E-Commerce Recall Using Incremental Swing Similarity Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The Swing algorithm for e-commerce recommendation systems is resource-intensive and time-consuming due to data bloating processes that involve traversing user behavior sequences and expanding item pairs, affecting recall efficiency.
Innovation Solution
Perform a full Swing calculation based on a full user behavior sequence to obtain a result, calculate first and second Swing similarities, fuse them, and update the previous day's similarity as a historical result, reducing the need to traverse and expand item pairs each time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If full-graph Swing calculation is performed by traversing all user behavior sequences and expanding all item pairs, then complete item similarity information is obtained, but memory consumption and calculation time increase significantly
Solution Approach 1:
The patent segments the full user behavior sequence into daily user behavior sequences, processing one day at a time. This divides the large-scale full-graph calculation into smaller, manageable daily calculations, reducing memory consumption while maintaining the completeness of similarity information through incremental updates.
Solution Approach 2:
The patent performs preliminary full-graph Swing calculation once to establish baseline item similarity relationships. Subsequently, only daily user behavior sequences are processed to update similarities, avoiding repeated full traversals. This preliminary action reduces ongoing computational resources while preserving information completeness.
2Measurement precision
If full-graph Swing calculation is performed by traversing all user behavior sequences and expanding all item pairs, then complete item similarity information is obtained, but calculation time increases significantly
Solution Approach 1:
The patent segments the calculation process into an initial full-graph calculation phase and subsequent daily incremental update phases. By dividing the computation into temporal segments (full graph once, then daily updates), the patent maintains information completeness while dramatically reducing repeated calculation time.
Solution Approach 2:
The patent performs the computationally intensive full-graph traversal as a preliminary one-time action to establish complete similarity relationships. Future recommendations leverage this pre-computed foundation with minimal daily updates, eliminating the need for repeated full traversals and reducing calculation time significantly.
3Reliability
If item pairs are expanded by item pairs clicked by users in the data bloating process, then comprehensive similarity relationships are captured, but resource consumption increases
Solution Approach 1:
The patent segments the item pair expansion process to occur only once during the initial full-graph calculation. Subsequent daily calculations work with a fixed set of item pairs derived from this initial expansion, maintaining relationship comprehensiveness while avoiding repeated resource-intensive expansion operations.
Solution Approach 2:
The patent performs the resource-intensive item pair expansion as a preliminary action during the initial full-graph calculation. This pre-expansion creates a comprehensive item pair foundation that is reused in subsequent daily calculations, capturing all similarity relationships upfront while minimizing ongoing resource consumption.
Data Source
AI summary
A recall method for an e-commerce recommendation system includes performing a full Swing calculation, based on a full user behavior sequence in the e-commerce recommendation system, to obtain a full Swing result, calculating first Swing similarity of a user behavior sequence of a previous day according to the full Swing result, performing a fusion of the first Swing similarity and second Swing similarity of a historical user behavior sequence, to obtain a target Swing result, and utilizing the target Swing result to update the first Swing similarity of the previous day, and using it as a historical Swing result of a next day, to realize a recall in the e-commerce recommendation system. By using the solution of the present application, resource consumption may be reduced and recall efficiency may be improved.