Get GenAI guide

Access HaxiTAG GenAI research content, trends and predictions.

Monday, May 13, 2024

Large-scale Language Models and Recommendation Search Systems: Technical Opinions and Practices of HaxiTAG

In the digital age, recommendation search systems have become an indispensable part of our daily lives. As a company focused on large-scale language models (LLMs) and recommendation search systems, HaxiTAG has proposed a series of key points to optimize system design and improve the efficiency of recommendation search systems. This article will provide a technical overview of these points and explore how to build an efficient recommendation search system based on HaxiTAG's practices.

Firstly, system design and model design play an important role in recommendation search systems. 

HaxiTAG believes that while model design is important, the goal of productization is related to the scenario and pre-solved problems. Therefore, the focus of problem analysis is the overall view of system design, and the focus of problem-solving becomes how to fully leverage the advantages of large models, large computing power, and big data in system integration, providing systematized solutions with higher energy efficiency ratios to achieve high-ROI innovative practices.

Secondly, the ability to process large-scale data in real-time and in batches is crucial for recommendation search systems. 

Unlike traditional recommendation and search systems, where large-scale data tagging work relies on offline or asynchronous computing architecture, LLM-based new development paradigms can shift from batch processing to real-time processing to cope with ever-changing user needs and preferences. Real-time systems can respond to user behavior faster, while batch processing is more efficient in terms of computation and resource management. One of HaxiTAG's customer application cases has achieved daily processing of tens of millions of business data, with data records exceeding 50 billion and growing rapidly every day.

In recommendation systems, candidate set retrieval (Retrieval) and ranking (Ranking) are two key processes. 

Candidate set retrieval is a fast but less accurate process that screens out a small number of candidate items from a large amount of relevant data. Ranking, on the other hand, is a slower but more accurate process that sorts the retrieved candidate items to determine the final content recommended to the user. To improve the experience and output accuracy, more factors are introduced in the candidate set processing for re-ranking.

Feature storage and embedding models are also important components of recommendation search systems. 

Feature storage is used to collect and organize user and item features, while embedding models convert these features into mathematical representations for similarity calculations. In candidate retrieval, Approximate Nearest Neighbors (ANN) is used to quickly find projects that are most relevant to user queries, Of course, other similarity algorithms such as cosine similarity can also be utilized, e.g.

The advantage of real-time recommendation systems is that they can provide more personalized and timely recommendations, especially when user needs change rapidly. Although real-time systems may be more expensive in terms of resources and maintenance, they are necessary in some cases to provide high-quality user experiences.

In the ranking stage, in addition to considering the interaction probability between users and items, business logic should also be considered, such as increasing the diversity of the recommendation list. When building a recommendation system from scratch, HaxiTAG recommends using simple models (such as Word2Vec or simple statistical methods) to create embedding models, then using approximate nearest neighbor search for candidate retrieval, and finally using methods such as logistic regression for ranking.

Finally, HaxiTAG points out that although the latest research results in academia are important, many production systems actually use more mature and stable technologies. Therefore, in practical applications, we should combine the actual situation and choose the appropriate technical solutions.

Through the above technical overview of HaxiTAG's key points, we can see that the design and implementation of recommendation search systems is a complex process that involves understanding user behavior, data processing, model selection, real-time computing, and business logic integration. The knowledge and technical insights provided by HaxiTAG are crucial for building efficient recommendation search systems. We hope that through the analysis and comments in this article, we can provide some inspiration and reference for readers.

Key Point Q&A:

  • What are the key considerations in designing and implementing a recommendation search system according to the discussed points?

The key considerations include system and model design, real-time and batch processing capabilities for large-scale data, candidate set retrieval and ranking, feature storage and embedding models, approximate nearest neighbor search, cost-efficiency trade-offs, personalization and diversity in ranking, MVP development approach, and the balance between academic research and industrial application.


  • How does the shift towards real-time processing in recommendation systems affect user experience and system efficiency?

Real-time processing enables more personalized and timely recommendations, especially crucial in situations with rapidly changing user demands. It allows for faster response to user behavior, enhancing user experience. However, it may involve higher operational costs and development time compared to batch processing, though it's deemed necessary for providing high-quality user experiences in certain scenarios.


  • What are the recommended steps for developing a recommendation system from scratch, as suggested by the speaker?

The speaker suggests starting with simple models like Word2Vec or basic statistical methods to create embedding models. Then, employing approximate nearest neighbor search for candidate retrieval, and finally using techniques like logistic regression for ranking. This approach emphasizes starting with a Minimum Viable Product (MVP) and gradually refining the system based on performance and user feedback.