Technical Summary: AI-Driven Client Acquisition in Social Media

Problem Statement

Despite the widespread adoption of artificial intelligence in marketing, particularly for customer retention and recommendation systems, there is a significant gap in research regarding the use of AI for automated client acquisition. Existing approaches often rely on manual targeting parameter adjustments or lack the flexibility to process the dynamic, unstructured, and imbalanced data typical of social media environments. Furthermore, comprehensive pipelines that integrate data extraction, target audience identification, content analysis, and automated content generation for new customer acquisition remain underexplored.

The primary challenge addressed is the automation of the full client acquisition cycle on the VKontakte (VK) social network. This involves:

1. Identifying potential target groups from vast, heterogeneous social data.
2. Classifying promotional content to understand engagement patterns.
3. Generating tailored advertising content to attract new users.
4. Overcoming severe class imbalance in promotional data and the "cold-start" problem without requiring massive labeled datasets or excessive computational resources.

Methodology

The authors propose a hybrid neural network architecture designed as a three-stage pipeline for automated client acquisition. The system integrates Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Large Language Models (LLM).

1. Data Extraction and Preprocessing

• Source: Data was collected from VKontakte groups and food delivery services using the VK API and mashup tools (Targethunter, Livedune).
• Process: A custom Python application handled HTTP requests to the REST API. To manage API rate limits and ensure fault tolerance, the system employed batch processing and exponential backoff.
• Privacy: An anonymization procedure was implemented to remove user identifiers (user_id) and strip private information, ensuring compliance with platform policies.
• Dataset Construction:
  • Module 1 (Group Classification): A binary classification dataset was created with 10,000 vectorized instances (5,000 target food delivery groups vs. 5,000 non-target groups). Hard negative examples (e.g., recipe aggregators) were added to improve robustness.
  • Module 2 (Post Classification): A five-class classification problem (Soft-sell, Discount, Gift, Cashback, On-time delivery) was formulated. Due to severe class imbalance (e.g., "Cashback" <3%), LLM-based data augmentation (using Gemma-4-31b-it) was employed to balance the training set to 1,000 examples per class. Validation and test sets remained strictly human-annotated.

2. Model Architecture and Training

The study evaluated and compared CNN, LSTM, and GRU architectures for classification tasks, alongside several LLMs for generation.

• Module 1: Target Group Identification

  • Best Model: A CNN architecture.
  • Configuration: Embedding layer (16-dim), three parallel Conv1D branches (kernel sizes 3, 7, 9 with 64 filters), ELU activation, Dropout (0.5), and Nadam optimizer (learning rate 1e-3).
  • Rationale: CNNs were found superior for extracting local features from short group descriptions.

• Module 2: Promotional Post Classification

  • Best Model: A CNN architecture.
  • Configuration: Embedding layer (48-dim), three parallel Conv1D branches (kernel sizes 3, 4, 5), ReLU activation, Dropout (0.5), and Nadam optimizer (learning rate 1e-2).
  • Rationale: Local lexical patterns in short posts were more predictive than long-range dependencies captured by RNNs.

• Module 3: Automated Content Generation

  • Approach: Fine-tuning of LLMs (YaGPT, Qwen2-7B, Llama3-8B, Suzume-Llama-3-8B) to generate advertising posts based on identified engagement patterns.
  • Selection: YaGPT was selected as the optimal model due to its high prompt adherence and optimization for Russian-language content, producing coherent texts with appropriate calls to action.

Key Results

The experimental evaluation demonstrated the efficacy of the proposed CNN-based architecture over recurrent models in this specific domain:

• Target Group Classification (Module 1): The CNN model achieved 98.8% accuracy, 98.81% precision, 98.8% recall, and an F1-score of 98.8%. The AUC reached 99.93%. The model successfully minimized false negatives, ensuring few target groups were missed.
• Post Classification (Module 2): The CNN model achieved 90.6% accuracy and an F1-score of 90.68% across five classes. The use of LLM-based augmentation successfully mitigated class imbalance, allowing for high recognition of minority classes (Cashback, On-time delivery) that were previously underrepresented.
• Content Generation (Module 3): The system successfully generated promotional content that aligned with specific strategies (e.g., discounts, gifts). YaGPT demonstrated the highest consistency in following prompts, generating texts with appropriate marketing elements (emojis, calls to action) without manual template construction.

Significance and Contributions

The paper claims several significant contributions to the field of AI-driven marketing and data mining:

1. Novel Pipeline for Client Acquisition: Unlike previous studies focusing solely on data extraction or sentiment analysis, this work presents a comprehensive, end-to-end pipeline for automating the client acquisition cycle, from data collection to content generation.
2. Efficiency of Lightweight Architectures: The study demonstrates that for short-text social media data, lightweight CNN architectures outperform more computationally expensive recurrent models (LSTM/GRU) and large transformer models in terms of both accuracy and deployment efficiency.
3. Handling Data Imbalance: The research validates the use of LLMs for data augmentation to resolve severe class imbalance in promotional datasets, enabling effective training without massive manual labeling efforts.
4. Practical Implementation: The system has been piloted by the "Inter Program" company for food delivery communities on VK, confirming the feasibility of reducing marketing costs and improving campaign effectiveness through automation.
5. Reproducibility: The authors provide the complete source code, preprocessing scripts, datasets, and model configurations via a public GitHub repository to ensure the reproducibility of the experiments.

Limitations and Future Work

The authors maintain a modest stance regarding the current scope:

• Platform Specificity: The current implementation is specific to VKontakte. Adapting the system to other platforms requires significant modifications to data collection modules due to differing API structures and rate limits.
• Human Oversight: While the generation module is automated, the authors note that human verification is still advisable before publishing generated content to ensure quality and relevance.
• Evaluation Scope: Experiments were conducted on a fixed train/validation/test split. Future work aims to evaluate temporal generalization using chronological data splits and to test the system's effectiveness in real-world customer acquisition campaigns across additional platforms.