Web Ad Classifier Data Balancing via Clustering and Oversampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Web advertising faces challenges in optimizing click-through and conversion rates due to imbalanced data sets, where few instances of positive data points (clicks or conversions) are overshadowed by numerous negative data points, leading to biased evaluation metrics and ineffective machine learning model training.
Innovation Solution
A method and apparatus that balance the data set by clustering negative data points using multiple algorithms to identify consensus-based representations, oversampling positive data points, and training a classifier to predict relevant advertisements for display, thereby optimizing click-through and conversion rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning models are trained on imbalanced advertisement data, then the model can be trained with available data, but the evaluation metrics become biased towards negative data points and the click-through rate optimization fails
Solution Approach 1:
The patent creates synthetic copies of positive data points through data generation techniques. It generates artificial advertisement-impression records that mimic the characteristics of successful conversions, thereby increasing the quantity of positive training samples without requiring additional real-world conversion events.
Solution Approach 2:
The patent modifies data representation parameters by creating clustered representations of negative data points. It groups similar negative samples into clusters and uses cluster centroids or representative samples, changing how negative data is parameterized to reduce its overwhelming numerical dominance while preserving informational content.
2Reliability
If all negative data points are used for training, then the model receives comprehensive negative examples, but the positive data points become overwhelmed and the model cannot learn effective patterns for conversion prediction
Solution Approach 1:
The patent extracts and isolates positive data points from the overwhelming majority of negative data. It separates the minority positive class into distinct processing streams, applying different clustering and sampling strategies specifically to positive samples to prevent them from being drowned out by negative data during training.
Solution Approach 2:
The patent segments the training data processing into distinct phases: first processing negative data points through clustering to identify representative patterns, then separately processing positive data points through oversampling and clustering, and finally combining them in balanced proportions for model training.
3Measurement precision
If data balancing techniques are applied to overcome class imbalance, then the model can learn from balanced data, but the device complexity increases due to multiple clustering algorithms and data processing steps
Solution Approach 1:
The patent merges multiple clustering algorithms (such as K-means, hierarchical clustering, and density-based clustering) into a unified data balancing framework. Instead of using each algorithm separately, it combines their strengths to create robust clustered representations that capture different aspects of the data distribution, reducing the need for complex individual processing pipelines.
Solution Approach 2:
The patent performs preliminary clustering and data balancing actions before the main model training process. By pre-processing the imbalanced data to create balanced training sets through clustering and oversampling, it eliminates the need for complex real-time balancing during training, simplifying the overall system architecture.
Data Source
AI summary
A method and apparatus for providing Web advertisements to online users is disclosed. A balanced set of negative data points and positive data points is derived from a log of Ad impressions and used to train a classifier. In response to an Ad request signal, a plurality of Ads is retrieved from a database. The Ad request signal indicates a request to provide an Ad for a slot available on a Web page associated with a website. The signal is provided in relation to an access of the Web page by an online user and includes information related to the online user. A choice of an Ad is predicted based on the information related to the online user and the plurality of Ads. The Ad is provided to a Web server to cause display of the Ad on the slot when the Web page is displayed to the online user.