Real-Time Ad Recommendation System Using Historical and Session Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital item advertisement systems fail to provide relevant recommendations to customers by relying solely on historical data or current user intent, leading to missed sales opportunities as customers are shown irrelevant items, causing them to abandon purchases.
Innovation Solution
Implement a system that generates item recommendations in real-time by combining historical user data with current user session data, incorporating temporal and trend data, and using machine learning models to balance past interactions with real-time user behavior, thereby providing personalized and relevant advertisements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If item advertisements are generated based solely on historical data, then the system can operate offline and process data efficiently, but the advertisements become irrelevant to current user intent and lead to lost sales opportunities
Solution Approach 1:
The system pre-processes and stores historical user data, item data, and feature embeddings offline before the user visits the website. This preliminary preparation allows the system to quickly generate relevant advertisements in real-time by combining pre-computed historical insights with current session data, resolving the contradiction between offline processing efficiency and real-time relevance.
Solution Approach 2:
The advertisement generation system dynamically adjusts its behavior by switching between offline pre-processing modes and online real-time inference modes. It combines static historical data with dynamic current session data, allowing the system to adapt to changing user intentions while maintaining computational efficiency through selective real-time processing.
2Loss of time
If item advertisements are generated based solely on current user session intent, then real-time relevance is improved, but the system loses context from historical purchases and user habits
Solution Approach 1:
The system merges offline historical data processing with online real-time inference by combining feature embeddings from past user behavior with current session data. This integration allows the system to maintain both historical context and real-time responsiveness, generating advertisements that reflect both past habits and current intentions simultaneously.
Solution Approach 2:
The system introduces an intermediary layer of feature embeddings and machine learning models that mediate between historical data and current session data. This intermediary component processes both offline historical patterns and online real-time signals, translating them into unified advertisement recommendations that balance past context with present intent.
3Reliability
If the system processes both historical data and real-time session data, then advertisement relevance improves, but computational complexity and processing time increase
Solution Approach 1:
The system segments the data processing workflow into distinct offline and online components. Historical data processing, feature embedding generation, and model training occur offline, while only the essential real-time inference and session data processing occur online. This segmentation reduces overall system complexity by isolating computationally intensive tasks from real-time requirements.
Solution Approach 2:
By performing data processing, feature extraction, and model inference offline in advance, the system prepares computational resources ahead of time. This preliminary action reduces the computational burden during real-time operation, allowing the system to handle both historical and current data without overwhelming processing complexity at the moment of advertisement generation.
4Reliability
If the system uses machine learning models for real-time inference, then advertisement relevance to user intent improves, but processing speed and computational resources are consumed
Solution Approach 1:
The system performs machine learning model training and feature embedding generation offline in advance, preparing the computational models before real-time inference is needed. This preliminary action allows the system to use complex machine learning algorithms for high-relevance advertisements while maintaining fast processing speeds during real-time operation, as the heavy computational work is already completed.
Solution Approach 2:
The system creates simplified representations of complex machine learning models in the form of pre-computed feature embeddings and trained model parameters. These copies allow real-time inference to proceed quickly using the pre-processed information, while the full complexity of the machine learning algorithms is encapsulated in the offline-trained models, resolving the contradiction between model complexity and processing speed.
Data Source
AI summary
This application relates to apparatus and methods for providing recommended items to advertise. In some examples, a computing device determines a first set of items for recommendation based on historical user data associated with a user, and a second set of items for recommendation based on real-time user session data for the user. The computing device may then determine a subset of the first set of items based on associated scores and a predetermined threshold number of first items that can be presented for optimal user interaction. The computing device may generate a set of item recommendations by combining the subset of the first set of items and at least one of the second set of items to present to the user as advertisements.