Uncertainty_Batch

Batch Selection for Multi-Label Classification: Guided by Dynamic Uncertainty and Label Correlations

Package:

python==3.7 skmultilearn==1.2.1 pytorch==1.11.0 numpy== 1.21.5

Usage:

run UncertainBatch.ipynb cell by cell

In detail, we can train on different datasets for the last cell:

give all dataset used in paper:

path_to_arff_files = ["emotions","scene","yeast", "Corel5k","rcv1subset1","rcv1subset2","rcv1subset3","yahoo-Business1","yahoo-Arts1","bibtex",'tmc2007','enron','cal500','LLOG-F']

give the label count of corresponding dataset:

label_counts = [6, 6,14,374,101,101,101,28,25,159,22,53,174,75]

give the feature retention ratio of the corresponding dataset:

select_feature=[1,1,1,1,0.02,0.02,0.02,0.05,0.05,1,0.01,1,1,1]

The warm epoch is set to 5 by default.

For $$\lambda_{1}$$ default:

$$ E[idx, j] = \frac{1}{2} \times meandiffs + \frac{1}{2} \times currententropy, \quad where \quad \lambda_{1} = \frac{1}{2} $$

key code:

1. def update_H(H, y_pred, ids, max_history_length=5):

y_pred_numpy = y_pred.detach().cpu().numpy() 

for i, idx in enumerate(ids):

    if idx not in H:
    
        H[idx] = deque(maxlen=max_history_length) 
        
    H[idx].append(y_pred_numpy[i])   
    
return H

max_history_length is the size of the sliding window, H is a queue used to store the latest sliding window size predictions

2. def update_E(H, E, ids, label_dim):

for idx in ids:

    current_predictions_history = np.array(H[idx])
    
    last_row_index = len(current_predictions_history) - 1
    
    for j in range(label_dim): 
    
        diffs = np.abs(np.diff(current_predictions_history[:, j]))
        
        mean_diffs = np.sum(diffs)/len(diffs)
        
        current_entropy = -1 / np.log(2) * (
            current_predictions_history[last_row_index][j] * np.log(current_predictions_history[last_row_index][j]) 
            + (1 - current_predictions_history[last_row_index][j]) * np.log(1 - current_predictions_history[last_row_index][j])
        )
        
        E[idx,j] = 1/2 * mean_diffs + 1/2 * current_entropy
        
return E

E is a two-dimensional array of n (number of samples) * q (number of labels), which updates the uncertainty of each sample and label in current epoch.

3. def update_U(E, U, epoch):

E[E > 1] = 1

if epoch >= 5:

    bins = np.array([0, 0.2, 0.4, 0.6, 0.8, 1.0])
    
    discrete_values = np.array([0.1, 0.3, 0.5, 0.7, 0.9])
    
    indices = np.digitize(E, bins) - 1
    
    indices = np.clip(indices, 0, len(discrete_values) - 1)
    
    E = discrete_values[indices]
    
else:

    I = np.ones((E.shape[1], E.shape[1]))
    
I = np.ones((E.shape[1], E.shape[1]))

U = np.dot(E, I)

w = np.sum(U, axis=1)

return w

Update the selection weight w through E

Comparison method:

analyse.ipynb includes all comparison methods,You can run the corresponding cell to obtain the results

Acknowledgement and Citation:

We are pleased that this paper has been accepted by AAAI2025 and will update it after its official publication

Contact:

If you have any questions or suggestions, feel free to contact me: zacqupt@gmail.com.