以類別感知之多任務學習框架增進序列推薦效果

Abstract

序列推薦系統的目的是根據使用者以往與項目互動的序列資訊，預測使用者可能感興趣的下個互動項目主動進行推薦。本論文提出可套用於 GRU/bi- GRU/Caser 類神經網路模型的類別感知之多任務學習框架，利用項目的類別特徵做為項目的上層資訊，輔助模型預測使用者未互動過的下個項目預測任務。此框架以同時學習預測下個互動類別及下個互動項目為目標，在項目模組中將項目互 動序列經過類神經網路學習到項目層級行為表示法，再融合類別模組所學習到的類別層級行為表示法預測使用者下個互動項目。本論文所提出的方法，分別在 Foursquare 及 MovieLens 兩種不同序列強度的資料集上進行實驗，預測命中率的評估結果顯示：本論文提出的類別感知多任務學習框架在預測使用者下個未互動過的項目，相較只以單任務類神經網路模型的效能，在 Foursquare 資料集 Hit@10 最高可提升10.73%；MovieLens 資料集 Hit@10最高可提升7.29%。

The goal of sequential recommendation system is to learn patterns by constructing a model from user-item historical interaction records to predict the next item that user will be interested. In this thesis, a multi-task learning framework was proposed, is applicable on GRU/bi-GRU/Caser neural network model to learn category/item prediction tasks simultaneously. The main idea of the framework is to apply the category feature of item sequence, i.e., the high-level concept of item, for improving the prediction of next item that user has not interacted with. The item-level behavior representation learned in the item module of the framework is fused with the category- level behavior representation learned in the category module to predict the next item. The experiments performed on two datasets, which contain data sequences with different levels of sequential intensity, demonstrated that the proposed category-aware multi-task learning framework could get better performance than thesingle-task learning neural network model. Our approach achieves 10.73% improvement than the baseline model on the Foursquare dataset and 7.29% improvement on the MovieLens dataset as well.