| Summary: | 碩士 === 國立政治大學 === 資訊科學系 === 106 === With the rise of online shopping, the fashion industry has flourished by the mode of internet marketing. Much research has begun to focus on the recommendation of apparel products, helping users to select appropriate items for saving the time and effort in choosing from a huge amount of products, which provides a convenient and fast way for users to get a stylish and good-looking dressing patterns. By aiming at the clothing compatibilities and popular elements as factors of fashion product recommendation, to recommend items that can be matched with each other for online shopping, not only enables people who are not good at wearing to obtain the suitable clothing matching suggestions with today’s fashion trend, but also can increase the profits of business while improving joint sales rate of goods.
This thesis proposes a clothing matching recommendation mechanism based on style and popularity, and implements the clothing recommendation system. First of all, we consider the influence of overall style on the collocation relationships. From a large number of fashion data, we capture the text information of all items in outfits, including the product title, category, description. Then, Latent Dirichlet Allocation (LDA) topic model is employed to calculate the theme distribution for each outfit, and the most representative theme is manually defined as the style of the outfit. In addition, in order to allow the system to effectively discriminate the item category of the input image, and to recommend matching items in heterogeneous categories according the style selected by the user, therefore, we use Convolutional Neural Network (CNN) to train two image classification models. The first one is the category classification model which can recognize the image is top or others. The second one is outfit compatibility classification model which automatically learn whether the item pairs are suitable matching between each other or not. When learning the concept of matching on item images, we utilize the image contents and user preference scores that indicate the popularity of the product as a measure of clothing compatibility.
Finally, we develop a web system for clothing matching recommendation, that provides a platform for users to select the style they like and upload the clothing image. This system can automatically recommend users the suitable clothing products, that are matching with input image in the same style and different categories.
|
|---|
