以元學習方法建構個人化類別感知序列推薦模型

Abstract

序列推薦的目的是根據使用者以往與項目互動的序列資訊，推薦使用者可能感興趣的下個互動項目。本論文擴展轉移元學習器的模型架構(MetaTL)，採用類別層級(Category-level)及項目層級(Item-level)兩個轉移元學習器(Transitional Meta-Learner)進行結合，稱為 CAI-TML 模型。利用類別層級轉移元學習器學習到使用者較一般性的類別轉移行為特徵，並輸入到項目層級轉移元學習器，以注意力機制影響使用者較個人化的項目轉移行為特徵表示，用來預測推薦的下個互動項目。本論文以 Foursquare Globalscale Check-in 資料集的使用者打卡序列進行實驗評估，實驗結果顯示：本論文所提出的CAI-TML 模型相較於 MetaTL 模型，在對下個互動項目推薦的第一名命中率效能提升比率為 10.2%，項目類別的命中率效能提升 23.8%。此外，對於冷啟動使用者及推薦使用者未曾互動過的項目等特殊情況， CAI-TML 模型亦較 MetaTL 模型發揮更佳的推薦效果。The task of sequential recommendation system is to learn patterns from user-item historical interaction records to recommend the next item that user may be interested. In this thesis, we extended the framework of Meta Transitional Learning (Meta TL) model to combine a category-level transitional meta-learner and an item-level transitional meta-learner, which is called the CAI-TML (Category-Aware Item-level Transitional Meta Learner). The model learns the general features of user behaviors from category-level behaviors through the category-level transitional meta-learner. Then the obtained feature representation of a user is inputted into the item-level transitional meta-learner to influence the result of obtaining personalized features from item-level behaviors through an attention mechanism, which is used to predict the next interactive items for recommendation. The experiments were performed on the Foursquare Global-scale Check-in Dataset. The results show that, the proposed CAI-TML model improves the performance of prediction on top one item hit rate and category hit rate 10.2% and 23.8%, respectively, than the ones of MetaTL. Moreover, in the test cases of cold-start users or cases recommending the items never occurred in user history behavior, the CAITML model performs better than the MetaTL model more significantly.