Sequential Recommendation Algorithm Based On Deep Learning

With the emergence of new forms of data such as the internet,social networks,and mobile applications,sequential recommendation algorithms have received increasing attention.Sequential recommendation algorithms can analyze users’ historical behavior sequences,uncover their preferences and behavior patterns,and provide personalized recommendation services based on this information.However,when cold-start users join,sequential recommendation methods cannot accurately model user preferences due to the lack of sufficient historical interaction data,known as the cold-start problem of sequential recommendation.Existing solutions include content-based recommendation,popularity-based recommendation,and collaborative filtering-based recommendation methods.Among them,content-based recommendation can utilize product feature information to recommend similar products,popularity-based recommendation recommends popular products to users according to certain rules,while collaborative filtering-based recommendation can utilize users’ historical behavior data and similarity between other users to make recommendations.In addition,there are some novel solutions,such as reinforcement learning and deep learning methods.These methods can continuously improve recommendation performance by learning user feedback information.However,these methods typically require more data and computational resources,and have higher implementation difficulty.Recently,researchers have used meta-learning to leverage the advantages of few-shot learning in addressing coldstart user recommendation as a new task.By using prior experience gained from training tasks,the model can quickly adapt to new tasks and generate personalized recommendation lists based on user preferences.However,existing meta-learning-based solutions for addressing cold-start problems in sequential recommendation still have the following shortcomings:(1)Many meta-learning-based methods require auxiliary information during training or focus on mining the correlation between items within a sequence and modeling user interests in a static manner,ignoring the influence of item correlations between sequences.(2)Many approaches require auxiliary information when modeling user preferences,and their performance is affected when this information is missing.(3)Most methods only model single-type interactions such as click,and do not take into multi-type user-item relationships.(4)Most of the existing solutions for addressing cold-start problems cannot be integrated with deployed recommendation algorithms,and redeploying the recommendation system would incur significant costs,making the proposed algorithms difficult to apply in actual production.To address these issues,this thesis leverages the idea of meta-learning and combines it with graph neural networks to innovate on existing research and design better models for addressing cold-start problems.The main contributions of this thesis are as follows:(1)Due to the fact that many meta-learning based methods require auxiliary information during training or focus on mining the relevance of items within a sequence and model user interests statically,they ignore the impact of item correlations between sequences.The performance of these methods are affected when auxiliary information is missing or when user preferences change rapidly in the short term.We propose a meta-learning and graph transfer learning sequential recommendation model for cold-start users.It is based on a user-item bipartite graph constructed from the user interaction sequence and uses graph neural networks to model the high-order relationships between items in the sequence.To dynamically model user interests,we design a sequence encoding module to capture the transition relationships between items in the sequence and generate specific preference representations for each user.To adapt the cold-start recommendation task,we use meta-learning to train the model and optimize initial model parameters by learning common behavioral patterns from users with rich interaction data.This enables the model to accurately model the interest preferences of cold-start users and generate personalized recommendation lists with only one or a few gradient updates.(2)In cold start recommendation tasks,purchase behavior data is extremely scarce,and a recommendation model based solely on purchase behavior is difficult to provide high-quality recommendations for new users without the history of purchase records.If other easily occurring behaviors,such as clicks or browsing,can be fully utilized,it can enrich user and item feature information,help the model to model user preferences,and provide high-quality recommendation content to solve the cold start problem.A meta-learning based multi-behavior sequence recommendation model is proposed.The model defines the initial embedding as the basic feature and the features learned from each behavior as the behavior feature.It constructs a graph consisting of item nodes and multiple behavior edges,and uses graph convolutional networks to mine the behavior feature information between items.To capture the behavior features of each user,a user behavior representation layer is constructed,combined with a multi-behavior attention mechanism,and dynamically models the user’s interaction sequence while using behavior feature information to generate preference representations for cold-start users.In order to adapt to the multi-behavior recommendation task in the cold-start scenario,the model is trained by using meta-learning method.In the training stage,different training tasks are divided according to the target behavior,and the user behavior layer parameters are updated through inner loops.In the outer loop,all task information is aggregated to update all model parameters,enabling the model to perform multi-behavior recommendation tasks for cold-start users with just one or a few gradient descents.(3)Although meta-learning methods can help models overcome the cold start problem,many existing meta-learning-based model algorithms cannot be combined with the recommendation algorithms already used by online platforms and require the deployment of a new recommendation system.To this end,this paper designs a general framework MCSR for cold start based on the meta-learning method,which includes a graph embedder and a metatrainer.To address the problem of data sparsity,this method constructs a graph from user interactions,and aggregates neighbor information to obtain user and item embedding,and provides accurate embedding representations of users and items for existing sequence recommendation algorithms.During model training and updating,the meta-trainer is used to update the framework parameters,enabling existing recommendation algorithms to quickly update and adapt to cold-start recommendation tasks.