根據過去的文獻研究，大部分的 Next-item 預測問題都專注於預測使用者可能感興趣的下一個項目。雖然有些研究會利用時間資訊來幫助預測使用者下一個互動的項目，或是預測使用者可能互動的項目及時間間隔，但目前尚未有研究同時預測使用者可能感興趣的下一項目、時間間隔及互動持續時間。然而，在大量的先前順序推薦模型研究中，圖形神經網路(GNN)被發現能夠充分納入整體資訊，增強資訊編碼的完整性，從而提高下一個項目預測的準確度。因此，本研究提出了一種名為 GSMRecIT 的模型，利用 GNN 圖神經網路對用戶序列進行嵌入處理。同時該模型使用 Transformer 以及注意力網路來決定不同序列資料的權重，以捕捉用戶的長期和短期偏好。並針對使用者感興趣的項目、時間間隔和互動持續時間進行 Top-N 預測。實驗結果顯示，本研究的方法在項目、時間間隔、互動持續時間三個方面的預測任務中都提高了預測的準確率。;Based on existing literature, most studies on next-item prediction focus on predicting the next item of interest to users. While some research incorporates temporal information to predict the next user interaction item or both the item and the time interval until the next interaction, there is currently no study that simultaneously predicts the next item, the time interval, and the duration of the interaction. However, extensive research on sequential recommendation models has shown that Graph Neural Networks (GNNs) effectively integrate overall information and enhance the integrity of information encoding, thereby improving the accuracy of next-item prediction. Therefore, this study introduces GSMRecIT, a novel model that leverages GNNs for embedding processing on user sequences. Additionally, the model incorporates Transformer and attention networks to determine the weights of different sequence data, capturing users′ long-term and short-term preferences. Moreover, the model provides top-N predictions for items, time intervals, and interaction durations that users are interested in. Experimental results demonstrate that the proposed approach significantly enhances the prediction accuracy across all three aspects: item, interval and duration.
