Research On Collaborative Filtering Recommendation Algorithm Based On Neural Network Ensemble And User Preference Model

With the development of Internet technology and the flourish of e-commerce, network data has grew with the speed of index level, which has leaded to users spending a large amount of time to search information and commodities they want. Our society has stepped into an era of information overload nowadays. In response to the needs of times, recommended system emerged. Its main task is to push potentially wanted information resources for users from massive amounts of information resources, thus to alleviate the pressure of information retrieval. In the present application, the collaborative filtering algorithm is undoubtedly the greatest achievement, while during its enhancement, it also inevitably encounters many obstacles that data sparseness problem is one of the major difficult problems.As for the sparse data issue, the thesis is based on the users' interests, uses the current users' data to construct a preference model of user to predict the score of ungraded items, fills predicted data to the user ratings matrix. However, it brings some difficulties on user preference modeling as the description of users' preferences exist the problem of ambiguity and uncertainty. Thus it needs to introduce machine learning approach to building user preference model. Neural network ensemble algorithm has a good generalization ability, it is a hot topic in the field of machine learning currently, which can be used to simulate the user's preferences. But confronting of the complexity of the user preference, neural network ensemble algorithm also possesses various deficiencies. In this situation, the author firstly proposes an improved thought on the traditional neural network ensemble algorithm, puts forward an algorithm which is based on differential evolution of negatively correlated neural network ensemble so as to improve its generalization ability. Secondly, the author combines with the improved algorithm and the existing user data to build a user preference model, and finally uses the constructed preference model to predict the score of ungraded items, fills the predicted data to the user rating matrix, and for avoiding the problem of filling over that may arise, the calculation of similarity also makes an improvement.The thought of based on the basic idea of differential evolution of negatively correlated neural network ensemble algorithm is that in order to increase the differences of integration individual in generation of them, it introduces an algorithm of negative correlation learning to parallel train member networks of the integration; in the conclusions generated, the author uses the good optimization ability of differential evolution algorithm to optimize weighting coefficients of the member networks. Through simulation of experiments, and compared with the improved algorithm and other algorithms, the result shows that the algorithm has performed better on whether the generalization performance or robustness.The thought of based on the basic idea of differential evolutionary neural network ensemble of user preference model is that it makes full use of the item characteristic property to build the project feature vectors, and utilizes the project feature mapping vector and user preference to build a user preference model, and uses the proposed differential evolution of negatively correlated neural network ensemble algorithm to simulate user's interests. The result of experiment shows that the proposed differential evolution of negatively correlated neural network ensemble algorithm can simulate the users' preferences well, so it also can forecast the score of ungraded items well.The thought of based on the basic idea of collaborative filtering recommendation algorithm of user preference model is that it uses the constructed preference model to predict the score of ungraded items to fill the predicted data in user rating matrix, thus to construct a pseudo user rating matrix. During using the pseudo user rating matrix to calculate the similarity degree of user, to avoid the problem of filling over issues that might arise, the author only selects a part of the project to calculate. Through the test of MovieLens data set, it shows that the improved algorithm has a better performance than traditional collaborative filtering algorithms.