Design And Implementation Of Music Recommendation System Based On Multi-Task Learning And User Behavior Sequence

Modeling multiple targets is a common need for recommendation systems.For example,a user may generate various behavior feedbacks such as collecting,downloading,sharing,etc.after listening to a song.Different behavior feedbacks represent that the song meets the needs of users at different levels,and each behavior feedback represents different meanings for different users.A recommendation algorithm that only performs single-objective optimization for a certain type of behavior feedback is difficult to fully meet the needs of different users.Therefore,it is necessary to design a recommendation algorithm to perform multi-objective optimization learning for different types of user behavior feedback.In recent years,multi-task learning has become a standard method to solve such needs.Most of the multi-task recommendation models proposed at present focus on the use of non-serialized input features,but in actual music recommendation scenarios,the input data is usually serialized,and implicit feedback data such as user behavior logs are easier to obtain than explicit feedback data such as ratings.Modeling the user behavior sequence as an explicit sequence representation can make the multi-task model incorporate time dependence to more accurately predict future user behavior.Therefore,music recommendation based on multi-task learning and user behavior sequence is an important but not yet deeply explored issue.This paper proposes a new multi-objective recommendation model MOGA(Multi-Objective GRU with Attention).Based on the multi-task learning and attention mechanism,this model expands the ability of the recurrent neural network GRU to extract multiple behavior sequence information at the same time,and improves the accuracy of simultaneously giving multiple target behavior prediction values with one model.For the first time,the multi-task learning method SNR subnet routing algorithm is combined with GRU to realize the ability to extract multiple behavior sequence information at the same time.And the attention mechanism is designed to weight the input information according to the correlation with the current target behavior,so as to alleviate the problem of low frequency target behavior data being overwhelmed.In addition,the model supports real-time updates of user interests.Finally,this paper designs and implements a private FM music recommendation system.Based on open source computing frameworks such as Hadoop,Storm,Keras,it implements the user's multiple behavior sequence acquisition module,model offline training module,real-time update module,and real-time recommendation module,and applies the MOGA model proposed in this paper to this system.