基于改进的深度Q网络结构的商品推荐模型

传统推荐方法存在数据稀疏和特征识别差等问题，为了解决这些问题，根据隐式反馈构建具有时序性的正负反馈数据集。由于正负反馈数据集和商品购买具有强时序性特征，引入长短期记忆（LSTM）网络作为模型构件。考虑用户自身特征和用户动作选择回报由不同的输入数据决定，对竞争架构的深度Q网络进行改进，融合用户正负反馈和商品购买时序性，设计了基于改进的深度Q网络结构的商品推荐模型。模型对正负反馈数据进行区分性训练，对商品购买的时序性特征进行提取。在Retailrocket数据集上，与因子分解机（FM）模型、W&D模型和协同过滤（CF）模型中表现最好的相比，所提模型的准确率、召回率、平均准确率（MAP）和归一化折损累计增益（NDCG）分别提高了158.42%、89.81%、95.00%和67.57%。同时，使用DBGD作为探索方法，改善了推荐商品多样性低的缺陷。

Commodity recommendation model based on improved deep Q network structure

Traditional recommendation methods have problems such as data sparsity and poor feature recognition. To solve these problems, positive and negative feedback datasets with time-series property were constructed according to implicit feedback. Since positive and negative feedback datasets and commodity purchases have strong time-series feature, Long Short-Term Memory (LSTM) network was introduced as the component of the model. Considering that the user’s own characteristics and action selection returns are determined by different input data, the deep Q network based on competitive architecture was improved: integrating the user positive and negative feedback and the time-series features of commodity purchases, a commodity recommendation model based on the improved deep Q network structure was designed. In the model, the positive and negative feedback data were trained differently, and the time-series features of the commodity purchases were extracted. On the Retailrocket dataset, compared with the best performance among the Factorization Machine (FM) model, W&D (Wide & Deep learning) and Collaborative Filtering (CF) models, the proposed model has the precision, recall, Mean Average Precision (MAP) and Normalized Discounted Cumulative Gain (NDCG) increased by 158.42%, 89.81%, 95.00% and 65.67%. At the same time, DBGD (Dueling Bandit Gradient Descent) was used as the exploration method, so as to improve the low diversity problem of recommended commodities.