隨著線上時尚零售蓬勃發展，客戶和商家對於服裝推薦的需求也逐漸增長，推薦的商品越能精準得滿足顧客喜好越能提升商家的銷售量。過去服裝推薦領域研究多著重在單品之間搭配程度計算，盡可能讓機器模擬人類的審美觀。然而服裝的審美其實包含許多主觀成分且因人而異，因此在穿搭推薦上不應該忽視個人化偏好因素。個人化資訊的提取過往多仰賴於用戶評分矩陣，並依據協同過濾相關方法預測用戶潛在喜好項目，屬於使用明確資訊的個人化推薦方法。在大數據時代下，科學家試圖探索從其他歷史資料提取有意義的內隱資訊幫助推薦，本研究使用結構化用戶物品Meta-path並搭配深度神經網路，試圖提取內隱資訊幫助個人化偏好的建立。 本研究提出之架構GPA-BPR(General Compatibility and Personal Preference with Co-Attention mechanism) 主要由兩模組建立而成，分別是服裝單品匹配模組和個人化偏好模組。服裝單品匹配模組依據單品其圖像及文字資訊做匹配度計算，並在圖像特徵提取建立實驗找出最適合方法。個人化偏好模組透過用戶單品網路建立之Meta-path，搭配深度網路提取重要之內隱資訊，相較於過往研究有助於提取更多的個人化資訊。 為探討本研究架構中各模組之必要性，本研究建立消融研究並使用真實世界資料集iQon3000作為訓練和驗證之依據，本研究採用AUC及MRR作為模型評估指標，探討訓練模型之判別能力以及推薦排序效果。從實驗結果可以得知本研究提出之複合架構在運用共同注意力機制深度網路可以有效地從Meta-path提取重要特徵，提升模型效能並且在iQon3000資料集上勝過過往研究，同時提高可解釋性。 ;With the huge economic value of online fashion retail, the demand for clothing recommendations by customers has gradually increased. The more accurately the recommended products can meet the preferences of customers, the more the merchants can improve their sales. Most existing research in the field of clothing recommendation focused on the calculation of general compatibility for the items, in order to allow the machine to simulate human’s perspective of aesthetics as much as possible. However, the aesthetics of clothing can be highly subjective, so personal preferences should not be ignored in outfit recommendations. In the past, the extraction of personalized preference relied on explicit information such as user rating matrices using collaborative filtering methods to predict the user′s preferences. In the era of Big Data, scientists are trying to extract meaningful implicit information from other historical data. This study uses structured user-item meta-paths and a deep neural network to extract implicit information to tackle the problem of personalized compatibility. The proposed framework, GPA-BPR (General Compatibility and Personal Preference with Co-Attention mechanism), comprising of two important modules, namely the general compatibility module and the personal preference module. The general compatibility module is mainly based on the image and text information of the item for matching calculation. Also, the image feature extraction experiment is established to find the most suitable method. The personal preference module can be better than previous studies in terms of personalization through the analysis of the user-item meta-path. Moreover, we use a large real-world dataset, iQon3000 for training and evaluation. We use AUC and MRR as model evaluation metrics to investigate the performance of the trained model and the recommended ranking effect. Extensive experiments have demonstrated the proposed architecture can effectively improve the representations for meta-path based context by using the co-attention mechanism, which further improves model performance and outperform previous studies on iQon3000 dataset and has potentially good interpretability.
