Research And Application Of Attribute Inference Technique Based On Neural Network

Personalized recommendation system because of the enormous economic benefits is used extensively in various business website. However, there are still many problems which restrict its development. The current personalized recommendation system is based on the correlation ship between users and products, that is, it predicts the interestingness of users to products according to the existing consuming behavior or the evaluations and recommend personally. This makes the existing personalized recommendation systems will be inevitably constrained by cold start-up, data sparseness, expansibility, real-time and other issues.Through analysis the personalized recommendation system, and the principle, problems and reasons existing in the current several primary recommended technology, This thesis investigates the root reason that caused users to spend. In essence, the user attributes determine their purchase behaviors, so if companies can know the user attributes and recommend products according to their attributes, the recommendation will be more targeted and precise.Study the relationship between users and products from the view of causal relationship. This thesis puts forward the concept of attribute inference, which is to make the personalized recommendation by user attributes to solve the existing problems. The attribute inference which is based on neural network uses the neural network algorithm to explore users'transaction record, construct the relational model between the user attributes and products, and then infer the user attributes. Construction of the causal relationship table between users and products, and apply to a personalized recommendation process.This thesis has implemented the inference of the single user attribute through an experiment, and has simply implemented the personalized product recommendations for users as well. The first step, attribute inference, is to use transaction record constructing the relational model between user attributes and products, moreover, to infer the user attributes according to selected products. The second step, single attribute-based personalized recommendation, is to use the user attribute training model of the original data to recommend products to customers and also to calculate the accuracy of recommended results. The third step, inference attribute-based personalized recommendation, is to use the user attributes training model inferred in the first step to recommend products to customers, and to calculate the accuracy of recommended results as well. The final step is to compare and analyze the accuracy of recommended results of the second step and the third step. Whereas, when the user attributes are only 68.02% correct, the accuracy of recommended results based on inference attribute-based personalized recommendation is still able to achieve 19.92%, still maintaining good recommendation efficiency.Due to the stability and universal applicability of the causal relationship the recommendation based on attribute inference supplies a good solution to common problems of the existing personalized recommendation system, and the association rules may be explained as well. The attribute inference offers a new thinking for data mining and personalized recommendation...