Cross-Domain Recommendation Training With Real Hard Negative Samples
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing cross-domain recommendation (CDR) models suffer from low accuracy due to the focus on randomly extracted negative samples, leading to ineffective information recommendation in target domains.
Innovation Solution
A method that fuses interaction features from source and target domains, filters hard negative samples based on similarity scores, and uses time-sensitive interaction behavior and cluster centers to identify real hard negative samples for training, enhancing the CDR model's accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If randomly extracted negative samples are used for training, then the training process is simple and fast, but the recommendation accuracy in target domain is low
Solution Approach 1:
The patent applies local quality by differentiating the treatment of negative samples based on their characteristics. Instead of uniformly treating all negative samples the same, the system identifies and processes hard negative samples (HNS) separately from easy negative samples. HNS are those with high similarity scores to positive samples, requiring special attention during training. This selective processing improves recommendation accuracy while maintaining training efficiency.
Solution Approach 2:
The patent segments the negative samples into different categories: hard negative samples (HNS) and easy negative samples. This segmentation allows the training process to focus computational resources on the most challenging samples that contribute most to improving recommendation accuracy, while still incorporating easier samples for comprehensive training. The segmentation is achieved through similarity score thresholding between positive and negative samples.
2Device complexity
If feature-level cross-domain correlations of negative samples are focused on, then the training process is simplified, but the precision of information recommendation in target domain is affected
Solution Approach 1:
The patent inverts the conventional approach by not just focusing on feature-level correlations but instead examining sample-level characteristics. Instead of treating all negative samples uniformly at the feature level, the system inverts the approach by identifying individual hard negative samples that require special attention. This inversion from feature-level to sample-level processing significantly improves recommendation precision.
Solution Approach 2:
The patent implements feedback mechanisms during training by continuously evaluating the similarity scores between positive and negative samples. The feedback loop identifies hard negative samples that are difficult to distinguish from positive samples, and these samples receive enhanced weighting or special processing. This feedback-driven approach ensures that the training process focuses on the most informative samples for improving recommendation precision.
3Measurement precision
If hard negative samples are identified and processed separately, then recommendation accuracy is improved, but the processing time and computational resources increase
Solution Approach 1:
The patent applies partial action by not processing all negative samples with the same intensity. Instead, it processes only the hard negative samples (those with high similarity scores to positive samples) with enhanced attention and computational resources. Easy negative samples are processed with standard treatment. This partial application of intensive processing to only the necessary samples improves accuracy while minimizing unnecessary computational time.
Solution Approach 2:
The patent changes the processing parameters dynamically based on sample characteristics. Hard negative samples receive higher weights, more iterations, or specialized processing parameters compared to easy negative samples. This parameter adaptation allows the system to allocate computational resources efficiently, focusing intensity on samples that most impact recommendation accuracy while reducing processing time for less critical samples.
Data Source
AI summary
In a method, a plurality of recommended items in a target domain is obtained. A first interaction feature of a sample object in a source domain is fused with a second interaction feature of the sample object in the target domain to obtain a fused interaction feature. Similarity scores between the fused interaction feature and each of the plurality of recommended items are determined. A plurality of hard negative samples (HNSs) is filtered from the plurality of recommended items based on the similarity scores. The plurality of HNSs is combined into a candidate recommended item set. A third interaction feature is fused with a fourth interaction feature to obtain a transfer interaction feature. A plurality of real hard negative samples (RHNSs) is filtered from the plurality of HNSs based on similarity scores between the transfer interaction feature and each of the plurality of HNSs.