Multi-Objective ML Training via Feedback Loops
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning algorithms face inaccuracies and inefficiencies due to insufficient or non-representative training data, leading to issues such as overprediction or underprediction, particularly in applications like automated trading and search engine marketing, where outdated models can result in overpayments and misclassification of new data points.
Innovation Solution
A system and method for training multi-objective machine learning algorithms that generate a combined objective using user-provided objectives and constraints, calculate return per click for advertisement campaigns, determine bids based on predictive algorithms, and iteratively update models using feedback loops to ensure accuracy and relevance without requiring new training data or pre-trained models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If more training data is gathered to improve model accuracy, then the accuracy and reliability of machine learning algorithms improve, but the time required for training increases significantly
Solution Approach 1:
The system performs preliminary actions by gathering diverse data sources (clickstream data, contextual data, conversion data) and pre-processing them into structured formats before actual model training. This includes creating feature engineering pipelines, data validation frameworks, and pre-computing statistical metrics that prepare the data in advance, reducing the time needed during actual model training while maintaining high accuracy
Solution Approach 2:
The system creates synthetic training data through data augmentation techniques, generating copies and variations of existing data samples. This includes creating synthetic clickstream patterns, simulated user behaviors, and artificially generated contextual scenarios that expand the training dataset without requiring additional real-world data collection time
2Reliability
If pre-trained models are purchased to improve model performance, then the accuracy and reliability improve, but the cost increases and the models become outdated quickly
Solution Approach 1:
The system enables self-service through automated model retraining and updating mechanisms that continuously ingest new data sources (real-time clickstream data, updated contextual information, recent conversion patterns) and automatically retrain models without external intervention. This includes automated model monitoring, performance degradation detection, and scheduled retraining pipelines that keep models current with the latest data patterns
Solution Approach 2:
The system implements feedback loops where model predictions are continuously evaluated against actual outcomes (click-through rates, conversion rates, cost-per-click data). This feedback is used to identify prediction errors, update training datasets with new examples, and trigger model retraining when performance thresholds are breached, ensuring models adapt to changing patterns
3Reliability
If training data is purchased from vendors to improve model accuracy, then the quality and representativeness of training data improve, but the cost increases and the data becomes outdated quickly
Solution Approach 1:
The system implements multi-functionality by integrating multiple data collection channels (clickstream data from websites, contextual data from user profiles, conversion data from e-commerce platforms, demographic data from third-party sources) into a single unified training data framework. This diverse data ecosystem provides continuous fresh data from multiple sources simultaneously, eliminating reliance on single vendor datasets and ensuring both high quality and currency
Data Source
AI summary
Systems and methods including one or more processors and one or more non-transitory storage devices storing computing instructions configured to run on the one or more processors and perform: receiving one or more objectives and one or more constraints from a user electronic device of a user; generating a combined objective using the one or more objectives; calculating, using the combined objective, a return per click for an advertisement campaign; determining one or more bids based on (a) the return per click for the advertisement campaign and (b) a return on advertising spend for the advertisement campaign; estimating, using a first predictive algorithm and the one or more bids, an average cost per click; estimating an expected number of clicks for the advertisement campaign based on the average cost per click; determining a total amount used of the one or more constraints for the one or more bids based on the expected number of clicks for the advertisement campaign; and when the total amount used of the one or more constraints for the one or more bids exceeds at least one of the one or more constraints, lowering the one or more bids. Other embodiments are disclosed herein.