.linalg.norm(self.C[:,candidate_ind], axis=0)**2.
            diag_terms = self.gamma2 + self.C.diagonal()[candidate_ind]
        else:
            Cavk = self.C @ self.V[candidate_ind,:].T
            col_norms = np.linalg.norm(Cavk, axis=0)**2.
            diag_terms = (self.gamma2 + np.array([np.inner(self.V[k,:], Cavk[:, i]) for i,k in enumerate(candidate_ind)]))
            
        return col_norms / diag_terms

    def update(self, query_ind, query_labels):
        for k in query_ind:
            if self.storage == 'full':
                self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
            else:
                vk = self.V[k]
                Cavk = self.C @ vk
                ip = np.inner(vk, Cavk)
                self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 

        return 




class sigma_opt(acquisition_function):
    """Sigma Optimization
    ===================

    Active learning algorithm that selects points that minimizes the sum of the associated entries in the covariance matrix.

    Examples
    --------
    ```py
    import graphlearning.active_learning as al
    import graphlearning as gl
    import numpy as np
    import matplotlib.pyplot as plt
    import sklearn.datasets as datasets

    X,labels = datasets.make_moons(n_samples=500,noise=0.1)
    W = gl.weightmatrix.knn(X,10)
    train_ind = gl.trainsets.generate(labels, rate=5)
    plt.scatter(X[:,0],X[:,1], c=labels)
    plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
    plt.show()

    # compute initial, low-rank (spectral truncation) covariance matrix 
    evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
    C = np.diag(1. / (evals + 1e-11))
    AL = gl.active_learning.active_learner(model, gl.active_learning.sigma_opt, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

    for i in range(10):
        query_points = AL.select_queries() # return this iteration's newly chosen points
        query_labels = y[query_points] # simulate the human in the loop process
        AL.update(query_points, query_labels) # update the active_learning object's labeled set

        # plot
        plt.scatter(X[:,0],X[:,1], c=y)
        plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
        plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
        plt.show()
        print(AL.labeled_ind)
        print(AL.labels)
    ```

    Reference
    ---------
    [1] Ma, Y., Garnett, R., and Schneider, J., “Σ-optimality for active learning on Gaussian random fields,”
    in [Advances in Neural Information Processing Systems 26 ], Burges, C. J. C., Bottou, L., Welling, M.,
    Ghahramani, Z., and Weinberger, K. Q., eds., 2751–2759, Curran Associates, Inc. (2013).
    """
    ## this only handles the FULL C computation or the spectral truncation
    def __init__(self, C, V=None, gamma2=0.1**2.):
        assert (C.shape[0] == C.shape[1]) or (V is not None)
        self.C = C.copy()
        self.V = V
        self.gamma2 = gamma2
        if self.V is None:
            self.storage = 'full'
        else:
            self.storage = 'trunc'
        
        
    def compute(self, u, candidate_ind):
        if self.storage == 'full':
            col_sums = np.sum(self.C[:, candidate_ind], axis=0)**2.
            diag_terms = self.gamma2 + self.C.diagonal()[candidate_ind]
        else:
            Cavk = self.C @ self.V[candidate_ind,:].T
            col_sums = np.sum(Cavk, axis=0)**2.
            diag_terms = (self.gamma2 + np.array([np.inner(self.V[k,:], Cavk[:, i]) for i,k in enumerate(candidate_ind)]))
            
        return col_sums/ diag_terms


    def update(self, query_ind, query_labels):
        for k in query_ind:
            if self.storage == 'full':
                self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
            else:
                vk = self.V[k]
                Cavk = self.C @ vk
                ip = np.inner(vk, Cavk)
                self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 

        return



class model_change(acquisition_function):
    """Model Change
    ===================

    Active learning algorithm that selects points that will produce the greatest change in the model.

    Examples
    --------
    ```py
    import graphlearning.active_learning as al
    import graphlearning as gl
    import numpy as np
    import matplotlib.pyplot as plt
    import sklearn.datasets as datasets

    X,labels = datasets.make_moons(n_samples=500,noise=0.1)
    W = gl.weightmatrix.knn(X,10)
    train_ind = gl.trainsets.generate(labels, rate=5)
    plt.scatter(X[:,0],X[:,1], c=labels)
    plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
    plt.show()

    # compute initial, low-rank (spectral truncation) covariance matrix 
    evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
    C = np.diag(1. / (evals + 1e-11))
    AL = gl.active_learning.active_learner(model, gl.active_learning.model_change, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

    for i in range(10):
        query_points = AL.select_queries() # return this iteration's newly chosen points
        query_labels = y[query_points] # simulate the human in the loop process
        AL.update(query_points, query_labels) # update the active_learning object's labeled set

        # plot
        plt.scatter(X[:,0],X[:,1], c=y)
        plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
        plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
        plt.show()
        print(AL.labeled_ind)
        print(AL.labels)
    ```

    Reference
    ---------
    [1] Miller, K. and Bertozzi, A. L., “Model-change active learning in graph-based semi-supervised learning,”
    (Oct. 2021). arXiv: 2110.07739.

    [2] Karzand, M. and Nowak, R. D., “Maximin active learning in overparameterized model classes,” IEEE
    Journal on Selected Areas in Information Theory 1, 167–177 (May 2020).
    """
    def __init__(self, C, V=None, gamma2=0.1**2., unc_method='smallest_margin'):
        assert (C.shape[0] == C.shape[1]) or (V is not None)
        self.C = C.copy()
        self.V = V
        self.gamma2 = gamma2
        self.unc_sampling = unc_sampling(unc_method=unc_method)
        if self.V is None:
            self.storage = 'full'
        else:
            self.storage = 'trunc'
        
    def compute(self, u, candidate_ind):
        unc_terms = self.unc_sampling.compute(u, candidate_ind)
        if self.storage == 'full':
            col_norms = np.linalg.norm(self.C, axis=0)
            diag_terms = self.gamma2 + self.C.diagonal()
        else:
            Cavk = self.C @ self.V[candidate_ind,:].T
            col_norms = np.linalg.norm(Cavk, axis=0)
            diag_terms = (self.gamma2 + np.array([np.inner(self.V[k,:], Cavk[:, i]) for i,k in enumerate(candidate_ind)]))
        return unc_terms * col_norms / diag_terms  


    def update(self, query_ind, query_labels):
        for k in query_ind:
            if self.storage == 'full':
                self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
            else:
                vk = self.V[k]
                Cavk = self.C @ vk
                ip = np.inner(vk, Cavk)
                self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 
        return
        

class model_change_var_opt(acquisition_function):
    """Model Change Variance Optimization
    ===================

    Active learning algorithm that is a combination of Model Change and Variance Optimization.

    Examples
    --------
    ```py
    import graphlearning.active_learning as al
    import graphlearning as gl
    import numpy as np
    import matplotlib.pyplot as plt
    import sklearn.datasets as datasets

    X,labels = datasets.make_moons(n_samples=500,noise=0.1)
    W = gl.weightmatrix.knn(X,10)
    train_ind = gl.trainsets.generate(labels, rate=5)
    plt.scatter(X[:,0],X[:,1], c=labels)
    plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
    plt.show()

    # compute initial, low-rank (spectral truncation) covariance matrix 
    evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
    C = np.diag(1. / (evals + 1e-11))
    AL = gl.active_learning.active_learner(model, gl.active_learning.model_change_var_opt, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

    for i in range(10):
        query_points = AL.select_queries() # return this iteration's newly chosen points
        query_labels = y[query_points] # simulate the human in the loop process
        AL.update(query_points, query_labels) # update the active_learning object's labeled set

        # plot
        plt.scatter(X[:,0],X[:,1], c=y)
        plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
        plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
        plt.show()
        print(AL.labeled_ind)
        print(AL.labels)
    ```

    Reference
    ---------
    [1] Ji, M. and Han, J., “A variance minimization criterion to active learning on graphs,” in [Artificial Intelligence
    and Statistics ], 556–564 (Mar. 2012).

    [2] Miller, K. and Bertozzi, A. L., “Model-change active learning in graph-based semi-supervised learning,”
    (Oct. 2021). arXiv: 2110.07739.

    [3] Karzand, M. and Nowak, R. D., “Maximin active learning in overparameterized model classes,” IEEE
    Journal on Selected Areas in Information Theory 1, 167–177 (May 2020).
    """
    def __init__(self, C, V=None, gamma2=0.1**2., unc_method='smallest_margin'):
        assert (C.shape[0] == C.shape[1]) or (V is not None)
        self.C = C.copy()
        self.V = V
        self.gamma2 = gamma2
        self.unc_sampling = unc_sampling(unc_method=unc_method)
        if self.V is None:
            self.storage = 'full'
        else:
            self.storage = 'trunc'
        
    def compute(self, u, candidate_ind):
        unc_terms = self.unc_sampling.compute(u, candidate_ind)
        if self.storage == 'full':
            col_norms = np.linalg.norm(self.C, axis=0)**2.
            diag_terms = self.gamma2 + self.C.diagonal()
        else:
            Cavk = self.C @ self.V[candidate_ind,:].T
            col_norms = np.linalg.norm(Cavk, axis=0)**2.
            diag_terms = (self.gamma2 + np.array([np.inner(self.V[k,:], Cavk[:, i]) for i,k in enumerate(candidate_ind)]))
        return unc_terms * col_norms / diag_terms  


    def update(self, query_ind, query_labels):
        for k in query_ind:
            if self.storage == 'full':
                self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
            else:
                vk = self.V[k]
                Cavk = self.C @ vk
                ip = np.inner(vk, Cavk)
                self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 
        return

Classes

class acquisition_function

Acquisition Function

Object that computes a measure of ''utility'' for labeling nodes that are currently unlabeled. Users can define their own acqusition functions to inherit from this class as follows:

import graphlearning as gl

class new_acq_func(gl.acquisition_function):
    def __init__(self, arg1=None):
        self.arg1 = arg1           # any arguments that are passed into this acquisition function are given 
                                   # as kwargs in active_learner instantiation

    def compute(self, u, candidate_ind):
        vals = ...                 # compute the acquisition function so that larger values are more desired
        return vals  

Expand source code

class acquisition_function:  
    """Acquisition Function
    ======
    
    Object that computes a measure of ''utility'' for labeling nodes that are currently unlabeled. Users can define their own acqusition functions to inherit from this class as follows:
    
    ```py
    import graphlearning as gl
    
    class new_acq_func(gl.acquisition_function):
        def __init__(self, arg1=None):
            self.arg1 = arg1           # any arguments that are passed into this acquisition function are given 
                                       # as kwargs in active_learner instantiation

        def compute(self, u, candidate_ind):
            vals = ...                 # compute the acquisition function so that larger values are more desired
            return vals  
    
    ```
    
    
    """
    @abstractmethod
    def compute(self, u, candidate_ind):
        """Internal Compute Acquisition Function Values Function
        ======

        Internal function that any acquisition function object must override. 

        Parameters
        ----------
        u : numpy array
            score matrix from GSSL classifier. 
        candidate_ind : numpy array (or list)
            (sub)set of indices on which to compute the acquisition function

        Returns
        -------
        acquisition_values : numpy array, float
            acquisition function values
        """
        raise NotImplementedError("Must override compute")

    @abstractmethod
    def update(self, query_ind, query_labels):
        return 

Subclasses

• model_change
• model_change_var_opt
• sigma_opt
• unc_sampling
• var_opt

Methods

def compute(self, u, candidate_ind)

Internal Compute Acquisition Function Values Function

Internal function that any acquisition function object must override.

Parameters

u : numpy array

score matrix from GSSL classifier.

candidate_ind : numpy array (or list)

(sub)set of indices on which to compute the acquisition function

Returns

acquisition_values : numpy array, float

acquisition function values

Expand source code

@abstractmethod
def compute(self, u, candidate_ind):
    """Internal Compute Acquisition Function Values Function
    ======

    Internal function that any acquisition function object must override. 

    Parameters
    ----------
    u : numpy array
        score matrix from GSSL classifier. 
    candidate_ind : numpy array (or list)
        (sub)set of indices on which to compute the acquisition function

    Returns
    -------
    acquisition_values : numpy array, float
        acquisition function values
    """
    raise NotImplementedError("Must override compute")

def update(self, query_ind, query_labels)

Expand source code

@abstractmethod
def update(self, query_ind, query_labels):
    return 

class active_learner (model, acq_function, labeled_ind, labels, policy='max', **kwargs)

Expand source code

class active_learner:
    def __init__(self, model, acq_function, labeled_ind, labels, policy='max', **kwargs):
        self.model = model
        self.labeled_ind = labeled_ind.copy()
        self.labels = labels.copy()
        self.acq_function = acq_function(**kwargs)
        self.acq_function.update(labeled_ind, labels)
        self.policy = policy
        self.u = self.model.fit(self.labeled_ind, self.labels) # initialize the ssl model on the initially labeled data
        self.n = self.model.graph.num_nodes
        self.all_inds = np.arange(self.n)
        self.unlabeled_ind = np.setdiff1d(self.all_inds, self.labeled_ind)
        self.printed_warning = False
        

    def select_queries(self, batch_size=1, policy=None, candidate_ind='full', rand_frac=0.1, return_acq_vals=False, prop_gamma=1.0, 
                       allow_repeat=False):
        if policy is None:
            policy = self.policy
        
        if isinstance(candidate_ind, np.ndarray):
            if (candidate_ind.min() < 0) or (candidate_ind.max() > self.n):
                raise ValueError(f"candidate_ind must have integer values between 0 and {self.n}")
        elif candidate_ind == 'full':
            if allow_repeat:
                candidate_ind = np.arange(self.all_inds)
            else:
                candidate_ind = np.setdiff1d(self.all_inds, self.labeled_ind)
        elif (candidate_ind == 'rand') and (rand_frac>0 and rand_frac<1):
            if allow_repeat:
                candidate_ind = np.random.choice(self.all_inds, size=int(rand_frac * self.n), replace=False)
            else:
                candidate_ind = np.random.choice(self.unlabeled_ind, size=int(rand_frac * len(self.unlabeled_ind)), replace=False)
        else:
            raise ValueError("Invalid input for candidate_ind")
        
        acq_vals = self.acq_function.compute(self.u, candidate_ind)
        
        if policy == 'max':
            query_ind = candidate_ind[(-acq_vals).argsort()[:batch_size]]
        elif policy == 'prop':
            probs = np.exp(prop_gamma*(acq_vals - acq_vals.max()))
            probs /= probs.sum()
            query_ind = np.random.choice(candidate_ind, batch_size, p=probs)
        else:
            query_ind = policy(candidate_ind, acq_vals, batch_size) # user-defined policy

        if return_acq_vals:
            return query_ind, acq_vals
        return query_ind
    

    def update(self, query_ind, query_labels):
        if np.intersect1d(query_ind, self.labeled_ind).size > 0 and not self.printed_warning:
            print("WARNING: Having multiple observations at a single node detected")
            self.printed_warning = True
        self.labeled_ind = np.append(self.labeled_ind, query_ind)
        self.labels = np.append(self.labels, query_labels)
        self.u = self.model.fit(self.labeled_ind, self.labels)
        self.unlabeled_ind = np.setdiff1d(self.all_inds, self.labeled_ind)
        self.acq_function.update(query_ind, query_labels)
        return

Methods

def select_queries(self, batch_size=1, policy=None, candidate_ind='full', rand_frac=0.1, return_acq_vals=False, prop_gamma=1.0, allow_repeat=False)

Expand source code

def select_queries(self, batch_size=1, policy=None, candidate_ind='full', rand_frac=0.1, return_acq_vals=False, prop_gamma=1.0, 
                   allow_repeat=False):
    if policy is None:
        policy = self.policy
    
    if isinstance(candidate_ind, np.ndarray):
        if (candidate_ind.min() < 0) or (candidate_ind.max() > self.n):
            raise ValueError(f"candidate_ind must have integer values between 0 and {self.n}")
    elif candidate_ind == 'full':
        if allow_repeat:
            candidate_ind = np.arange(self.all_inds)
        else:
            candidate_ind = np.setdiff1d(self.all_inds, self.labeled_ind)
    elif (candidate_ind == 'rand') and (rand_frac>0 and rand_frac<1):
        if allow_repeat:
            candidate_ind = np.random.choice(self.all_inds, size=int(rand_frac * self.n), replace=False)
        else:
            candidate_ind = np.random.choice(self.unlabeled_ind, size=int(rand_frac * len(self.unlabeled_ind)), replace=False)
    else:
        raise ValueError("Invalid input for candidate_ind")
    
    acq_vals = self.acq_function.compute(self.u, candidate_ind)
    
    if policy == 'max':
        query_ind = candidate_ind[(-acq_vals).argsort()[:batch_size]]
    elif policy == 'prop':
        probs = np.exp(prop_gamma*(acq_vals - acq_vals.max()))
        probs /= probs.sum()
        query_ind = np.random.choice(candidate_ind, batch_size, p=probs)
    else:
        query_ind = policy(candidate_ind, acq_vals, batch_size) # user-defined policy

    if return_acq_vals:
        return query_ind, acq_vals
    return query_ind

def update(self, query_ind, query_labels)

Expand source code

def update(self, query_ind, query_labels):
    if np.intersect1d(query_ind, self.labeled_ind).size > 0 and not self.printed_warning:
        print("WARNING: Having multiple observations at a single node detected")
        self.printed_warning = True
    self.labeled_ind = np.append(self.labeled_ind, query_ind)
    self.labels = np.append(self.labels, query_labels)
    self.u = self.model.fit(self.labeled_ind, self.labels)
    self.unlabeled_ind = np.setdiff1d(self.all_inds, self.labeled_ind)
    self.acq_function.update(query_ind, query_labels)
    return

class model_change (C, V=None, gamma2=0.010000000000000002, unc_method='smallest_margin')

Model Change

Active learning algorithm that selects points that will produce the greatest change in the model.

Examples

import graphlearning.active_learning as al
import graphlearning as gl
import numpy as np
import matplotlib.pyplot as plt
import sklearn.datasets as datasets

X,labels = datasets.make_moons(n_samples=500,noise=0.1)
W = gl.weightmatrix.knn(X,10)
train_ind = gl.trainsets.generate(labels, rate=5)
plt.scatter(X[:,0],X[:,1], c=labels)
plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
plt.show()

# compute initial, low-rank (spectral truncation) covariance matrix 
evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
C = np.diag(1. / (evals + 1e-11))
AL = gl.active_learning.active_learner(model, gl.active_learning.model_change, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

for i in range(10):
    query_points = AL.select_queries() # return this iteration's newly chosen points
    query_labels = y[query_points] # simulate the human in the loop process
    AL.update(query_points, query_labels) # update the active_learning object's labeled set

    # plot
    plt.scatter(X[:,0],X[:,1], c=y)
    plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
    plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
    plt.show()
    print(AL.labeled_ind)
    print(AL.labels)

Reference

[1] Miller, K. and Bertozzi, A. L., “Model-change active learning in graph-based semi-supervised learning,” (Oct. 2021). arXiv: 2110.07739.

[2] Karzand, M. and Nowak, R. D., “Maximin active learning in overparameterized model classes,” IEEE Journal on Selected Areas in Information Theory 1, 167–177 (May 2020).

Expand source code

class model_change(acquisition_function):
    """Model Change
    ===================

    Active learning algorithm that selects points that will produce the greatest change in the model.

    Examples
    --------
    ```py
    import graphlearning.active_learning as al
    import graphlearning as gl
    import numpy as np
    import matplotlib.pyplot as plt
    import sklearn.datasets as datasets

    X,labels = datasets.make_moons(n_samples=500,noise=0.1)
    W = gl.weightmatrix.knn(X,10)
    train_ind = gl.trainsets.generate(labels, rate=5)
    plt.scatter(X[:,0],X[:,1], c=labels)
    plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
    plt.show()

    # compute initial, low-rank (spectral truncation) covariance matrix 
    evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
    C = np.diag(1. / (evals + 1e-11))
    AL = gl.active_learning.active_learner(model, gl.active_learning.model_change, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

    for i in range(10):
        query_points = AL.select_queries() # return this iteration's newly chosen points
        query_labels = y[query_points] # simulate the human in the loop process
        AL.update(query_points, query_labels) # update the active_learning object's labeled set

        # plot
        plt.scatter(X[:,0],X[:,1], c=y)
        plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
        plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
        plt.show()
        print(AL.labeled_ind)
        print(AL.labels)
    ```

    Reference
    ---------
    [1] Miller, K. and Bertozzi, A. L., “Model-change active learning in graph-based semi-supervised learning,”
    (Oct. 2021). arXiv: 2110.07739.

    [2] Karzand, M. and Nowak, R. D., “Maximin active learning in overparameterized model classes,” IEEE
    Journal on Selected Areas in Information Theory 1, 167–177 (May 2020).
    """
    def __init__(self, C, V=None, gamma2=0.1**2., unc_method='smallest_margin'):
        assert (C.shape[0] == C.shape[1]) or (V is not None)
        self.C = C.copy()
        self.V = V
        self.gamma2 = gamma2
        self.unc_sampling = unc_sampling(unc_method=unc_method)
        if self.V is None:
            self.storage = 'full'
        else:
            self.storage = 'trunc'
        
    def compute(self, u, candidate_ind):
        unc_terms = self.unc_sampling.compute(u, candidate_ind)
        if self.storage == 'full':
            col_norms = np.linalg.norm(self.C, axis=0)
            diag_terms = self.gamma2 + self.C.diagonal()
        else:
            Cavk = self.C @ self.V[candidate_ind,:].T
            col_norms = np.linalg.norm(Cavk, axis=0)
            diag_terms = (self.gamma2 + np.array([np.inner(self.V[k,:], Cavk[:, i]) for i,k in enumerate(candidate_ind)]))
        return unc_terms * col_norms / diag_terms  


    def update(self, query_ind, query_labels):
        for k in query_ind:
            if self.storage == 'full':
                self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
            else:
                vk = self.V[k]
                Cavk = self.C @ vk
                ip = np.inner(vk, Cavk)
                self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 
        return

Ancestors

• acquisition_function

Methods

def update(self, query_ind, query_labels)

Expand source code

def update(self, query_ind, query_labels):
    for k in query_ind:
        if self.storage == 'full':
            self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
        else:
            vk = self.V[k]
            Cavk = self.C @ vk
            ip = np.inner(vk, Cavk)
            self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 
    return

Inherited members

• acquisition_function:
  • compute

class model_change_var_opt (C, V=None, gamma2=0.010000000000000002, unc_method='smallest_margin')

Model Change Variance Optimization

Active learning algorithm that is a combination of Model Change and Variance Optimization.

Examples

import graphlearning.active_learning as al
import graphlearning as gl
import numpy as np
import matplotlib.pyplot as plt
import sklearn.datasets as datasets

X,labels = datasets.make_moons(n_samples=500,noise=0.1)
W = gl.weightmatrix.knn(X,10)
train_ind = gl.trainsets.generate(labels, rate=5)
plt.scatter(X[:,0],X[:,1], c=labels)
plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
plt.show()

# compute initial, low-rank (spectral truncation) covariance matrix 
evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
C = np.diag(1. / (evals + 1e-11))
AL = gl.active_learning.active_learner(model, gl.active_learning.model_change_var_opt, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

for i in range(10):
    query_points = AL.select_queries() # return this iteration's newly chosen points
    query_labels = y[query_points] # simulate the human in the loop process
    AL.update(query_points, query_labels) # update the active_learning object's labeled set

    # plot
    plt.scatter(X[:,0],X[:,1], c=y)
    plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
    plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
    plt.show()
    print(AL.labeled_ind)
    print(AL.labels)

Reference

[1] Ji, M. and Han, J., “A variance minimization criterion to active learning on graphs,” in [Artificial Intelligence and Statistics ], 556–564 (Mar. 2012).

[2] Miller, K. and Bertozzi, A. L., “Model-change active learning in graph-based semi-supervised learning,” (Oct. 2021). arXiv: 2110.07739.

[3] Karzand, M. and Nowak, R. D., “Maximin active learning in overparameterized model classes,” IEEE Journal on Selected Areas in Information Theory 1, 167–177 (May 2020).

Expand source code

class model_change_var_opt(acquisition_function):
    """Model Change Variance Optimization
    ===================

    Active learning algorithm that is a combination of Model Change and Variance Optimization.

    Examples
    --------
    ```py
    import graphlearning.active_learning as al
    import graphlearning as gl
    import numpy as np
    import matplotlib.pyplot as plt
    import sklearn.datasets as datasets

    X,labels = datasets.make_moons(n_samples=500,noise=0.1)
    W = gl.weightmatrix.knn(X,10)
    train_ind = gl.trainsets.generate(labels, rate=5)
    plt.scatter(X[:,0],X[:,1], c=labels)
    plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
    plt.show()

    # compute initial, low-rank (spectral truncation) covariance matrix 
    evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
    C = np.diag(1. / (evals + 1e-11))
    AL = gl.active_learning.active_learner(model, gl.active_learning.model_change_var_opt, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

    for i in range(10):
        query_points = AL.select_queries() # return this iteration's newly chosen points
        query_labels = y[query_points] # simulate the human in the loop process
        AL.update(query_points, query_labels) # update the active_learning object's labeled set

        # plot
        plt.scatter(X[:,0],X[:,1], c=y)
        plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
        plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
        plt.show()
        print(AL.labeled_ind)
        print(AL.labels)
    ```

    Reference
    ---------
    [1] Ji, M. and Han, J., “A variance minimization criterion to active learning on graphs,” in [Artificial Intelligence
    and Statistics ], 556–564 (Mar. 2012).

    [2] Miller, K. and Bertozzi, A. L., “Model-change active learning in graph-based semi-supervised learning,”
    (Oct. 2021). arXiv: 2110.07739.

    [3] Karzand, M. and Nowak, R. D., “Maximin active learning in overparameterized model classes,” IEEE
    Journal on Selected Areas in Information Theory 1, 167–177 (May 2020).
    """
    def __init__(self, C, V=None, gamma2=0.1**2., unc_method='smallest_margin'):
        assert (C.shape[0] == C.shape[1]) or (V is not None)
        self.C = C.copy()
        self.V = V
        self.gamma2 = gamma2
        self.unc_sampling = unc_sampling(unc_method=unc_method)
        if self.V is None:
            self.storage = 'full'
        else:
            self.storage = 'trunc'
        
    def compute(self, u, candidate_ind):
        unc_terms = self.unc_sampling.compute(u, candidate_ind)
        if self.storage == 'full':
            col_norms = np.linalg.norm(self.C, axis=0)**2.
            diag_terms = self.gamma2 + self.C.diagonal()
        else:
            Cavk = self.C @ self.V[candidate_ind,:].T
            col_norms = np.linalg.norm(Cavk, axis=0)**2.
            diag_terms = (self.gamma2 + np.array([np.inner(self.V[k,:], Cavk[:, i]) for i,k in enumerate(candidate_ind)]))
        return unc_terms * col_norms / diag_terms  


    def update(self, query_ind, query_labels):
        for k in query_ind:
            if self.storage == 'full':
                self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
            else:
                vk = self.V[k]
                Cavk = self.C @ vk
                ip = np.inner(vk, Cavk)
                self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 
        return

Ancestors

• acquisition_function

Methods

def update(self, query_ind, query_labels)

Expand source code

def update(self, query_ind, query_labels):
    for k in query_ind:
        if self.storage == 'full':
            self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
        else:
            vk = self.V[k]
            Cavk = self.C @ vk
            ip = np.inner(vk, Cavk)
            self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 
    return

Inherited members

• acquisition_function:
  • compute

class sigma_opt (C, V=None, gamma2=0.010000000000000002)

Sigma Optimization

Active learning algorithm that selects points that minimizes the sum of the associated entries in the covariance matrix.

Examples

import graphlearning.active_learning as al
import graphlearning as gl
import numpy as np
import matplotlib.pyplot as plt
import sklearn.datasets as datasets

X,labels = datasets.make_moons(n_samples=500,noise=0.1)
W = gl.weightmatrix.knn(X,10)
train_ind = gl.trainsets.generate(labels, rate=5)
plt.scatter(X[:,0],X[:,1], c=labels)
plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
plt.show()

# compute initial, low-rank (spectral truncation) covariance matrix 
evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
C = np.diag(1. / (evals + 1e-11))
AL = gl.active_learning.active_learner(model, gl.active_learning.sigma_opt, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

for i in range(10):
    query_points = AL.select_queries() # return this iteration's newly chosen points
    query_labels = y[query_points] # simulate the human in the loop process
    AL.update(query_points, query_labels) # update the active_learning object's labeled set

    # plot
    plt.scatter(X[:,0],X[:,1], c=y)
    plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
    plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
    plt.show()
    print(AL.labeled_ind)
    print(AL.labels)

Reference

[1] Ma, Y., Garnett, R., and Schneider, J., “Σ-optimality for active learning on Gaussian random fields,” in [Advances in Neural Information Processing Systems 26 ], Burges, C. J. C., Bottou, L., Welling, M., Ghahramani, Z., and Weinberger, K. Q., eds., 2751–2759, Curran Associates, Inc. (2013).

Expand source code

class sigma_opt(acquisition_function):
    """Sigma Optimization
    ===================

    Active learning algorithm that selects points that minimizes the sum of the associated entries in the covariance matrix.

    Examples
    --------
    ```py
    import graphlearning.active_learning as al
    import graphlearning as gl
    import numpy as np
    import matplotlib.pyplot as plt
    import sklearn.datasets as datasets

    X,labels = datasets.make_moons(n_samples=500,noise=0.1)
    W = gl.weightmatrix.knn(X,10)
    train_ind = gl.trainsets.generate(labels, rate=5)
    plt.scatter(X[:,0],X[:,1], c=labels)
    plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
    plt.show()

    # compute initial, low-rank (spectral truncation) covariance matrix 
    evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
    C = np.diag(1. / (evals + 1e-11))
    AL = gl.active_learning.active_learner(model, gl.active_learning.sigma_opt, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

    for i in range(10):
        query_points = AL.select_queries() # return this iteration's newly chosen points
        query_labels = y[query_points] # simulate the human in the loop process
        AL.update(query_points, query_labels) # update the active_learning object's labeled set

        # plot
        plt.scatter(X[:,0],X[:,1], c=y)
        plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
        plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
        plt.show()
        print(AL.labeled_ind)
        print(AL.labels)
    ```

    Reference
    ---------
    [1] Ma, Y., Garnett, R., and Schneider, J., “Σ-optimality for active learning on Gaussian random fields,”
    in [Advances in Neural Information Processing Systems 26 ], Burges, C. J. C., Bottou, L., Welling, M.,
    Ghahramani, Z., and Weinberger, K. Q., eds., 2751–2759, Curran Associates, Inc. (2013).
    """
    ## this only handles the FULL C computation or the spectral truncation
    def __init__(self, C, V=None, gamma2=0.1**2.):
        assert (C.shape[0] == C.shape[1]) or (V is not None)
        self.C = C.copy()
        self.V = V
        self.gamma2 = gamma2
        if self.V is None:
            self.storage = 'full'
        else:
            self.storage = 'trunc'
        
        
    def compute(self, u, candidate_ind):
        if self.storage == 'full':
            col_sums = np.sum(self.C[:, candidate_ind], axis=0)**2.
            diag_terms = self.gamma2 + self.C.diagonal()[candidate_ind]
        else:
            Cavk = self.C @ self.V[candidate_ind,:].T
            col_sums = np.sum(Cavk, axis=0)**2.
            diag_terms = (self.gamma2 + np.array([np.inner(self.V[k,:], Cavk[:, i]) for i,k in enumerate(candidate_ind)]))
            
        return col_sums/ diag_terms


    def update(self, query_ind, query_labels):
        for k in query_ind:
            if self.storage == 'full':
                self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
            else:
                vk = self.V[k]
                Cavk = self.C @ vk
                ip = np.inner(vk, Cavk)
                self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 

        return

Ancestors

• acquisition_function

Methods

def update(self, query_ind, query_labels)

Expand source code

def update(self, query_ind, query_labels):
    for k in query_ind:
        if self.storage == 'full':
            self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
        else:
            vk = self.V[k]
            Cavk = self.C @ vk
            ip = np.inner(vk, Cavk)
            self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 

    return

Inherited members

• acquisition_function:
  • compute

class unc_sampling (unc_method='smallest_margin')

Uncertainty Sampling

Active learning algorithm that selects points that the classifier is most uncertain of.

Examples

import graphlearning.active_learning as al
import graphlearning as gl
import numpy as np
import matplotlib.pyplot as plt
import sklearn.datasets as datasets

X,labels = datasets.make_moons(n_samples=500,noise=0.1)
W = gl.weightmatrix.knn(X,10)
train_ind = gl.trainsets.generate(labels, rate=5)
plt.scatter(X[:,0],X[:,1], c=labels)
plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
plt.show()


model = gl.ssl.laplace(W)
AL = gl.active_learning.active_learner(model, gl.active_learning.unc_sampling, train_ind, y[train_ind])

for i in range(10):
    query_points = AL.select_queries() # return this iteration's newly chosen points
    query_labels = y[query_points] # simulate the human in the loop process
    AL.update(query_points, query_labels) # update the active_learning object's labeled set

    # plot
    plt.scatter(X[:,0],X[:,1], c=y)
    plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
    plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
    plt.show()
    print(AL.labeled_ind)
    print(AL.labels)

Reference

[1] Settles, B., [Active Learning], vol. 6, Morgan & Claypool Publishers LLC (June 2012).

Expand source code

class unc_sampling(acquisition_function):
    """Uncertainty Sampling
    ===================

    Active learning algorithm that selects points that the classifier is most uncertain of.

    Examples
    --------
    ```py
    import graphlearning.active_learning as al
    import graphlearning as gl
    import numpy as np
    import matplotlib.pyplot as plt
    import sklearn.datasets as datasets

    X,labels = datasets.make_moons(n_samples=500,noise=0.1)
    W = gl.weightmatrix.knn(X,10)
    train_ind = gl.trainsets.generate(labels, rate=5)
    plt.scatter(X[:,0],X[:,1], c=labels)
    plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
    plt.show()

    
    model = gl.ssl.laplace(W)
    AL = gl.active_learning.active_learner(model, gl.active_learning.unc_sampling, train_ind, y[train_ind])

    for i in range(10):
        query_points = AL.select_queries() # return this iteration's newly chosen points
        query_labels = y[query_points] # simulate the human in the loop process
        AL.update(query_points, query_labels) # update the active_learning object's labeled set

        # plot
        plt.scatter(X[:,0],X[:,1], c=y)
        plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
        plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
        plt.show()
        print(AL.labeled_ind)
        print(AL.labels)
    ```

    Reference
    ---------
    [1] Settles, B., [Active Learning], vol. 6, Morgan & Claypool Publishers LLC (June 2012).
    """
    def __init__(self, unc_method='smallest_margin'):
        self.unc_method = unc_method

    def compute(self, u, candidate_ind):
        if self.unc_method == "norm":
            u_probs = softmax(u[candidate_ind], axis=1)
            one_hot_predicted_labels = np.eye(u.shape[1])[np.argmax(u[candidate_ind], axis=1)]
            unc_terms = np.linalg.norm((u_probs - one_hot_predicted_labels), axis=1)
        elif self.unc_method == "entropy":
            u_probs = softmax(u[candidate_ind], axis=1)
            unc_terms = np.max(u_probs, axis=1) - np.sum(u_probs*np.log(u_probs +.00001), axis=1)
        elif self.unc_method == "least_confidence":
            unc_terms = np.ones((u[candidate_ind].shape[0],)) - np.max(u[candidate_ind], axis=1)
        elif self.unc_method == "smallest_margin":
            u_sort = np.sort(u[candidate_ind])
            unc_terms = 1.-(u_sort[:,-1] - u_sort[:,-2])
        elif self.unc_method == "largest_margin":
            u_sort = np.sort(u[candidate_ind])
            unc_terms = 1.-(u_sort[:,-1] - u_sort[:,0])
        elif self.unc_method == "unc_2norm":
            unc_terms = 1. - np.linalg.norm(u[candidate_ind], axis=1)
        return unc_terms

Ancestors

• acquisition_function

Inherited members

• acquisition_function:
  • compute

class var_opt (C, V=None, gamma2=0.010000000000000002)

Variance Optimization

Active learning algorithm that selects points that minimizes the variance of the distribution of unlabeled nodes.

Examples

import graphlearning.active_learning as al
import graphlearning as gl
import numpy as np
import matplotlib.pyplot as plt
import sklearn.datasets as datasets

X,labels = datasets.make_moons(n_samples=500,noise=0.1)
W = gl.weightmatrix.knn(X,10)
train_ind = gl.trainsets.generate(labels, rate=5)
plt.scatter(X[:,0],X[:,1], c=labels)
plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
plt.show()

# compute initial, low-rank (spectral truncation) covariance matrix 
evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
C = np.diag(1. / (evals + 1e-11))
AL = gl.active_learning.active_learner(model, gl.active_learning.var_opt, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

for i in range(10):
    query_points = AL.select_queries() # return this iteration's newly chosen points
    query_labels = y[query_points] # simulate the human in the loop process
    AL.update(query_points, query_labels) # update the active_learning object's labeled set

    # plot
    plt.scatter(X[:,0],X[:,1], c=y)
    plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
    plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
    plt.show()
    print(AL.labeled_ind)
    print(AL.labels)

Reference

[1] Ji, M. and Han, J., “A variance minimization criterion to active learning on graphs,” in [Artificial Intelligence and Statistics ], 556–564 (Mar. 2012).

Expand source code

class var_opt(acquisition_function):
    """Variance Optimization
    ===================

    Active learning algorithm that selects points that minimizes the variance of the distribution of unlabeled nodes.

    Examples
    --------
    ```py
    import graphlearning.active_learning as al
    import graphlearning as gl
    import numpy as np
    import matplotlib.pyplot as plt
    import sklearn.datasets as datasets

    X,labels = datasets.make_moons(n_samples=500,noise=0.1)
    W = gl.weightmatrix.knn(X,10)
    train_ind = gl.trainsets.generate(labels, rate=5)
    plt.scatter(X[:,0],X[:,1], c=labels)
    plt.scatter(X[train_ind,0],X[train_ind,1], c='r')
    plt.show()

    # compute initial, low-rank (spectral truncation) covariance matrix 
    evals, evecs = model.graph.eigen_decomp(normalization='normalized', k=50)
    C = np.diag(1. / (evals + 1e-11))
    AL = gl.active_learning.active_learner(model, gl.active_learning.var_opt, train_ind, y[train_ind], C=C.copy(), V=evecs.copy())

    for i in range(10):
        query_points = AL.select_queries() # return this iteration's newly chosen points
        query_labels = y[query_points] # simulate the human in the loop process
        AL.update(query_points, query_labels) # update the active_learning object's labeled set

        # plot
        plt.scatter(X[:,0],X[:,1], c=y)
        plt.scatter(X[AL.labeled_ind,0],X[AL.labeled_ind,1], c='r')
        plt.scatter(X[query_points,0],X[query_points,1], c='r', marker='*', s=200, edgecolors='k', linewidths=1)
        plt.show()
        print(AL.labeled_ind)
        print(AL.labels)
    ```

    Reference
    ---------
    [1] Ji, M. and Han, J., “A variance minimization criterion to active learning on graphs,” in [Artificial Intelligence
    and Statistics ], 556–564 (Mar. 2012).
    """

    ## this only handles the FULL C computation or the spectral truncation
    def __init__(self, C, V=None, gamma2=0.1**2.):
        assert (C.shape[0] == C.shape[1]) or (V is not None)
        self.C = C.copy()
        self.V = V
        self.gamma2 = gamma2
        if self.V is None:
            self.storage = 'full'
        else:
            self.storage = 'trunc'
        
        
    def compute(self, u, candidate_ind):
        if self.storage == 'full':
            col_norms = np.linalg.norm(self.C[:,candidate_ind], axis=0)**2.
            diag_terms = self.gamma2 + self.C.diagonal()[candidate_ind]
        else:
            Cavk = self.C @ self.V[candidate_ind,:].T
            col_norms = np.linalg.norm(Cavk, axis=0)**2.
            diag_terms = (self.gamma2 + np.array([np.inner(self.V[k,:], Cavk[:, i]) for i,k in enumerate(candidate_ind)]))
            
        return col_norms / diag_terms

    def update(self, query_ind, query_labels):
        for k in query_ind:
            if self.storage == 'full':
                self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
            else:
                vk = self.V[k]
                Cavk = self.C @ vk
                ip = np.inner(vk, Cavk)
                self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 

        return 

Ancestors

• acquisition_function

Methods

def update(self, query_ind, query_labels)

Expand source code

def update(self, query_ind, query_labels):
    for k in query_ind:
        if self.storage == 'full':
            self.C -= np.outer(self.C[:,k], self.C[:,k]) / (self.gamma2 + self.C[k,k])
        else:
            vk = self.V[k]
            Cavk = self.C @ vk
            ip = np.inner(vk, Cavk)
            self.C -= np.outer(Cavk, Cavk)/(self.gamma2 + ip) 

    return 

Inherited members

• acquisition_function:
  • compute