A fashion product recommendation based on adaptive VPKNN-NET algorithm without fuzzy similar image

Area of Science:

• Computer Science
• Artificial Intelligence
• E-commerce Technology

Background:

• Recommender systems are crucial for e-commerce, aiding users in large product catalogs.
• Visually driven sectors like fashion present challenges for traditional keyword-based recommendation systems.
• Subjective style preferences are difficult for conventional methods to capture effectively.

Purpose of the Study:

• To propose a novel fashion recommendation framework utilizing an Adaptive VPKNN-net algorithm.
• To enhance the assessment of visual similarity in fashion items.
• To address limitations of existing systems in subjective style preference detection.

Main Methods:

• Deep visual feature extraction using a pre-trained VGG16 Convolutional Neural Network (CNN).
• Dimensionality reduction via Principal Component Analysis (PCA).
• A modified K-Nearest Neighbors (KNN) algorithm integrating Euclidean and cosine similarity metrics.

Main Results:

• The proposed system achieved high accuracy at 98.69%.
• Demonstrated lower Root Mean Square Error (RMSE) of 0.8213 and Mean Absolute Error (MAE) of 0.6045 compared to baseline models.
• Outperformed Random Forest, Support Vector Machine (SVM), and standard KNN.

Conclusions:

• The Adaptive VPKNN-net framework significantly enhances precision, interpretability, and efficiency in visual fashion recommendations.
• Effectively overcomes limitations of fuzzy similarity models.
• Provides a scalable solution for visually oriented e-commerce, particularly for cold-start and low-data conditions.