【T112017-数据工程和技术分会场】用于图像标记的应用深度学习-旅行推荐引擎应用_27页_17mb

Pragmatic Deep Learning for Image Labelling - Travel Recommendation Engine

Core Content

This document outlines the application of pragmatic deep learning techniques for image labelling in the context of a travel recommendation engine. The goal is to enhance the user experience by leveraging image data to improve recommendation accuracy and relevance, especially for highly temporary and expensive travel products.

Main Points

• Image Labelling for Recommendation: Image labelling is used to extract meaningful features from travel images, which are then integrated into the recommendation system to provide more personalized and context-aware suggestions.
• Deep Learning for Dummies: The document addresses common issues faced by data scientists when using deep learning, such as lack of GPUs, expertise, and labelled data.
• Transfer Learning: This approach is used to utilize pre-trained models, which helps in reducing training time and resource requirements.
• Pre-trained Models: Available models from the community (e.g., Places and SUN databases) are used for image classification and feature extraction.
• Image Content Detection: Techniques like NMF matrix factorization are used to detect and score topics in images, enabling better understanding and categorization of visual content.
• Iterative Building of Recommender System: The system is built incrementally, starting with basic collaborative filtering and content-based recommendations, then incorporating image data to refine and enhance the results.
• Results: The integration of image information significantly improves the performance of the recommendation system, especially in scenarios where visual context is crucial, such as for sun and beach destinations.
• BI on Steroids: Image labelling is seen as a powerful tool for business intelligence, allowing for more nuanced and accurate insights from visual data.

Key Information

• Image Labeling: The process involves using deep learning models to classify and extract features from images, which are then used to improve the recommendation system.
• Pre-trained Models: These are used to overcome the challenge of limited labelled data and to speed up the model training process.
• Transfer Learning: Enables the use of existing knowledge from large datasets to improve the performance of the recommendation system on smaller, specific datasets.
• NMF Matrix Factorization: A technique for dimensionality reduction and feature balancing, enhancing the interpretability and effectiveness of image-based recommendations.
• Recommendation Engine: The system is designed to handle temporary sales and expensive products by incorporating image features, which provide more context than traditional text-based approaches.
• Dataiku DSS: A data science software platform used for the entire process, from data preparation to model building and monitoring.
• Real-time Scoring: The system is capable of providing real-time predictions based on image data, improving user engagement and conversion rates.

Specificities of the Recommendation System

• Time-sensitive Sales: The system accounts for seasonal and event-based promotions (e.g., Christmas, ski season).
• Expensive Products: These require careful handling due to fewer recurrent buyers, emphasizing the importance of visual appeal and context.
• Visual Context: Images provide more context than text, especially for impulsive purchases like travel.
• Complementary and Redundant Information: The system uses both types of information to refine recommendations and improve accuracy.

Results and Outcomes

• Visits: Images with visual features like "pool" and "sun & beach" are displayed more effectively.
• Recommendations: The system recommends destinations based on image features, improving user engagement and conversion.
• User Engagement Metrics: The document highlights the importance of tracking metrics such as clicks and average time spent on the page.
• Performance Evaluation: The system is evaluated iteratively, ensuring continuous improvement and adaptation.

Conclusion

• Iterative Data Science: The recommendation system is built in an iterative manner, starting simple and growing with each step.
• Deep Learning: The use of pre-trained models and transfer learning makes deep learning accessible to non-experts.
• Future Steps: The system is planned to be customized further for specific destinations and products, enhancing its adaptability and effectiveness.

Key Figures

• Client Base: 18 million clients.
• Sales Opened: Hundreds of sales opened every day.
• Image Importance: Sale images are critical for the recommendation engine, influencing user decisions and engagement.

Design and Workflow

• Prepare: Load and prepare the data using Dataiku DSS.
• Analyse: Visualize and share the work to ensure clarity and collaboration.
• Model: Build the recommendation model using image features and transfer learning.
• Monitor: Continuously monitor the performance of the model and the system.
• Score: Provide real-time scoring and predictions based on image data, enhancing the user experience.

What's Next

• Customization: The system will be further customized to handle specific destinations like Kenya, Prague, Berlin, and Cambodia, improving its relevance and effectiveness for diverse travel products.