Multitask Transfer Learning for Promotional Revenue Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current recommender systems do not effectively utilize multitask learning to enhance promotion recommendations on a per-user basis, failing to meaningfully improve recommendations for promotions offered to specific users.
Innovation Solution
A computer-implemented method that trains a base model using customer records associated with items across multiple categories, tunes the model via transfer learning for specific promotion tasks, and combines relevancy predictions with advertising revenue to estimate overall revenue, enabling personalized promotion recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional recommender systems are used for promotion recommendations, then the system structure remains simple, but the promotion effectiveness and revenue optimization are insufficient
Solution Approach 1:
The patent applies multitask learning to enable a single recommendation system to simultaneously perform multiple functions: predicting customer purchase probability, estimating promotion effectiveness, and optimizing revenue. The neural network model is trained to handle multiple tasks (purchase prediction, promotion response prediction, revenue optimization) within a unified architecture, allowing the system to generate personalized promotion recommendations that balance customer trust and total revenue without requiring separate specialized systems for each function
2Productivity
If personalized promotion recommendations are generated for each user, then promotion effectiveness increases, but the computational complexity and processing time increase significantly
Solution Approach 1:
The system performs preliminary actions by pre-training a neural network model on historical customer data and promotion effectiveness information before actual promotion deployment. The model learns patterns from past data and pre-computes relevant features and embeddings. During live promotion generation, the pre-trained model quickly infers personalized recommendations without requiring real-time training, significantly reducing processing time while maintaining high promotion effectiveness through personalized recommendations
3Measurement precision
If transfer learning is applied to tune the base model for specific promotion tasks, then the accuracy of promotion recommendations improves, but the model training and tuning process becomes more complex
Solution Approach 1:
The patent segments the model training process into distinct phases: first training a base neural network model on general customer behavior and purchase patterns, then applying transfer learning to fine-tune the model for specific promotion tasks. This segmentation allows the system to leverage pre-learned general knowledge while adapting to task-specific requirements, improving recommendation accuracy without requiring complete retraining from scratch and managing complexity through structured training stages
4Productivity
If the system balances retail revenue and promotional revenue optimization, then overall revenue increases, but the optimization process becomes more complex
Solution Approach 1:
The patent merges the optimization of retail revenue and promotional revenue into a single unified optimization process. The neural network model simultaneously predicts both retail purchase probability and promotion effectiveness, and the recommendation system optimizes both revenue streams together rather than separately. This merging approach allows the system to find synergistic opportunities where promotional activities enhance both immediate sales and long-term customer value, increasing total revenue while managing optimization complexity through integrated rather than separate optimization processes
Data Source
AI summary
Multitask learning is applied to predict a customer's propensity to purchase an item within a particular category of items. Then, the network is tuned using transfer learning for a specific promotional campaign. Retail revenue and promotional revenue are jointly optimized, conditioned on customer trust. Accordingly, a particular promotional program may be selected that is specific to the user.