AI, Emotions and Representativeness: Investigating the Performance of Convolutional Neural Networks in Facial Recognition by Gender, Ethnicity and Age

doi:10.22167/2675-6528-202602787

Article

September 30, 2026

AI, Emotions and Representativeness: Investigating the Performance of Convolutional Neural Networks in Facial Recognition by Gender, Ethnicity and Age

Gender, Emotions and Representativeness: Investigating the Performance of Convolutional Neural Networks in Facial Recognition by Gender, Ethnicity and Age

Cíntia Cristina Soares; Diego Pedroso dos Santos

DOI: 10.22167/2675-6528-202602787

Article derived from a Course Conclusion Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

Deep Learning models can exhibit algorithmic bias in facial emotion recognition, impacting the representativeness of expressions from different cultures. The study evaluated four convolutional neural network (CNN) algorithms – SimpleCNN, VGG-16, ResNet-50, and MobileNet-V2 – and proposed strategies to mitigate this bias. A quantitative experimental survey was conducted, using the RAF-DB databases for training and validation, and FairFace for bias evaluation. Performance and fairness metrics were applied, considering gender, ethnicity, and age, and the mitigation strategies Class Weights, Undersampling, and Hybrid were tested. The results indicated that SimpleCNN achieved the best overall performance. In the bias analysis, an overprediction of the “Sad” emotion and significant disparities by gender, ethnicity, and age were observed, especially in children. The Class Weights strategy proved to be the most effective in reducing the prediction of “Sad” and gender disparity, although it increased age disparity. It was concluded that the work contributed to the development of a more equitable emotion recognition model, but highlighted technical limitations related to dataset balancing in computer vision.

Keywords: Deep learning; Fairness; Facial emotion recognition; Algorithmic bias.

1. Introduction

Understanding human emotions represents a complex challenge for artificial intelligence models. The same person can express different emotions through varied facial patterns, such as crying, which can indicate both happiness and sadness. This duality suggests that classification using a multi-label, probability-based approach may be more effective than assigning a single emotional class. According to Barsoum et al. (2016), the use of image banks with label distribution demonstrates greater efficiency compared to the classification of a single emotion per image.

The complexity of emotional recognition is amplified by the vast diversity of social groups, which encompasses differences in race, age, ethnicity, and gender. Deep Learning models, although promising, can exhibit algorithmic bias, compromising the accuracy of emotion recognition. This requires models to not only learn facial patterns but also identify the origin of these patterns across different demographic groups.

The participation of minority social groups in the development of artificial intelligence models is limited, especially in the creation of datasets for training in natural language processing and facial recognition (Castaneda et al., 2022). This under-representation can lead to high-risk decisions by these technologies, exacerbating negative social impacts and perpetuating inequalities.

Given this scenario, the creation of more balanced and equitable image banks is fundamental to reducing algorithmic bias and developing fairer technologies (Akter et al., 2021; Barsotti and Koçer, 2024). The recognition and interpretation of emotions are challenged by cultural diversity and emotional subjectivity, and even advanced solutions still present limitations (Wang et al., 2023). The absence of diversity in datasets is a cause of systematic errors, disproportionately affecting racial minorities and women, which makes transparency and representativeness in these systems indispensable (Kim et al., 2024).

Deepening the understanding of the influence of algorithmic bias on the accuracy of emotion recognition models and identifying effective strategies to develop more inclusive and fair systems is, therefore, essential. Emotion recognition is relevant in various areas, but algorithmic bias can lead to misinterpretations and negative impacts on automated decisions. The lack of fairness in the representation of groups compromises the reliability of these technologies. Thus, this study sought to contribute to bias mitigation, proposing an artificial intelligence model that is fairer, more representative, and diverse in classifying facial expressions of gender, ethnicity, and age groups, using Deep Learning architectures to improve fairness and accuracy in emotion recognition.

2. Material and Methods

This study was characterized by an experimental, quantitative research. Image manipulation and database diversity were employed for numerical evaluations, aiming at the detection of algorithmic bias in both the models and the datasets used for emotion recognition. For image processing, “Deep Learning” techniques were adopted, particularly Convolutional Neural Networks (CNNs), due to their efficiency in identifying visual patterns and their robust performance in the training, validation, and testing stages.

It was considered that emotional recognition models may present systematic confusions between certain emotion classes, given the subtleties in facial expressions. It was investigated whether these limitations manifested unequally among distinct demographic groups. To this end, after training and validating the models, an analysis with confusion matrices was conducted to determine if the model erroneously classified emotions, reaffirming algorithmic bias.

Additionally, the “FairFace” database was used for the neutral testing phase of the models. The objective was to verify the main hypothesis of this research: the detection and quantification of algorithmic bias in emotion recognition tasks, especially among different ethno-racial groups (Kärkkäinen and Joo, 2019).

For data collection, the widely studied RAF-DB (Real-world Affective Faces Database) was used to ensure representativeness in the final set and mitigate possible algorithmic biases (Li & Deng, 2019). The initial set consisted of 25,158 images of faces in an uncontrolled environment, divided into seven basic emotions: “angry”, “disgust”, “fear”, “happy”, “neutral”, “sad”, and “surprise”. The images were annotated by approximately 40 independent raters, who determined the final “label” by majority vote.

An additional processing step was applied to ensure data quality, using a “Multi-task Cascaded Convolutional Networks” (MTCNN) (Zhang et al., 2016) facial detector with a confidence threshold of 90%. This detection removed partially cropped faces, low-quality images, multiple faces, or incorrect classifications. After filtering, the final image bank totaled 15,480 images, centered and standardized to 224×224 pixels.

The dataset split was performed stratified by class, allocating 70% of the images for training (10,829 images), 10% for validation (1,552 images), and 20% for testing (3,099 images). This process occurred in two stages: first, 20% was set aside for testing, and the remaining set was divided into 87.5% for training and 12.5% for validation, ensuring the final 70/10/20 proportion.

During the conversion of images into tensors and the application of pre-processing, the training set was reduced by 22.9%, totaling 8,345 images, due to the removal of inconsistencies in normalization or processing. The validation (1,152 images) and test (3,099 images) sets remained unchanged, resulting in a final database of 12,996 images in tensor format, distributed in proportions of 64.2% for training, 11.9% for validation, and 23.8% for testing.

The “FairFace” dataset (Kärkkäinen & Joo, 2021) was used exclusively during the neutral testing phase of the models, with 10,954 images. This dataset has clear labels for ethnicity, age, and gender, distributed among: (i) Gender: “Female” (47.1%) and “Male” (52.9%); (ii) Ethnicity: 7 groups (“Black”, “East Asian”, “Indian”, “Latino”/”Hispanic”, “Middle Eastern”, “Southeast Asian”, “White”); (iii) Age: 9 age ranges (0-2, 3-9, 10-19, 20-29, 30-39, 40-49, 50-59, 60-69, 70+).

The image preprocessing consisted of essential steps to ensure that the data were prepared for the “machine learning” model training and evaluation pipelines. Initially, the database was divided into training and testing; subsequently, the training set was subdivided into training and validation, with the prediction data used separately to evaluate algorithmic bias.

To mitigate the natural imbalance of the image set, a balancing with a hybrid strategy was applied, based on three levels of sample quantity per class. Over-represented classes (> 2,500 images) were subjected to reduction techniques to 1,500 images, while intermediate (700-2,500 images) and under-represented classes (< 700 images) were kept entirely. This strategy preserved 50% of the original data.

Next, “data augmentation” was performed, a technique for artificially expanding the dataset through transformations on the original images, also by class. The augmentation reached approximately 2,000 images per class, with factor variations between 1.3 and 4.9 times, totaling 13,944 balanced training images. The transformations included random rotation (±15°), horizontal flip (50% probability), zoom, brightness and contrast adjustment (±20°), saturation (±20°), random crop with resize (scale 0.8-1.0), and affine transformations (translation ±10%). To avoid excessive distortions, the transformations respected a maximum limit of five per original image, randomly and combined.

The images were standardized to dimensions of 224×224 pixels, following an input pattern for modern CNNs. They were converted into tensors with normalization on a scale of 0 to 1, adjusting the color channels to “Red”, “Green”, “Blue” (RGB). The transformations were applied using the “Image” modules from the “PIL” library and “transforms” from the “torchvision” library. The “data augmentation” process was applied exclusively to the balanced training sets, aiming to reduce “overfitting” and improve model generalization.

The present study adopted the Python language with the “PyTorch” framework (Paszke et al., 2019), in addition to auxiliary libraries such as “torchvision”, “PIL”, “NumPy”, Pandas, and “Matplotlib” for data preprocessing and visualization. To mitigate the limitation of emotion-labeled databases, models were trained with “transfer learning” and pre-trained weights on “ImageNet” (Wang et al., 2023). Custom classification was performed by replacing the final layer (“CustomHead”) of the architectures, adapting it to the seven classes of the problem.

Four types of CNN architectures were used: SimpleCNN (“baseline”), ResNet-50, MobileNet-V2, and VGG-16, all with weights pre-trained on “ImageNet”, except for SimpleCNN. The Simple CNN has two convolution layers adjusted for three channels (RGB), followed by dense layers, batch normalization (Ioffe & Szegedy, n.d.), “ReLU” activation functions, and “pooling” layers, to evaluate learning with low complexity (Krizhevsky et al., 2017; Lecun et al., 1998).

ResNet-50 offers lower computational cost and good generalization capability, based on the work “Deep Residual Learning for Image Recognition” (He et al., 2016; PyTorch, 2017b). The VGG-16 model is characterized by being deep, with 16 layers and 3×3 convolutional filters, capturing hierarchical “features” consistently (Simonyan et al., 2014; PyTorch, 2017c). MobileNet-V3, created for application on mobile devices or those with limited resources, uses depthwise separable convolutions, reducing the number of parameters and computational operations (Howard et al., 2019; PyTorch, 2017a).

To ensure organization, traceability, and reproducibility in experiments, standardized parameters were defined through configuration files. The configurations applied to all training sessions included 50 epochs, “Stochastic Gradient Descent” (SGD) optimizers for Simple CNN and “Adam” for models with “transfer learning” (Hospodarskyy et al., 2024), and “CrossEntropyLoss” as the loss function (Bhagat and Bakariya, 2025). Additionally, “Early stopping” with a patience of seven epochs, “Learning Rate” of 0.001 for Simple CNN and 0.001 and 0.0001 for other models, “Batch Size” of 64, and “Momentum” of 0.9 were used.

Experiments were conducted under different scenarios, evaluating the models’ performance according to the strategies of balancing, partitioning, and applying “data augmentation”. The experiments followed the “Holdout” approaches and, immediately after, “K-Fold” validation. Firstly, the four architectures were trained with the balanced training set, in addition to being validated with the unbalanced validation set.

The best performing model during validation was selected and submitted to “K-Fold” cross-validation (k=5) with the complete training image bank, to obtain the final statistics. The trained models were saved and evaluated on the neutral test set, allowing for final generalization evaluation.

To validate the efficiency of the models regarding learning and the capacity for recognizing emotional patterns, analyses of quantitative metrics and visualizations of the graphs generated during training and validation were performed (Mukhlif et al., 2022; Nurunnabi et al., 2022; Plested and Gedeon, 2022). The main metrics analyzed were Accuracy, F1-Score per class, Precision, Recall, Training and Validation Loss, Confusion Matrix, and Fairness.

For the calculation of the metrics, the “Scikit-Learn” library was used, highlighting the functions “accuracy_score”, “f1_score”, “confusion_matrix”, and “ConfusionMatrixDisplay”. The combination of these metrics offered a comprehensive view of performance and supported a critical analysis of the models’ robustness. To identify recurring error patterns among the emotional classes, confusion matrices were analyzed, which revealed that subtleties in facial expressions can lead to ambiguous interpretations.

As the main objective, the presence of algorithmic bias related to demographic markers, especially ethnic-racial groups, was investigated. To verify this issue, the “FairFace” image bank was used as a neutral test set, as it has clear labels for ethnicity, age, and gender. This “dataset” was used exclusively in the prediction phase, ensuring that the models did not view the images during training and validation.

To quantify possible algorithmic inequalities, two “fairness” metrics were applied (Dulani, 2023; Lee, 2023). The first, Demographic Parity, investigated whether the rate of positive predictions is similar between different demographic groups, expressed by the equation P(Ŷ = 1 |A = 0) ≈ P(Ŷ = 1|A = 1), where A represents the sensitive attribute (ethnicity) and Ŷ the model’s prediction.

The second metric, “Equalized Opportunity”, assessed whether the true positive rate is consistent across groups, ensuring that individuals from different groups with the same true condition had similar probabilities of correct classification, calculated by the equation P(Ŷ = 1 |Y = 1,A = 0) ≈ P(Ŷ = 1|Y = 1, A = 1), where A represents the sensitive attribute, Ŷ the model’s prediction, and Y the true condition. These comparisons allowed for the identification of disparities in model performance.

After preliminary analyses of algorithmic bias, three strategies were applied to mitigate bias within the database. The first was “Class Weights”, which assigned differentiated weights to classes according to frequency and error patterns, using “CrossEntropyLoss” as the “weight” parameter in “PyTorch”. The second strategy was “Undersampling”, which limited the maximum number per class within the training set to 2000 samples, performing stratified random selection.

The third strategy, “Hybrid” (Híbrida), combined the two previous techniques to maximize the benefits of each approach. The strategies were performed only on the chosen model that obtained the best performance, keeping the hyperparameters constant, and resulting performance and “fairness” metrics for the analysis of algorithmic equity.

In line with the topic “Materials and Methods”, a GitHub repository was created, which details each algorithm developed that is of great importance to the project and complements the training with the proposed models. The ‘tcc_mba’ repository is available at: https://github.com/crisCintia/tcc_mba.git.

3. Results and Discussion

The performance evaluation of convolutional neural network (CNN) architectures in facial emotion recognition revealed significant differences. Four models were compared: SimpleCNN, VGG-16, ResNet-50, and MobileNet-V2, using the RAF-DB dataset, which contained 4,960 images for testing. The results indicated that SimpleCNN achieved the best overall performance, reaching an accuracy of 70.7% and a Macro F1-Score of 62.03%. This model outperformed the other architectures, which presented lower accuracies and F1-Scores, such as VGG-16 with 66.5% accuracy and 56.51% Macro F1-Score, ResNet-50 with 51.1% accuracy and 40.83% Macro F1-Score, and MobileNet-V2 with 50.2% accuracy and 39.43% Macro F1-Score.

The superiority of SimpleCNN, which was trained from scratch with approximately 500 thousand trainable parameters, can be attributed to its ability to learn the dataset’s specific patterns more effectively, resulting in less “overfitting”. In contrast, the VGG-16, ResNet-50, and MobileNet-V2 models, which employed “transfer learning” techniques with weights pre-trained on ImageNet, showed inferior performance. This difference is explained by “domain mismatch”, where the transferred low-level features were insufficient to capture the subtleties of facial expressions, and by the freezing of 90% of the “backbone”, which limited the learning of emotion-specific patterns. Deeper models, such as ResNet-50 (50 layers) and MobileNet-V2 (53 layers), may suffer from “overfitting” on small to medium-sized databases like RAF-DB, as pointed out by Ahmad et al. (2025) and Ashwini et al. (2023).

The analysis of the F1-Score by emotion, a crucial metric for imbalanced datasets, highlighted that the “Disgust” emotion was the most difficult to classify, with an average F1 of 18.8% among the models. In contrast, the “Happy” emotion proved to be the easiest to classify, with an average F1 of 72.8%. SimpleCNN demonstrated superior performance, recognizing six out of the seven emotions with higher accuracy and tying with VGG-16 only in the classification of the “Angry” emotion, with an F1-Score of 0.572. This discrepancy in performance between classes, especially between “Happy” and “Disgust”, is consistent with the imbalance of the RAF-DB dataset, where the class with the largest sample representation tends to be better predicted.

Deepening the analysis of error patterns, row-normalized confusion matrices revealed the classes most frequently confused by each model. For SimpleCNN, the highest accuracies occurred for the “Happy” emotion, with 902 correct predictions (79.2% recall), and “Neutral”, with 532 correct predictions (76.1% recall). The main confusions identified were the “Sad” emotion being confused with “Neutral” in 106 cases (22.4%), “Happy” with “Neutral” in 89 cases (7.8%), and “Disgust” with “Neutral” in 38 cases (19.9%). These patterns indicate that, despite theoretically distinct facial features, the models presented inherent limitations in classifying emotions with subtleties.

The VGG-16, ResNet-50, and MobileNet-V2 models also exhibited confusion patterns. VGG-16, for example, had 846 correct predictions for “Happy” (74.3% recall) and 562 for “Neutral” (80.4% recall), but confused “Happy” with “Neutral” in 182 cases (16%) and “Sad” with “Neutral” in 125 cases (26.4%). ResNet-50 and MobileNet-V2 showed lower total correct predictions and more pronounced confusions, such as “Happy” with “Neutral” in 236 cases (20.7%) for ResNet-50 and in 282 cases (24.8%) for MobileNet-V2. These universal confusions, present in all models, suggest that the overlap of facial “features”, such as the absence of a smile and low intensity between “Sad” and “Neutral”, or subtle smiles in “Happy” interpreted as “Neutral”, hinder precise classification (Abbas & Chalup, 2019; Ashwin & Biswas, 2024).

The difficulty in classifying the emotion “Disgust” was multifaceted. Firstly, the expression of “Disgust” is subtle, involving a wrinkling of the nose and a raising of the upper lip. Secondly, its sample representation in the RAF-DB dataset was low, with approximately 717 training examples compared to 4,772 for “Happy”. Finally, the similarity of its facial tensions with the emotion “Angry” contributed to the confusion. The high variability of the F1-Score for “Disgust” among models (0.424 for SimpleCNN versus 0.044 for ResNet-50) confirmed the difficulty in its generalized classification, a finding consistent with the literature addressing the complexity of facial expressions and the limitations of databases (Xu et al., 2020).

In contrast, the “Happy” emotion was easier to classify due to its distinct and pronounced facial features, such as the smile, lip curvature, and cheek elevation. This resulted in F1-Scores consistently above 0.60 across all models. The SimpleCNN presented a directional classification pattern, where “Sad” was confused with “Neutral” and vice versa, while the “Transfer Learning” based models exhibited more balanced bidirectional confusions, as in ResNet-50, which confused “Neutral” with “Sad” in 121 cases and “Sad” with “Neutral” in 118 cases. This behavior suggests that SimpleCNN, despite its good overall performance, may have struggled to learn robust features for these expressions due to possible overfitting on a limited-size dataset.

To validate the robustness of the SimpleCNN model, a K-Fold cross-validation with k=5 was performed, using the balanced and augmented RAF-DB dataset. The results demonstrated an average accuracy of 69.9% and an average F1-Score of 60.3%, values very close to those obtained in the “Holdout” training (70.7% accuracy and 62.03% F1-Score). The low standard deviations observed, 1.24% for accuracy and 2.45% for F1-Score, with coefficients of variation of 1.7% and 4.1% respectively, confirmed the stability and robustness of the model. This indicated that SimpleCNN was not excessively dependent on a specific training sample, validating its choice for the algorithmic bias analysis in the FairFace dataset.

The analysis of algorithmic bias, using the FairFace dataset, which has annotations for ethnicity, age, and gender, but not for emotions, allowed us to investigate whether classification patterns manifested unequally among different demographic groups. An overprediction of the “Sad” emotion was observed, representing 40.2% of the overall predictions, a value significantly higher than the expected prediction of 14.3% for a random distribution among seven classes. In contrast, the “Disgust” emotion was underpredicted, with only 1.1% of classifications, evidencing a failure of the model to identify this emotion.

Regarding gender bias, disparities were relatively low, with a maximum difference of 4.5% in the prediction of the “Neutral” emotion. It was noted that women were more frequently classified as “Sad” (41.5%), while men were more classified as “Neutral” (31.5%). The analysis of ethnic bias revealed moderate disparities, with a maximum difference of 6.3% between the “Sad” and “Neutral” emotions. Latino/Hispanic and Middle Eastern individuals were more classified as “Sad”, with 42.5% and 36.1% respectively. Indians (33.9%) and Latino/Hispanics (27.6%) were more classified as “Neutral”. These findings suggest a sensitivity of the model to facial morphological variations, resulting in high error rates for individuals with darker skin tones and non-Caucasians (Kim et al., 2021; Singhal et al., 2025).

The age bias showed the largest disparities, being particularly pronounced in children aged 3 to 9 years, who were classified as “Sad” in 46.7% of cases. In comparison, seniors over 70 years old were classified as “Sad” in 37.3% of cases. This trend indicates that the model interpreted children’s facial expressions as sadness in most situations, which can be explained by the difference in children’s facial patterns compared to adults, as discussed by Y. Huang et al. (2024). The systematic overprediction of the “Sad” emotion (40.2%) was observed across all analyzed demographic groups, reinforcing the presence of an inherent bias in the model or the training dataset.

To mitigate algorithmic bias, three strategies were tested: “Class Weights”, “Undersampling”, and “Hybrid”. The “Class Weights” strategy proved to be the most effective, promoting a 26% reduction in the over-prediction of the “Sad” emotion, decreasing from 40.21% to 29.70%. Furthermore, this technique reduced gender disparity by 51% (from 4.52% to 2.19%) and ethnic disparity by 10% (from 6.31% to 5.69%). Although there was a small “trade-off” in overall performance, with a decrease of 1.1 percentage points in accuracy and 2.4 percentage points in F1-Score, the gains in bias reduction were significant. However, age disparity increased by 82% (from 9.4% to 17.1%), indicating the need for more specific approaches for this attribute.

In contrast, the “Undersampling” technique resulted in a 54% increase in the overprediction of the “Sad” emotion, raising it to 61.98%, and worsened the performance and “fairness” metrics. This behavior can be explained by the reduction of the training set, which can lead to insufficient learning of patterns, causing the model to converge to the majority classes and increase bias, according to studies by Buda et al. (2018) and Fernando & Tsokos (2022). The “Hybrid” strategy, which combined “Class Weights” and “Undersampling”, presented intermediate results. Although it reduced the overprediction of “Sad” by 21% and gender disparity by 49%, there was a 29% increase in ethnic disparity and a 97% increase in age disparity, suggesting that the interaction of the techniques may have generated local “overfitting”, harming generalization to specific demographic subgroups.

The comparative analysis between the baseline model and the bias mitigation strategies revealed that “Class Weights” offered the best balance, maintaining performance close to the original model while providing substantial gains in reducing gender bias and the over-prediction of “Sad”. These results align with the literature demonstrating the effectiveness of weight adjustment techniques in imbalanced datasets (Fernando & Tsokos, 2022). However, age bias, especially concerning the classification of children as “Sad”, persisted and even increased across all strategies, highlighting that infant facial characteristics represent a critical challenge for emotion recognition models, which are predominantly trained with adult data.

In summary, the study demonstrated that SimpleCNN outperformed pre-trained models in facial emotion recognition, but revealed a significant algorithmic bias, characterized by the overprediction of the “Sad” emotion and notable disparities concerning age, especially in children. The “Class Weights” strategy proved to be the most effective in reducing gender bias and the excessive prediction of “Sad”, although it increased age disparity. These findings highlight the importance of considering fairness in the development of artificial intelligence models and point to the need for future research focused on dataset balancing and intersectional approaches to mitigate specific demographic biases.

4. Conclusion

The study sought to evaluate the performance of convolutional neural network algorithms in facial emotion recognition and to propose strategies for mitigating algorithmic bias related to gender, ethnicity, and age. It was found that SimpleCNN, an architecture trained from scratch, achieved the best overall performance, with an accuracy of 70.7% and a Macro F1-Score of 62.03%, outperforming pre-trained models that exhibited “domain mismatch” and “overfitting”. However, bias analysis revealed an over-prediction of the “Sad” emotion in 40.2% of classifications and an under-prediction of “Disgust”, with notable disparities by age, especially in children. The “Class Weights” strategy proved to be the most effective in reducing the over-prediction of “Sad” by 26% and gender disparity by 51%, even with a small “trade-off” in overall performance. These findings contribute to the development of more equitable and representative artificial intelligence models in facial expression classification.

Despite the advances, the study highlighted technical limitations related to dataset balancing in computer vision, as age bias, particularly the classification of children as “Sad”, persisted and even increased with the tested mitigation strategies. The lack of age uniformity in the RAF-DB dataset and the imbalance of emotional classes represented significant challenges. It is suggested that future research investigate bias mitigation strategies in a multidimensional way and focus on creating more balanced emotion datasets, incorporating intersectional approaches to reduce more specific demographic biases, such as age and ethnic-racial bias.

Bibliographic References

Akter, S.; Dwivedi, Y.; Biswas, K.; Michael, K.; Bandara, R.; Sajib, S. 2021. Addressing algorithmic bias in Al-driven customer management. Journal of Global Information Management, 29 (6): 1–27. Disponível em: https://doi.org/10.4018/JGIM.20211101.oa3.

Barsotti, F.; Koçer, R. G. 202

Barsoum, E.; Zhang, C.; Ferrer, C. C.; & Zhang, Z. 2016. Training deep networks for facial expression recognition with crowd-sourced label distribution. In: Proceedings of the 18th ACM International Conference on Multimodal Interaction, 2016, Tokyo, Japão, Anais p. 279-283. Disponível em: https://doi.org/10.1145/2993148.2993165.

Castaneda, J.; Jover, A.; Calvet, L.; Yanes, S.; Juan, A. A.; Sainz, M. 2022. Dealing with gender bias issues in data-algorithmic processes: a social-statistical perspective. Algorithms, 15(9): 303. Disponível em: https://doi.org/10.3390/a15090303.

Article originating from the Final Course Work of the Specialization in Data Science and Analytics of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform

Most recent

October 22, 2026

Challenges and Possibilities of the First School Year of Children with ASD for Teaching Work

October 22, 2026

Consumer Journey in the Decision to Purchase Infoproducts in the Digital Environment

October 22, 2026

Psychosocial Risks and Mental Health at Work in a Banking Institution

October 22, 2026

Digital Observatory of the Amazon: Data Science Applied to the Analysis of Social, Economic, Environmental, and Health Indicators

October 22, 2026

Continuous Learning as a Factor of Organizational Growth in the Perception of Collaborators

October 22, 2026

Statistical Regression in Microgrids for Operational Performance Estimation: Campusgrid Case

October 22, 2026

Forecast of Aggregate Default in the Brazilian Financial System by Sarimax Model with Macroeconomic Variables

October 22, 2026

Stakeholder Engagement Factors in International Technology Transfer to a Public Hospital

October 22, 2026

Combating Commercial Electric Energy Losses Using a Time Series Model

The inclusion of children with Autism Spectrum Disorder (ASD) in Early Childhood Education presents significant challenges, given the difficulty of adapting to new environments and the intensity of stimuli. The objective was to analyze the teaching performance in Early Childhood Education in the initial schooling process of children with ASD. A case study was conducted, using narrative interview as a data production technique. The results revealed that the teaching performance was characterized by mediation and affection practices, with the flexibility of routine and the use of playful resources as essential strategies for the child’s development and sensory organization. The importance of the school support professional for the success of the pedagogical work was also highlighted. The child’s integral development was promoted through intentional teaching actions that prioritized environmental adaptation, strengthening of bonds, and attention to the specific needs of ASD. As a contribution, a guide of pedagogical strategies aimed at welcoming children with ASD in public daycare centers was developed.

Keywords: Early Childhood Education; School Inclusion; Teacher Narratives; Inclusive Pedagogical Practices; Autism Spectrum Disorder.

October 22, 2026

Consumer Journey in the Decision to Purchase Infoproducts in the Digital Environment

The expansion of e-commerce and technological transformation have driven the need to understand consumer behavior regarding infoproducts, whose perceived experience has become crucial for professional development. The study aimed to analyze the factors that influenced the purchase decision of infoproducts in the digital environment, considering the stages of the consumer journey. An applied, quantitative, and descriptive research was conducted through a field survey with 100 consumers from Serra Gaúcha. Data collection occurred via a structured questionnaire, using non-probabilistic sampling. The results revealed a qualified audience with strong “mobile-first” behavior. It was found that the availability of detailed information and social approval were the main attraction triggers in the pre-purchase phase. In the transactional and subsequent stages, the minimization of operational friction, the agility of immediate access, and system security proved crucial for conversion and satisfaction, although gaps in human support efficiency were identified. Retention was based on the trust generated by positive past experiences, overcoming the price factor. It was concluded that the infoproduct consumer journey operates as an integrated system, requiring alignment between technological infrastructure, content curation, and support humanization to ensure engagement.

Keywords: E-commerce; Consumer behavior; Consumer experience; Loyalty; Virtual business.

October 22, 2026

Psychosocial Risks and Mental Health at Work in a Banking Institution

Mental health in the workplace is a topic of growing relevance, impacting professional development and employee well-being. A study was conducted to identify aspects and situations that can influence absenteeism levels in a banking institution. The research involved 108 participants and used a structured questionnaire, based on the short version of COPSOQ II by Gonçalves (2021), which contained 45 closed Likert scale questions and two open-ended questions. Data were collected via an online platform and analyzed with a descriptive and quantitative approach. The results revealed a scenario that indicated the need for adjustments in the distribution of activities and in managers’ perception of behaviors and signs of disorders. Important evidence was identified in the participants’ open-ended responses. It was concluded that the study contributed to promoting the perception of a safe environment for mental health in the organization.

Keywords: Absences due to disorders; Professional development; Operational leadership; Medical leave; Mental health.

October 22, 2026

Digital Observatory of the Amazon: Data Science Applied to the Analysis of Social, Economic, Environmental, and Health Indicators

The state of Amazonas faced historical challenges in consolidating the Unified Health System (SUS), stemming from geographical barriers, socioeconomic inequalities, and structural limitations. This study described the development and implementation of a health information geoportal, designed to support monitoring, evaluation, and planning in the Amazonian context. A descriptive case study approach was adopted, based on documentary analysis and secondary data, covering the mapping of functionalities, technical infrastructure, and validation steps of the tool. The Portal was developed with open technologies and integrated multiple national databases, offering interactive dashboards, thematic maps, and a digital document repository. Among the results, highlighted were the georeferenced visualization of key indicators, the possibility of temporal comparative analyses, and the use of the platform in reviewing health plans and redirecting strategic actions. The portal can represent an advance in digital governance and evidence-based management, with potential for replication in other highly complex territories, provided it is accompanied by investments in technological infrastructure, professional training, and inter-institutional coordination.

Keywords: Amazon; Health Evaluation; Digital Health; Health Information Systems.

October 22, 2026

Continuous Learning as a Factor of Organizational Growth in the Perception of Collaborators

The impact of continuous learning and knowledge management on organizational growth was studied in a context of globalization and constant changes. The study’s objective was to analyze employees’ perception of learning culture and its importance for personal and professional development. Data were obtained through an anonymous survey, with a quantitative approach, using a form with “survey” type questions applied to 87 professionals from the technology, health, education, commerce, and services sectors. The results indicated that employees had a positive perception of the contribution of continuous learning to personal and organizational growth, recognizing its benefits for the development of new skills, innovation, and competitiveness. However, it was observed that the incentive for internal knowledge exchange and the offering of practical training still presented moderate levels of agreement. It was concluded that continuous learning significantly contributes to individual and organizational development, suggesting that companies should develop strategic policies for learning focused on new skills, aiming for greater competitive advantage and innovation.

Keywords: Continuous learning; Self-development; Organizational growth; Knowledge Management; Innovation.

October 22, 2026

Statistical Regression in Microgrids for Operational Performance Estimation: Campusgrid Case

A multiple statistical regression model was proposed to estimate the operational performance of UNICAMP’s CampusGrid microgrid, based on internal electrical quantities. The microgrid integrates photovoltaic generation, batteries, and a gas generator. Energy management system data were collected at five-minute intervals over one month. The Operational Reliability Index, defined as the ratio between supplied and demanded energy, was established as the dependent variable. The model was adjusted for the islanded operation mode, evaluating the impact of demand, state of charge, solar generation, and supplied energy. The results revealed high explanatory power, indicating that operational reliability depends on the interaction between supply and demand. It was observed that the nighttime period concentrated the highest rate of partial supply, attributed to the reduction in solar generation. It was concluded that multiple regression offers an interpretable and useful framework for microgrid monitoring and control strategy support.

Keywords: Battery storage; Photovoltaic generation; Reliability Index; Islanded operation; Energy Management System.

October 22, 2026

Forecast of Aggregate Default in the Brazilian Financial System by Sarimax Model with Macroeconomic Variables

Aggregate non-payment in the Brazilian financial system is a crucial indicator of credit risk, the prediction of which is regulated by CMN Resolutions n.° 4.557/2017 and 4.966/2021 for the prospective measurement of expected losses. Aggregate non-payment was modeled and predicted using the SARIMAX model, projecting its trajectory until December 2026 under optimistic, base, and pessimistic scenarios. 164 monthly observations (July/2012 to February/2026) and macroeconomic variables such as unemployment, Selic, IPCA, IBC-Br, and household debt were used. The specification with contemporaneous variables was selected based on information criteria and coefficient stability, and residual diagnostics confirmed statistical adequacy, with a MAPE of 3.12%. Historical non-payment ranged between 4.33% and 7.64%. The estimation identified unemployment (β = 0.2534; p < 0.001) and Selic (β = 0.2128; p < 0.001) as the main positive determinants, with strong temporal persistence (AR(1) = 0.8965). Selic showed increasing correlation with non-payment (r = 0.62 at lag 12), while unemployment demonstrated a contemporaneous effect (r = 0.33). Projections indicated non-payment between 6.30% and 7.98% in December 2026, with a range of 1.68 percentage points, highlighting direct implications for provisioning in financial institutions.

Keywords: Unemployment; Predictive modeling; Monetary policy; Loss provisioning; Time series.

October 22, 2026

Stakeholder Engagement Factors in International Technology Transfer to a Public Hospital

This study described the critical factors perceived for stakeholder engagement in international technology transfer in a public hospital, with a focus on planning. The objective was to describe these conditioning factors, according to the perception of professionals. A single case study was conducted, applying a questionnaire to 25 professionals and using content analysis, complemented by descriptive tabulations. The main obstacles identified were resistance to change (72%), funding (64%), lack of communication (60%), and bureaucracy (40%). Only 28% of respondents recognized the existence of adequate tools for stakeholder management, while 48% denied it and 24% did not know. Exploratory cross-tabulations indicated variations in the assessment of influence according to time of practice, but did not establish causal relationships. Lack of engagement was associated with delays, low adherence, waste of resources, communication failures, compromised knowledge absorption, and risks to patient safety. The conclusions were limited to the perceptions of the researched group and supported preliminary recommendations for intersectoral communication and coordination.

Keywords: Case study; Project Management; Health innovation; Stakeholders; Public sector.

October 22, 2026

Combating Commercial Electric Energy Losses Using a Time Series Model

Commercial losses of electric energy represented one of the main challenges for distributors, directly impacting economic sustainability and supply quality. The efficiency of time series models, focusing on exponential smoothing techniques, was evaluated in forecasting billed consumption of consumer units, to support the identification of deviations associated with non-technical losses. Historical data of monthly consumption from an electric energy distributor in Brazil were used, segmented by behavior profiles (stationarity, trend, and seasonality). Simple exponential smoothing models, Holt, and Holt-Winters (additive and multiplicative) were applied, compared to a traditional moving average. The Mean Absolute Percentage Error (MAPE) was the evaluation criterion. Additionally, the ETS approach was employed for automatic model selection and residual diagnostic tests, such as Ljung-Box and autocorrelation function analysis. The results demonstrated that more robust models showed a greater capacity for adaptation to the series’ characteristics and lower error, with a highlight on superior performance in scenarios with trend. For seasonal series, the ETS model showed a greater capacity for adaptation to transitions. It was concluded that the use of more structured time series models contributed significantly to the improvement of forecast accuracy and offered relevant analytical support for the identification of commercial losses and the optimization of operational and commercial processes in the electric sector.

Keywords: Energy distribution; MAPE; Optimization; Consumption forecasting; Exponential smoothing.

Economy

October 19, 2026

End of the 6×1 scale: what do the numbers allow us to conclude?

Following the promulgation of the end of the 6×1 shift system, the debate on the shorter workday moves to implementation and requires distinguishing social benefits from economic costs. The end of the 6×1 shift system transitioned from legislative discussion to a constitutional change in Brazil. On October 16, 2026, the National Congress promulgated Constitutional Amendment No. 141/2026, originating from Proposal […]

Artificial Intelligence

October 16, 2026

Why study if AI exists?

When answers become easy, learning comes to mean more than just arriving at them. Why do I need to learn this if I can ask an artificial intelligence? This type of question has accompanied education for decades. Why learn an operation if a calculator exists? Why memorize information if it can be found […]

Tax Management

October 15, 2026

Pink tax and the gender of price

How different marketing treatments for men and women are incorporated by consumption taxation Imagine a shelf on which two razors of the same brand and intended for the same purpose are displayed side by side: one has dark packaging and promises strength, precision, and efficiency; the other comes in shades of pink or […]

Health

October 13, 2026

The missing link in health transformation

Why is innovation not enough to drive changes in organizations and health systems Digital transformation has become a strategic priority for health systems worldwide. Electronic health records, artificial intelligence, telemedicine, remote monitoring, and data analysis are now part of the strategic agenda of governments and health organizations. Paradoxically, […]

Marketing

October 09, 2026

Bitcoin or pets: what Artificial Intelligence would not see

A real-life consumer intelligence case shows what gets left out of the brief A few years ago, an American insurance company hired a specialized consumer intelligence firm to answer three questions: The insurer was also considering launching paid insurance in bitcoin. At the time, the consultancy’s workflow was as follows: the analyst created a query […]

School Management

October 06, 2026

The school is not a casino: when trading cards become currency in the playground

Why, for certain students, does school manage to offer so few experiences of achievement? The World Cup is over. Brazil left the competition, the school holidays came, the final was played and the football calendar moved on. It would be reasonable to imagine that, with the end of the tournament and the return to classes, […] would also begin

Energy Transition

September 30, 2026

Renewable energies: the Brazilian advantage

Brazil needs to invest in the capacity to transform natural resources into innovation, productivity, and development. Talking about energy is talking about development. Behind industry, transportation, services, cities, and increasingly artificial intelligence, there is an energy infrastructure that sustains the economy. Understanding a country’s energy matrix is understanding […]

Data Analysis

September 30, 2026

When the data changes, has reality changed?

An analysis of almost 20 million teaching positions-year between 2020 and 2025 shows how changes in population composition and record production can alter the interpretation of historical series Between 2020 and 2025, the number of formal teaching employment positions identified in RAIS (Annual Relation of Social Information), from the Ministry of Labor, […]

Artificial Intelligence

September 28, 2026

Artificial intelligence and the trends

Why do so many people join in on the fun with ChatGPT? I was never the trend person. I still am not. Some pass through my timeline, I see them, I find them interesting but not to the point of trying them out. This was the case, inclusive, with some recent waves of images made by artificial intelligence, like the 80s trend, the one that […]

Tax Management

September 01, 2026

Tax reform and estate planning in the countryside

Efficiency in managing taxes facilitates governance and allows for the organization of family succession. Rural wealth organization has become a strategic measure in light of the tax reform in Brazil, which came into effect this year. The planning of activities and rural wealth management, through tools such as the transition from an Individual to […]

Renata de Gaspari Valdejão Almeida

Artificial Intelligence

August 05, 2026

Inside the company and off the radar: the risks of Shadow AI

According to Microsoft, 75% of professionals already use artificial intelligence at work without company approval. The use of generative artificial intelligence is already part of the corporate routine. Tools like ChatGPT, Copilot, and other platforms help professionals produce texts, analyze data, create presentations, and accelerate tasks in practically all sectors. But this advance […]

Renata de Gaspari Valdejão Almeida

Strategy

July 02, 2026

The future as a competence

Academic interest reinforces the projection of scenarios as an essential tool for decision-making. While some companies are still trying to get the next step right, others are already working with multiple possible futures — and preparing for all of them. But after all, what does it mean to “think about the future”? Is it a trend, a strategy, or a new competence that […]

Renata de Gaspari Valdejão Almeida

Innovation

Technology

May 21, 2026

Agro SP Corridor: innovation, institutional articulation and future

Pecege commands the executive operation of the 400 km ecosystem activities that became known as the Brazilian agro Silicon Valley In 2025, in addition to taking over the management of the Piracicaba Technology Park, the Pecege Institute was also designated to command the executive operation of the State Agribusiness Innovation Corridor activities, the so-called […]

Education

April 27, 2026

How to turn a TCC into a business

New management of the Piracicaba Technology Park offers consulting to MBA graduates The new project director of the Piracicaba Technology Park (PTP), Alexandre Barreto, often says that one of the greatest rewards he has ever had in life was seeing his MBA USP/Esalq students structure companies from their final projects […]

Technology

March 20, 2026

The human repertoire in the age of AI

Critical thinking and creativity are irreplaceable in the work of a marketing professional Artificial Intelligence (AI) is capable of incredible feats, but it does not yet possess the judgment, empathy, intuition, and strategic creativity necessary to do the job of a good marketing professional. In this area, it is important to know how to use the tool to automate […]

Renata de Gaspari Valdejão Almeida

Agribusiness

Innovation

February 23, 2026

ESG as a growth strategy in a cooperative in RS

Santa Clara balances social and environmental responsibility with innovation Founded in 1911 by Italian immigrants in Carlos Barbosa, in the Serra Gaúcha region, Cooperativa Santa Clara is an example that it is possible to balance economic growth with social and environmental responsibility. From a small cheese production company, the organization has evolved into a […]

Renata de Gaspari Valdejão Almeida

Agribusiness

Innovation

January 22, 2026

Brazil free of Foot-and-Mouth Disease: what changes for Brazilian beef?

With a herd of over 200 million head of cattle, the country is currently the leader in exports. Brazil recently received, for the first time, the certification of country free of foot -and-mouth disease without vaccination, conferred by the World Organisation for Animal Health (WOAH). The seal strengthens the country’s position as a reliable supplier of meat for the […]

Renata de Gaspari Valdejão Almeida

Innovation

December 09, 2025

Innovation: the new cycle of the Piracicaba Technology Park

Instituto Pecege assumes management, betting on the creation of new businesses Starting in 2026, Instituto Pecege will assume the management of the Piracicaba Technology Park (PTP). Among its responsibilities will be the administration, operation, maintenance, and expansion of the park, in addition to fostering activities in innovation, entrepreneurship, and science and technology. The choice was made […]

Renata de Gaspari Valdejão Almeida

July 06, 2026

Large-Scale Architecture: DSR-SCRUM Change Management

Mônica Alessandra Guerios; Maria do Carmo Assis Todorov Abstract generated by ResumeAI, an artificial intelligence solution developed by Instituto Pecege focused on synthesis and writing. Architecture and urbanism are configured as essential bases in an increasingly complex and challenging urban space, where the intrinsic connection between social, environmental, and technological issues […]