Machine Learning Auction Parameter Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining reserve prices in online ad auctions lead to heavy server loads and are prone to market manipulation, resulting in inefficient equilibrium states.
Innovation Solution
A method using machine learning algorithms to predict optimal auction parameters by analyzing interaction history data from two portions of users, allowing for offline training and application of optimal values to reduce server load and prevent market manipulation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If reserve price is determined as a function of advertiser's bid or history of bids, then the auction parameter can be adjusted dynamically, but advertisers can negatively influence the market and intentionally decrease the reserve price, resulting in bad market equilibrium states
Solution Approach 1:
The patent introduces a machine learning model as an intermediary between bid data and reserve price determination. The model processes historical bid data and advertiser characteristics to generate reserve price recommendations, preventing direct manipulation while maintaining adaptability. The model acts as a mediator that translates raw bid information into equitable pricing decisions based on learned patterns rather than direct advertiser influence.
Solution Approach 2:
The system performs preliminary training of the machine learning model offline using historical auction data before actual auctions occur. This preliminary action establishes the model's understanding of market patterns and equilibrium states in advance, allowing it to make reliable reserve price determinations during live auctions without real-time computation that could be manipulated.
2Speed
If online algorithms are used for calculating reserve price in real-time, then the auction can be processed quickly, but heavy server load occurs
Solution Approach 1:
The machine learning model is trained offline in advance using historical auction data, performing the computationally intensive work before real-time auctions. This preliminary training phase captures market patterns and equilibrium relationships, allowing the model to make rapid reserve price recommendations during live auctions without requiring heavy real-time computation.
Solution Approach 2:
The system creates a simplified computational model that replicates the complex relationships learned from historical data. Instead of recalculating reserve prices from raw bid data in real-time, the system uses the trained model's learned patterns to quickly determine appropriate reserve prices, effectively copying the essential market dynamics into a lightweight computational form.
Data Source
AI summary
There is disclosed a method of determining an optimal value of an auction parameter for a digital object. The method comprises using: an indication of a digital object, an auction parameter associated with the digital object and an environment feature at the respective moment of time to execute an offline training of a machine learning algorithm to predict an optimal value of auction parameters for a plurality of digital objects, the plurality of digital objects being associated with the interaction history of the first portion of users. The method further comprises applying the machine learning algorithm to determine a first optimal value of an auction parameter for a plurality of digital objects associated with the second portion of users and using such determined value for determining a digital object being relevant to the request from a user from the second portion of users.