Clustering WEB Advertising Based On Multi-feature Fusion

With the development of Internet,web advertising has become an important way for network service providers to earn profit and the useful platform for traditional companys to promote their brand and product.It is meaningful to mine knowledge and information,which is hidden in the massive amounts of Web advertisement,by clustering.And it is useful to business competition and economic projection.After the study of web advertising data which has several features,we observe that multi-feature fusion can be more comprehensive to describe the web advertising object than single-feature.Therefore,this thesis studies the problem of clustering web advertisement based on multi-feature fusion.It is mainly to complete the following tasks:(1)Analyzing the characteristics of web advertisement,summing up a number of features and collecting the web advertisement dataset.The dataset,which we collect,include two text features and one image feature.For the purpose of extracting more useful information,a fuzzy matching method is presented to reduce the sparsity of the text data,and four methods are used to extract image information;(2)Usually,few features decide whether two clusters is seperated.In order to let these features have more influence,the discriminative subspace kmeans-type clustering(Dkmeans)algorithm is proposed.This algorithm considers the influence of both intra-cluster compactness and inter-cluster separation.Different to traditional weighting kmeans-type algorithms,a 3-order tensor is constructed to evaluate the importance of different feature in order to integrate the aforementioned two types of information.The proof of the theorems and experimental results in several public data sets is given.Experimental results corroborate that our algorithm outperforms the state-of-art kmeans-type clustering algorithms;(3)Through different feature fusion experiments which based on the different combination of web advertising features,it observes that the combination of all features has best clustering result among all combinations.It confirms that clustering through multi-feature fusion can improve the clustering effect of web advertising.