Research On Session-Aware Recommendation Problem Based On Multi-user Shared Accounts

Traditional recommender systems usually construct the user's portrait based on the user's identity and long-term historical interactions on the website,so as to fully explore the user's preference and make personalized recommendation.These methods usually focus on user's static long-term preference and decompose user's historical interaction sequences into multiple static user-item records,thus ignore user's preference shift over time.In addition,in some scenarios,user's identity may not be available.For example,some users take privacy seriously,and refuse to provide their identity information,or access the application system anonymously.In this scenario,traditional recommender systems that rely on user identity will not be applicable.An effective solution is to use the session as the basic unit of recommendation.User's interactions within the time that the user logs in and out of the system are organized into a session.The goal of session-based recommendation aims to predict the next interaction for each independent session.Session-aware recommendation is a special form of session-based recommendation,which considers both the user's shortterm behavior in the current session and the long-term preference implied in the historical sessions,and predicts the user's next interaction based on the historical sessions and the current session,so as to achieve better recommendation results.Though existing methods have achieved promising results in respective application fields,they still have drawbacks in some aspects.One the one hand,most existing deep learning methods model a session as a sequence,which only model single-way transition relationships between consecutive items,and neglect the complex transitions between items that are far away in the sequence.Besides,they also ignore the influence of other auxiliary information such as item dwell time.One the other hand,a single account is usually regarded as a single user by default,where the scenario of multiple users sharing the same account is ignored.Therefore,we propose a multi-user identification network named MISS for shared-account session-aware recommendation,which can identify different latent users and the current user based on the historical sessions and current session,so as to make personalized recommendation for the current user.MISS consists of two core components.One is the Dwell Graph Neural Network(DGNN),which builds all sessions of each account into a weighted directed graph.Item dwell time that reflects user's interest is also incorporated into the graph.Then we use the graph neural network to capture the complex item transition relationships in the dwell session graph and generate the feature vectors of all sessions.The other is a Multiuser Identification Network(MIN),which identifies different latent users under shared account and makes personalized recommendations to current user.First,we assume that each account is shared by M latent users.For each user,we use the self-attention mechanism to calculate the probability of each historical session belonging to this user.We sum up the historical session vectors to obtain the long-term preference feature of this user.Then,by comparing current session with these latent user feature vectors,we calculate the probability of each latent user being the current user and obtain the current user's long-term preference feature.Finally,we concatenate user's local interest in the current session and general preference to generate the final current user vector,and then the personalized recommendation task for the current user is completed.The main works and contributions of this paper are summarized as follows:1.A new task of shared-account session-aware recommendation is introduced,and a novel model MISS is proposed to solve this problem.MISS can capture complex item transition relationships in the historical and current sessions,and identify different latent users and current user in the shared account,thus to make personalized recommendations for the current user and improve the recommendation performance.2.A dwell graph neural network(DGNN)is proposed to extract session features,and a multi-user identification network(MIN)is proposed based on self-attention mechanism.DGNN models account sessions as graph,and incorporates item dwell time into model,which jumps out of the traditional sequential modeling idea so as to better capture the complex transition relationship between items.Based on the interest preference extracted from historical and current sessions by DGNN,MIN utilizes the self-attention mechanism to identify different latent users and current user under the shared account,and captures both the long-term preference and local interest of current user,so as to make real personalized recommendation to the current user.3.Two data sets with shared-account characteristics are constructed based on the watching logs collected from Hisense Internet TV platform,and extensive experiments are conducted on the two data sets.The experimental results show that the MISS model performs better than all baselines in two evaluation metrics of HR and MRR.In addition,we design an ablation study to verify the performance of the two core components,and analyze the influence of the hyper-parameter M on the recommendation results.