Research On The Technology Of Spot Recommendation Based On The Model Of The Relationship Between "The User And The Spot"

Personalized recommendation task models the relationship between users and goods by mining user information and goods information, thus providing recommendation for users according to the relationship modeled. The complex Internet environment has generated a mass of multi-source heterogeneous data, makes the discovery and expression of the relationship between users and goods a big difficulty.In consideration of the above, against the background of tourist attraction recommendation, our thesis designs and implements a personalized tourist attraction recommendation system based on multi-source heterogeneous data. In order to integrate multi-source heterogeneous data, the system respectively models user information and spots information so that the data of different sources and types can be separated and used. Since the integration of multi-source heterogeneous data will bring about much noise, sparse learning method is adopted to model the spots information, playing the role of selection for the characteristics of data after integration. Compared with traditional sparse learning method, our method adds homogeneity coefficient regularization terms based on users' prior knowledge and then sets up the preference relationship between users and tourist attractions through the history grades the users gave to the spots, thus more efficiently eliminating the noise in multi-source heterogeneous data integration and enhancing its capacity in characteristic selection.The main research contents of the thesis are showed as follows:First, the thesis generally reviews the related studies in the field of recommendation systems, specifically defines the relationship of “user-spot” according to the actual situation of tourist attraction recommendation, and definitely lists the tasks of recommendation and the main problem to be solved. Besides, it models and implements the system in accordance with the following steps:(1) A design scheme of personalized spots recommendation is given in the thesis. It briefly introduces the composition of system modeling from the logical aspects, namely spots information modeling, user information modeling and recommended spots sorting, and discusses their respective functions and the problems to be solved. In addition, it introduces the data sources and working processes from the angle of implementation, and meanwhile presents the mapping relationship between the logic module and the implementation module.(2) This thesis proposes a method of spots information modeling, which takes spots information as the center and makes modeling for it on the basis of coefficient learning method through the integration results of multi-source heterogeneous data that are the texts describing the spots. The information about the history grades that the users gave is introduced into the modeling by the method of homogeneity coefficient regularization terms to reflect the degree of users' preference to different spots. And the preference degree is taken as the output of spots information modeling.(3) In the thesis, the method to model user information is proposed, which takes users' interest as the center, puts forward several assumptions about the reasons why the relation “user-spot” comes into being. Then, in accordance with the above assumptions and combining with users' prior knowledge, the relation “user-spot” is characteristically represented from several angles by the means of probability statistics variable, normalization and compression function. The characteristic representation is taken as the output of user information modeling.(4) The method to sort the recommended spots is proposed in this thesis. The method proposed takes the output combination of spot information modeling and user information modeling as module input, and ordinal regression plays a central role in the module, and combines with the specific geographic location information of spots to select and finally generate a collection of candidate spots. Besides, to make sure the sorting list can be normally generated, a sorting strategy is formulated as well, which makes sorting for spots according to the results predicted by ordinal regression and the features of the relation “user-spot”.In the end, the thesis verifies the validity of the spots information modeling and the recommendation system on the basis of the datasets from 2014 TREC CS Track. Experiment results show that the spots information modeling with sparse learning and homogeneity coefficient regularization terms as its core is better than the traditional method in sorting; while the recommendation performance of the whole system boasts a remarkable advantage over the systems in the competition of 2014 TREC CS Track.