Deep Neural Language Model for Query Clustering in Automated Bidding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The sparsity of historical advertisement records with identical attributes leads to insufficient statistically significant predictions of advertisement performance, particularly due to the vast dimension of search queries, resulting in inaccurate estimation of advertisement effectiveness and bidding adjustments.
Innovation Solution
Employing data aggregation through deep learning techniques, such as clustering search queries in a dense vector space using a deep learning neural network language model to group semantically similar queries, allowing for the use of aggregated historical performance records for estimating advertisement effectiveness and adjusting bids.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If historical advertisement records are used for prediction, then prediction accuracy is improved, but the sparsity of records with identical attributes worsens due to the vast dimension of search queries
Solution Approach 1:
The patent merges semantically similar search queries into unified query groups using deep learning embedding techniques. By transforming individual queries into vector representations and clustering those with similar embeddings, the system combines historical performance data across multiple queries, thereby increasing the quantity of usable historical records while maintaining prediction accuracy.
2Measurement precision
If data aggregation through deep learning is employed, then prediction accuracy for larger portions of advertisement traffic is improved, but system complexity increases
Solution Approach 1:
The patent introduces deep learning embedding models as an intermediary layer between raw search queries and the advertisement bidding system. This intermediary transforms high-dimensional, sparse query data into dense, continuous vector representations, enabling efficient similarity computation and data aggregation while managing system complexity through standardized processing pipelines.
3Measurement precision
If individual search queries are analyzed separately, then query-specific precision is maintained, but the overall prediction reliability deteriorates due to data sparsity
Solution Approach 1:
The patent combines data from multiple semantically similar queries by aggregating their historical performance records after embedding transformation. This merging approach maintains the distinct characteristics of individual queries while pooling sufficient data across the query group to achieve statistically reliable predictions, thereby simultaneously preserving query-specific precision and improving overall reliability.
Data Source
AI summary
Systems, devices, and methods are disclosed for predicting potential effectiveness of query-triggered internet advertisements received from different web page publishers using a deep learning neural network language model for clustering queries, and for automatically adjusting bids for advertisements by advertisers based on the predicted potential effectiveness. Using query-clusters rather than queries for adjusting bids for advertisements allows for more accurate and more consistent bidding strategy despite of sparsity in historical advertisement performance data, higher return on investments for the advertisers, and higher revenue for the publishers of the advertisements.