Recommendation Technology And System Implementation Of Conversational Music Recommender System

As one of the most popular products,music has become a significant and unique spice of everyday life.It has many characteristics,such as low consumption cost,various types and the diversity of tastes,which make that it is difficult to response users in time by the traditional offline music recommendation.But the current interactive methods are very limited.In recent years,conversational recommender system is a kind of innovative research topic and attracts great attentions.In this scenario,system can recommend products in conversation,and allows users richer expressions.In this thesis,we will integrate music recommendation into the dialog system,and give a detailed analysis,research and implementation.Different from delay update and implicit feedbacks in the common music recommendation scenarios,conversational music recommendation can capture users' preferences timely and accurately,and obtain feedbacks actively,which can slove the diversity of user tastes.This thesis introduces Conversational Music Recommender System,including music knowledge,appropriate conversational music recommendation algorithms and system's design and implementation.Main study and contributions are as follows:· the construction of Music Knowledge Graph(MKG)This thesis investigates and constructs a music-domain knowledge graph,which includes not only typical music entities like the song,album and artist,but also music genre to help recommendations.MKG adopts a top-down construction manner from the beginning,integrating the music content from multiple data sources,like Chinese general knowledge graph.There are over millions of entities and ten-million of relations in the final MKG.MKG is stored in the Neo4 j.· the appropriate conversational music online recommendation algorithms This thesis divides the conversational music recommendation into two phases: capture user real-time requirements and online recommendation.For the first phase,we adopt reinforcement learning(RL)to solve the sequential conversation decisionmaking problem for system ask-user response interaction mode.The recommender adopts the interactive,real-time and exploratory recommendation algorithms– banditbased methods.Due to the lack of conversational music recommendation data,this thesis uses offline data to simulate the conversational music recommendation users,and evaluates the performance of the above two phases' algorithms.· the design and implementation of the conversational online music recommender system Based on the previous MKG and conversational online recommendation algorithms,this thesis designs and implements a conversational online music recommender system.For music recommendation,it includes not only the common music recommendation but also emotion-based recommendation.Besides,the system also provides memory,Q&A and chat for functional completeness and richness.The remembered user portrait can assist the no-constraint music recommendation.The system is implemented in Python,and a We Chat based service has been deployed for volunteers already.