Recommendation Algorithm With Session-based Model

With the development of the Internet and the arrival of the era of big data,film and television information,electronic goods and other items are increasingly rich.The recommendation system can extract valuable information from the user's history log to quickly find their favorite items and provide recommendation results for users.The problem of recommendation system can be distinguished from the feedback of users to recommendation results,which can be divided into behavior prediction and score prediction.Aiming at these two problems,this paper proposes a fusion interest recommendation algorithm based on time series model and a scoring model based on product quality.In the behavior prediction,the recommendation system will give the recommendation result to the user at the next time step,such as which movie he may watch and which song he may listen to.The more accurate the distribution of user interest captured by the recommendation algorithm,the better the recommendation effect.In the past,most of the algorithms only considered one type of user's interest,either the static interest of users was considered while the dynamic interest was ignored,or only the dynamic interest of users was extracted without combining the static interest.The two interests of users are complementary.Based on this,we propose a hybrid interest model,which integrates users' static interest and dynamic interest.Firstly,the user's unique hot code is used as the original expression to extract the user's static interest.At the same time,the consumption history of each user is input into the cyclic neural network in time sequence,and then the attention layer is used to extract the dynamic interest of users.Finally,we use adaptive mechanism to fuse the extracted user interests and get the final user interest expression.We implemented the algorithm in Movielens and TaoBao datasets.Compared with the existing algorithms,our model achieves better results on Recall@N and Merr@N indicators.In addition,we also designed experiments to verify the validity of our model.In the rating prediction,the recommendation system predicts the user's rating of other items,and then gives recommendations according to the rating level.Users have both positive and negative ratings on items,and the reason why users give different ratings is the quality of items.Therefore,for rating prediction,the recommendation system not only captures the user's interest distribution,but also obtains the quality information of the items.Existing recommendation algorithms such as collaborative filtering recommendation often only consider the static interest of users,while time-series recommendation algorithm based on cyclic neural network ignores the quality information of items,which restricts the recommendation effect.In order to solve the above problems,based on our proposed hybrid interest model,this paper further proposes a rating model based on the quality of goods.By inputting the user's rating matrix,we can learn the relevant information between goods and objects,and then combine it with the integrated interest model,so that we can not only capture the user's interest distribution,but also extract the quality information of goods.We validate the model on the Movielens dataset.Compared with the existing algorithms,our model improves both on MAE and RMSE.At the same time,we designed relevant experiments to analyze the impact of hidden space feature dimension and other factors on the model.