| In recent years,mobile phones have become indispensable portable devices for people.Nowadays,applications in mobile phones cover all areas of life,and people’s clothing,food,housing and transportation are increasingly intelligent.In all kinds of applications,recommendation system is widely used and relied on because it is suitable for various scenarios,and plays a decisive role in daily life.The development of all walks of life makes it more difficult for people to make a choice when faced with homogeneous goods and services.The existence of recommendation system can dig out personal preferences from massive data,so as to give reasonable "suggestions" when faced with choices.This thesis takes the typical catering recommendation in the field of recommendation as the research object,conducts a series of processing on the real data set,deeply studies the dining behavior of users and understands their interests and needs,aiming at optimizing the catering recommendation system to improve the life efficiency of users.By mining historical behaviors from massive data,on the one hand,the universal laws of objects can be mined;on the other hand,combined with the user’s personal identification,the behavior habits and hobbies of individuals can be analyzed,which helps the recommendation system to "understand" information and scenes more intelligently.In the mass historical data,the more intuitive processing method is to mine the user’s personal information and candidate single product information,further processing the user’s personal information to generate a personal "portrait",the user "portrait" records the user’s behavior habits and hobbies;In-depth mining of candidate product information is conducive to "define" the product through multiple dimensions,and establish "contact" with a wider range of user groups,so as to be recalled and ranked more accurately.However,users’ interests are not invariable.The original method mines the "state" of users at a certain point or time period in the past,which cannot effectively track the evolution of the state.Meanwhile,the expectation of the industry for the recommendation system is more and more focused on timeliness,and a sufficiently intelligent recommendation algorithm can instantly perceive what users want.Traditional methods based on portrait mining can not achieve this goal well.In this thesis,we study the performance optimization of recommendation system based on time sequence characteristics,and propose a Double Attention Units structure based on depth model.First,in the daily recommendation system,the sequence features are divided into two categories: CTR sequence of candidate sets and behavior sequence.The CTR sequence of candidate set refers to the sequence of candidate items clicked by the user history(usually represented by ID).The introduction of candidate sequence is a kind of idea to predict the "dependent variable" based on the historical "dependent variable".The sequence of behavior refers to the user’s historical operation behavior.For example,in the restaurant recommendation scenario,the navigation distance between the candidate restaurant and the user,while strongly correlated with the "dependent variable",also has a certain implied change rule according to the time order.It is an idea of predicting the "dependent variable" based on the "independent variable".In this thesis,two Attention unit structures are used to trace the two kinds of sequences to construct a mechanism to enhance the weight extensibility.Meanwhile,Position Embedding is introduced so that the position information can be expressed in the model.Experimental results show that the proposed method is more effective than traditional deep learning and traditional sequence prediction methods.Secondly,considering the importance of geographical location in life recommendation algorithms,the logical relationship between users and points of interest is often not a simple mapping from users to "points",but an indirect mapping from users to "surfaces" + "surfaces" to "points",because the mapping from users to "points" in real scenarios often has a one-tomany relationship,which is not easy to be mined and modeled,and has a high error rate.However,users to "face" and "face" to point can form the unique mapping between users and points of interest to the maximum extent.Based on this point,this thesis proposes the active business circle mining algorithm,introduces the concept of virtual business circle,and changes the traditional geographical location application mode from simple distance calculation to the calculation of business circle center distance based on the core idea of clustering.Experimental results show that this method can improve the application of geographical location and the overall effect. |
