NONLINEAR DIMENSIONALITY REDUCTION FOR LOOKALIKE AUDIENCE DETECTION USING MANIFOLD LEARNING AND AUTOENCODER-BASED REPRESENTATIONS

Abstract

Identifying users with similar behavioral characteristics is a critical task in modern targeted advertising and customer analytics systems. High-dimensional tabular datasets describing user activity often contain complex nonlinear relationships that cannot be effectively captured by traditional linear dimensionality reduction techniques. This study investigates representation learning approaches for constructing scalable look-alike audience detection systems using large-scale telecommunications data. Classical dimensionality reduction techniques, including Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE), are first analyzed as baseline methods for exploring the structure of high-dimensional data. While PCA performs linear projections that preserve global variance and t-SNE reveals local neighborhood structures through nonlinear embedding, these methods are primarily designed for visualization and exploratory analysis and do not provide scalable parametric mappings for new data samples. To address these limitations, a representation learning framework based on autoencoders is proposed for generating compact latent embeddings of users. The model is trained on a large-scale anonymized telecommunications dataset containing behavioral, demographic, device-related, and service usage attributes. Embeddings are learned for multiple feature entities and concatenated into a unified user representation that integrates heterogeneous behavioral information. User similarity is then computed using cosine similarity in the latent space, enabling efficient identification of look-alike audiences. The proposed system is evaluated using clustering metrics and multiple independent validation tasks with external target variables to ensure unbiased performance estimation. Experimental results demonstrate that autoencoder-based embeddings produce a more structured latent space and improve both similarity-based retrieval and downstream classification performance compared to classical dimensionality reduction techniques. The findings highlight the effectiveness of deep representation learning for high-dimensional tabular data in real-world recommendation and targeted advertising systems.