Research And Application Of Top-N Personalized Recommendation Technology Based On Deep Learning

The flourishing of information technology is a prelude to the early 21^st century,which boosts the evolution of economic model towards the combination of reality and virtualization.Especially,the ubiquitous Internet provides an online promotion platform for the entity industry,like catering,news and commodity transaction,as well as a global-scale highway for the dissemination and sharing of information.And people become information consumers in virtual platforms due to the convenience.In such circumstance,the amount of information soars up exponentially and intricate data disrupt the links between vendors and consumers on various online platforms which are unfavorable for business revenues and user experience.Therefore,it's imperative to put forward a kind of stable and powerful Top-N recommendation technology.And nowadays,the potential service values and economic benefits behind it have attracted a swarm of scholars to the relevant research.Recommendation algorithm is the core of the custom service,which delivers the Top-N recommendation service according to the predication and ranking of user ratings or user behavior commonly.However,easily-implemented traditional ones with simple structures would result in significant errors while predicting user ratings and behavior,which get stuck in the performance bottleneck.Deep learning has gradually dominated the development of recommendation algorithm in recent years,based on which hybrid models spring out and bring about remarkable promotion on the quality of Top-N personalized recommendation in large-scale scenarios,whereas they are lack of the deliberation about merchant benefits.Even worse,complicated hybrid models put tremendous pressure on the system response pf service.Thus this paper absorbs deep learning technical framework and delves into novel recommendation models for Top-N personalized recommendation,based on user ratings and their dynamic behaviors,in order to improve the service experience of users.What's more,taking the feasibility and economic benefit of real scenarios into consideration,this paper comes up with a complete design of the recommender system,with load optimization and recommendation algorithms as the key components,expecting to make a win-win service situation for users and merchants.The main research works and innovative achievements are summarized as below:(1)Puts forward a novel model called Attention-based Collaborative Convolutional Network,abbreviated as ACCN,which is adequate to reduce the error and cost of user rating prediction for Top-N recommendation.It enables ratings in history as side information alternatively,and allocates user attention to each attribute of the designated item through the designed attention mechanism,after which 2D-row CNNs realize the interactions among high-order feature groups.This paper simulates the recommendation scenarios on two real world data sets,relatively large Movie Lens-1M and large-highly-sparse Niconico respectively,which prove that the proposed ACCN reaches improvements of more than 12.61%for RMSE and 18.45%for MAPE,and saves 2%?31%training time cost,in comparison with PMF,CFN and NFM.(2)Proposes a novel prediction algorithm on dynamic features called Multi-Temporal-Scale Transformer(MTS Transformer)at first,for the goal of improving the service dynamics and personalization of Top-N recommendation,with LSTMs as multi-period encoders and concurrent transformers as decoders,increasing the time span,efficiency and accuracy of feature prediction.Based on the ACCN introduced in clause(1),this paper further proposes a MTS-Transformer-boosted Dynamic ACCN for recommendation(MTST ACCDN).With AUC,HR@N and NDCG@N as assessment metrics,the recommendation simulation is implemented on aforementioned data sets again,which verifies that designed MTST ACCDN is superior to RRN,NCF,x Deep FM and CFM in terms of user behavior prediction in context and Top-N personalized recommendation.(3)Aiming at the applicability of Top-N custom service,a complete design of the recommender system with load optimization is proposed.The scheme consists of multiple recall routes based on first-order user feature statistics,rule-based filters and the quantile-based fusion for multiple recommendation lists,etc.It arranges MTST ACCDN in parallel with the transferred GBDT-LR to output the hybrid Top-N recommendation.On the basis of the simulation in clause(2),this paper sets coverage and diversity as additional economic metrics.And the ablation experiments indicate that the designed components have expected effects in which systematic coverage and diversity increase minimally by 1.52%and 10.4%respectively at the expense of the loss in service quality ranging from 1.1%to 4.3%.This architecture also helps to alleviate the recommendation load by 25.2%and 42.4%separately in two scenarios.So the recommender system in paper reaches the well-rounded performance in recommendation service and economic benefits.