Research On Consumer Flow Forecast Of Offline Merchant Based On Dynamic Weight Fusion Model

The transformation and upgrading of the retail industry has been a focus of the market in recent years.With the continuous development of big data and machine learning,it is possible for us to combine these two industries to achieve smart retail sales.Precisely predicting the offline store's consumer traffic(that is,those customers who come to the store and consume,hereinafter referred to as the consumer traffic),can help the merchants to implement better operational services,and can also provide suggestions in the supply-chain and warehousing areas.And it can also provide effective advice to reduce the cost of businesses and improve the user's experience.In the existing research,there are some studies using machine learning to analyze and forecast the sales and traffic of the merchant.However,these studies mainly focus on the forecast of e-commerce sales.Due to the differences between the e-commerce model and offline business model,the characteristics that affect the store sales or customer traffic are not consistent,such as e-commerce model does not contain the restaurant business type,the offline business hours are more fixed and so on.Therefore,to better solve the consumer traffic forecasting problem of offline merchants,this paper first analyzes the data distribution and characteristics of the corresponding scenarios and preprocesses the datasets.Based on this and the understanding of the consumer traffic,a variety of practical features are extracted.After that,the classic time series analysis model,linear regression,gradient boosting tree,multi-layer neural network and other methods were used to analyze it.To achieve better prediction results,this paper also studies ensemble learning.ensemble learning learns a number of different models and integrates them.In order to combine ensemble learning with the consumer traffic scenes,this paper proposes a fusion model based on dynamic weights.Compared to the single model and the traditional fusion method,the new model does not give each model a static weight.The dynamic weights will vary according to the performance of corresponding model.Therefore,each sample is given different weights related to the models.And a pre-training model is used to determine the weight of each feature,which avoids the defects of the general fusion method and improves the prediction accuracy of the model.In the end,this paper compares the proposed model and other models on Alibaba's dataset.Through the full analysis of the experimental results,we find that the accuracy and robustness in the proposed model has been improved,which proves the effectiveness of the model,and the practical significance of this article.