A Deep Embedding Clustering Algorithm Considering Preservation Of Initial Clustering Structure And Its Application

Data clustering,which can obtain the inherent relationship and laws of data from unlabeled data,is one of the foundations of data analysis.k-means clustering algorithm is widely used in data clustering because of its simplicity and efficiency.However,the clustering results of k-means are easily effected by the selection of initial centers.What's more,the performance of traditional k-means clustering algorithm is not satisfactory when dealing with the clustering problem of imbalanced data sets or linearly inseparable data sets.This paper proposed firstly proposes an improved global k-means multi-prototype clustering algorithm.Secondly,traditional clustering algorithms often face the problem of the curse of dimensionality when dealing with high-dimensional data,while deep learning algorithms can effectively capture low-dimensional features of high-dimensional data.Therefore,combining the clustering algorithm with deep learning,this paper proposes a deep embedding clustering algorithm considering the preservation of the initial clustering structure.Finally,the proposed deep embedding clustering algorithm is applied into the personalized recommendation problem of financial products.The main work and contributions of this paper are summarized as follows:(1)An improved global k-means based multi-prototype clustering algorithm(IGKM-MPC)is proposed.Although the GKM algorithm overcomes the defect that the clustering results of the k-means algorithm are easily affected by the selection of the initial center points,at the same time,the global optimization strategy of GKM also makes it have a higher time complexity.Therefore,an improved global k-means(IGKM)is proposed in this paper.IGKM has lower time complexity while ensuring clustering performance.In addition,in order to solve the clustering problem of linearly inseparable and imbalanced datasets,we propose an improved global k-means based multi-prototype clustering algorithm(IGKM-MPC)by combining the proposed IGKM algorithm with a multi-prototype clustering algorithm.Through experimental simulation,the effectiveness and superiority of the proposed algorithm are verified.(2)A deep embedding clustering algorithm considering the preservation of the initial clustering structure is proposed.Based on the auto-encoder structure,deep embedding clustering(DEC)algorithm performs feature learning and clustering at the same time,which solved the problem that traditional clustering algorithms face the curse of dimensionality when dealing with high-dimensional data.But the algorithm only considers the clustering loss,which destroys the data structure during training.Therefore,this paper proposes a deep embedding clustering algorithm that considers the preservation of the initial clustering structure.The algorithm introduces the initial clustering structure preservation loss function,and comprehensively considers the clustering loss,reconstruction loss and clustering structure preservation loss.In addition,in order to make the algorithm obtain better initial cluster centers,the proposed IGKM algorithm is used as the initialization method.Through experimental simulation,the effectiveness and superiority of the proposed algorithm are verified.(3)The proposed deep clustering algorithm is used for personalized recommendation of banking financial products.First of all,the background of banking financial product marketing is introduced.Based on the characteristics of high dimensionality and strong sparsity of bank data,this paper proposes a personalized recommendation algorithm combining deep embedding clustering considering the initial clustering structure and collaborative filtering.The algorithm can effectively deal with high-dimensional user or commodity data.Then,the effectiveness of the algorithm is verified by experiments on numerical data.Finally,the algorithm is used in the personalized recommendation of real bank financial products.Compared with the recommendation results based on k-means clustering algorithm,the recommendation algorithm based on deep clustering has achieved better recommendation results.The algorithms proposed in this paper effectively overcome the shortcomings of traditional clustering algorithms,such as unstable results and poor ability to handle high-dimensional or linear inseparable data.Simulation experiments verify the effectiveness and superiority of the proposed algorithms.At the same time,the application of the proposed algorithm in the personalized recommendation of financial products further reflects the application value of the method in this paper.In future research,the application of the method proposed in this paper in the fields of customer portrait and risk assessment will be explored.