Design And Implementation Of Cross-Platfrom Spam App Review Detection System

The application market is one of the most important application distribution channels.Reviews in the app market will directly affect users'perception of the app,which in turn affects the likelihood that users will download the app.In order to increase app downloads,some bad app developers will hire sailors to post spam reviews on the platform.This greatly affects the objectivity and fairness of the reviews in the platform.In recent years,spam review detection has received widespread attention,and various recognition features and models have been applied by researchers to detect spam review.However,the existing research work has some limitations.First of all,in terms of features,existing work usually only performs feature extraction in a single app store,ignoring the behavioral differences caused by brush review behavior in different app stores.Secondly,in terms of models,existing models do not combine features with deep learning methods.The method based on shallow learning needs to extract features based on prior knowledge.However,this method relies on the completeness of features.It is easy to misreport under the condition of incomplete feature selection,and deep learning methods can automatically extract data from data.However it requires more training data to extract more comprehensive features.Incorporating features into deep learning methods enables them to focus on extracting features beyond the artificial expert knowledge,thereby learning more comprehensive features with fewer training samples.This paper proposes an improved deep learning method.On the one hand,this method makes full use of the prior knowledge accumulated by researchers,selects highly recognizable review features,on the other hand,integrates deep learning methods in automatically extracting semantic features of text the advantages.(1)In terms of features,in view of the problem that the current features are not comprehensive enough,multiple discriminative features are defined from the three dimensions of application,review,and user.In the application dimension,multiple cross-store comparison features are extracted;in the review dimension,some statistical features hat are difficult for a neural network to learn are filtered and retained;in the user dimension,existing features are expanded.These features describe the review behavior from different angles,which can effectively help the neural network model to learn and detect spam review from angles other than text semantics.(2)In terms of models,this paper proposes an improved deep learning model that incorporates set features,and incorporates features into the parameter matrix of the neural network to more accurately model the relationship between review text and applications and users.The scheme is divided into three steps:First,the feature and word vectors are fused using an improved parametric attention mechanism to obtain a fused feature vector.Then,the vector is input to the bidirectional long-short memory neural network to learn the sequence relationship,and the keyposition information is selected based on the attention mechanism to form a feature vector representing the entire sentence.Finally,the feature vector is input to the fully connected layer for spam detection.Existing feature fusion work usually combines features and text vectors into a classifier,and it is difficult to learn the fine-grained association between features and local words and phrases.However,this solution injects features into the weight matrix of the model and establishes the association between the features and the review text,which can effectively improve the effect of feature fusion.(3)Based on the above work,this paper designs and implements a cross-store spam review detection prototype system.First,according to the specifications of the software design document,the requirements analysis,summary design and detailed design of the spam review detection system were carried out.Then,based on the design,the system was implemented in a modular form.Finally,the system was tested thoroughly to verify its detection accuracy,stability and operating efficiency.Experimental results show that the method proposed in this paper has higher recognition accuracy than other existing methods,and can meet the needs of various application platforms for automatic recognition of false reviews.