Multi-label NLP: An Analysis of Class Imbalance and Loss Function Approaches

Multi-label NLP: An Evaluation of Class Imbalance and Loss Perform Approaches – KDnuggets

Multi-label NLP refers back to the activity of assigning a number of labels to a given textual content enter, somewhat than only one label. In conventional NLP duties, similar to textual content classification or sentiment evaluation, every enter is usually assigned a single label based mostly on its content material. Nevertheless, in lots of real-world situations, a bit of textual content can belong to a number of classes or categorical a number of sentiments concurrently.

Multi-label NLP is essential as a result of it permits us to seize extra nuanced and sophisticated data from textual content knowledge. For instance, within the area of buyer suggestions evaluation, a buyer evaluation could categorical each optimistic and unfavourable sentiments on the identical time, or it could contact upon a number of facets of a services or products. By assigning a number of labels to such inputs, we are able to achieve a extra complete understanding of the shopper’s suggestions and take extra focused actions to deal with their considerations.

This text delves right into a noteworthy case of Provectus use of multi-label NLP.



A shopper approached us with a request to assist them automate labeling paperwork of a sure sort. At first look, the duty gave the impression to be easy and simply solved. Nevertheless, as we labored on the case, we encountered a dataset with inconsistent annotations. Although our buyer had confronted challenges with various class numbers and adjustments of their evaluation staff over time, they’d invested vital efforts into creating a various dataset with a spread of annotations. Whereas there existed some imbalances and uncertainties within the labels, this dataset offered a beneficial alternative for evaluation and additional exploration.


Lets take a better take a look at the dataset, discover the metrics and our strategy, and recap how Provectus solved the issue of multi-label textual content classification.



The dataset has 14,354 observations, with 124 distinctive courses (labels). Our activity is to assign one or a number of courses to each statement.

Desk 1 supplies descriptive statistics for the dataset.

On common, we’ve about two courses per statement, with a mean of 261 totally different texts describing a single class.


Desk 1: Dataset Statistic


In Determine 1, we see the distribution of courses within the prime graph, and we’ve a sure variety of HEAD labels with the best frequency of prevalence within the dataset. Additionally observe that almost all of courses have a low frequency of prevalence.


Multi-label NLP: An Analysis of Class Imbalance and Loss Function Approaches


Within the backside graph we see that there’s frequent overlap between the courses which might be greatest represented within the dataset, and the courses which have low significance.

We modified the method of splitting the dataset into practice/val/take a look at units. As an alternative of utilizing a standard methodology, we’ve employed iterative stratification, to supply a well-balanced distribution of proof of label relations. For that, we used Scikit Multi-learn

from skmultilearn.model_selection import iterative_train_test_split

mlb = MultiLabelBinarizer()

def balanced_split(df, mlb, test_size=0.5):
    ind = np.expand_dims(np.arange(len(df)), axis=1)
    labels = mlb.remodel(df["tag"])
    ind_train, _, ind_test, _ = iterative_train_test_split(
        ind, labels, test_size
    return df.iloc[ind_train[:, 0]], df.iloc[ind_test[:, 0]]

df_train, df_tmp = balanced_split(df, test_size=0.4)
df_val, df_test = balanced_split(df_tmp, test_size=0.5)


We obtained the next distribution:

  1. The coaching dataset accommodates 60% of the information and covers all 124 labels
  2. The validation dataset accommodates 20% of the information and covers all 124 labels
  3. The take a look at dataset accommodates 20% of the information and covers all 124 labels



Multi-label classification is a sort of supervised machine studying algorithm that enables us to assign a number of labels to a single knowledge pattern. It differs from binary classification the place the mannequin predicts solely two classes, and multi-class classification the place the mannequin predicts just one out of a number of courses for a pattern.

Analysis metrics for multi-label classification efficiency are inherently totally different from these utilized in multi-class (or binary) classification as a result of inherent variations of the classification downside. Extra detailed data could be discovered on Wikipedia.

We chosen metrics which might be most fitted for us:

  1. Precision measures the proportion of true optimistic predictions among the many complete optimistic predictions made by the mannequin.
  2. Recall measures the proportion of true optimistic predictions amongst all precise optimistic samples.
  3. F1-score is the harmonic imply of precision and recall, which helps to revive steadiness between the 2.
  4. Hamming loss is the fraction of labels which might be incorrectly predicted

We additionally observe the variety of predicted labels within the set outlined as depend for labels, for which we obtain an F1 rating > 0.



Multi-Label Classification is a sort of supervised studying downside the place a single occasion or instance could be related to a number of labels or classifications, versus conventional single-label classification, the place every occasion is just related to a single class label.

To resolve multi-label classification issues, there are two principal classes of strategies:

  1. Downside transformation strategies
  2. Algorithm adaptation strategies

Downside transformation strategies allow us to remodel multi-label classification duties into a number of single-label classification duties. For instance, the Binary Relevance (BR) baseline strategy treats each label as a separate binary classification downside. On this case, the multi-label downside is reworked into a number of single-label issues.

Algorithm adaptation strategies modify the algorithms themselves to deal with multi-label knowledge natively, with out remodeling the duty into a number of single-label classification duties. An instance of this strategy is the BERT mannequin, which is a pre-trained transformer-based language mannequin that may be fine-tuned for varied NLP duties, together with multi-label textual content classification. BERT is designed to deal with multi-label knowledge straight, with out the necessity for downside transformation.

Within the context of utilizing BERT for multi-label textual content classification, the usual strategy is to make use of Binary Cross-Entropy (BCE) loss because the loss operate. BCE loss is a generally used loss operate for binary classification issues and could be simply prolonged to deal with multi-label classification issues by computing the loss for every label independently, after which summing the losses. On this case, the BCE loss operate measures the error between predicted possibilities and true labels, the place predicted possibilities are obtained from the ultimate sigmoid activation layer within the BERT mannequin.

Now, let’s take a better take a look at Determine 2 beneath.


Multi-label NLP: An Analysis of Class Imbalance and Loss Function Approaches
Determine 2. Metrics for baseline fashions


The graph on the left exhibits a comparability of metrics for a baseline: BERT and baseline: ML. Thus, it may be seen that for baseline: BERT, the F1 and Recall scores are roughly 1.5 instances greater, whereas the Precision for baseline: ML is 2 instances greater than that of mannequin 1. By analyzing the general share of predicted courses proven on the fitting, we see that baseline: BERT predicted courses greater than 10 instances that of baseline: ML.

As a result of the utmost end result for the baseline: BERT is lower than 50% of all courses, the outcomes are fairly discouraging. Lets work out the best way to enhance these outcomes.



Primarily based on the excellent article “Balancing Strategies for Multi-label Textual content Classification with Lengthy-Tailed Class Distribution”, we discovered that distribution-balanced loss will be the most fitted strategy for us.


Distribution-balanced loss


Distribution-balanced loss is a method utilized in multi-label textual content classification issues to deal with imbalances at school distribution. In these issues, some courses have a a lot greater frequency of prevalence in comparison with others, leading to mannequin bias towards these extra frequent courses.

To handle this difficulty, distribution-balanced loss goals to steadiness the contribution of every pattern within the loss operate. That is achieved by re-weighting the lack of every pattern based mostly on the inverse of its frequency of prevalence within the dataset. By doing so, the contribution of much less frequent courses is elevated, and the contribution of extra frequent courses is decreased, thus balancing the general class distribution.

This method has been proven to be efficient in bettering the efficiency of fashions on long-tailed class distribution issues. By decreasing the impression of frequent courses and growing the impression of rare courses, the mannequin is ready to higher seize patterns within the knowledge and produce extra balanced predictions.

Multi-label NLP: An Analysis of Class Imbalance and Loss Function Approaches
Implementation of Resample Class


import torch
import torch.nn as nn
import torch.nn.useful as F
import numpy as np

class ResampleLoss(nn.Module):
    def __init__(
        focal=dict(focal=True, alpha=0.5, gamma=2),
        map_param=dict(alpha=10.0, beta=0.2, gamma=0.1),
        CB_loss=dict(CB_beta=0.9, CB_mode="average_w"),
        logit_reg=dict(neg_scale=5.0, init_bias=0.1),
        tremendous(ResampleLoss, self).__init__()
        assert (use_sigmoid is True) or (partial is False)
        self.use_sigmoid = use_sigmoid
        self.partial = partial
        self.loss_weight = loss_weight = discount
        if self.use_sigmoid:
            if self.partial:
                self.cls_criterion = partial_cross_entropy
                self.cls_criterion = binary_cross_entropy
            self.cls_criterion = cross_entropy
        # reweighting operate
        self.reweight_func = reweight_func
        # normalization (non-compulsory)
        self.weight_norm = weight_norm
        # focal loss params
        self.focal = focal["focal"]
        self.gamma = focal["gamma"]
        self.alpha = focal["alpha"]
        # mapping operate params
        self.map_alpha = map_param["alpha"]
        self.map_beta = map_param["beta"]
        self.map_gamma = map_param["gamma"]
        # CB loss params (non-compulsory)
        self.CB_beta = CB_loss["CB_beta"]
        self.CB_mode = CB_loss["CB_mode"]
        self.class_freq = (
        self.num_classes = self.class_freq.form[0]
        self.train_num = train_num  # solely was once divided by class_freq
        # regularization params
        self.logit_reg = logit_reg
        self.neg_scale = (
            logit_reg["neg_scale"] if "neg_scale" in logit_reg else 1.0
        init_bias = (
            logit_reg["init_bias"] if "init_bias" in logit_reg else 0.0
        self.init_bias = (
            -torch.log(self.train_num / self.class_freq - 1) * init_bias
        self.freq_inv = (
            torch.ones(self.class_freq.form).cuda() / self.class_freq
        self.propotion_inv = self.train_num / self.class_freq

    def ahead(
        assert reduction_override in (None, "none", "imply", "sum")
        discount = (
            reduction_override if reduction_override else
        weight = self.reweight_functions(label)
        cls_score, weight = self.logit_reg_functions(
            label.float(), cls_score, weight
        if self.focal:
            logpt = self.cls_criterion(
            # pt is sigmoid(logit) for pos or sigmoid(-logit) for neg
            pt = torch.exp(-logpt)
            wtloss = self.cls_criterion(
                cls_score, label.float(), weight=weight, discount="none"
            alpha_t = torch.the place(label == 1, self.alpha, 1 - self.alpha)
            loss = alpha_t * ((1 - pt) ** self.gamma) * wtloss
            loss = reduce_loss(loss, discount)
            loss = self.cls_criterion(
                cls_score, label.float(), weight, discount=discount
        loss = self.loss_weight * loss
        return loss

    def reweight_functions(self, label):
        if self.reweight_func is None:
            return None
        elif self.reweight_func in ["inv", "sqrt_inv"]:
            weight = self.RW_weight(label.float())
        elif self.reweight_func in "rebalance":
            weight = self.rebalance_weight(label.float())
        elif self.reweight_func in "CB":
            weight = self.CB_weight(label.float())
            return None
        if self.weight_norm shouldn't be None:
            if "by_instance" in self.weight_norm:
                max_by_instance, _ = torch.max(weight, dim=-1, keepdim=True)
                weight = weight / max_by_instance
            elif "by_batch" in self.weight_norm:
                weight = weight / torch.max(weight)
        return weight

    def logit_reg_functions(self, labels, logits, weight=None):
        if not self.logit_reg:
            return logits, weight
        if "init_bias" in self.logit_reg:
            logits += self.init_bias
        if "neg_scale" in self.logit_reg:
            logits = logits * (1 - labels) * self.neg_scale + logits * labels
            if weight shouldn't be None:
                weight = (
                    weight / self.neg_scale * (1 - labels) + weight * labels
        return logits, weight

    def rebalance_weight(self, gt_labels):
        repeat_rate = torch.sum(
            gt_labels.float() * self.freq_inv, dim=1, keepdim=True
        pos_weight = (
            self.freq_inv.clone().detach().unsqueeze(0) / repeat_rate
        # pos and neg are equally handled
        weight = (
            torch.sigmoid(self.map_beta * (pos_weight - self.map_gamma))
            + self.map_alpha
        return weight

    def CB_weight(self, gt_labels):
        if "by_class" in self.CB_mode:
            weight = (
                torch.tensor((1 - self.CB_beta)).cuda()
                / (1 - torch.pow(self.CB_beta, self.class_freq)).cuda()
        elif "average_n" in self.CB_mode:
            avg_n = torch.sum(
                gt_labels * self.class_freq, dim=1, keepdim=True
            ) / torch.sum(gt_labels, dim=1, keepdim=True)
            weight = (
                torch.tensor((1 - self.CB_beta)).cuda()
                / (1 - torch.pow(self.CB_beta, avg_n)).cuda()
        elif "average_w" in self.CB_mode:
            weight_ = (
                torch.tensor((1 - self.CB_beta)).cuda()
                / (1 - torch.pow(self.CB_beta, self.class_freq)).cuda()
            weight = torch.sum(
                gt_labels * weight_, dim=1, keepdim=True
            ) / torch.sum(gt_labels, dim=1, keepdim=True)
        elif "min_n" in self.CB_mode:
            min_n, _ = torch.min(
                gt_labels * self.class_freq + (1 - gt_labels) * 100000,
            weight = (
                torch.tensor((1 - self.CB_beta)).cuda()
                / (1 - torch.pow(self.CB_beta, min_n)).cuda()
            increase NameError
        return weight

    def RW_weight(self, gt_labels, by_class=True):
        if "sqrt" in self.reweight_func:
            weight = torch.sqrt(self.propotion_inv)
            weight = self.propotion_inv
        if not by_class:
            sum_ = torch.sum(weight * gt_labels, dim=1, keepdim=True)
            weight = sum_ / torch.sum(gt_labels, dim=1, keepdim=True)
        return weight

def reduce_loss(loss, discount):
    """Cut back loss as specified.
        loss (Tensor): Elementwise loss tensor.
        discount (str): Choices are "none", "imply" and "sum".
        Tensor: Decreased loss tensor.
    reduction_enum = F._Reduction.get_enum(discount)
    # none: 0, elementwise_mean:1, sum: 2
    if reduction_enum == 0:
        return loss
    elif reduction_enum == 1:
        return loss.imply()
    elif reduction_enum == 2:
        return loss.sum()

def weight_reduce_loss(loss, weight=None, discount="imply", avg_factor=None):
    """Apply element-wise weight and cut back loss.
        loss (Tensor): Ingredient-wise loss.
        weight (Tensor): Ingredient-wise weights.
        discount (str): Similar as built-in losses of PyTorch.
        avg_factor (float): Avarage issue when computing the imply of losses.
        Tensor: Processed loss values.
    # if weight is specified, apply element-wise weight
    if weight shouldn't be None:
        loss = loss * weight
    # if avg_factor shouldn't be specified, simply cut back the loss
    if avg_factor is None:
        loss = reduce_loss(loss, discount)
        # if discount is imply, then common the loss by avg_factor
        if discount == "imply":
            loss = loss.sum() / avg_factor
        # if discount is 'none', then do nothing, in any other case increase an error
        elif discount != "none":
            increase ValueError(
                'avg_factor can't be used with discount="sum"'
    return loss

def binary_cross_entropy(
    pred, label, weight=None, discount="imply", avg_factor=None
    # weighted element-wise losses
    if weight shouldn't be None:
        weight = weight.float()
    loss = F.binary_cross_entropy_with_logits(
        pred, label.float(), weight, discount="none"
    loss = weight_reduce_loss(
        loss, discount=discount, avg_factor=avg_factor

    return loss



loss_func = ResampleLoss(
    focal=dict(focal=True, alpha=0.5, gamma=2),
    logit_reg=dict(init_bias=0.05, neg_scale=2.0),
    map_param=dict(alpha=0.1, beta=10.0, gamma=0.405),
class_freq - record of frequencies for every class,
train_num - measurement of practice dataset


By carefully investigating the dataset, we’ve concluded that the parameter Equation

= 0.405.


Threshold tuning


One other step in bettering our mannequin was the method of tuning the edge, each within the coaching stage, and within the validation and testing phases. We calculated the dependencies of metrics similar to f1-score, precision, and recall on the edge stage, and we chosen the edge based mostly on the best metric rating. Under you may see the operate implementation of this course of.

Optimization of the F1 rating by tuning the edge:

def optimise_f1_score(true_labels: np.ndarray, pred_labels: np.ndarray):
    best_med_th = 0.5
    true_bools = [tl == 1 for tl in true_labels]
    micro_thresholds = (np.array(vary(-45, 15)) / 100) + best_med_th
    f1_results, pre_results, recall_results = [], [], []
    for th in micro_thresholds:
        pred_bools = [pl > th for pl in pred_labels]
        test_f1 = f1_score(true_bools, pred_bools, common="micro", zero_division=0)
        test_precision = precision_score(
            true_bools, pred_bools, common="micro", zero_division=0
        test_recall = recall_score(
            true_bools, pred_bools, common="micro", zero_division=0
        best_f1_idx = np.argmax(f1_results)
    return micro_thresholds[best_f1_idx]


Analysis and comparability with baseline


These approaches allowed us to coach a brand new mannequin and procure the next end result, which is in comparison with the baseline: BERT in Determine 3 beneath.


Multi-label NLP: An Analysis of Class Imbalance and Loss Function Approaches
Determine 3. Comparability metrics by baseline and newer strategy.


By evaluating the metrics which might be related for classification, we see a major enhance in efficiency measures virtually by 5-6 instances:

The F1 rating elevated from 12% 55%, whereas Precision elevated from 9% 59% and Recall elevated from 15% 51%.

With the adjustments proven in the fitting graph in Determine 3, we are able to now predict 80% of the courses.


Slices of courses


We divided our labels into 4 teams: HEAD, MEDIUM, TAIL, and ZERO. Every group accommodates labels with an analogous quantity of supporting knowledge observations.

As seen in Determine 4, the distributions of the teams are distinct. The rose field (HEAD) has a negatively skewed distribution, the middlebox (MEDIUM) has a positively skewed distribution, and the inexperienced field (TAIL) seems to have a standard distribution.

All teams even have outliers, that are factors outdoors the whiskers within the field plot. The HEAD group has a serious impression on a MAJOR class.

Moreover, we’ve recognized a separate group named “ZERO” which accommodates labels that the mannequin was unable to be taught and can’t acknowledge as a result of minimal variety of occurrences within the dataset (lower than 3% of all observations).


Multi-label NLP: An Analysis of Class Imbalance and Loss Function Approaches
Determine 4. Label counts vs. teams


Desk 2 supplies details about metrics per every group of labels for the take a look at subset of knowledge.


Multi-label NLP: An Analysis of Class Imbalance and Loss Function Approaches
Desk 2. Metrics per group.


  1. The HEAD group accommodates 21 labels with a mean of 112 help observations per label. This group is impacted by outliers and, attributable to its excessive illustration within the dataset, its metrics are excessive: F1 – 73%, Precision – 71%, Recall – 75%.
  2. The MEDIUM group consists of 44 labels with a mean help of 67 observations, which is roughly two instances decrease than the HEAD group. The metrics for this group are anticipated to lower by 50%: F1 – 52%, Precision – 56%, Recall – 51%.
  3. The TAIL group has the biggest variety of courses, however all are poorly represented within the dataset, with a mean of 40 help observations per label. Consequently, the metrics drop considerably: F1 – 21%, Precision – 32%, Recall – 16%.
  4. The ZERO group consists of courses that the mannequin can’t acknowledge in any respect, probably attributable to their low prevalence within the dataset. Every of the 24 labels on this group has a mean of seven help observations.

Determine 5 visualizes the data introduced in Desk 2, offering a visible illustration of the metrics per group of labels.


Multi-label NLP: An Analysis of Class Imbalance and Loss Function Approaches
Determine 5. Metrics vs. label teams. All ZERO values = 0.



On this complete article, we’ve demonstrated {that a} seemingly easy activity of multi-label textual content classification could be difficult when conventional strategies are utilized. We’ve proposed the usage of distribution-balancing loss capabilities to deal with the problem of sophistication imbalance.

We’ve in contrast the efficiency of our proposed strategy to the basic methodology, and evaluated it utilizing real-world enterprise metrics. The outcomes display that using loss capabilities to deal with class imbalances and label co-occurrences supply a viable resolution for multi-label textual content classification.

The proposed use case highlights the significance of contemplating totally different approaches and strategies when coping with multi-label textual content classification, and the potential advantages of distribution-balancing loss capabilities in addressing class imbalances.

If you’re dealing with an analogous difficulty and in search of to streamline doc processing operations inside your group, please contact me or the Provectus staff. We might be completely happy to help you to find extra environment friendly strategies for automating your processes.

Oleksii Babych is a Machine Studying Engineer at Provectus. With a background in physics, he possesses glorious analytical and math abilities, and has gained beneficial expertise by means of scientific analysis and worldwide convention shows, together with SPIE Photonics West. Oleksii focuses on creating end-to-end, large-scale AI/ML options for healthcare and fintech industries. He’s concerned in each stage of the ML growth life cycle, from figuring out enterprise issues to deploying and operating manufacturing ML fashions.

Rinat Akhmetov is the ML Answer Architect at Provectus. With a stable sensible background in Machine Studying (particularly in Laptop Imaginative and prescient), Rinat is a nerd, knowledge fanatic, software program engineer, and workaholic whose second greatest ardour is programming. At Provectus, Rinat is in command of the invention and proof of idea phases, and leads the execution of advanced AI initiatives.

Leave a Reply

Your email address will not be published. Required fields are marked *