Research On Purchase Tree Spectral Clustering Algorithm Based On Customer Trading Data

The rapid development of the Internet in the era of big data has brought the market economy to a state of incandescent competition.The demands of new generation consumers on services are also more meticulous.How to formulate effective marketing strategies has become a problem that companies need to think about.Although there are many methods available today for customer segmentation,most of them emphasize numerical calculations rather than business goals.The customer transaction data is sales data of the retail company and accumulates the big data of the customer's daily shopping transaction.In order to generate marketing value for these customer historical data,they are processed using clustering.Customer transaction data is a collection of customer purchasing data recorded by a retail company.Early statistical variables were collected for convenience and ease of use,and became the initial research method.With the rapid increase in customer purchase transaction data,new research turned to the use of specific product variables such as purchased goods.Although there are already some methods of data clustering,these methods are time consuming and cannot handle a large number of transaction records.Most work uses hierarchical clustering algorithms and cannot be extended to large-scale transaction data.PurTreeClust clustering algorithm uses the purchase tree structure to cluster customer transaction big data,but there are still two problems.First of all,it is difficult to adjust the hierarchical weights in the purchase tree distance,and secondly,the clustering performance is not optimized.This paper proposes two new methods to solve the problem of PurTreeClust:First,this paper proposes a local PurTree Spectral Clustering(LPS)method for the difficulty of adjusting the level weight of the purchase tree.This article uses the weighted purchase tree distance to measure the difference between the two purchase trees.In the clustering process,this new method automatically learns the data similarity matrix from the local distance and hierarchical weights,and applies an iterative optimization algorithm to optimize the new model.Second,this paper proposes a two-level subspace weighted spectral clustering(TSW)method for the difficulty of adjusting the hierarchical weights of the purchasing tree and optimizing the clustering performance.In order to better reconstruct the cluster structure hidden in the customer transaction data,learn a group of sparse node weights to represent a few important leaf nodes and reduce the complexity of the distance calculation.The new method learns an adaptive similarity matrix from the hierarchical weights and sparse node weights in the purchased tree subspace distance.Because it is difficult to set the appropriate parameters manually,regularization parameters and iterative optimization methods are used to optimize the new model.The two new methods presented in this paper are experimentally compared on real data sets.The first LPS method is superior to the four commonly used clustering methods.The second TSW is superior to the six commonly used clustering methods and is superior to the first LPS method.According to the experimental results of this paper,the effectiveness of the two new methods in customer transaction data clustering is verified and has certain application value.