Ekstraksi Fitur Dengan Convolutional Neural Network Dan Rekomendasi Fashion Menggunakan Algoritma K-Nearest Neighbours

Abstrak

The rapid growth of the fashion industry on e-commerce platforms means that fashion can be obtained easily by various consumer segments. Consumer segmentation can be represented in every search for the type of fashion that is desired, but searches for types of fashion in e-commerce are carried out using searches based on string keywords so that consumer segmentation based on fashion characteristics is difficult to do. Fashion is an object that is easily recognized visually so image-based search is very necessary on e-commerce platforms to select fashion based on consumer segmentation. Implementation of image-based search can be done with content-based fashion recommendations with k-nearest neighbor (KNN) to approach fashion features to fashion image input by consumers with each data feature being extracted into a convolution layer in the convolutional neural network (CNN) model and Histogram oriented gradient (HOG) can be evaluated with top-n accuracy against the Resnet, GoogLeNet, VGG, and HOG models with the performance of each model compared so that an accuracy of 93% can be obtained on GoogLeNet with KNN as the best model in fashion recommendations. The approach between feature fashion is based on the label results from the classification process into convolution and fully connected layers in convolutional neural networks (CNN). It can be evaluated using evaluation matrices for the Resnet, GoogLeNet, VGG models with each model's performance compared so that a value can be obtained. accuracy of 99%, precision of 100%, recall of 99%, f1-score of 99% on VGG as the best model for identifying fashion types.

Keywords: Fashion, Ekstraksi Feature, Sistem Rekomendasi, Arsitektur CNN, HOG, KNN, Evaluation Matrices, Top-n accuracy