Query-Ad Ranking via Semantic Embeddings and Factorization Machines
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current internet advertising systems struggle to effectively match advertisements with new or tail queries that lack historical data or user feedback, leading to difficulties in displaying relevant ads for uncommon search terms.
Innovation Solution
A method and system that classify search queries based on intent, enhance them with semantic information, and use a factorization machine to rank advertisements, incorporating a knowledge graph and ad-query indexing to improve matching and ranking of ads for new queries.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional matching algorithms using dictionary lookup and index based retrieval are used, then matching speed is maintained, but the system cannot match new/tail queries with bidterms and associated advertisements
Solution Approach 1:
The system performs preliminary actions by pre-computing and storing query embeddings and ad embeddings in advance. These pre-computed representations are stored in memory, allowing the system to quickly retrieve and compare embeddings for new queries without requiring historical data or complex real-time computations.
Solution Approach 2:
The patent introduces embedding vectors as an intermediary representation between queries and advertisements. Instead of directly matching queries to bidterms using traditional methods, the system transforms both into semantic embeddings and computes similarity in the embedding space, enabling matching of new queries through semantic similarity rather than exact keyword matching.
2Reliability
If semantic relevance matching using word2vec similarity score is used, then ad relevance is improved, but the system still cannot effectively match tail queries without historical data
Solution Approach 1:
The system changes the parameter representation from traditional keyword-based matching to embedding-based semantic representation. By transforming queries and advertisements into dense vector embeddings that capture semantic meaning, the system can measure similarity through vector operations rather than keyword overlap, enabling effective matching even when historical data is unavailable.
3Speed
If offline matching methods with strict latency constraints are used, then response time is controlled, but the system lacks the flexibility to adapt to new queries
Solution Approach 1:
The system performs computationally intensive embedding computations in advance and stores the results. When a new query arrives, the system only needs to retrieve pre-computed embeddings and perform simple similarity calculations, maintaining fast response times while being able to handle novel queries that traditional offline methods would miss.
Data Source
AI summary
A system and method for ranking query-advertisement combinations is disclosed. Embodiments use an online component to enhance and rank query ad combinations. The query ad combination is then reranked with a trained factorization machine. The subsequent list of ranked query-ad combinations is then output. The output may be to an auction for determine ad-query combinations having the greatest expected revenue.