Session-Based Recommendation Models for Long-Tail Bias Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Recommender Systems (RS) tend to recommend more popular items instead of relevant long-tail items, and existing methods fail to address biases arising during the training of deep learning models, ignoring skewed class distributions and conformity biases, leading to inaccurate recommendations.
Innovation Solution
A processor-implemented method using deconfounding training of a neural network model with causal graphs to mitigate popularity and conformity biases by identifying and removing indirect effects, normalizing item and session embeddings, and applying softmax functions to obtain relevance scores.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If existing recommendation methods focus on popular items based on data distribution, then the system can leverage abundant interaction data, but it fails to recommend relevant long-tail items and introduces popularity bias
Solution Approach 1:
The patent applies L2 normalization to item embeddings, transforming the parameter space of embedding vectors. This normalization changes the scale and distribution of embedding parameters, allowing the model to equally weigh popular and long-tail items during similarity computation, thereby mitigating popularity bias while utilizing all available training data
Solution Approach 2:
The patent introduces session embeddings as an intermediary representation that captures user intent and context. By computing similarity between normalized item embeddings and session embeddings rather than directly using item popularity, the system mediates the recommendation process to favor relevance over popularity
2Reliability
If re-ranking is applied as a post-processing step, then popularity bias can be partially mitigated, but the skewed class distribution during training is ignored leading to suboptimal recommendations
Solution Approach 1:
The patent performs L2 normalization of item embeddings during the training phase rather than as a post-processing step. This preliminary action ensures that the normalization effect is integrated into the learning process itself, allowing the model to learn from equally weighted item representations from the beginning, avoiding the need for complex re-ranking procedures
3Reliability
If prior information about item popularity is used to balance data, then bias can be controlled, but such prior information is not always available
Solution Approach 1:
The patent makes the recommendation system self-balancing by automatically normalizing item embeddings based on their learned representations rather than requiring external prior information about popularity. The L2 normalization is computed directly from the embedding magnitudes, allowing the system to adapt to any item distribution without needing pre-knowledge of popularity statistics
Data Source
Figure 1
Figure 2
Figure 3A~3B
AI summary
Recommender Systems (RS) tend to recommend more popular items instead of the relevant long-tail items. Mitigating such popularity bias is crucial to ensure that less popular but relevant items are recommended. System described herein analyses popularity bias in session-based RS obtained via deep learning (DL) models. DL models trained on historical user-item interactions in session logs (having long-tailed item-click distributions) tend to amplify popularity bias. To understand source of this bias amplification, potential sources of bias at data-generation stage (user-item interactions captured as session logs) and model training stage are considered by the system for recommendation wherein popularity of item has causal effect on user-item interactions via conformity bias, and item ranking from models via biased training process due to class imbalance. While most existing approaches address only one of these effects, a comprehensive causal inference framework is implemented by present disclosure that identifies and mitigates effects at both stages.