Research On Trust Recommendation Based On Deep Deterministic Policy Gradient Algorithm

Collaborative filtering algorithms can solve information overload,but they have data sparse and cold start problems.Furthermore,the existing researches show that the recommendation algorithms combined with trust are beneficial to alleviate the data sparsity and improve the interpretability of the recommendation results.However,most of the previous studies assumed that the trust between users was fixed,ignoring its changes in the interaction process.In addition,using a fixed trust value for calculation may cause the recommendation results to gradually move away from the actual needs of users.Most existing recommendation models based on deep learning regard the recommendation process as static,resulting in a fixed recommendation policy,which cannot adapt to the dynamic changes of user interests in real time.The reinforcement learning algorithms can interact with the environment through trial and error,and use the environmental feedback to obtain the maximum reward to achieve the optimal behavior policy.In addition,they can model the user-item interaction process and adjust the recommendation policy flexibly.Deep reinforcement learning algorithms combine the advantages of the two,and are usually used to build models on high-dimensional,continuous action spaces,and can actively learn real-time feedback from users to learn user preferences and adjust policies in time.Therefore,in view of the insufficient researches on trust dynamics in the recommendation process and the applicability of deep reinforcement learning algorithms,a trust recommendation algorithm DDPG-TR based on deep deterministic policy gradients(DDPG)is proposed,which can dynamically update the trust value between users according to the user interaction experience during the recommendation process,and capture the dynamic change of user preference by using the designed state representation module.Firstly,the algorithm uses the embedding method to pre-train,and generates user and item vectors to form the input of the algorithm,and designs a state representation module to model the user's state and obtain user preferences.Then,the algorithm recommends the item to the user based on the current state.After the user receives the recommended item,it integrates the trust and similarity relationships in the social network to predict the rating of the recommended item.Finally,the algorithm calculates the difference between the actual score and the expected score to update the trust value between users,and regard the trust difference as the reward at this moment and input to the network for training and updating.The experiments show that the DDPG-TR algorithm can provide more accurate recommendation results and better interpretability compared with other algorithms.Finally,an attention mechanism is introduced based on the DDPG-TR algorithm,and the DDPGTRA algorithm is proposed.This algorithm adds a layer of attention network to the user state module,which can obtain the user's attention to different items and improve the interpretability of recommendation results.The experimental results show that the DDPGTRA algorithm is better than the comparison algorithms and improves the recommendation accuracy of the DDPG-TR algorithm.