Risk song detection method and device, storage medium and electronic equipment

Abstract

The invention relates to a risk song detection method and device, a storage medium and electronic equipment, and relates to the technical field of machine learning, and the method comprises the steps: obtaining the current behavior data and the current label data of a to-be-detected song, and obtaining a risk song according to the current behavior data and the current label data; constructing an original multi-dimensional feature vector of the song to be detected; performing feature mining processing on the original multi-dimensional feature vector based on a preset deep learning network to obtain an intermediate feature vector with a first dimension, and performing dimension reduction processing on the intermediate feature vector with the first dimension to obtain a binary state vector; and performing clustering processing on the to-be-detected songs based on the binary state vector to obtain a grouping clustering result of the to-be-detected songs, and determining whether the to-be-detected songs belong to high-risk songs or not according to the grouping clustering result. The accuracy of the detection result of the high-risk song is improved.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of machine learning technology. More specifically, the embodiments of the present disclosure relate to a method for detecting risky songs, a device for detecting risky songs, a computer-readable storage medium, and an electronic device. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the disclosure that are recited in the claims. No description herein is admitted to be prior art by inclusion in this section.

[0003] The existing risky song detection method can be implemented based on the supervised machine learning model. Specifically, in the actual application process, people with fraudulent behavior and people without fraudulent behavior can be combined as sample data, and by selecting corresponding features, a supervised machine learning method is used to build a model to identify risky songs.

[0004] However, this method can only identify risky songs of existing patterns, but cannot identify risky songs of new patterns, which results in a low accuracy rate in the identification results of risky songs. Summary of the invention

[0005] However, in the prior art, on the one hand, the use of supervised machine learning methods to build models to identify risky songs has the problem of low accuracy of recognition results; on the other hand, the use of unsupervised machine learning methods to identify risky songs cannot accurately describe group information and also has the problem of low accuracy of recognition results.

[0006] To this end, an improved method for identifying risky songs is urgently needed to construct an original multidimensional feature vector of the song to be detected based on the current behavior data and the current label data; then, based on a preset deep learning network, feature mining is performed on the original multidimensional feature vector to obtain an intermediate feature vector with a first dimension, and the intermediate feature vector with the first dimension is reduced in dimension to obtain a binary state vector; finally, clustering is performed on the songs to be detected based on the binary state vector to obtain grouping clustering results of the songs to be detected, and whether the songs to be detected are high-risk songs is determined based on the grouping clustering results, thereby improving the accuracy of the recognition results of risky songs.

[0007] In this context, embodiments of the present disclosure are intended to provide a method for detecting risky songs, an apparatus for detecting risky songs, a computer-readable storage medium, and an electronic device.

[0008] According to one aspect of the present disclosure, a method for detecting risky songs is provided, comprising:

[0009] Acquire current behavior data and current label data of the song to be detected, and construct an original multi-dimensional feature vector of the song to be detected according to the current behavior data and the current label data;

[0010] Performing feature mining processing on the original multidimensional feature vector based on a preset deep learning network to obtain an intermediate feature vector with a first dimension, and performing dimensionality reduction processing on the intermediate feature vector with the first dimension to obtain a binary state vector;

[0011] The songs to be detected are clustered based on the binary state vector to obtain grouping and clustering results of the songs to be detected, and whether the songs to be detected are high-risk songs is determined according to the grouping and clustering results.

[0012] In an exemplary embodiment of the present disclosure, the preset deep learning network includes a multi-layer randomly generated neural network, and the multi-layer randomly generated neural network includes a first layer of randomly generated neural network, a second layer of randomly generated neural network, ..., a Z-1th layer of randomly generated neural network, and a Zth layer of randomly generated neural network;

[0013] Among them, based on a preset deep learning network, feature mining processing is performed on the original multidimensional feature vector to obtain an intermediate feature vector with a first dimension, and dimension reduction processing is performed on the intermediate feature vector with the first dimension to obtain a binary state vector, including:

[0014] Performing feature mining processing on the original multidimensional feature vector based on the first layer of randomly generated neural network, the second layer of randomly generated neural network, ..., and the Z-1th layer of randomly generated neural network to obtain an intermediate feature vector with a first dimension;

[0015] Based on the Z-th layer randomly generated neural network, dimension reduction processing is performed on the intermediate feature vector with the first dimension to obtain a binary state vector.

[0016] In an exemplary embodiment of the present disclosure, feature mining processing is performed on the original multidimensional feature vector based on the first layer of randomly generated neural network, the second layer of randomly generated neural network, ..., and the Z-1 layer of randomly generated neural network to obtain an intermediate feature vector having a first dimension, including:

[0017] The original multidimensional feature vector is received based on the first visible layer in the first randomly generated neural network, and feature extraction is performed on the original multidimensional feature vector based on the first hidden layer in the first randomly generated neural network to obtain a first feature extraction result; wherein the feature dimension of the first feature extraction result is higher than the feature dimension of the original multidimensional feature vector;

[0018] Inputting the first feature extraction result into the second layer of randomly generated neural network to obtain a second feature extraction result; wherein the feature dimension of the second feature extraction result is higher than the feature dimension of the first feature extraction result;

[0019] The feature extraction process is cycled sequentially to obtain the Z-2th feature extraction result, and the Z-2th feature extraction result is input into the Z-1th layer of randomly generated neural network to obtain an intermediate feature vector with the first dimension.

[0020] In an exemplary embodiment of the present disclosure, the preset deep learning network includes a deep belief network, and the randomly generated neural network includes a restricted Boltzmann machine.

[0021] In an exemplary embodiment of the present disclosure, clustering the songs to be detected based on the binary state vector to obtain grouping clustering results of the songs to be detected includes:

[0022] Traversing the binary state vectors in sequence, extracting songs to be detected having the same binary state vector; wherein the binary state vector is a vector whose eigenvalue of the intermediate eigenvector having the first dimension is 1 or 0;

[0023] The songs to be detected having the same binary state vector are clustered to obtain grouping clustering results of the songs to be detected.

[0024] In an exemplary embodiment of the present disclosure, determining whether the song to be detected is a high-risk song according to the grouping clustering result includes:

[0025] Determine the current risk flow and the current total flow of songs for each song to be detected in the grouping clustering result, and determine the sum of the risk flow and the sum of the total flow of songs for the grouping clustering result according to the current risk flow and the current total flow of songs for each song to be detected;

[0026] Determine the fraud traffic ratio of the grouping clustering result according to the risk traffic sum and the song traffic sum, and judge whether the fraud traffic ratio is greater than a preset threshold;

[0027] If the fraudulent traffic proportion is greater than a preset threshold, it is determined that the song to be detected in the grouping clustering result is a high-risk song.

[0028] In an exemplary embodiment of the present disclosure, the method for detecting risky songs further includes:

[0029] If the songs to be detected are high-risk songs, a high-risk song list is generated according to the songs to be detected in the grouping and clustering results;

[0030] The high-risk song list is pushed to the terminal device corresponding to the application operation and maintenance personnel, so that the application operation and maintenance personnel can process the high-risk songs in the high-risk song list.

[0031] In an exemplary embodiment of the present disclosure, the current behavior data includes at least one of the number of plays, the play duration, and the number of users who play the song to be detected within a preset time period; the current label data includes the song popularity level label and / or song style label of the song to be detected.

[0032] In an exemplary embodiment of the present disclosure, the preset deep learning network is trained in the following manner:

[0033] Acquire historical detected songs, and construct historical song feature vectors based on historical behavior data and historical label data of the historical detected songs;

[0034] Performing forward training on the deep learning network to be trained based on the historical song feature vectors to obtain forward propagation results, and constructing a loss function according to the forward propagation results and the actual song labels of the historical detected songs;

[0035] Adjusting the bias parameters and weight parameters in the deep learning network to be trained based on the loss function until the deep learning network to be trained is in a convergence state;

[0036] After determining that the deep learning network is in a converged state, reverse fine-tuning is performed on the deep learning network in the converged state based on the forward propagation result after the converged state to obtain the preset deep learning network.

[0037] In an exemplary embodiment of the present disclosure, forward training is performed on a deep learning network to be trained based on the historical song feature vector to obtain a forward propagation result, including:

[0038] Inputting the historical song feature vector into a first layer of a randomly generated neural network in a deep learning network to be trained to obtain a first output result;

[0039] Inputting the first output result into a second layer of randomly generated neural network in the deep learning network to be trained to obtain a second output result;

[0040] The above input process is looped sequentially to obtain the forward propagation result corresponding to the Z-th layer of the randomly generated neural network in the deep learning network to be trained.

[0041] In an exemplary embodiment of the present disclosure, reverse fine-tuning is performed on the deep learning network in a converged state based on the forward propagation result in a converged state to obtain the preset deep learning network, including:

[0042] Inputting the forward propagation result after convergence into the Z-1th layer of the randomly generated neural network in the deep learning network after convergence to obtain the first prediction result;

[0043] Inputting the first prediction result into the Z-2th layer of the randomly generated neural network in the deep learning network after convergence to obtain a second prediction result;

[0044] The above input process is looped sequentially to obtain the preset deep learning network.

[0045] According to one aspect of the present disclosure, a device for detecting risky songs is provided, comprising:

[0046] A multi-dimensional feature vector construction module, used to obtain current behavior data and current label data of the song to be detected, and to construct an original multi-dimensional feature vector of the song to be detected based on the current behavior data and current label data;

[0047] A state vector determination module, configured to perform feature mining processing on the original multidimensional feature vector based on a preset deep learning network to obtain an intermediate feature vector having a first dimension, and perform dimensionality reduction processing on the intermediate feature vector having the first dimension to obtain a binary state vector;

[0048] The risk song detection module is used to cluster the songs to be detected based on the binary state vector to obtain the grouping clustering results of the songs to be detected, and determine whether the songs to be detected are high-risk songs according to the grouping clustering results.

[0049] In an exemplary embodiment of the present disclosure, the preset deep learning network includes a multi-layer randomly generated neural network, and the multi-layer randomly generated neural network includes a first layer of randomly generated neural network, a second layer of randomly generated neural network, ..., a Z-1th layer of randomly generated neural network, and a Zth layer of randomly generated neural network;

[0050] Among them, based on a preset deep learning network, feature mining processing is performed on the original multidimensional feature vector to obtain an intermediate feature vector with a first dimension, and dimension reduction processing is performed on the intermediate feature vector with the first dimension to obtain a binary state vector, including:

[0051] Performing feature mining processing on the original multidimensional feature vector based on the first layer of randomly generated neural network, the second layer of randomly generated neural network, ..., and the Z-1th layer of randomly generated neural network to obtain an intermediate feature vector with a first dimension;

[0052] Based on the Z-th layer randomly generated neural network, dimension reduction processing is performed on the intermediate feature vector with the first dimension to obtain a binary state vector.

[0053] In an exemplary embodiment of the present disclosure, feature mining processing is performed on the original multidimensional feature vector based on the first layer of randomly generated neural network, the second layer of randomly generated neural network, ..., and the Z-1 layer of randomly generated neural network to obtain an intermediate feature vector having a first dimension, including:

[0054] The original multidimensional feature vector is received based on the first visible layer in the first randomly generated neural network, and feature extraction is performed on the original multidimensional feature vector based on the first hidden layer in the first randomly generated neural network to obtain a first feature extraction result; wherein the feature dimension of the first feature extraction result is higher than the feature dimension of the original multidimensional feature vector;

[0055] Inputting the first feature extraction result into the second layer of randomly generated neural network to obtain a second feature extraction result; wherein the feature dimension of the second feature extraction result is higher than the feature dimension of the first feature extraction result;

[0056] The feature extraction process is cycled sequentially to obtain the Z-2th feature extraction result, and the Z-2th feature extraction result is input into the Z-1th layer of randomly generated neural network to obtain an intermediate feature vector with the first dimension.

[0057] In an exemplary embodiment of the present disclosure, the preset deep learning network includes a deep belief network, and the randomly generated neural network includes a restricted Boltzmann machine.

[0058] In an exemplary embodiment of the present disclosure, clustering the songs to be detected based on the binary state vector to obtain grouping clustering results of the songs to be detected includes:

[0059] Traversing the binary state vectors in sequence, extracting songs to be detected having the same binary state vector; wherein the binary state vector is a vector whose eigenvalue of the intermediate eigenvector having the first dimension is 1 or 0;

[0060] The songs to be detected having the same binary state vector are clustered to obtain grouping clustering results of the songs to be detected.

[0061] In an exemplary embodiment of the present disclosure, determining whether the song to be detected is a high-risk song according to the grouping clustering result includes:

[0062] Determine the current risk flow and the current total flow of songs for each song to be detected in the grouping clustering result, and determine the sum of the risk flow and the sum of the total flow of songs for the grouping clustering result according to the current risk flow and the current total flow of songs for each song to be detected;

[0063] Determine the fraud traffic ratio of the grouping clustering result according to the risk traffic sum and the song traffic sum, and judge whether the fraud traffic ratio is greater than a preset threshold;

[0064] If the fraudulent traffic proportion is greater than a preset threshold, it is determined that the song to be detected in the grouping clustering result is a high-risk song.

[0065] In an exemplary embodiment of the present disclosure, the risk song detection device further includes:

[0066] A high-risk song list generating module, used for generating a high-risk song list according to the songs to be detected in the grouping and clustering results if the songs to be detected belong to high-risk songs;

[0067] The high-risk song processing module is used to push the high-risk song list to the terminal device corresponding to the application operation and maintenance personnel, so that the application operation and maintenance personnel can process the high-risk songs in the high-risk song list.

[0068] In an exemplary embodiment of the present disclosure, the current behavior data includes at least one of the number of plays, the play duration, and the number of users who play the song to be detected within a preset time period; the current label data includes the song popularity level label and / or song style label of the song to be detected.

[0069] In an exemplary embodiment of the present disclosure, the preset deep learning network is trained in the following manner:

[0070] Acquire historical detected songs, and construct historical song feature vectors based on historical behavior data and historical label data of the historical detected songs;

[0071] Performing forward training on the deep learning network to be trained based on the historical song feature vectors to obtain forward propagation results, and constructing a loss function according to the forward propagation results and the actual song labels of the historical detected songs;

[0072] Adjusting the bias parameters and weight parameters in the deep learning network to be trained based on the loss function until the deep learning network to be trained is in a convergence state;

[0073] After determining that the deep learning network is in a converged state, reverse fine-tuning is performed on the deep learning network in the converged state based on the forward propagation result after the converged state to obtain the preset deep learning network.

[0074] In an exemplary embodiment of the present disclosure, forward training is performed on a deep learning network to be trained based on the historical song feature vector to obtain a forward propagation result, including:

[0075] Inputting the historical song feature vector into a first layer of a randomly generated neural network in a deep learning network to be trained to obtain a first output result;

[0076] Inputting the first output result into a second layer of randomly generated neural network in the deep learning network to be trained to obtain a second output result;

[0077] The above input process is looped sequentially to obtain the forward propagation result corresponding to the Z-th layer of the randomly generated neural network in the deep learning network to be trained.

[0078] In an exemplary embodiment of the present disclosure, reverse fine-tuning is performed on the deep learning network in a converged state based on the forward propagation result in a converged state to obtain the preset deep learning network, including:

[0079] Inputting the forward propagation result after convergence into the Z-1th layer of the randomly generated neural network in the deep learning network after convergence to obtain the first prediction result;

[0080] Inputting the first prediction result into the Z-2th layer of the randomly generated neural network in the deep learning network after convergence to obtain a second prediction result;

[0081] The above input process is looped sequentially to obtain the preset deep learning network.

[0082] According to one aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the method for detecting risky songs described in any one of the above items is implemented.

[0083] According to one aspect of the present disclosure, there is provided an electronic device, including:

[0084] Processor; and

[0085] A memory, configured to store executable instructions of the processor;

[0086] Wherein, the processor is configured to execute any of the above-mentioned risk song detection methods by executing the executable instructions.

[0087] According to the risk song detection method and risk song detection device of the embodiment of the present disclosure, the current behavior data and current label data of the song to be detected can be obtained, and the original multidimensional feature vector of the song to be detected can be constructed according to the current behavior data and the current label data; then, the original multidimensional feature vector is subjected to feature mining processing based on the preset deep learning network to obtain an intermediate feature vector with a first dimension, and the intermediate feature vector with the first dimension is subjected to dimensionality reduction processing to obtain a binary state vector; finally, the song to be detected is clustered based on the binary state vector to obtain a grouping clustering result of the song to be detected, and it is determined whether the song to be detected belongs to a high-risk song according to the grouping clustering result, without the need to perform risk song identification based on supervised machine learning or unsupervised machine learning, thereby significantly reducing the problem of low recognition result accuracy caused by the inability to identify risk songs of new patterns through supervised machine learning, and reducing the problem of low recognition result accuracy caused by the inability to accurately describe group information through unsupervised machine learning, thereby bringing a better experience to users. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which:

[0089] Figure 1 The following is a schematic diagram showing an example process of a method for detecting risky songs according to an example embodiment of the present disclosure;

[0090] Figure 2 A schematic diagram of a structure example of a preset deep learning network according to an exemplary embodiment of the present disclosure is shown;

[0091] Figure 3 A schematic diagram showing an example structure of a randomly generated neural network in a preset deep learning network according to an example embodiment of the present disclosure is shown;

[0092] Figure 4 A flowchart of a training process of a preset deep learning network according to an exemplary embodiment of the present disclosure is schematically shown;

[0093] Figure 5 A schematic diagram of a scenario example of a training process of a preset deep learning network according to an exemplary embodiment of the present disclosure is shown;

[0094] Figure 6 A block diagram schematically shows a device for detecting risky songs according to an exemplary embodiment of the present disclosure;

[0095] Figure 7 A computer program product for storing a risk song detection method according to an exemplary embodiment of the present disclosure is schematically shown;

[0096] Figure 8 An electronic device for implementing a method for detecting risky songs according to an exemplary embodiment of the present disclosure is schematically shown.

[0097] In the drawings, the same or corresponding reference numerals represent the same or corresponding parts. DETAILED DESCRIPTION

[0098] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0099] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0100] According to an embodiment of the present disclosure, a method for detecting risky songs, a device for detecting risky songs, a computer-readable storage medium, and an electronic device are proposed.

[0101] In this document, any number of elements in the drawings is used for illustration rather than limitation, and any naming is used only for distinction and does not have any limiting meaning.

[0102] The principle and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure. SUMMARY OF THE INVENTION

[0104] The applicant has discovered that with the development of Internet technology, a black industry chain has emerged that defrauds platforms by falsifying data to earn revenue. In the cloud music scenario, the black industry chain uses cheating technologies such as scripts or plug-ins to control batch accounts to play, collect, comment on specific target songs and other types of behavioral data. The song business data generated by such behavior is fraudulent traffic, and songs with concentrated fraudulent traffic can be considered risky songs. At the same time, the black industry chain uses the above-mentioned fraudulent traffic to obtain traffic settlement funds from the cloud music platform, which will cause financial losses to the cloud music platform. At the same time, fraudulent traffic will have an adverse impact on the song community and charts.

[0105] In the face of the above-mentioned problem of fraudulent traffic, in the relevant technical solutions, risky songs can be identified in the following ways: one is to identify risky songs based on rule strategies and supervised machine learning models; the other is to identify risky songs by using unsupervised learning algorithms to train detection models. Among them:

[0106] For the rule-based strategy method, it can transform the experience and knowledge of risk control experts into fraud prevention rules and match them through rule-based strategies; at the same time, for the supervised machine learning model-based method, it can combine people with fraudulent behavior and people without fraudulent behavior as sample data, select corresponding features, and use supervised machine learning methods to build a model to identify fraud risks; however, although this method is more effective; however, in the context of the Internet, due to the rapid development of black industry technology, the fraud methods and technologies in the cloud music business scenario are also constantly updated. The methods based on rule strategies and supervised machine learning models are models obtained based on existing fraud patterns and can only identify existing fraud patterns. At the same time, under this route, by analyzing the collected black sample data, the features that can effectively distinguish anomalies and determine the corresponding feature thresholds, effective business rules or strategies are formulated to identify fraudulent traffic. Furthermore, supervised machine learning can be used to learn the collected labeled positive and negative samples to generate an effective fraud identification model, and the generated fraud model can be used to detect business traffic and identify fraudulent traffic. However, this method requires a certain amount of labeled black sample data, and collecting this data requires a certain cost. Moreover, this method can only identify the black production methods marked in the sample, and cannot learn the fraud methods that are not identified by the existing labels.

[0107] On the other hand, for the method of using unsupervised learning algorithms to train detection models, it can use the K-Means algorithm or density-based clustering algorithm to directly cluster the data and divide the data into several groups (without dimensionality reduction). This clustering algorithm is easily affected by data noise points, and its essence is metric learning (metric-learning) based on similarity. It is necessary to manually define the distance between samples based on experience in advance. For data with high-dimensional features, it is difficult to manually determine the appropriate similarity measurement method. The most commonly used feature dimensionality reduction method is principal component analysis (PCA). However, PCA is suitable for linear data that obeys Gaussian distribution. The data in practical applications is basically nonlinear, so in practical applications, PCA cannot achieve the expected dimensionality reduction effect or even fails. Therefore, in the existing methods, for nonlinear high-dimensional feature data in practical applications, whether it is direct clustering or dimensionality reduction through PCA, it is ultimately impossible to accurately describe group information, thereby affecting the rationality of fraudulent transaction identification; and, since the K-Means algorithm is an unsupervised algorithm based on distance, it is necessary to define the distance between samples and then determine whether they belong to the same category by calculating the sample distance size; however, it is impossible to simply define the appropriate distance for real multidimensional data, and thus the K-Means method may not be able to give the correct clustering result, resulting in certain false positives.

[0108] In view of this, an example embodiment of the present disclosure provides a method for identifying risky songs. The method for identifying risky songs is a method for clustering high-risk songs of fraudulent traffic in a big data scenario based on DBN. This method does not require defining sample distances, so it has higher accuracy than clustering methods based on distance or density similarity, and can also be applied in high-dimensional nonlinear scenarios. It has better applicability than the dimensionality reduction clustering method of principal component analysis.

[0109] After introducing the basic principles of the present disclosure, various non-limiting embodiments of the present disclosure are described in detail below.

[0110] Exemplary Methods

[0111] The exemplary embodiment of the present disclosure first provides a method for identifying risky songs, which can be run on a server, server cluster or cloud server, etc. Of course, those skilled in the art can also run the method of the present disclosure on other platforms as required, which is not specifically limited in the exemplary embodiment. Figure 1 As shown, the risk song identification method may include the following steps:

[0112] Step S110. Acquire current behavior data and current label data of the song to be detected, and construct an original multi-dimensional feature vector of the song to be detected based on the current behavior data and current label data;

[0113] Step S120. Perform feature mining processing on the original multidimensional feature vector based on a preset deep learning network to obtain an intermediate feature vector with a first dimension, and perform dimensionality reduction processing on the intermediate feature vector with the first dimension to obtain a binary state vector;

[0114] Step S130. Cluster the songs to be detected based on the binary state vector to obtain grouping clustering results of the songs to be detected, and determine whether the songs to be detected are high-risk songs based on the grouping clustering results.

[0115] In the above-mentioned risk song identification method, the current behavior data and current label data of the song to be detected can be obtained, and the original multidimensional feature vector of the song to be detected can be constructed according to the current behavior data and the current label data; then, the original multidimensional feature vector is subjected to feature mining processing based on the preset deep learning network to obtain an intermediate feature vector with a first dimension, and the intermediate feature vector with the first dimension is subjected to dimensionality reduction processing to obtain a binary state vector; finally, the song to be detected is clustered based on the binary state vector to obtain a grouping clustering result of the song to be detected, and it is determined whether the song to be detected is a high-risk song according to the grouping clustering result, without the need to perform risk song identification based on supervised machine learning or unsupervised machine learning, thereby significantly reducing the problem of low recognition result accuracy caused by the inability to identify risk songs of new patterns through supervised machine learning, and reducing the problem of low recognition result accuracy caused by the inability to accurately describe group information through unsupervised machine learning, thereby bringing a better experience to users.

[0116] Hereinafter, the risk song identification method recorded in the exemplary embodiment of the present disclosure will be further explained and illustrated in conjunction with the accompanying drawings.

[0117] First, the technical implementation principle of the example embodiment of the present disclosure is explained and illustrated. In the cloud music scenario, each song will generate multi-dimensional behavior data and label data every day. For example, the number of songs played in a day, the length of time the songs are played in a day, etc. are behavior data; song ratings, song style labels, etc. are label data; at the same time, after collecting the behavior data and label data of each song, a multi-dimensional feature vector of the corresponding song can be constructed. Then, based on the above-mentioned multi-dimensional feature vector of the song, the DBN network is trained, and the multi-dimensional feature vector data of the sample song to be detected is calculated through the trained DBN neural network model, and the feature samples of the song to be detected are mapped into binary state vectors to achieve dimensionality reduction and clustering of the multi-dimensional feature vector of the song; further, the training song samples with binary state vectors of the same state are classified into the same group to achieve grouping and clustering of the songs. Furthermore, by counting the risk traffic and total song traffic identified by online strategies in each grouped clustered song, we can calculate the proportion of known fraudulent traffic in each song group, compare the fraudulent traffic proportion value of each song group with the preset threshold, and mark the group with fraudulent traffic proportion exceeding the threshold as a high-risk song group; at the same time, the songs in the corresponding high-risk song group will be given to business operations for analysis and judgment, and corresponding business penalties will be given. In addition, this method does not need to define sample distance, so it has higher accuracy than clustering methods based on distance or density similarity, and can also be applied in high-dimensional nonlinear scenarios, and has better applicability than the dimensionality reduction clustering method of principal component analysis.

[0118] Secondly, the preset deep learning network involved in the exemplary embodiment of the present disclosure is explained and illustrated. Specifically, the preset deep learning network recorded in the exemplary embodiment of the present disclosure may be a DBN (Deep Belief Networks); the preset deep learning network may include a multi-layer randomly generated neural network, and the multi-layer randomly generated neural network includes a first layer of randomly generated neural network, a second layer of randomly generated neural network, ..., a Z-1th layer of randomly generated neural network, and a Zth layer of randomly generated neural network; the specific model results can be referred to Figure 2 As shown; at the same time, based on this deep learning network, the input samples can be reduced in dimension and clustered.

[0119] In an exemplary embodiment, the randomly generated neural network described above may be a Restricted Boltzmann machine (RBM). Meanwhile, the RBM (Restricted Boltzmann machine) is a randomly generated neural network that can learn probability distribution through input data sets. The specific model structure can be referred to Figure 3At the same time, the structure of the RBM described here can include two layers: one is the visible layer v = {v1, v2, ..., v i ,...,v m}, configured with a bias term b, which is mainly responsible for receiving input data; the second is the hidden layer h = {h1,h2,...,h j ,...,h n}, configured with a bias term c, which is responsible for performing feature extraction tasks; at the same time, the connection between the two layers is bidirectional, the weight is marked by w, and there is no connection between the units in the layer. This unique structure enables RBM to efficiently learn the probability distribution of input data.

[0120] Furthermore, stacking multiple layers of RBMs provides an effective means for deep feature representation of data. Each layer of RBM is dedicated to extracting higher-order features from the data input; at the same time, with the processing of a series of RBM layers, the level of feature extraction is significantly improved. Subsequently, the data is introduced into one or more back propagation layers. These two parts together constitute a complete Figure 2 The DBN architecture shown.

[0121] The following will be combined Figure 4 Explain and illustrate the specific training process of the preset deep learning network. For details, refer to Figure 4 As shown, the specific training process of the preset deep learning network may include the following steps:

[0122] Step S410. Obtain historical detected songs, and construct a historical song feature vector based on the historical behavior data and historical tag data of the historical detected songs;

[0123] Step S420. Perform forward training on the deep learning network to be trained based on the historical song feature vector to obtain a forward propagation result, and construct a loss function based on the forward propagation result and the actual song label of the historical detected song; wherein the actual song label of the historical detected song recorded here refers to the binary feature value of the song marked as a risky song or a normal song in the above feature dimension;

[0124] Step S430. Adjust the bias parameters and weight parameters in the deep learning network to be trained based on the loss function until the deep learning network to be trained is in a convergence state;

[0125] Step S440. After determining that the deep learning network is in a converged state, reverse fine-tuning is performed on the deep learning network in the converged state based on the forward propagation result after the converged state to obtain the preset deep learning network.

[0126] In an example embodiment, forward training is performed on a deep learning network to be trained based on the historical song feature vector to obtain a forward propagation result, which can be achieved in the following manner: inputting the historical song feature vector into the first layer of randomly generated neural network in the deep learning network to be trained to obtain a first output result; inputting the first output result into the second layer of randomly generated neural network in the deep learning network to be trained to obtain a second output result; and repeating the above input process in sequence to obtain a forward propagation result corresponding to the Zth layer of randomly generated neural network in the deep learning network to be trained.

[0127] In an exemplary embodiment of the present disclosure, reverse fine-tuning of the deep learning network after the convergence state is performed based on the forward propagation result after the convergence state to obtain the preset deep learning network, which can be achieved in the following way: inputting the forward propagation result after the convergence state into the Z-1th layer randomly generated neural network in the deep learning network after the convergence state to obtain a first prediction result; inputting the first prediction result into the Z-2th layer randomly generated neural network in the deep learning network after the convergence state to obtain a second prediction result; and repeating the above input process in sequence to obtain the preset deep learning network.

[0128] Below, the specific training process of the preset deep learning network will be further explained and illustrated. Specifically, in the process of actual application, first, song sample data (that is, historically detected songs) are collected to construct historical song feature vectors of each historically detected song; wherein, in the process of constructing the historical song feature vector, the historical behavior data and historical label data of the historically detected song can be first obtained; at the same time, the historical behavior data recorded here can include the song play volume, song play duration, number of song play accounts, etc. of the historically detected song within a period of time (such as one day, one week, or one month, etc.); the historical label data recorded here can be the song heat level, song style label, etc. of the historically detected song; further, in the process of actual application, assuming that there are M historically detected songs, each historically detected song has N-dimensional feature data (that is, historical song feature vectors), these dimensions include: song play times, song play duration, number of song play accounts, song heat level, song style label, etc., and historical behavior data and historical label data of the song; for each historically detected song i (1≤i≤M), the following N-dimensional feature vector can be used to form its corresponding historical song feature vector, marked as: X i ={x1,x2,...,x N};

[0129] Secondly, the DBN network constructed based on the historical song feature vector constructed above is trained to obtain a trained DBN network model (that is, a preset deep learning network). The detailed processing process is as follows: (1) The prepared song sample data is used to train an RBM network with N (the N recorded here is consistent with the dimension N of the historical song feature vector) nodes as the input layer and T1 nodes as the output layer; wherein T1 is a specific positive integer; T1 is a value that can be adjusted and ultimately determined through the training effect; this network can ultimately convert the song feature vector into a T1 feature vector; (2) The feature vector trained by the RBM network with N nodes as the input layer and T1 nodes as the output layer is used as a sample to train an RBM network with T1 nodes as the input layer and T2 nodes as the output layer; wherein T2 is a specific positive integer, T2 is a value that can be adjusted and ultimately determined through the training effect; this network can ultimately convert the song feature vector into a T2 feature vector; (3) By analogy, a DBN network formed by stacking Z layers of RBM networks head to tail can be trained; wherein the last layer of RBM network selects a binary RBM network to ensure that the final result feature vector is a binary feature vector, that is, a feature vector in which the feature values ​​of each dimension of the vector are 1 or 0.

[0130] It should be noted here that, based on the unique network structure of DBN, its training process is divided into two parts: the initial stage and the fine-tuning stage. In the initial stage, an unsupervised greedy algorithm is used to pre-train DBN, aiming to convert the original data into high-level feature representations; at the same time, after each RBM layer training is completed, the state of its internal hidden layer is randomly used as the visible layer input of the next layer of RBM, and a deep abstract processing of the features is achieved through a layer-by-layer progressive approach; further, after the pre-training is completed, DBN enters the fine-tuning stage, and the entire network is carefully adjusted by introducing a back propagation layer; at the same time, this stage aims to optimize the DBN parameters to achieve a transition from a local optimal solution to a global optimal solution; finally, after three layers of RBM progressive training, DBN can extract deep feature signals from the original data, which not only improves the training efficiency of the model, but also enhances the performance and accuracy of DBN when processing complex data. For specific training scenarios, please refer to Figure 5 shown.

[0131] The following will be combined Figure 2-Figure 5 right Figure 1 The detection method of risky songs shown in the following is further explained and illustrated. Specifically:

[0132] In step S110, the current behavior data and the current label data of the song to be detected are obtained, and the original multi-dimensional feature vector of the song to be detected is constructed according to the current behavior data and the current label data.

[0133] In this example embodiment, first, the current behavior data and current label data of the song to be detected are obtained; wherein the song to be detected recorded here may be a song that the application system believes may be abnormal; for example, it suddenly has a large amount of traffic, or has similar traffic every day, etc., of course, it may be other situations, and this example does not impose special restrictions on this; at the same time, the current behavior data and current label data recorded here are generally consistent with the historical behavior data and historical label data recorded above, and also represent the song to be detected in a certain time period (such as one day, one week or one month, etc., this example does not impose special restrictions on this) song playback volume, song playback time, song playback account volume, song popularity level and song style label, etc.; secondly, according to the current behavior data and current label data, the original multidimensional feature vector of the song to be detected is constructed; wherein the original multidimensional feature vector recorded here may be the specific values ​​of data of different dimensions, such as the specific song playback volume, song playback time, song playback account volume, song popularity level and song style label, etc.

[0134] In step S120, feature mining is performed on the original multidimensional feature vector based on a preset deep learning network to obtain an intermediate feature vector with a first dimension, and dimension reduction is performed on the intermediate feature vector with the first dimension to obtain a binary state vector.

[0135] Specifically, the specific determination process of the binary state vector can be achieved in the following way: first, based on the first layer of randomly generated neural network, the second layer of randomly generated neural network, ..., the Z-1th layer of randomly generated neural network, the original multidimensional feature vector is subjected to feature mining processing to obtain an intermediate feature vector with a first dimension; secondly, based on the Zth layer of randomly generated neural network, the intermediate feature vector with a first dimension is subjected to dimensionality reduction processing to obtain a binary state vector.

[0136] In an exemplary embodiment, feature mining processing is performed on the original multidimensional feature vector based on the first layer of randomly generated neural network, the second layer of randomly generated neural network, ..., the Z-1th layer of randomly generated neural network to obtain an intermediate feature vector with a first dimension, which can be achieved in the following manner: first, the original multidimensional feature vector is received based on the first visible layer in the first layer of randomly generated neural network, and feature extraction is performed on the original multidimensional feature vector based on the first hidden layer in the first layer of randomly generated neural network to obtain a first feature extraction result; wherein the feature dimension of the first feature extraction result is higher than the feature dimension of the original multidimensional feature vector; secondly, the first feature extraction result is input into the second layer of randomly generated neural network to obtain a second feature extraction result; wherein the feature dimension of the second feature extraction result is higher than the feature dimension of the first feature extraction result; then, the feature extraction process is looped in sequence to obtain the Z-2th feature extraction result, and the Z-2th feature extraction result is input into the Z-1th layer of randomly generated neural network to obtain an intermediate feature vector with a first dimension.

[0137] It should be noted here that the binary state vector of the song to be detected is determined by the DBN network, so that the dimensionality reduction and clustering of the multi-dimensional feature vector of the song to be detected can be achieved; at the same time, by classifying the training song samples with binary state vectors of the same state into the same group, and then grouping and clustering the songs to be detected based on the group, the accuracy of the detection results of the songs to be detected can be improved.

[0138] In step S130, clustering processing is performed on the songs to be detected based on the binary state vector to obtain grouping clustering results of the songs to be detected, and whether the songs to be detected are high-risk songs is determined according to the grouping clustering results.

[0139] In this example embodiment, first, the songs to be detected are clustered based on the binary state vector to obtain the grouping clustering results of the songs to be detected; specifically, it can be achieved in the following way: traverse the binary state vector in sequence to extract the songs to be detected with the same binary state vector; wherein the binary state vector is a vector whose eigenvalue of the intermediate eigenvector with the first dimension is 1 or 0; cluster the songs to be detected with the same binary state vector to obtain the grouping clustering results of the songs to be detected. That is, in the actual application process, after obtaining the binary state eigenvector, the songs to be detected with the binary state eigenvector with the same eigenvalue result can be classified into one category, so the songs to be detected can be grouped and clustered to form various groups (grouping clustering results) in the end, and the songs in each grouping clustering result are the same binary state eigenvector after the trained DBN network is reduced in dimension.

[0140] Secondly, determine whether the song to be detected is a high-risk song based on the grouping clustering result; specifically, this can be achieved in the following ways: determine the current risk flow and the current total flow of songs for each song to be detected in the grouping clustering result, and determine the sum of the risk flow and the total flow of songs in the grouping clustering result based on the current risk flow and the current total flow of songs for each song to be detected; determine the fraud flow ratio of the grouping clustering result based on the risk flow ratio and the total flow of songs, and determine whether the fraud flow ratio is greater than a preset threshold; if the fraud flow ratio is greater than the preset threshold, determine that the song to be detected in the grouping clustering result is a high-risk song. That is, in the actual application process, the risk flow and the total flow of songs identified by the online strategy of the songs included in each grouping clustering result can be counted, and the known fraud flow ratio in each song group can be calculated, and the fraud flow ratio value of each song group is compared with the preset threshold, and the group with the fraud flow ratio exceeding the threshold is marked as a high-risk song group.

[0141] Furthermore, after obtaining the detection results of high-risk songs, the risk song detection method also includes: if the song to be detected is a high-risk song, a high-risk song list is generated according to the songs to be detected in the grouping and clustering results; the high-risk song list is pushed to the terminal device corresponding to the application operation and maintenance personnel, so that the application operation and maintenance personnel can process the high-risk songs in the high-risk song list. That is, in the actual application process, the songs in the corresponding high-risk song group will be given to the business operation for analysis and judgment, and corresponding business penalties will be given; at the same time, specific business penalty measures can be set according to the risk category and risk size, and this example does not impose special restrictions on this.

[0142] So far, the risk song detection method recorded in the exemplary embodiment of the present disclosure has been fully realized. Based on the above-mentioned contents, it can be known that the risk song detection method recorded in the exemplary embodiment of the present disclosure, on the one hand, does not need to define the sample distance, so it has higher accuracy than the clustering method based on distance or density similarity; on the other hand, it can also be applied in high-dimensional nonlinear scenarios, and has better applicability than the dimensionality reduction clustering method of principal component analysis. It can be used in high-dimensional nonlinear feature scenarios and has higher usability.

[0143] Exemplary Devices

[0144] After introducing the risk song detection method of the exemplary embodiment of the present disclosure, next, refer to Figure 6 The risky song detection device according to the exemplary embodiment of the present disclosure is explained and described.

[0145] Specifically, refer to Figure 6 As shown, the risk song detection device may include a multi-dimensional feature vector construction module 610, a state vector determination module 620 and a risk song detection module 630. Among them:

[0146] The multi-dimensional feature vector construction module 610 can be used to obtain the current behavior data and the current label data of the song to be detected, and construct the original multi-dimensional feature vector of the song to be detected according to the current behavior data and the current label data;

[0147] The state vector determination module 620 may be used to perform feature mining processing on the original multidimensional feature vector based on a preset deep learning network to obtain an intermediate feature vector having a first dimension, and perform dimensionality reduction processing on the intermediate feature vector having the first dimension to obtain a binary state vector;

[0148] The risk song detection module 630 can be used to cluster the songs to be detected based on the binary state vector to obtain the grouping clustering results of the songs to be detected, and determine whether the songs to be detected are high-risk songs based on the grouping clustering results.

[0149] In an exemplary embodiment of the present disclosure, the preset deep learning network includes a multi-layer randomly generated neural network, and the multi-layer randomly generated neural network includes a first layer of randomly generated neural network, a second layer of randomly generated neural network, ..., a Z-1th layer of randomly generated neural network and a Zth layer of randomly generated neural network; wherein, based on the preset deep learning network, the original multi-dimensional feature vector is subjected to feature mining processing to obtain an intermediate feature vector with a first dimension, and the intermediate feature vector with the first dimension is subjected to dimensionality reduction processing to obtain a binary state vector, including: based on the first layer of randomly generated neural network, the second layer of randomly generated neural network, ..., the Z-1th layer of randomly generated neural network, the original multi-dimensional feature vector is subjected to feature mining processing to obtain an intermediate feature vector with a first dimension; based on the Zth layer of randomly generated neural network, the intermediate feature vector with a first dimension is subjected to dimensionality reduction processing to obtain a binary state vector.

[0150] In an exemplary embodiment of the present disclosure, feature mining processing is performed on the original multidimensional feature vector based on the first layer of randomly generated neural network, the second layer of randomly generated neural network, ..., the Z-1th layer of randomly generated neural network to obtain an intermediate feature vector with a first dimension, including: receiving the original multidimensional feature vector based on the first visible layer in the first layer of randomly generated neural network, and performing feature extraction on the original multidimensional feature vector based on the first hidden layer in the first layer of randomly generated neural network to obtain a first feature extraction result; wherein the feature dimension of the first feature extraction result is higher than the feature dimension of the original multidimensional feature vector; inputting the first feature extraction result into the second layer of randomly generated neural network to obtain a second feature extraction result; wherein the feature dimension of the second feature extraction result is higher than the feature dimension of the first feature extraction result; looping the feature extraction process in sequence to obtain the Z-2th feature extraction result, and inputting the Z-2th feature extraction result into the Z-1th layer of randomly generated neural network to obtain an intermediate feature vector with a first dimension.

[0151] In an exemplary embodiment of the present disclosure, the preset deep learning network includes a deep belief network, and the randomly generated neural network includes a restricted Boltzmann machine.

[0152] In an exemplary embodiment of the present disclosure, clustering processing is performed on the songs to be detected based on the binary state vector to obtain a grouping clustering result of the songs to be detected, including: traversing the binary state vector in sequence to extract the songs to be detected with the same binary state vector; wherein the binary state vector is a vector whose eigenvalue of the intermediate eigenvector with the first dimension is 1 or 0; clustering the songs to be detected with the same binary state vector to obtain a grouping clustering result of the songs to be detected.

[0153] In an exemplary embodiment of the present disclosure, determining whether the song to be detected is a high-risk song is based on the grouping clustering result, including: determining the current risk flow and the current total song flow of each song to be detected in the grouping clustering result, and determining the total risk flow of the grouping clustering result and the total song flow based on the current risk flow and the current total song flow of each song to be detected; determining the fraudulent flow ratio of the grouping clustering result based on the risk flow ratio and the total song flow ratio, and judging whether the fraudulent flow ratio is greater than a preset threshold; if the fraudulent flow ratio is greater than the preset threshold, determining that the song to be detected in the grouping clustering result is a high-risk song.

[0154] In an exemplary embodiment of the present disclosure, the risk song detection device further includes:

[0155] A high-risk song list generating module, used for generating a high-risk song list according to the songs to be detected in the grouping and clustering results if the songs to be detected belong to high-risk songs;

[0156] The high-risk song processing module is used to push the high-risk song list to the terminal device corresponding to the application operation and maintenance personnel, so that the application operation and maintenance personnel can process the high-risk songs in the high-risk song list.

[0157] In an exemplary embodiment of the present disclosure, the current behavior data includes at least one of the number of plays, the play duration, and the number of users who play the song to be detected within a preset time period; the current label data includes the song popularity level label and / or song style label of the song to be detected.

[0158] In an exemplary embodiment of the present disclosure, the preset deep learning network is trained in the following manner: historical detected songs are obtained, and historical song feature vectors are constructed based on historical behavior data and historical label data of the historical detected songs; forward training is performed on the deep learning network to be trained based on the historical song feature vectors to obtain forward propagation results, and a loss function is constructed based on the forward propagation results and actual song labels of the historical detected songs; bias parameters and weight parameters in the deep learning network to be trained are adjusted based on the loss function until the deep learning network to be trained is in a convergence state; after determining that the deep learning network is in a convergence state, reverse fine-tuning is performed on the deep learning network in the convergence state based on the forward propagation results after the convergence state to obtain the preset deep learning network.

[0159] In an exemplary embodiment of the present disclosure, forward training is performed on a deep learning network to be trained based on the historical song feature vector to obtain a forward propagation result, including: inputting the historical song feature vector into the first layer of randomly generated neural network in the deep learning network to be trained to obtain a first output result; inputting the first output result into the second layer of randomly generated neural network in the deep learning network to be trained to obtain a second output result; and repeating the above input process in sequence to obtain a forward propagation result corresponding to the Zth layer of randomly generated neural network in the deep learning network to be trained.

[0160] In an exemplary embodiment of the present disclosure, the deep learning network in a converged state is reversely fine-tuned based on the forward propagation result after the convergence state to obtain the preset deep learning network, including: inputting the forward propagation result after the convergence state into the Z-1th layer randomly generated neural network in the deep learning network in the convergence state to obtain a first prediction result; inputting the first prediction result into the Z-2th layer randomly generated neural network in the deep learning network in the convergence state to obtain a second prediction result; and repeating the above input process in sequence to obtain the preset deep learning network.

[0161] Exemplary Storage Media

[0162] After introducing the risk song detection method and risk song detection device according to the exemplary embodiment of the present disclosure, next, refer to Figure 7 A storage medium according to an exemplary embodiment of the present disclosure is described.

[0163] refer to Figure 7 As shown, a program product 700 for implementing the above method according to an embodiment of the present disclosure is described, which can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto.

[0164] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0165] Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, wherein readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Readable signal media may also be any readable medium other than a readable storage medium.

[0166] Program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN).

[0167] Exemplary Electronic Devices

[0168] After introducing the storage medium of the exemplary embodiment of the present disclosure, next, reference is made to Figure 8 An electronic device according to an exemplary embodiment of the present disclosure is described.

[0169] Figure 8 The electronic device 800 shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0170] like Figure 8 As shown, the electronic device 800 is in the form of a general computing device. The components of the electronic device 800 may include, but are not limited to: the at least one processing unit 810, the at least one storage unit 820, a bus 830 connecting different system components (including the storage unit 820 and the processing unit 810), and a display unit 840.

[0171] The storage unit 820 stores program codes, which can be executed by the processing unit 810, so that the processing unit 810 performs the steps according to various exemplary embodiments of the present disclosure described in the above “Exemplary Method” section of this specification. For example, the processing unit 810 can perform the following steps: Figure 1 Steps S110-S130 shown in FIG.

[0172] The storage unit 820 may include a volatile storage unit, such as a random access storage unit (RAM) 8201 and / or a cache storage unit 8202, and may further include a read-only storage unit (ROM) 8203. The storage unit 820 may also include a program / utility 8204 having a set (at least one) of program modules 8205, such program modules 8205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination thereof may include the implementation of a network environment.

[0173] The bus 830 may include a data bus, an address bus, and a control bus.

[0174] The electronic device 800 may also communicate with one or more external devices 900 (e.g., keyboards, pointing devices, Bluetooth devices, etc.) through an input / output (I / O) interface 850. Furthermore, the electronic device 800 may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) through a network adapter 860. As shown, the network adapter 860 communicates with other modules of the electronic device 800 through a bus 830. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 800, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0175] It should be noted that, although several units / modules or subunits / modules of the device are mentioned in the above detailed description, this division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.

[0176] In addition, although the operations of the disclosed method are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0177] Although the spirit and principle of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A method for detecting risky songs, characterized in that: include: Acquire current behavior data and current label data of the song to be detected, and construct an original multi-dimensional feature vector of the song to be detected according to the current behavior data and the current label data; Performing feature mining processing on the original multidimensional feature vector based on a preset deep learning network to obtain an intermediate feature vector with a first dimension, and performing dimensionality reduction processing on the intermediate feature vector with the first dimension to obtain a binary state vector; The songs to be detected are clustered based on the binary state vector to obtain grouping and clustering results of the songs to be detected, and whether the songs to be detected are high-risk songs is determined according to the grouping and clustering results.

2. The method for detecting risky songs according to claim 1, characterized in that: The preset deep learning network includes a multi-layer randomly generated neural network, and the multi-layer randomly generated neural network includes a first layer of randomly generated neural network, a second layer of randomly generated neural network, ..., a Z-1th layer of randomly generated neural network, and a Zth layer of randomly generated neural network; Among them, based on a preset deep learning network, feature mining processing is performed on the original multidimensional feature vector to obtain an intermediate feature vector with a first dimension, and dimension reduction processing is performed on the intermediate feature vector with the first dimension to obtain a binary state vector, including: Performing feature mining processing on the original multidimensional feature vector based on the first layer of randomly generated neural network, the second layer of randomly generated neural network, ..., and the Z-1th layer of randomly generated neural network to obtain an intermediate feature vector with a first dimension; Based on the Z-th layer randomly generated neural network, dimension reduction processing is performed on the intermediate feature vector with the first dimension to obtain a binary state vector.

3. The method for detecting risky songs according to claim 2, characterized in that: Based on the first layer of randomly generated neural network, the second layer of randomly generated neural network, ..., the Z-1th layer of randomly generated neural network, the original multidimensional feature vector is subjected to feature mining processing to obtain an intermediate feature vector having a first dimension, including: The original multidimensional feature vector is received based on the first visible layer in the first randomly generated neural network, and feature extraction is performed on the original multidimensional feature vector based on the first hidden layer in the first randomly generated neural network to obtain a first feature extraction result; wherein the feature dimension of the first feature extraction result is higher than the feature dimension of the original multidimensional feature vector; Inputting the first feature extraction result into the second layer of randomly generated neural network to obtain a second feature extraction result; wherein the feature dimension of the second feature extraction result is higher than the feature dimension of the first feature extraction result; The feature extraction process is cycled sequentially to obtain the Z-2th feature extraction result, and the Z-2th feature extraction result is input into the Z-1th layer of randomly generated neural network to obtain an intermediate feature vector with the first dimension.

4. The method for detecting risky songs according to claim 2 or 3, characterized in that: The preset deep learning network includes a deep belief network, and the randomly generated neural network includes a restricted Boltzmann machine.

5. The method for detecting risky songs according to claim 1, characterized in that: The songs to be detected are clustered based on the binary state vector to obtain grouping clustering results of the songs to be detected, including: Traversing the binary state vectors in sequence, extracting songs to be detected having the same binary state vector; wherein the binary state vector is a vector whose eigenvalue of the intermediate eigenvector having the first dimension is 1 or 0; The songs to be detected having the same binary state vector are clustered to obtain grouping clustering results of the songs to be detected.

6. The method for detecting risky songs according to claim 1, characterized in that: Determining whether the song to be detected is a high-risk song according to the grouping and clustering result includes: Determine the current risk flow and the current total flow of songs for each song to be detected in the grouping clustering result, and determine the sum of the risk flow and the sum of the total flow of songs for the grouping clustering result according to the current risk flow and the current total flow of songs for each song to be detected; Determine the fraud traffic ratio of the grouping clustering result according to the risk traffic sum and the song traffic sum, and judge whether the fraud traffic ratio is greater than a preset threshold; If the fraudulent traffic proportion is greater than a preset threshold, it is determined that the song to be detected in the grouping clustering result is a high-risk song.

7. The method for detecting risky songs according to claim 1, characterized in that: The method for detecting risky songs further includes: If the songs to be detected are high-risk songs, a high-risk song list is generated according to the songs to be detected in the grouping and clustering results; The high-risk song list is pushed to the terminal device corresponding to the application operation and maintenance personnel, so that the application operation and maintenance personnel can process the high-risk songs in the high-risk song list.

8. A risk song detection device, characterized in that: include: A multi-dimensional feature vector construction module, used to obtain current behavior data and current label data of the song to be detected, and to construct an original multi-dimensional feature vector of the song to be detected based on the current behavior data and current label data; A state vector determination module, configured to perform feature mining processing on the original multidimensional feature vector based on a preset deep learning network to obtain an intermediate feature vector having a first dimension, and perform dimensionality reduction processing on the intermediate feature vector having the first dimension to obtain a binary state vector; The risk song detection module is used to cluster the songs to be detected based on the binary state vector to obtain the grouping clustering results of the songs to be detected, and determine whether the songs to be detected are high-risk songs according to the grouping clustering results.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for detecting risky songs according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: include: processor; as well as A memory, configured to store executable instructions of the processor; Wherein, the processor is configured to execute the risk song detection method described in any one of claims 1-7 by executing the executable instructions.

Images