Hongrui Zhao

• Search Algorithm Design & Optimization: Led the optimization of search algorithms, improving intent recognition, dictionary construction, and recall strategies, significantly enhancing search relevance across multiple use cases.
• Data-Driven Strategy Optimization: Conducted in-depth data analysis to identify performance gaps in search algorithms, refining strategies to improve search relevance and boosting click-through rates.
• Feature Engineering & Search Ranking: Optimized search ranking by adjusting similarity metrics, refining matching strategies, and conducting user behavior analysis to enhance the overall search experience.
• Search Log Analysis: Developed tools to analyze search logs, built relevance evaluation bots, and performed bad-case analysis to fine-tune models for continuous improvement.

Key Projects:

Clustering and Matching for Companies

Objective:

• Improve Company matching by clustering similar companies based on their features to enhance search relevance.

Contributions:

• Developed vector representations for Companies using TF-IDF on company names and introductions, and incorporated Job functions to capture key features.
• Applied clustering algorithms to group Companies with similar attributes, reducing the candidate set for large language model (LLM) grouping.
• Implemented a real-time matching mechanism using edit distance to improve similarity between search queries and Companies.

Results:

• Achieved a 34% improvement in search relevance, reducing irrelevant Company matches.
• Increased click-through rate by 8% in supplemental job search scenarios, improving the efficiency of search results.

Technologies Used:

• Python, Scikit-learn, TensorFlow, LLM, TF-IDF, Clustering algorithms.

Personalized Search Query Recommendation

Objective:

• Develop a personalized search query recommendation system to enhance user search efficiency by suggesting more relevant queries based on user behavior.

Contributions:

• Collaborated with product managers to identify data sources and constructed a batch of candidate search queries for personalization.
• Developed a personalized ranking system using collaborative filtering and semantic similarity algorithms to improve the relevance of suggested queries.
• Created APIs with Flask for seamless integration into the existing search infrastructure, enabling real-time personalized suggestions.

Results:

• Successfully launched the personalized search query recommendation feature, significantly enhancing user search efficiency and relevance.

Technologies Used:

• Python, Collaborative Filtering, Content-Based Filtering, TensorFlow, Flask.

Search Dictionary Reconstruction

Objective:

• Address performance issues in the cold-start phase of the search system by reconstructing the search dictionary to better handle new search scenarios and improve query relevance.

Contributions:

• Conducted a comprehensive analysis of existing dictionaries, identifying gaps and areas needing improvement for better domain-specific coverage.
• Reconstructed domain-specific dictionaries using LLMs to generate specialized terms, synonyms, and skills, enhancing query matching.
• Performed dictionary cleaning and enhancement using bad-case analysis, improving intent recognition accuracy.

Results:

• Increased search relevance for domain-specific queries, reducing irrelevant results and improving user experience during the cold-start phase.
• Reduced occurrences of empty search results by expanding the dictionary's coverage and adaptability to new queries.

Technologies Used:

• Python, LLMs, Data analysis.

• Search Pre-training Optimization: Focused on improving the relevance of long-tail SPUs in product search by enhancing the pre-training process of a BERT-based relevance classification model.
• Algorithm Improvement: Researched and refined BERT-based relevance classification algorithms, addressing gaps in performance for less common product queries through data-driven analysis and custom pre-training tasks.
• Data Analysis & Pre-training Task Design: Analyzed search cases to identify low-relevance instances for long-tail SPUs, and designed a custom BERT pre-training task to specifically target these cases, improving search result accuracy for niche products.
• Corpus Creation & Data Processing: Processed and cleaned large datasets using Pyspark and HiveSQL to create a specialized corpus for model pre-training, ensuring it accurately represented the target search domain.
• Model Training & Optimization: Trained and fine-tuned a BERT model using TensorFlow, optimizing it for better performance on long-tail items, significantly improving search relevance for rare product queries.

Key Result:

• Reduced bad-case search results to 5%, drastically improving long-tail search relevance as measured by offline accuracy evaluations.

Technologies Used: Python, Pyspark, HiveSQL, TensorFlow, BERT, Data Preprocessing.