Research On Commodity Sales Forecast Model Based On Gaussian Process Regression Under Large-scale Data With Uncertain Demand

As machine learning brings more and more revenue to enterprises in the prediction or rec-ommendation field,in recent years,the research on sales forecasting issues has been continuouslystudied in depth,especially the proposal and development of new retail,which makes FMCG mer-chandise sales forecasting research very high-profile.Traditional forecasting uses deterministicpoint forecasting methods to predict FMCG merchandise sales,however,due to the randomnessand non-linear characteristics of demand,it is impossible to show what may happen at the forecastmoment and the confidence level of its appearance.Thus it increases the risk of decision-makingrelying on the points prediction results of FMCG sales.Although Gaussian Process Regression(GPR)can solve this problem,it is difficult to apply to large-scale data sets due to the limitationof computing performance.In view of the shortcomings of the deterministic prediction method ofFMCG sales,this paper proposes to use the Gaussian process to build the model.According to thehistorical sales data of commodities,extract the feature vectors that affect the sales volume,andcombine the sliding window of the sample point selection strategy based on the kernel distance withthe Gaussian process regression method to overcome the Gaussian process regression Computingbottlenecks on large-scale data.This model considers various factors that affect sales,and bases onthe randomness and basis of the factors affecting the sales of FMCG merchandise.We treat FMCGmerchandise sales as a random process under the influence of multiple influencing factors and pre-dict them.Its prediction result has probabilistic significance.While ensuring the accuracy of theforecast,it can also give the confidence level of the forecast,which is conducive to evaluating thedecision-making risk that depends on the forecast results,and provides more scientific guidancefor the business enterprise's decision-making.The work done in this article is as follows:Firstly,it briefly describes the research status of Gaussian process regression and forecastingbased on retail sales data,and introduces related technologies in retail data preprocessing.Andfor the specific application problems based on FMCG retail sales estimation,this paper performsfeature extraction based on correlation coefficients and lifting tree models to reduce the number ofproportional features as much as possible,thereby reducing the dimension of the input data.Secondly,based on the idea of Gaussian process regression and sliding window,a probabilityprediction model of retail product sales based on the sliding window Gaussian process under large-scale data is proposed.The sales volume is regarded as a random process.The distribution of the function is defined by Gaussian process regression.And the sample point selection strategy based on the kernel distance is used to transform the conventional sliding window,and then the two are combined to improve the kernel calculation in the traditional Gaussian Process Regression model Performance issues.The model can quickly adapt to the new observation data,and can evaluate the uncertainty of the prediction,and solve the limitation of the ability of the midpoint forecast of sales forecasting to express the uncertainty and the performance of the nuclear computing limits.Finally,we made a specific study on the Gaussian process regression method and made some improvements to achieve it.We applied the method to the actual scenario and got some results,and compared it with the common machine learning methods.Experiments and experimental results show that the Gaussian process regression method has certain effectiveness in dealing with this type of problem,indicating that the model in this paper overcomes the limitation of too large samples in the calculation of kernel functions,and has better performance and compute efficiency.