Research And Application Of The Ordered Clustering For Multi-criteria Decision Data

Multi-criteria decision making method is one of the important methods in the field of operations research.Under the background of multi-criteria decision making,most of the data is multi-criteria and unlabeled.In some scenario decision makers are more concerned about which objects belong to the same cluster,and the objects in some clusters are significantly better evaluated than those in other clusters,the classical clustering algorithm can not meet this kind of needs of decision makers,it only divide the data into serveral clusters,but there is no prioritized relationship between any two clusters.This thesis proposes a new ordered clustering algorithm based on PROMETHEE and classical K-Medoids clustering algorithm.Concretely,a new similarity measurement formula is proposed by combining the linear preference function of PROMETHEE,and a new objective function is proposed by improving the net outranking flow of PROMETHEE.The ordered clustering algorithm can consider the relationship between different criteria in the process of clustering,and there is ranking relationship of priority between any two clusters in the final result.On the other hand,under the background of multi-criteria decision making,users have multi-criteria scores for commodities,and the traditional recommendation method only consider the users' rating on a single criterion of the product,so the experimental effect of using the traditional single criterion recommendation method is not effective enough for multi-criteria decision data.This thesis proposes a new multi-criteria recommendation algorithm which combines the multi-criteria decision making method PROMETHEE and the ordered clustering algorithm.The method uses multiple linear regression method of machine learning to learn preferences of different criterias about different users,and uses the ordered clustering algorithm to divide different type of users effectively,which is better than the classical clustering algorithm.And it uses non-negative matrix factorization method on each criterion to predict user's ratings for products in the corresponding criteria.Finally,the products are ranked Top-N by PROMETHEE method and recommended to relevant users.Based on the HDI(Human Development Index)dataset,this thesis uses ordered clustering algorithm proposed in this thesis to make the contrast and analysis to the other ordered clustering algorithm proposed in recent years,the results show that the distribution of data in this thesis is more reasonable and effective,and it increases by about 1 percentage point on the Silhouette coefficient and decreases by about 1 to 2 percentage points on the Davies-Bouldin Index.Then this thesis uses the proposed multi-criteria recommendation algorithm on Trip Advisor's multi-criteria scoring dataset,and makes experimental comparison and analysis with the five multi-criteria recommendation algorithms proposed in recent years,both RMSE and MAE indexes are decreased by about 2 percentage points.