Session-Based Recommendation Using Gated Graph Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing recommendation systems face accuracy issues due to over-fitting caused by one-hot coding, which reduces the effectiveness of item recommendations.
Innovation Solution
A session-based recommendation method using a pre-trained model that constructs a directed session graph, inputs it into a gated graph neural network to generate item embedding vectors, determines dynamic user preferences, and calculates prediction scores to recommend items, while employing label distribution calculations and loss functions to optimize the model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If one-hot coding is used to represent items in the recommendation system, then the implementation is simple and straightforward, but it leads to over-fitting of the model and reduces recommendation accuracy
Solution Approach 1:
The patent transforms item representation from discrete one-hot encoding to continuous embedding vectors. Item embeddings are learned parameters that capture semantic relationships between items, changing the parameter space from sparse binary vectors to dense continuous vectors. This resolves the contradiction by maintaining implementation simplicity while dramatically improving recommendation accuracy through better feature representation.
Solution Approach 2:
The patent introduces graph neural networks as an intermediary between item data and the recommendation model. The GNN processes item relationships through graph structures, acting as a mediator that transforms raw item data into enriched embeddings that capture contextual relationships. This intermediary layer prevents over-fitting by learning robust representations while maintaining system simplicity.
2Device complexity
If traditional recommendation models are used without session-based approaches, then the model complexity is lower, but the ability to capture dynamic user preferences and session context is insufficient
Solution Approach 1:
The patent implements dynamic preference capture through session-based modeling where user preferences evolve over time within sessions. The model dynamically updates user state based on sequential interactions, allowing preferences to change adaptively. This dynamic approach significantly improves preference capture accuracy while the modular architecture keeps implementation complexity manageable.
Solution Approach 2:
The patent segments the recommendation problem into distinct components: session representation, item representation, preference modeling, and prediction. By dividing the complex task into manageable segments with specialized processing for each, the system achieves high reliability in preference capture while maintaining overall model complexity at acceptable levels.
3Ease of manufacture
If static user profiles are used for recommendation, then the system is easier to implement and maintain, but it cannot adapt to changing user preferences and behaviors
Solution Approach 1:
The patent performs preliminary embedding learning for items and users before the actual recommendation process. These pre-trained embeddings capture fundamental patterns in user behavior and item characteristics. This preliminary action creates a solid foundation that is easy to maintain while enabling the system to rapidly adapt to changing preferences through session-based updates without retraining the entire model.
Solution Approach 2:
The session-based model automatically updates user preferences based on observed interactions without requiring external intervention or manual profile updates. The system self-adjusts by processing session data through the neural network, maintaining adaptability while keeping the system simple to operate and maintain.
Data Source
AI summary
A session-based recommendation method and device according to one or more embodiments of this disclosure are provided, which use a pre-trained recommendation model to perform item recommend. The method includes following contents: a directed session graph is constructed according to a session to be predicted; the directed session graph is then input into a gated graph neural network which outputs the item embedding vector; a user's dynamic preference is determined according to a user's current preference and a first long-term preference, the current preference is an item embedding vector of a last item in the session and the first long-term preference is determined according to the item embedding vector and an importance score of the item; a prediction score of a respective item is determined according to the dynamic preference and the item embedding vector; and a recommended item is output according to the prediction score of the respective item.