Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Netflix Artwork Personalization via LLM Post-training

Hyunji Nam, Sejoon Oh, Emma Kong, Yesu Feng, Moumita Bhattacharya

TL;DR

The paper tackles personalized artwork selection for Netflix titles by post-training LLMs to account for diverse user tastes, predicting which among multiple artwork captions a user will prefer. It represents artworks as captions and leverages SFT and DPO with LoRA, including reasoning augmentation from a larger model to craft high-quality training data. On a dataset with 110K training examples and 5K held-out user-title pairs, post-trained Llama 3.1 8B models achieve $3-5\%$ IPS improvements over the Netflix production model, indicating the viability of fine-grained, per-user artwork personalization. The work analyzes output-format sensitivity, model-size effects, and training-size effects, and suggests future directions such as direct image embeddings and extending personalization to other content components like synopses and trailers.

Abstract

Large language models (LLMs) have demonstrated success in various applications of user recommendation and personalization across e-commerce and entertainment. On many entertainment platforms such as Netflix, users typically interact with a wide range of titles, each represented by an artwork. Since users have diverse preferences, an artwork that appeals to one type of user may not resonate with another with different preferences. Given this user heterogeneity, our work explores the novel problem of personalized artwork recommendations according to diverse user preferences. Similar to the multi-dimensional nature of users' tastes, titles contain different themes and tones that may appeal to different viewers. For example, the same title might feature both heartfelt family drama and intense action scenes. Users who prefer romantic content may like the artwork emphasizing emotional warmth between the characters, while those who prefer action thrillers may find high-intensity action scenes more intriguing. Rather than a one-size-fits-all approach, we conduct post-training of pre-trained LLMs to make personalized artwork recommendations, selecting the most preferred visual representation of a title for each user and thereby improving user satisfaction and engagement. Our experimental results with Llama 3.1 8B models (trained on a dataset of 110K data points and evaluated on 5K held-out user-title pairs) show that the post-trained LLMs achieve 3-5\% improvements over the Netflix production model, suggesting a promising direction for granular personalized recommendations using LLMs.

Netflix Artwork Personalization via LLM Post-training

TL;DR

The paper tackles personalized artwork selection for Netflix titles by post-training LLMs to account for diverse user tastes, predicting which among multiple artwork captions a user will prefer. It represents artworks as captions and leverages SFT and DPO with LoRA, including reasoning augmentation from a larger model to craft high-quality training data. On a dataset with 110K training examples and 5K held-out user-title pairs, post-trained Llama 3.1 8B models achieve IPS improvements over the Netflix production model, indicating the viability of fine-grained, per-user artwork personalization. The work analyzes output-format sensitivity, model-size effects, and training-size effects, and suggests future directions such as direct image embeddings and extending personalization to other content components like synopses and trailers.

Abstract

Large language models (LLMs) have demonstrated success in various applications of user recommendation and personalization across e-commerce and entertainment. On many entertainment platforms such as Netflix, users typically interact with a wide range of titles, each represented by an artwork. Since users have diverse preferences, an artwork that appeals to one type of user may not resonate with another with different preferences. Given this user heterogeneity, our work explores the novel problem of personalized artwork recommendations according to diverse user preferences. Similar to the multi-dimensional nature of users' tastes, titles contain different themes and tones that may appeal to different viewers. For example, the same title might feature both heartfelt family drama and intense action scenes. Users who prefer romantic content may like the artwork emphasizing emotional warmth between the characters, while those who prefer action thrillers may find high-intensity action scenes more intriguing. Rather than a one-size-fits-all approach, we conduct post-training of pre-trained LLMs to make personalized artwork recommendations, selecting the most preferred visual representation of a title for each user and thereby improving user satisfaction and engagement. Our experimental results with Llama 3.1 8B models (trained on a dataset of 110K data points and evaluated on 5K held-out user-title pairs) show that the post-trained LLMs achieve 3-5\% improvements over the Netflix production model, suggesting a promising direction for granular personalized recommendations using LLMs.
Paper Structure (16 sections, 2 equations, 2 figures, 4 tables)

This paper contains 16 sections, 2 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. LLMs for personalized recommendations
  4. LLM post-training
  5. Automated image captioning
  6. Problem setup
  7. Data composition
  8. Evaluation metric
  9. Method
  10. Problem formulation
  11. Post-training method
  12. Experiment & Results
  13. How do post-trained LLMs compare to the production model?
  14. How sensitive is the model to number versus text based prediction?
  15. How does post-training impact different base models?
  16. ...and 1 more sections

Figures (2)

  • Figure 1: Different components of user experience (e.g., artworks, synopsis, trailers) can be personalized in addition to the title.
  • Figure 2: A breakdown of accuracy across different ground truth labels to compare the performance of the models outputting the artwork option number versus the artwork caption in full text. The x-axis shows the ground truth label (artwork option number) and the y-axis shows the model's prediction accuracy for samples with a particular ground truth label. Although the average performance across all examples is higher for the number prediction (blue) compared to the text caption prediction (orange), the breakdown by the ground truth label suggests that number prediction performs poorly for cases with higher numbers for the selected artwork.