Research And Application Of Two Personalized Recommendation Algorithms

Along with the universality of the Internet and the fast fierce development of the electronic commerce, personalized recommendation system also becomes an important research contents in the electronic commerce domain gradually and gets more and more concern of researcher and electronic commerce business enterprise. The application of re- commendation system provides a more valid method for resolving the problem of information overload in the electronic commerce. Recommend- ation system simulates the store salesperson to provide merchandise recommendation for the customer, help the customer to find out the mer- chandise which is needed, thus make sure the customer to complete the purchase process smoothly, therefore can reserve a customer effectively, raise the sale quantity of the electronic commerce system; Company can also through the recommendation system to establish the relationship between customer and company, keep the contact with customer, raise the customer's loyalty.The general workflow of the recommendation system is: collects the evaluate information through customer's behavior, then processes the in- formation, uses personalized recommend technique to recommend resource to customer, finally shows the result to the customer. Some recommendation systems has feedback process, allow customer to do an evaluation to the recommendation, and return the feedback to recommend system, then system adjusts itself in order to recommend for the customer better. Network education resource management system (NERMS) is a Jilin province science and technology development project, the main target is to realize the effective organization and management to numerous of the network education resource (examination paper, document, video, course- ware etc 12 major types), to convenient efficiently share and obtain the network education resource, thus speed the development of the network education resources and promote the development of network education. NERMS is designed to be a intelligent and personalized website for the goal of the system. The intelligence turns to mean that the system can speculate a customer's related information according to his behavior; the personalized turns to mean that the system can provide different service to the different customer according to the customer's related information. Therefore, we did a series research and work about making website to intelligent and per- sonalized, the personalized recommendation technology is one of them.This thesis introduced the workflow and technology of personalized recommendation system, then researched and applied two kinds of per- sonalized recommendation algorithms, and tested their functions. These two kinds of algorithms are the connection rule algorithm based on two stages count and the algorithm combines PRM model and collaborative filter recommendation. Among them the first algorithm uses two stages count frame and applies the connection rule which distributes according to the perpendicular data efficiently discovers algorithm to carry on the excavation of connection rule; the second algorithm improves traditional collaborative filter and combines the PRM model. Finally, applies these algorithms to NERMS to recommend the resource for the customer.Firstly, this thesis uses Two-Stage-Count Frame, and apply the efficient algorithm for discovering association rules based on vertical data layout to carry out the algorithm TSCF-FD. The recommendation which is based on the customer association usually digs out all the customer association from the rating table, recommends the resource that all customers liked to the current customer in the customer association conclusion with certain recommendation degree. Because the general rating table has the high dimensions and sparsity, and in the electronic commerce website, the merchandise which a customer once evaluated(downloaded or purchased) is seldom opposite to the whole resource on the website, so we can use current customer to restrict the form of the rule which digged, also can use current customer to restrict the division to the rating table before digging, so this thesis uses Two-Stage-Count Frame, and apply the efficient algorithm for discovering association rules based on vertical data layout to carry out the algorithm TSCF-FD. This thesis did a performance test comparison between ASARM and TSCF-FD, as a result, TSCF-FD can acquire better timespace function.Secondly, this thesis combined the collaborative filtering based customer grade with PRM, carried out PRM-CF algorithm. The quality of the collaborative filtering recommendation technology, depends on the quantity of the items which evaluated by the customer. The more quantity, the more accurate to reflect the customer's interest by the rating data, the quality also immediately rises. Whereas, the quantity of the items which evaluated by the customer is less and the recommend quality also lower, when there is no rating data, the collaborative filtering recommendation technology can't recommend. This thesis combined the collaborative filtering recommendation technology with the PRM recommendation technology to resolve the problem of low quality of low class customer recommend. The recommendation which based on the PRM model, can well make use of customer's information, item information, the customer's rating information to the item; Its one-rank characteristic, makes itself don't depend on particular customer or particular item. This thesis combined the collaborative filtering based customer grade with PRM, carried out PRM-CF algorithm, and did a performance test comparison with traditional collaborative filtering, as a result, because of less numbers of candidate neighbors, the efficiency of algorithm obviously raises, but the influence of recommend quality is smaller, combined it with PRM, the recommend quality has raised.Finally, this thesis applied these two kinds of algorithms to NERMS to recommend resource for the customers, obtained better effecience.These two algorithms are carried out based on the connection rule, the collaborative filtering, and PRM, there are lots of mature recommend te- chnology in currently personalized recommendation area. These technique also can improved by combining with each other, these work still need to be complete. There are still some problems in this thesis: we can try other algorithms to carry out the Two-Stage-Count Frame; the digging of the connection rule still has the possibility of improvement; PRM can also lead other attribution of customer or resource in order to make the PRM more concrete etc. These problems still need to be carried on an improvement in aftertime's work.