Recommendation Prediction Module Snapshot Archive Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The quality of recommendations in two-step recommendation systems suffers when the main prediction module is trained on a training set that includes events for which features were not available at the time of the event, leading to biased over-trust and decreased performance compared to one-step approaches, especially when features become available post-event.
Innovation Solution
Implementing a two-step recommendation approach where user-nonspecific features are calculated and stored in snapshot archives prior to the event, and user-specific features are generated in real-time, using different training sets for each prediction module to mitigate over-fitting and biased over-trust, with user-nonspecific features retrieved from snapshot archives and user-specific features available at the time of the event for training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the main prediction module is trained on a training set that includes events for which features were not available at the time of the event, then the system can utilize more comprehensive data for training, but the quality of recommendations deteriorates due to biased over-trust and decreased performance
Solution Approach 1:
The patent applies preliminary action by calculating and storing user-nonspecific features in snapshot archives before the events occur. This allows the main prediction module to be trained only on features that were actually available at the time of each event, preventing biased over-trust while still utilizing comprehensive training data. The snapshot archives capture the state of user-nonspecific features at specific points in time, ensuring temporal consistency between features and events during training.
2Use of energy by moving object
If user-nonsspecific features are calculated and stored in snapshot archives prior to events, then computational power requirements are reduced, but the system complexity increases due to archive management
Solution Approach 1:
The patent applies segmentation by dividing features into two categories: user-nonspecific features (stored in snapshot archives) and user-specific features (calculated in real-time). This segmentation allows computationally intensive user-nonsspecific features to be pre-calculated and stored, reducing real-time computational requirements. The snapshot archive structure organizes data by time points, making retrieval efficient despite the added complexity of archive management.
3Reliability
If different training sets are used for each prediction module in a two-step recommendation system, then over-fitting is mitigated, but the data preparation complexity increases
Solution Approach 1:
The patent applies local quality by creating different training sets tailored to each prediction module's specific needs. The first prediction module (main module) is trained on a training set comprising events with their corresponding user-nonspecific features from snapshot archives and user-specific features available at event time. The second prediction module (auxiliary module) is trained on a different training set focusing on user-nonspecific features. This differentiated approach prevents over-fitting by ensuring each module learns from appropriately scoped data, while the systematic use of snapshot archives manages the complexity of data preparation.
Data Source
AI summary
There is disclosed a method for generating a content recommendation for a given user of a recommendation system. The method comprises: receiving a request for the content recommendation; responsive to the request generating a set of content recommendations for the given user, the generating being executed by a prediction module of the recommendation server, the prediction module having been trained using a training set of training events, such that for each given training event from the training set of training events: at least one user-nonspecific feature is used as a first input parameter for the prediction module training, the at least one user-nonspecific feature having been retrieved from a latest version of a snapshot archive available at a time of the given training event occurring; and at least one user-specific feature is used as a second input parameter for the prediction module training.