Recommendation Model Training with Sliding Windows for User Preference Balance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional approaches for training recommendation models using fixed window sampling fail to capture long-term user preferences, leading to suboptimal recommendations and increased computational costs when attempting to include historical data, while focusing exclusively on recent interactions.
Innovation Solution
A method involving generating fixed and sliding window samples to train a recommendation model, allowing for a broader range of user interactions, capturing both short-term and long-term preferences without increasing model size or computational cost.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If fixed window sampling is used to focus on recent user interactions, then short-term user behavior is prioritized for immediate recommendations, but long-term user preferences and historical interaction patterns are lost
Solution Approach 1:
The patent divides the user interaction history into multiple segments using sliding windows of different sizes (e.g., 100, 500, 1000 interactions). This segmentation allows the system to process and analyze both recent and historical interactions separately, preventing information loss while maintaining computational efficiency. Each window segment can be processed independently and combined to form a comprehensive user profile.
Solution Approach 2:
The patent implements dynamic window sampling where the window size and position are adjusted based on the user's interaction history and preferences. The sliding window dynamically shifts through the interaction sequence, allowing the system to adaptively balance between recent and historical data. This dynamic approach enables the system to prioritize recent interactions when needed while still capturing long-term patterns when appropriate.
2Loss of information
If the fixed window size is increased to include more historical interactions, then long-term user preferences are captured, but model size and computational cost increase
Solution Approach 1:
Instead of processing the entire interaction history as a single large window, the patent segments it into multiple smaller windows of manageable sizes. This segmentation reduces the computational burden on the model while still capturing long-term patterns through the collective analysis of multiple segments. The model processes smaller, more manageable data chunks rather than one large complex input.
Solution Approach 2:
The patent uses partial action by sampling only a portion of the user interaction history at a time through sliding windows, rather than processing the complete history. This partial approach allows the model to capture essential long-term patterns without the excessive computational cost of processing all historical data simultaneously. The system strategically selects representative samples that provide sufficient information for accurate recommendations.
3Device complexity
If only the most recent user interactions are used for training, then computational cost is minimized, but recommendation accuracy decreases due to inability to capture long-term preferences
Solution Approach 1:
The patent employs dynamic window sampling that adapts the focus between recent and historical interactions based on the user's behavior patterns. The sliding window dynamically adjusts its position and size to capture both short-term and long-term preferences, improving recommendation accuracy without consistently processing the maximum amount of historical data. This dynamic adaptation optimizes the balance between computational cost and accuracy.
Solution Approach 2:
The patent changes the parameters of the sampling window (size, position, step) to optimize the balance between computational cost and recommendation accuracy. By adjusting these parameters, the system can capture long-term preferences when needed while minimizing processing costs during normal operation. The parameter changes allow the model to achieve high accuracy without always processing the complete interaction history.
Data Source
AI summary
Techniques for training a machine learning model to generate one or more first recommendations include generating, based on user interaction data, a plurality of fixed window samples, generating, based on the user interaction data, a plurality of sliding window samples, and performing, based on the plurality of fixed window samples and the plurality of sliding window samples, one or more training operations to generate a trained machine learning model to generate the one or more first recommendations.