Research On Recommendation System Algorithm Based On Factorization Machines And Deep Learning

With the booming development of the Internet,services such as online shopping,online advertising and online video have become inseparable from people’s lives.In order to solve the problem of information overload and improve the efficiency of resource distribution,the recommendation system has become the product of this era.Click-through Rate(CTR)prediction is the core task of the recommendation system,and many models have been proposed to achieve this goal.These models aim to model the interactions between features and improve the performance of prediction models by using low-order or high-order cross features.At present,the models based on factorization machines mainly pool the sum of all the cross features into a scalar,which cannot make full use of the fine-grained information on the dimension of cross features.Although some models propose different ways to learn the interactions between features than Hadamard Product and inner product,this significantly increases the computational time complexity of these models.Many CTR prediction models based on deep learning have been proposed,but the focus of these models is usually only how to effectively model feature interactions,and few researchers pay attention to how to optimize the embedding vectors that represent the initial feature.Given the above problems,this thesis has conducted the following studies in the field of CTR prediction:(1)By analyzing the existing mainstream CTR prediction models,a general CTR prediction framework including seven modules was proposed.Most CTR prediction models can be unified under this framework.(2)Some second-order feature interaction neural networks were designed to solve the problem that some models based on factorization machines can not effectively utilize the dimension information of cross features.All the cross features are concatenated together and then transmitted to the neural network to make full use of the fine-grained information on the dimension of cross feature vectors.In order to optimize the initial embedding vectors of features,an input-aware neural factorization machine was proposed to weigh the same features of different instances by combining the inputaware strategy.(3)A Product-enhanced Network(Pe Net)was proposed to focus on the initial embedding vectors of features at the feature field level,feature dimension level and global feature bit level.Pe Net improves the embedding representation of features with Squeeze,Excitation and Scaling operations.In order to balance the accuracy of CTR prediction model and the cost of calculation time,some product-enhanced factorization machines were designed based on Pe Net and the factorization machines.The multiplication operation was introduced between the initial embedding vectors of feature and the embedding vectors optimized by Pe Net,and some product-enhanced neural network models were designed.These two types of models were integrated to further improve the performance of the recommendation system.