Abstract
Predictive modeling in social network advertising often suffers from data scarcity and class imbalance, leading to suboptimal classifier performance. This study evaluates the efficacy of generative data augmentation techniques—specifically Gaussian Mixture Models (GMMs), Variational Autoencoders (VAEs), and Generative Adversarial Networks (GANs)—in enriching small datasets. By synthesizing 200 additional user records for a 400-user dataset, we demonstrate significant improvements across various classification models. Our results show that GAN-based augmentation achieved the highest overall performance boost, increasing the Area Under the Curve (AUC) from 0.84 to 0.94.
Introduction
In digital marketing, the ability to accurately predict user responses to advertisements is critical for optimizing budget allocation and enhancing user experience. However, real-world social network datasets are frequently limited in size or exhibit heavy class imbalance, where positive responses (clicks) are rare compared to negative ones. Such "thin" data often results in machine learning models that fail to generalize effectively to unseen data. This research investigates whether manufacturing high-quality synthetic data can mitigate these issues and produce more robust predictive classifiers.
Methodology
The study employs a systematic three-phase framework consisting of Data Augmentation, Training, and Testing.
Data Augmentation Techniques
Three generative approaches were compared to expand the initial training set:
- Gaussian Mixture Models (GMMs): A statistical method modeling data as a mixture of multiple Gaussian distributions.
- Variational Autoencoders (VAEs): A deep learning architecture that learns a latent representation of the data to sample new instances.
- Generative Adversarial Networks (GANs): A framework where a generator (implemented as a VAE in this study for stability) competes against a discriminator to produce realistic synthetic data.
Experimental Setup
The primary dataset comprised 400 user records reflecting social network behavior. After standard preprocessing (cleaning, normalization, and encoding), each generative model produced 200 synthetic instances, increasing the training pool by 50%. Six classifiers—Decision Trees, K-Nearest Neighbors (KNN), Logistic Regression, Support Vector Machines (SVM-Rbf and SVM-Linear), and Dense Neural Networks—were then trained and evaluated on a held-out portion of the original real-world data.
Results and Discussion
The integration of synthetic data led to substantial quantitative improvements across nearly all evaluated metrics.
Performance Gains
As shown in the ROC analysis, the baseline model (no boost) achieved an AUC of 0.84. Data augmentation significantly improved this metric, with GAN-based augmentation reaching an AUC of 0.94. Specific model improvements included:
- Decision Trees: Accuracy rose from 0.85 to 0.94 with GAN augmentation.
- Dense Neural Networks: Accuracy increased from 0.88 to 0.93 using VAEs.
- Class Imbalance Mitigation: Logistic Regression, which initially failed to predict the minority class (F1 score of 0), saw its F1 score rise to 0.36 after augmentation.
Comparative Insights
The study revealed that different augmentation methods favor different classifier architectures. VAEs were particularly effective for Dense Neural Networks, likely due to their ability to capture complex underlying distributions in structured data. Conversely, GANs provided the best boost for logic-based models like Decision Trees, potentially because the adversarial training process better defines the sharp decision boundaries these models utilize.
Conclusion
This research confirms that generative data augmentation is a viable strategy for overcoming data scarcity in social network ad prediction. While GANs and VAEs provided the most significant gains, the choice of augmentation method should be tailored to the specific classifier being used. Future work should explore the scalability of these techniques across diverse social media platforms and investigate the ethical implications of synthetic data regarding fairness and bias in automated advertising systems.
Summary updated