Temporal Vector Embedding for Time-Aware Recommendation Stability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current collaborative filtering solutions in e-commerce fail to account for shifts in customer interest over time, leading to inaccurate item recommendations as they either focus on average customer interest or require short time cutoffs, resulting in the loss of historic user interest data.
Innovation Solution
A system that generates personalized item recommendations using a trained prediction model incorporating temporal information embedded into a finite-dimensional vector space, which captures and models user interactions and time values to provide time-aware recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Stability of the object's composition
If collaborative filtering solutions focus on average or long-term customer interest, then recommendation stability is improved, but adaptability to shifts in customer interest deteriorates
Solution Approach 1:
The patent segments customer interest into multiple temporal components: long-term interest and short-term interest. This is achieved by dividing the interaction sequence into different time windows, where long-term interest captures stable preferences and short-term interest captures recent behavioral patterns. The segmented approach allows the system to maintain stability from long-term data while adapting to changes through short-term data.
Solution Approach 2:
The patent implements dynamic weighting of temporal components that adapts to individual user behavior patterns. The model dynamically adjusts the contribution of long-term versus short-term interest based on user-specific characteristics and interaction patterns, enabling the system to be both stable and adaptive simultaneously.
2Adaptability or versatility
If sequential modeling solutions use short time cutoffs to account for shifts in customer interest, then adaptability is improved, but loss of historic user interest data increases
Solution Approach 1:
The patent segments the time window into multiple components rather than using a single short cutoff. By dividing interactions into long-term and short-term segments, the system preserves historic data for long-term interest modeling while using recent data for short-term interest modeling, thus avoiding information loss while maintaining adaptability.
Solution Approach 2:
The patent adds a temporal dimension to the analysis by considering interactions at different time scales simultaneously. Instead of flattening time into a single sequence with a cutoff, the model introduces multiple time window dimensions, allowing historic and recent interactions to coexist and contribute differently to recommendations.
3Loss of information
If collaborative filtering solutions use long time windows to capture historic interest, then information completeness is improved, but responsiveness to current preferences deteriorates
Solution Approach 1:
The patent segments the temporal information into distinct long-term and short-term components. The long-term component captures complete historic interest patterns, while the short-term component captures current preferences. This segmentation allows the system to maintain information completeness while remaining responsive to current user needs.
Solution Approach 2:
The patent applies different quality characteristics to different temporal segments. Long-term interest data is processed to capture stable, comprehensive user preferences, while short-term interest data is processed to capture dynamic, current preferences. Each segment is optimized for its specific temporal characteristic, improving overall system performance.
Data Source
AI summary
System and method for generating a ranked list are disclosed. A plurality of prior interactions for a first customer are received by a computing device. Each of the prior interactions includes a product interaction and time. A ranked list of item recommendations is generated based on the plurality of prior interactions. The ranked list of item recommendations is generated by a trained prediction model trained using temporal information embedded into a finite-dimensional vector space. The ranked list of item recommendations is output by the computing device.