Researches On Methods Of Financial Sales Oriented Personalized Recommendation Methods

Financial sales have attracted much attention of financial enterprises. Compared with other domains, there exists a big gap in financial domain in utilizing non-traditional techniques such as data mining to solve the problems of financial sales. However, with the rapid progress of application of information technology, such as web bank, data warehouse, it has become a hot topic to study how to use data mining technology to solve the problems of financial sales.One of key problems of financial sales is to provide customers the financial products and services meeting their personalized requirements. In financial domain, with the rapid development of technology and financial transactions, the types of financial products and the amount of financial customer transaction information grow very fast. It leads to that customers need to spend a lot of time to find their interested products. As a technology to solve information overload problem, personalized recommendation system constructs interest model according to historical behavior data of customers and recommends their interested information. As the most powerful information filtering means, recommendation system can also be introduced into financial domain for financial product recommendation.The kernel of recommendation system is its recommendation approach. The most widely studied are content based recommendation system and collaborative filtering recommendation system. The dissertation will study the financial sales oriented personalized recommendation approaches from content based and collaborative filtering angles.In content based personalized financial product recommendation, content corresponds to the information of the products that customers purchase. In collaborative filtering personalized recommendation, product recommendation is performed according to the customers having similar product interests. To realize content based personalized recommendation, the dissertation proposes average value constraint based sequential pattern mining method, and varying value oriented sequential pattern mining method respectively according to the difference of the weight associated with the item of sequence. To realize collaborative filtering based personalized recommendation, the dissertation proposes customer segmentation based personalized recommendation and personalized recommendation considering interest drift.The main research work and originalities are as follows:1) Studies average value constraint based sequential pattern mining method to solve the problem of content based personalized recommendation that the recommended financial products should satisfy the constraints such as customer earnings. In financial sales, to recommend products satisfying the personalized requirements of customers, often much attention is paid on the patterns of purchasing financial product, especially those patterns with high customer contribution or customer earnings. Sequential pattern mining technique can be used to predict the financial products to be purchased in the future. To this end, we regard each purchased product as"content"in content based personalized recommendation, so as to recommend the products possibly bought customers on the basis of the behavior pattern similarity, i.e. content similarity. Meanwhile, average value constraint is used to express the requirements on customer contribution or customer gains that should be satisfied by sequential patterns. The dissertation proposes the concept of satisfiability aiming at average value constraint. Then we design corresponding pruning strategy based on satisfiability. Finally, we propose average value constraint based frequent sequential pattern mining algorithm MPAC. Experimental results on the dataset generated by IBM standard synthetic data generator show that our proposed pruning strategy is effective, and MAPC algorithm has good performance.2) Aiming at the problem that the value associated with a certain recommended product usually changes in content based personalized recommendation, the dissertation studies sequential pattern mining method suitable for dealing with sequence data with varying values. In financial product recommendation, the values associated with the recommended products usually change. For example, the difference of the purchasing amount or buying time makes the earnings be different. All the currently available algorithms cannot deal with varying value sequence database. To this end, we first propose ACV, an Aggregate Constraint with Varying value items to denote the constraint that the aggregate feature of sequential patterns should satisfy. Then we design an algorithm utilizing ACV constraints to prune useless sequential patterns. The algorithm partitions initial sequence database into several sequence information vectors and then mines the sequential patterns satisfying the given ACV constraints. Experimental results on both synthetic and real data show that our pruning strategy is effective to reduce the number of candidate sequential patterns to be detected, and thus improves mining efficiency.3) Aiming at the problem of identifying customers having similar interests for collaborative filtering based personalized recommendation, we study the community mining based financial customer segmentation methods. An approach of community mining based financial customer segmentation is proposed. The proposed community mining based approach can build customer social network structure through customer feature similarity or behavior similarity. Experimental results on benchmark dataset and financial customer data show that our proposed community mining based customer segmentation approach is effective.4) Studies how to segment financial customers into multiple different interest preference groups, and performs association analysis in each different customer groups to realize collaborative filtering based personalized financial product recommendation. Association rule mining is a traditional method to solve such problem. However, the number of financial customers easily reaches tens of thousands, and possibly everyday a large number of transaction data are generated. If association analysis is directly performed on these data, two kinds of problems might arise. On the one hand, it might need too high cost of time and space. On the other hand, it is highly possible to make the mined association rules lack the effectiveness when directly mining association rule on customer groups with big difference, and thus greatly lowering the effect of cross-sale and personalized service. To solve such problem, the dissertation proposes a personalized financial product recommendation method based on customer segmentation. The method first segments customers into a number of customer groups having different interests according to customer interest and behavior characteristics. Then it performs association analysis on each different customer group. The method not only can reduce the required space and time cost, but also can make the mined association rule be more effective and thus improve the cross-sales and personalized recommendation effect.5) Aiming at the problem that the recommendation results could be influenced by the product interest change with the time of financial customers, we study the method of personalized recommendation by considering interest drift. It is well known that user interests often vary with the time change in many applications due to the influence of various factors, and we call such phenomenon customer interest drift. To this end, the dissertation proposes a graph structured recommendation algorithm considering customer interest drift. The algorithm first maps a rating sequence for a customer to an interest sequence, and weights the rated items through interest sequence. Secondly, it adopts resource allocation based bi-partite projection algorithm to construct the association graph between items. Lastly, the proposed algorithm uses interest weighted rated items to construct user feature vectors and generates recommendation sequence through random walk with restart on item association graph. The experimental results on Movielens and financial customer dataset show that our proposed algorithm has better performance than that of the available typical algorithms.