Books Resources Recommendation Algorithm Based On Factorization Machine I-FM

With the development of computer technology and information technology,the phenomenon of "Information Overload" has been appeared.Huge amounts of information need to be recommend to the potential users with interest in timely and accurately.However,current recommendation system inevitably face the data sparsity problem,compared with the common machine learning model,the factorization machine that appeared in 2010 has a better adaptability in this situation.While,facing another problem: the cold-start situation,factorization machine still need to be improved.This paper presents an interactive factorization machine model.With an interactive term integrated into the factorization machine,it has the ability to learn parse data setting in cold-start condition.The interaction term is designed for calculating similarity of text.Instead of the collaborative filtering need to accumulate a certain number of users and the information of size characteristics similarity in an indirect way,interaction term can extract text keywords directly and calculate the similarity between two or more text documents,and return the result into the factorization machine.Designed a mechanism to control the intensity of the interaction,the degree of involvement in factorization machine are defined by an Interaction Intensity Factor that controlling the interaction terms.Interaction Intensity Factor can be timely intervene in operation of interaction terms in accordance with the severity of the cold-start training environment,even shutting down interaction term.Experiments show that the interactive factorization machine with a better learning ability was more commonly used than SVM in sparse data sets.Even in the cold-start environment,this interactive factorization machine can have a relatively higher predicting accuracy.The mainly innovations of this paper are as follows:(1)In this paper,an i-FM model,which can integrate a content-based interactive term into FM,has been illustrated.The i-FM not only adapt to train sparse data sets but also can adapt to operate in cold-start conditions.(2)A new memory attenuation interaction intensity factor is designed.It is a decreasing function for the number of new features,it also detects the appearance of new features to control the interaction term to enable and shutting down at the right time.The memory attenuation interaction intensity factor partly or completely replace cross terms in the FM to train new features under cold-start environment and timely control the degree of interactions term to involved in FM(3)The last idea introduced in this paper is the comparison between i-FM performance under stochastic gradient descent method and that under alternating least squares method.The comparison proves that Alternating Least Squares method to the training effect of i-FM is comparably better.The derivation process and algorithm in the model is complete.The lemma and theorem in this paper were given mathematical proofs.