結合購買時間間隔資訊之類別感知序列推薦系統

Abstract

序列推薦系統的目的是以用戶過往購買紀錄所形成的序列，預測用戶在下一次購買時有興趣的商品。過往的推薦系統由於僅考慮商品項目層級的關聯性，因此忽略了商品類別層級的購買間隔資訊。本論文擴展 TiSASRec 模型，提出了兩個可將類別資訊加入到用戶購買紀錄中的模型: TiSASRecC 及TiSASRecDual，TiSASRecC 模型採用將商品項目嵌入向量表示法與商品類別嵌入向量表示法接合，使類別資訊直接融入到商品表示法中輔助模型學習用戶行為表示法。TiSASRecDual 模型則以類別層級子網路先單獨學習商品類別結合時間間隔的用戶行為特徵，再以學習出的子分類層級特徵，間接影響項目層級子網路中用戶行為表示法的學習。在 Amazon 五個不同性質的類別資料集上之實驗結果顯示:加入類別資訊對序列推薦任務有所提升。本研究提出的兩個模型，在用戶序列長度取樣減短時比 TiSASRec 模型有更佳的推薦效果，而透過將用戶多個類別的購買紀錄形成的序列對模型進行訓練及預測，對單一類別的商品的推薦效果也比 TiSASRec 有更明顯的增進效果。The purpose of a sequential recommendation system is to predict the product that the user is interested in the next purchase based on the sequence of the user's past purchasing records. Most of the previous studies on sequential recommendation systems only considered patterns of item level and ignored information of category level. This thesis extends the TiSASRec model and proposes two models, named TiSASRecC and TiSASRecDual, respectively, to effectively combine category information into the representation of user purchase records. TiSASRecC model concatenates the embeddings of item and its category to make the category information integrated intothe representation of purchase record explicitly to learn better representation of user behavior. TiSASRecDual model uses a category-level subnetwork to learn the patterns of product categories combined with time intervals of purchases. The category-level representation of user behavior implicitly affects the representation of item level. The experiments on the datasets of Amazon with various product categories show that using category information additionally improves the hit ratio of the sequential recommendation tasks. The proposed two models’ better performances than the TiSASRec model when the user sequence length sampling is shortened. Besides, by comparing to using records of a single category, after combining purchase records of the same user on multiple product categories for training and testing, the hit ratios of the proposed models have more significant improvement than TiSASRec.