基于注意力循環(huán)神經(jīng)網(wǎng)絡(luò)的聯(lián)合深度推薦模型

摘" 要： 為了向用戶推薦符合興趣偏好的項目，設(shè)計一種基于注意力循環(huán)神經(jīng)網(wǎng)絡(luò)的聯(lián)合深度推薦模型。將雙層注意力機制設(shè)置于網(wǎng)絡(luò)中，該模型由五個部分構(gòu)成，在輸入層中生成聯(lián)合深度推薦模型的輸入矩陣，通過序列編碼層對項目評論文本語義展開正向和反向編碼，獲得隱藏狀態(tài)輸出，并將其輸入雙層注意力機制中，提取項目特征，利用全連接層提取用戶偏好特征。在預(yù)測層中建立項目與用戶的交互模型，獲得項目評分，為用戶推薦高評分的項目。為了提高模型精度，加權(quán)融合MSE損失函數(shù)、CE損失函數(shù)和RK損失函數(shù)建立組合損失函數(shù)，對深度聯(lián)合訓(xùn)練模型展開訓(xùn)練，提高模型的推薦性能。仿真結(jié)果表明，所提方法具有良好的推薦效果，能夠適應(yīng)不斷變化的市場需求和用戶行為。

關(guān)鍵詞： 雙層注意力機制； 循環(huán)神經(jīng)網(wǎng)絡(luò)； 用戶偏好； 組合損失函數(shù)； 交互模型； 聯(lián)合深度推薦模型

中圖分類號： TN711?34； TP183" " " " " " " " " " 文獻標識碼： A" " " " " " " " " " "文章編號： 1004?373X（2025）01?0080?05

Joint deep recommendation model based on attention recurrent neural network

GUO Dongpo1， HE Bin2， ZHANG Mingyan3， DUAN Chao3

（1. Jianghan University， Wuhan 430056， China; 2. Central China Normal University， Wuhan 430079， China;

3. Zhejiang Provincial Key Laboratory of Intelligent Education Technology and Application， Zhejiang Normal University， Jinhua 321004， China）

Abstract： A joint deep recommendation model based on attention recurrent neural network is designed to recommend projects that meet user interests and preferences. The attention mechanism with double layers is set in the network. The designed model consists of five parts. The input matrix of the joint deep recommendation model is generated in the input layer. By the sequence coding layer， the semantics of the project comment text is encoded forward and backward to obtain the hidden state output. The hidden state output is input into the attention mechanism with double layers to extract the project features. The fully?connected layer is used to extract user preference features. An interaction model between projects and users is established in the prediction layer， so as to obtain project ratings and recommend high?rated projects for users. In order to improve the accuracy of the model， the combined loss function is established based on the weighted integration of MSE loss function， CE loss function and RK loss function. The deep joint training model is trained to improve the recommendation performance of the model. The simulation results show that the proposed method has good recommendation effect， so it can adapt to the changing market demand and user behavior.

Keywords： attention mechanism with double layers; recurrent neural network; user preference; combined loss function; interaction model; joint deep recommendation model

0" 引" 言

由于信息過載問題逐漸嚴重，專家和學(xué)者加大了對推薦系統(tǒng)的研究力度，推薦系統(tǒng)的主要作用是分析用戶的偏好特征，根據(jù)分析結(jié)果為用戶推薦相關(guān)項目。推薦系統(tǒng)通過采集與分析用戶的偏好與交互數(shù)據(jù)對項目評分，根據(jù)評分結(jié)果生成推薦列表，實現(xiàn)項目推薦，目前推薦系統(tǒng)已經(jīng)應(yīng)用在很多網(wǎng)絡(luò)服務(wù)平臺中，實現(xiàn)了個性化服務(wù)[1]，在此背景下研究深度聯(lián)合推薦模型具有重要意義。

文獻[2]方法通過單層神經(jīng)網(wǎng)絡(luò)融合節(jié)點特征，并通過影響因子調(diào)整實體聚合權(quán)重，在用戶與實體特征上展開評分預(yù)測和推薦列表生成，但聚合權(quán)重調(diào)整結(jié)果與節(jié)點特征不符，導(dǎo)致推薦結(jié)果的MAE較高。……