An L₁-and-L₂-Norm-Oriented Latent Factor Model for Recommender Systems.

Area of Science:

• Computer Science
• Data Science
• Artificial Intelligence

Background:

• Recommender systems (RS) excel at filtering information from high-dimensional and sparse (HiDS) data.
• Latent factor (LF) models are a popular approach for implementing RS.
• Existing LF models often use a single distance metric (e.g., L2 norm) for loss, potentially overlooking data characteristics captured by other metrics (e.g., L1 norm).

Purpose of the Study:

• To propose a novel LF model that integrates both L1 and L2 norms to enhance the description of HiDS data, particularly in the presence of outliers.
• To improve the prediction accuracy of recommender systems when dealing with complex datasets.

Main Methods:

• Development of an L1- and L2-norm-oriented LF model ([Formula: see text]).
• The model aggregates the robustness of the L1 norm and the stability of the L2 norm in its loss function.
• Adaptive adjustment of the weights for L1 and L2 norms within the loss function to precisely describe HiDS data.

Main Results:

• The proposed [Formula: see text] model demonstrates significantly superior prediction accuracy for missing data in HiDS datasets compared to state-of-the-art methods.
• Experimental results on nine real-world HiDS datasets validate the model's effectiveness.
• The computational efficiency of the [Formula: see text] model is comparable to existing efficient LF models.

Conclusions:

• The [Formula: see text] model effectively addresses the limitations of single-norm loss functions in LF-based recommender systems.
• Its ability to handle outliers and describe HiDS data makes it a promising solution for real-world applications.
• The model offers improved accuracy and comparable efficiency, highlighting its potential for practical use in recommender systems.