臺灣博碩士論文加值系統

近年來，透過網路能隨時讓消費者在網路上直接購買各種服飾商品，所帶來的便利性使得網路購物成為人們的主要消費模式，而後續又衍生出穿衣搭配推薦的熱潮。本文針對穿衣搭配目前現有的推薦技術進行了研究，並提出了一種能夠使用在網路購物平台上的符合用戶個性化風格的搭配推薦方法。
本研究使用深度學習技術與框架，實作一套以風格向量空間為基礎的個人化推薦系統，本系統使用公開的 50 萬張服飾圖片資料集、34 萬筆專家提供的搭配集數據及 1000 萬用戶歷史購買數據，本文的主要內容包括以下幾個方面：首先使用資料庫中多張服飾單品之圖片集合轉換為「特徵向量空間」，再使用此特徵向量空間加上專家推薦集數據訓練出「風格向量空間」，最後一個階段使用餘弦相似度演算法，在訓練後的模型實做個性化推薦系統，在風格向量空間內去找出與用戶風格相近的其他使用者，利用關連使用者的歷史購買紀錄來做個性化的服裝推薦，在最後實驗中，約有 70% 的使用者滿意本系統推薦的結果。茲證明本推薦系統在基於風格的推薦上有一定的成效。

In recent years, customers can buy all kinds of fashion products online at any time. Fashion online shopping becomes more and more popular due to it's convenience. Therefore, the clothing recommendation system has become more important for those online shopping web- sites as well. We study the exciting fashion recommendation system and proposes a clothing recommendation system based on personalized collaborative filtering and style vector space.
In this study, we uses deep learning techniques to implement a personalized recommendation system based on style vector space. This system totally uses 500,000 clothing pictures, 340,000 clothing matching data provided by experts and 10 million online shopping history. First, we train the "Feature Vector Space" by multiple clothing pictures. Second, we extend the "Feature Vector Space" to the "Style Vector Space" by using fashion match sets. After that, we can find the users who have similar shopping preference by using cosine similarity in the style vector space. Finally, we can recommend user shopping item based on his/her related users and their shopping history. In our experiment, we find about 70% users are satisfied with our recommendation system.

中文摘要 I
ABSTRACT II
目錄 III
圖目錄 IV
表目錄 V
一、 緒論 1
　1.1 研究動機 1
　1.2 研究目的 1
　1.3 研究架構 2
二、 相關文獻 4
　2.1 類神經網路模型 4
　2.2 深度學習 5
　2.3 卷積神經網路 8
　　2.3.1 VGG-16 卷積神經網路 12
　2.4 自然語言處理 13
　　2.4.1 Word2Vec 詞向量 14
　　2.4.2 CBOW 與 Skip-gram 模型 16
　　2.4.3 Hierarchical Softmax 與 Negative Sampling 抽樣率 17
　2.5 推薦技術概述 18
　　2.5.1 基於人口統計學的推薦（Demographic-based Recommendation） 19
　　2.5.2 基於內容化推薦（Content-based） 20
　　2.5.3 基於協同過濾推薦（Collaborative filtering） 21
　2.6 KNN 分類演算法（K-nearest Neighbors Algorithm） 23
三、 系統需求分析與設計 26
　3.1 名詞定義 26
　3.2 需求分析 27
　3.3 訓練數據描述 30
四、 實驗 32
　4.1 實驗的開發環境 32
　4.2 實驗流程及步驟 32
　　4.2.1 基於 CNN 的特徵向量空間的模型訓練 33
　　4.2.2 基於 Word2Vec 的風格向量空間的模型訓練 37
　　4.2.3 模型測試 - 餘弦相似度 40
　　4.2.4 基於風格向量空間的個性化協同過濾服裝推薦系統 44
五、 實驗結果與結論 48
　5.1 實驗結果 48
　5.2 結論及未來展望 51
六、 參考文獻 52
附錄 54

[1] Hanbit Lee, Jinseok Seol, Sang-goo Lee,“Style2Vec: Representation Learning for Fashion Items from Style Sets”, ARXIV, eprint arXiv:1708.04014, 08/2017.
[2] Yang Hu, Xi Yi, Larry S. Davis, “Collaborative Fashion Recommendation: A Functional Tensor Factorization Approach”, ACM, Brisbane, New York, NY, USA, Pages 129-138, 10/2015.
[3] Yong-Siang Shih, Kai-Yueh Chang, Hsuan-Tien Lin, Min Sun, “Compatibility Family Learning for Item Recommendation and Generation”, ARXIV, eprint arXiv:1712.01262, 12/2017.
[4] Qiang Liu, Shu Wu, Liang Wang, “DeepStyle: Learning User Preferences for Visual Recommendation”, ACM, Shinjuku, New York, NY, USA, Pages 841-844, 11/2017.
[5] Karen Simonyan, Andrew Zisserman, “Very Deep Convolutional Networks for Large-Scale Image Recognition”, ARXIV, eprint arXiv:1409.1556, 09/2014.
[6] Tomas Mikolov, Kai Chen, Greg Corrado, Jeffrey Dean,“Efficient Estimation of Word Representations in Vector Space”, ARXIV, eprint arXiv:1301.3781, 01/2013.
[7] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, Jeffrey Dean, “Distributed Representations of Words and Phrases and their Compositionality”,ARXIV, eprint arXiv:1310.4546, 10/2013.
[8] Yoav Goldberg, Omer Levy, “word2vec Explained: deriving Mikolov et al.'snegative-sampling word-embedding method”, ARXIV, eprint arXiv:1402.3722, 02/2014.
[9] Xin Rong, “word2vec Parameter Learning Explained”, ARXIV, eprint arXiv:1411.2738, 11/2014.
[10] Laurens van der Maaten, GeoffreyHinton, “VisualizingDatausingt-SNE”, Journel of machine learning research, 9(Nov):2579--2605, 2008.
[11] Cosine_Similarity, 取自http://mlwiki.org/index.php/Cosine_Similarity#Cosine_Similarity
[12] Machine Learning :: Cosine Similarity for Vector Space Models, 取自 http://blog.christianperone.com/2013/09/machine-learning-cosine-similarity-for-vector-space-models-part-iii/
[13] Python for Data Analysis Part 28: Logistic Regression, 取自 http://hamelg.blogspot.com/2015/11/python-for-data-analysis-part-28.html
[14] Activation function, 取自 https://en.wikipedia.org/wiki/Activation_function
[15] Activation Functions: Neural Networks, 取自 https://towardsdatascience.com/activation-functions-neural-networks-1cbd9f8d91d6
[16] CS231n Convolutional Neural Networks for Visual Recognitio, 取自 http://cs231n.github.io/neural-networks-1/
[17] An Intuitive Explanation of Convolutional Neural Networks, 取自 https://ujjwalkarn.me/2016/08/11/intuitive-explanation-convnets/
[18] sklearn.manifold.TSNE, 取自 http://scikit-learn.org/stable/modules/generated/sklearn.manifold.TSNE.html
[19] Deep Convolutional Neural Networks with transfer learning for computer vision-baseddata-driven pavement distress detection, 取自 https://www.researchgate.net/figure/A-s chematic-of-the-VGG-16-Deep-Convolutional-Neural-Network-DCNN-architecture-tr ained_fig2_319952138
[20] 機器學習(4) 類神經網路 Neural network, 取自 http://mropengate.blogspot.com/2015 /06/ch15-4-neural-network.html
[21] Neural Network 的概念探討, 取自 https://ithelp.ithome.com.tw/articles/10191528?sc =iThelpR
[22] 資料分析&機器學習:卷積神經網絡介紹(Convolutional Neural Network), 取自 https://medium.com/@yehjames/
[23] 計算機視覺與卷積神經網路, 取自 https://www.readhouse.net/articles/160795770/
[24] 激活函數, 取自https://feisky.xyz/machine-learning/neural-networks/active.html
[25] T-SNE 完整筆記, 取自 http://www.datakit.cn/blog/2017/02/05/t_sne_full.html
[26] [Machine Learning] kNN分類演算法 , 取自 https://medium.com/@NorthBei/machine-learning-knn%E5%88%86%E9%A1%9E%E6%BC%94%E7%AE%97%E6%B3%9 5-b3e9b5aea8df