Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Optimizing Recommendations using Fine-Tuned LLMs

Prabhdeep Cheema, Erhan Guven

TL;DR

This work tackles the challenge of personalized, expressive conversational search for media recommendations by generating synthetic, knowledge-graph–enriched prompts with large language models. It adopts cost-efficient LoRA/QLoRA fine-tuning on lightweight LLMs to handle complex queries with sub-500 ms latency while maintaining accuracy. Empirical results demonstrate strong gains in entity extraction and intent classification via Macro-F1 improvements over baselines, illustrating the viability of synthetic data for retrieval tasks. The approach promises broad applicability to conversational recommender systems and suggests avenues for future robustness and cross-domain deployment.

Abstract

As digital media platforms strive to meet evolving user expectations, delivering highly personalized and intuitive movies and media recommendations has become essential for attracting and retaining audiences. Traditional systems often rely on keyword-based search and recommendation techniques, which limit users to specific keywords and a combination of keywords. This paper proposes an approach that generates synthetic datasets by modeling real-world user interactions, creating complex chat-style data reflective of diverse preferences. This allows users to express more information with complex preferences, such as mood, plot details, and thematic elements, in addition to conventional criteria like genre, title, and actor-based searches. In today's search space, users cannot write queries like ``Looking for a fantasy movie featuring dire wolves, ideally set in a harsh frozen world with themes of loyalty and survival.'' Building on these contributions, we evaluate synthetic datasets for diversity and effectiveness in training and benchmarking models, particularly in areas often absent from traditional datasets. This approach enhances personalization and accuracy by enabling expressive and natural user queries. It establishes a foundation for the next generation of conversational AI-driven search and recommendation systems in digital entertainment.

Optimizing Recommendations using Fine-Tuned LLMs

TL;DR

This work tackles the challenge of personalized, expressive conversational search for media recommendations by generating synthetic, knowledge-graph–enriched prompts with large language models. It adopts cost-efficient LoRA/QLoRA fine-tuning on lightweight LLMs to handle complex queries with sub-500 ms latency while maintaining accuracy. Empirical results demonstrate strong gains in entity extraction and intent classification via Macro-F1 improvements over baselines, illustrating the viability of synthetic data for retrieval tasks. The approach promises broad applicability to conversational recommender systems and suggests avenues for future robustness and cross-domain deployment.

Abstract

As digital media platforms strive to meet evolving user expectations, delivering highly personalized and intuitive movies and media recommendations has become essential for attracting and retaining audiences. Traditional systems often rely on keyword-based search and recommendation techniques, which limit users to specific keywords and a combination of keywords. This paper proposes an approach that generates synthetic datasets by modeling real-world user interactions, creating complex chat-style data reflective of diverse preferences. This allows users to express more information with complex preferences, such as mood, plot details, and thematic elements, in addition to conventional criteria like genre, title, and actor-based searches. In today's search space, users cannot write queries like ``Looking for a fantasy movie featuring dire wolves, ideally set in a harsh frozen world with themes of loyalty and survival.'' Building on these contributions, we evaluate synthetic datasets for diversity and effectiveness in training and benchmarking models, particularly in areas often absent from traditional datasets. This approach enhances personalization and accuracy by enabling expressive and natural user queries. It establishes a foundation for the next generation of conversational AI-driven search and recommendation systems in digital entertainment.
Paper Structure (14 sections, 9 figures, 1 table)

This paper contains 14 sections, 9 figures, 1 table.

Table of Contents

  1. Introduction
  2. Objective and Contribution
  3. Related Work
  4. Method
  5. Datasets
  6. Fine-tuning
  7. Training pre-processing and Prompt Templates
  8. LoRA Configuration
  9. Model Training
  10. Evaluation Methodology
  11. Results
  12. Trade-offs in Synthetic Data Generation
  13. Broader Application Scenarios
  14. Conclusion and Future Work

Figures (9)

  • Figure 1: Lightweight models using knowledge distillation and pruning (meta_llama3.2_2024)
  • Figure 2: Speed Measured by Output Speed (tokens per second) (ArtificialAnalysis2024)
  • Figure 3: Quality based on Multi-task Language Understanding (MMLU) (ArtificialAnalysis2024)
  • Figure 4: LoRA only trains A and B (hu2021lora)
  • Figure 5: QLORA improves over LoRA by quantizing the transformer model to 4-bit precision and using paged optimizers to handle memory spikes (dettmers2023qlora)
  • ...and 4 more figures