A method, apparatus and equipment for predicting risks in communication services
By constructing a multilayer hidden Markov model and using Bayesian filters to process communication service data, the time-consuming and labor-intensive problem of communication service risk verification is solved, and efficient and secure risk prediction and analysis are achieved.
Patent Information
- Application Number
- CN202111333727.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-11
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2041-11-11
AI Technical Summary
In existing technologies, risk verification for communication services is time-consuming and labor-intensive, and is prone to the disaster of dimensionality, making it difficult to effectively uncover and verify the causes of risks.
By acquiring the communication service data of the users to be tested, extracting feature data and inputting it into a preset multilayer hidden Markov model for processing, and using correlation algorithms and Bayesian filters for risk prediction, a multilayer hidden Markov model is constructed to solve the problem of insufficient correlation in the business system.
It achieves time-saving, labor-saving, and secure risk prediction for communication services, improves the accuracy and efficiency of risk prediction, and can effectively mine multi-level risk information.
Smart Images

Figure CN116108609B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of communication technology, and specifically to a method, apparatus, and equipment for predicting communication service risks. Background Technology
[0002] With the development of modern communication and internet technologies, the penetration rate of mobile phones and the internet has rapidly increased, and people's communication methods have undergone tremendous changes. To meet the personalized needs of different user groups and maximize user experience, value-added services are generated. However, due to the complex relationships between these services, the causes of risks are difficult to verify. Identifying risks and their causes requires significant manpower and time. Current technologies primarily rely on manual risk verification for communication services, which is time-consuming, labor-intensive, and prone to problems such as dimensional disasters. Summary of the Invention
[0003] In view of the above problems, embodiments of the present invention are proposed to provide a communication service risk prediction method, apparatus and device that overcomes or at least partially solves the above problems.
[0004] According to one aspect of the present invention, a method for predicting communication service risks is provided, comprising:
[0005] Obtain the communication service data of the user under test;
[0006] Based on the communication service data of the user under test, obtain the feature data of the communication service data;
[0007] The feature data is input into a preset business risk prediction model for processing to obtain the risk prediction result; the preset business risk prediction model is obtained by training a multilayer hidden Markov model based on a training set.
[0008] According to another aspect of the present invention, a communication service risk prediction apparatus is provided, comprising:
[0009] The acquisition module is used to acquire the communication service data of the user under test.
[0010] The processing module is used to obtain feature data of the communication service data based on the communication service data of the user under test;
[0011] The output module is used to input the feature data into a preset business risk prediction model for processing to obtain the risk prediction result; the preset business risk prediction model is obtained by training a multilayer hidden Markov model based on a training set.
[0012] According to another aspect of the present invention, a computing device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus;
[0013] The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described communication service risk prediction method.
[0014] According to another aspect of the present invention, a computer storage medium is provided, the storage medium storing at least one executable instruction, the executable instruction causing a processor to perform an operation corresponding to the above-described communication service risk prediction method.
[0015] According to the solution provided in the above embodiments of the present invention, communication service data of the user to be tested is obtained; feature data of the communication service data is obtained based on the communication service data of the user to be tested; the feature data is input into a preset service risk prediction model for processing to obtain the risk prediction result; the preset service risk prediction model is obtained after training a multilayer hidden Markov model based on a training set, and the association relationship can be obtained using an association algorithm, thereby solving the problem of insufficient association tables stored in the service system and achieving the beneficial effects of time-saving, labor-saving and secure communication service risk prediction.
[0016] The above description is merely an overview of the technical solutions of the embodiments of the present invention. In order to better understand the technical means of the embodiments of the present invention and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of the present invention more obvious and understandable, specific implementation methods of the embodiments of the present invention are described below. Attached Figure Description
[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:
[0018] Figure 1 A flowchart of the communication service risk prediction method provided in an embodiment of the present invention is shown;
[0019] Figure 2 This diagram illustrates the process of constructing a multilayer hidden Markov model according to an embodiment of the present invention.
[0020] Figure 3 A flowchart illustrating a specific communication service risk prediction method provided in an embodiment of the present invention is shown.
[0021] Figure 4 A schematic diagram of the communication service risk prediction device provided in an embodiment of the present invention is shown;
[0022] Figure 5 This invention provides a schematic diagram of the structure of a specific communication service risk prediction device according to an embodiment of the invention.
[0023] Figure 6 A schematic diagram of the structure of a computing device provided in an embodiment of the present invention is shown. Detailed Implementation
[0024] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.
[0025] Figure 1 A flowchart of the communication service risk prediction method provided by an embodiment of the present invention is shown. Figure 1 As shown, the method includes the following steps:
[0026] Step 11: Obtain the communication service data of the user under test;
[0027] Step 12: Obtain the feature data of the communication service data based on the communication service data of the user to be tested;
[0028] Step 13: Input the feature data into the preset business risk prediction model for processing to obtain the risk prediction result; the preset business risk prediction model is obtained by training a multilayer hidden Markov model based on the training set.
[0029] In this embodiment, communication service data of the user under test is acquired; feature data of the communication service data is obtained based on the communication service data; the feature data is input into a preset service risk prediction model for processing to obtain the risk prediction result; the preset service risk prediction model is obtained by training a multilayer hidden Markov model based on a training set. This makes communication service risk prediction time-saving, labor-saving, and secure.
[0030] In embodiments of the present invention, the communication service risk mainly refers to the risk of revenue loss due to abnormal service implementation methods.
[0031] Each time a user completes a transaction, a corresponding transaction record is generated, recording the transaction time, channel, name of the transaction, type of transaction, and user code. Each user transaction also corresponds to a specific billing account, and a monthly billing record is generated, recording the billing account, amount, and account information. In addition, when a transaction is created, its attribute information is stored, including the transaction name, transaction code, transaction description, and associated business information. All of this information is stored in the business system.
[0032] In an optional embodiment of the present invention, step 11 may include:
[0033] Step 111: Obtain at least one piece of communication service data from the user under test, including business processing data, business billing data, and business attribute data.
[0034] In this embodiment, the business processing data, billing data, and business attribute data of the user under test are all stored in the business system. When obtaining the business processing data, billing data, and business attribute data of the user under test, the relevant data of prepaid, planned, package, and promotional businesses are mainly extracted, the data is re-encoded, and then the data is grouped and sorted by month and user. In this way, the actual implementation form of the business can be obtained from the business processing record information, the existing configuration content of the business can be obtained from the business attribute information, and the billing status of the business can be obtained from the billing record.
[0035] In another optional embodiment of the present invention, step 12 may include:
[0036] Step 121: Based on the communication service data of the user under test, obtain at least one of the following: text features, service association features, service time features, service attribute features, and user attribute features of the communication service data; wherein, the attribute features include: the price or preferential price corresponding to the service, the billing subject, the preferential method, and the billing method, etc.; the time features include: extracting the effective time and expiration time of the service handled by the user, etc.
[0037] Step 122: Combine at least two of the text features, business association features, business time features, business attribute features, and user attribute features to obtain the feature data of the communication service data.
[0038] In this embodiment, the aforementioned text features, business association features, business time features, business attribute features, and user attribute features are used as features of the sub-Hidden Markov Model to predict the risk of abnormal mutual exclusion settings. This is because the business description not only allows us to understand the designer's design intent but also provides a preliminary understanding of the business type and implementation method, thereby facilitating the verification of whether there are risks in the implementation method of the business and whether the implementation method of the business will conflict with the implementation of other businesses.
[0039] In step 122, the extraction and combination of each feature is illustrated by combining text features, business association features, and business attribute features, as detailed below:
[0040] The first combination method involves counting the number of prepaid services ordered by a user that are associated with the same value-added planning service, the number of value-added planning services ordered in the prepaid amount, whether the billing accounts of the promotional services included in these prepaid amounts are consistent with the billing accounts of the associated value-added planning services, and whether the discount amount of the promotional services included in these prepaid amounts is greater than the standard rate of the associated value-added planning service. This combination feature can be used as a feature of the first-level Layer 2 Hidden Markov Model. This sub-HMRV model is mainly used to predict the risks caused by abnormal combinations of prepaid services and value-added planning services.
[0041] The second combination method involves checking whether the billing methods for value-added plans and associated promotions subscribed to by users are consistent; and whether the standard rates for the plans and the discount amounts for the associated promotions are consistent. This combination feature can serve as a feature of the first-layer, third-level sub-Hidden Markov Model. This sub-model is primarily used to predict the risks arising from inconsistencies in the billing methods for value-added plans and promotional services.
[0042] The third approach involves considering that prepaid services typically offer users certain discounts. Some prepaid services require customers to subscribe to a specific service or for the service's fees to reach a certain threshold before subscription. In other words, certain restrictions must be met to subscribe to the service. This approach obtains the restrictions for users subscribing to prepaid services by acquiring the text features of sub-model 1. It then compares the attributes of other services subscribed to with the restrictions to see if any of these restrictions are met. These combined features are then used as the features for sub-model 4. This sub-model is primarily used to predict the risks caused by abnormal restrictions.
[0043] The fourth combination method compares whether there is a discrepancy between the birth and expiration times of value-added planning business and the birth and expiration times of its corresponding promotional business, and counts the frequency distribution of promotions, using this as a feature of sub-model No. 5.
[0044] In addition to the above features, it is also necessary to extract the attribute features of the user, including acquiring the user's user code, the city where the user is located, and the handling channel information, etc., as auxiliary features for judging various risks.
[0045] The above feature extraction and combination content is only part of the features, and other features are extracted and combined according to the above method.
[0046] In another optional embodiment of the present invention, in step 121, acquiring the text feature of the communication service data may include:
[0047] Step 1211: Acquire the service description of the communication service data according to the communication service data of the user to be tested;
[0048] Step 1212: Extract keywords from the service description using a word vector model to obtain keywords related to communication service risk judgment;
[0049] Step 1213: Classify the keywords to obtain the text feature of the communication service.
[0050] Specifically, first, Chinese word segmentation is performed on the text content of the service description. Chinese characters only have a certain meaning when combined with each other. Therefore, for Chinese text, word segmentation processing is required to split a sentence into several word groups. For example, "XX's service is really good" is split into "XX / 's / service / really good".
[0051] Secondly, Chinese word segmentation processing is performed on the text content to obtain processed segmented words. The processing method includes, but is not limited to, Jieba word segmentation. Jieba word segmentation has its own built-in dictionary, but it is still necessary to add terms related to mobile services to the existing dictionary content, including the names of currently effective prepaid, value-added planning, package-type and promotional services, etc., to update the dictionary. This can avoid some professional terms in the service field from being incorrectly segmented.
[0052] Then, stop words are removed and dimensionality reduction is performed on the processed segmented words, and then the Word2vector algorithm is used to establish a word embedding model to obtain the word vector model. Since the risk judgment words are mainly some nouns, verbs, adjective word groups, etc., the Jieba word segmentation algorithm can be used for part-of-speech tagging, and then the part-of-speech tagging is used to delete word groups other than nouns, verbs and adjectives to achieve the purposes of removing stop words and reducing dimensionality.
[0053] After obtaining the word vector model, several texts are randomly selected from the descriptions of various business types to extract risk judgment words. Then, using the trained word embedding model, all similar word groups are found, and businesses containing risk keywords are classified and labeled according to the keyword type, thereby obtaining the text features of the business. Specifically, the word vector model is used to find all word groups with the same meaning as "mutually exclusive," and based on these word groups, all businesses that need to be mutually exclusive with other businesses are selected. Mutually exclusive businesses are marked with mutual exclusion tags, and then the mutually exclusive items of these businesses are located according to the word groups, i.e., the types of mutually exclusive businesses. Finally, all businesses are classified and labeled according to the type of business item.
[0054] Finally, after obtaining the text features, it is determined whether the user has processed mutually exclusive transactions marked by the text features, and whether they have processed mutually exclusive transactions.
[0055] In this embodiment, different designers may use different phrases to represent the same business design meaning when writing business descriptions. For example, the phrases "mutually exclusive with XX business" and "cannot be ordered simultaneously with XX business" have the same meaning but are expressed differently. Word vector models can match a single word with all semantically similar words in the text. Therefore, based on the business description, word vector models can be used to obtain keywords for risk assessment, and these keywords can then be used to classify and label the business, thereby obtaining the textual features of the business.
[0056] In another optional embodiment of the present invention, step 121, obtaining the service association characteristics of the communication service data, may include:
[0057] Step 1214: Based on the communication service data of the user under test, obtain all services of the target user under test that have handled the associated services;
[0058] Step 1215: Obtain the support and credibility of all services for the target user under test;
[0059] Step 1216: Based on the support and credibility, obtain the business association table;
[0060] Step 1217: Based on the association table and the pre-stored association relationships of the associated services, obtain the business association characteristics of the user under test handling the business.
[0061] Specifically, the first step is to determine the set of business codes for the associated businesses, denoted as the business set. The business set includes: business A, business B, ..., business N, and each business code is a unique identifier for the business.
[0062] The second step is to obtain all target users who have processed the associated business based on the business code set.
[0063] The third step involves filtering out all target users who have completed the associated services, obtaining all services completed by these target users, and then removing free services that all users would complete from these service codes. This reduces the impact of noise.
[0064] The fourth step is to group the associated business transaction records based on all target users to be tested.
[0065] The fifth step is to obtain the support and credibility of all services of the target user under test using the support and credibility formula.
[0066] The sixth step is to filter out the associated business codes whose support and credibility both exceed the association threshold, enter them into the association table, and match them with their corresponding business names.
[0067] Step 7: Based on the association table and the pre-stored association relationships of the associated services, obtain the business association characteristics of the user under test handling the business.
[0068] Support is the frequency of an itemset in the dataset. It is calculated by dividing the number of times an itemset appears in a record by the total number of records in the dataset. The formula for support is: Support = (Number of occurrences of an itemset in a record) / (Number of occurrences of a records in the dataset).
[0069] Credibility is the number of times itemset A and itemset B appear together in the association rule {itemset A -> itemset B}, divided by the number of times itemset A appears. The formula is: Credibility
[0070] Where P(AB) is count(itemsset AB) / count(dataset), representing the degree of association between AB, and P(A) is count(itemsset A) / count(dataset);
[0071] In this embodiment, business association characteristics of actual business implementation are obtained based on the actual business processing situation, which can make up for the problem of insufficient association relationships stored in the business system.
[0072] In another optional embodiment of the present invention, in step 13, the multilayer hidden Markov model is trained through the following process:
[0073] Obtain feature data of the training business data in the training set;
[0074] The feature data of the training business data are normalized and vectorized to obtain the vectorized processing result;
[0075] The vectorization result is input into the first layer of the Hidden Markov Model for processing to obtain the first output result.
[0076] Based on the first output result, the parameters of the hidden Markov model are updated through a Bayesian filter to obtain the first layer of the trained hidden Markov model.
[0077] Based on the trained first-layer Hidden Markov Model, obtain the risk state sequence;
[0078] The optimal risk state sequence is input into the next layer of the Hidden Markov Model, and the above steps are repeated p times to obtain a trained p-layer Hidden Markov Model, where p is a positive integer.
[0079] like Figure 2 As shown in this embodiment, the first step is to obtain the feature data of the training business data in the training set, then normalize the feature data of the training business data in the training set, and then distribute the normalized processing results to each sub-model of the first layer of the multilayer hidden Markov model to obtain the features under each sub-model of the first layer of the multilayer hidden Markov model.
[0080] In the first step, each sub-model is used to predict a single risk scenario, where the risk state of each scenario is set to two states: risky and risk-free. Decomposing the multiple states of a single Hidden Markov Model into sub-Hidden Markov Models reduces the feature dimensionality.
[0081] The second step is to normalize and vectorize the features under each sub-model to obtain the vectorized processing results of the features under each sub-model, and the vectorized processing results are the observed values.
[0082] In the second step, the K-means algorithm can be used for vectorization, but it is not the only option. Different features (or different evaluation indicators) in the feature vector often have different dimensions and units, which can affect the results of data analysis. In order to eliminate the influence of dimensions between indicators, data normalization is also required.
[0083] The K-means algorithm described above, also known as K-average or K-means, where K is a constant that needs to be set beforehand, iterates through M unlabeled samples to form K clusters. The clustering process often uses the distance between samples as the metric. Specifically, the K-means algorithm uses Euclidean distance, calculated as follows: x (i) As the first sample, x (j)For the second sample, u is the variable and n is the number of samples; this step combines several scalar data groups of training business into a vector. Here, the scalar data are the feature data of the training business data, etc., and then the vector space is quantized as a whole, thereby compressing the data without losing much information.
[0084] The third step is to initialize the parameters of the first layer of the sub-Hidden Markov Model, input the vectorization result into the sub-Hidden Markov Model for training, and obtain the first output result.
[0085] In the third step, the parameters of the first layer of the sub-Hidden Markov Model are initialized, specifically including: setting the number of states of each sub-model; setting the initial state transition matrix A, observation probability matrix B, and initial state probability vector π; and confirming the number of sub-models and the number of layers of the entire multi-layer Hidden Markov Model.
[0086] Specifically, to further illustrate the principle of Hidden Markov Models, assume there are two hidden states, q1 and q2, and three observations, v1, v2, and v3. Then, the state transition matrix A is represented as: Among them, a ij =P(i t+1 =q j |i t =q i ), i = 1, 2; j = 1, 2; the a ij This indicates that at time t, the state is q. i Under the condition that the transition occurs at time t+1 to q j The probability of.
[0087] The observation probability matrix B is represented as follows: Among them, b j (k)=P(o t =v k |i t =q j ), k = 1, 2, 3; j = 1, 2; the b j (k) indicates that the state is q at time t. j Generate observations v under the condition of k The probability of.
[0088] The initial state probability vector π is expressed as: π = π i , where π i =P(i1=q i ), i = 1, 2; indicating that at the initial time t = 1, the state is q. i The probability of.
[0089] Here, λ represents the parameters of the Hidden Markov Model, i.e., λ = (A, B, π), where A, B, and π are called the three elements of the Hidden Markov Model.
[0090] Here, the vectorized result, i.e., the observed value O after vectorization, will be used. p The sequence is input into the first layer of the Hidden Markov Model (HMM). The observation probabilities are estimated using both forward and backward algorithms, and the parameters of the HMM are trained using the Baum-Welch algorithm. The computational flow of the forward algorithm is as follows:
[0091] Given a Hidden Markov Model λ = (A, B, π), the observation sequence up to time t is defined as o1, o2, o3, ... o t And the state is q i The probability of this is the forward probability, denoted as: a t (i) = P(o1, o2, o3, ... o t i t =q i |λ), and then recursively obtain the forward probability a. t (i) and the probability of the observed sequence P(O|λ).
[0092] To facilitate understanding, the following example of a dice-rolling game will be used to illustrate how a Hidden Markov Model recursively calculates the forward probability 'a'. t (i) and the observed sequence probability P(O|λ), are as follows:
[0093] Given the probability of choosing each type of die, the numbers rolled in the first t dice rolls are o1, o2, o3, ... o t And the die used in the t-th round is q. i The probability is a. t (i) = P(o1, o2, o3, ... o t i t =q i |λ), the first step in the positive probability calculation is to calculate the initial probability: a t=1 (i)=a1(i)=π i b i (o1), (=1, 2, ..., N), indicates that the first roll of the die was q. i The probability of rolling an O1, where π i This indicates that the dice q will be used at the start. i The probability of.
[0094] The recursive process for the positive probability is as follows:
[0095]
[0096] This indicates that the die q is used in the (t+1)th time. j Roll out o1, o2, o3, ... o t+1 The probability of.
[0097] The 'a' in the square brackets of the above formula t (i) represents the t-th use of the die q i And the numbers rolled are o1, o2, o3, ... o t The probability, a t (i)a ij This indicates that the numbers rolled in the first t throws are o1, o2, o3, ... o t The probability of using the die q on the (t+1)th time is related to the probability of using the die q on the t+1th time. j The probability of.
[0098] Since there are N types of dice in the t-th round, then q is used in the (t+1)-th round. j In the previous instance, i.e., the tth instance, there were N possible dice.
[0099] Finally, the formula is used.
[0100]
[0101] Calculate the probability of the observed sequence.
[0102] The formula for positive probability can be used to calculate a. T (i) indicates that the die used in the Tth time is q. i The resulting observation sequence is O = (o1, o2, o3, ... o2). T ), because q i Since there are N possible sums, the summation result is taken as the final result.
[0103] The calculation process of the reverse algorithm is as follows:
[0104] Given a hidden Markov model λ = (A, B, π), define the state q at time t. i Under the given conditions, the observation sequence from time t+1 to time T is o t+1 o t+2 o t+3 ,...o T The probability of is the reverse probability, denoted as β. t (i)=P(o t+1 o t+2 o t+3 ,...o T |i t =q i ,λ), where the die used in the t-th round is q. i , then β t (i) indicates that the number of observation points from the (t+1)th to the Tth time is o. t+1 o t+2 o t+3 ,...o TThe probability of.
[0105] First, through β T (i) = 1 to perform initial value calculation, given that the die used in the last round was q. i The probability that the value can be any value after the Tth time is 1.
[0106] The recursive formula is then:
[0107]
[0108] Finally obtained
[0109]
[0110] The specific solution process of the Baum-Welch algorithm is as follows:
[0111] Step A, Initialization, Model Parameters, for n=0, select b j (k) (0) , The model λ is obtained (0) =(A (0) B (0) ,π (0) ).
[0112] Step B involves recursively applying the formulas for n = 1, 2, 3, ...
[0113] Specifically,
[0114]
[0115]
[0116] Step C, terminate the recursion, and give the condition for stopping the iteration. Generally, for smaller positive numbers ε1 and ε2, if ||θ (i+1) -θ (i) ||<ε1 or||Q(θ) (i+1) θ (i) )-Q(θ (i) θ (i) If || < ε2, then the iteration stops, and the model λ is obtained. (n+1) =(A (n+1) B (n+1) , π (n+1) ).
[0117] The formulas for each parameter in step B are derived from the EM algorithm, and the derivation process is as follows:
[0118] The goal is to learn the parameters of a Hidden Markov Model (HMM) λ = (A, B, π). Treating the observed sequence data as observed data O and the state sequence data as unobservable latent data I, the HMM is actually a probabilistic model containing latent variables: "Formula 1" P(O|λ) = ∑ I P(O|I,λ)P(I|λ),
[0119] Step a, transform the probability function into a log-likelihood function: log P(O|λ).
[0120] Step b, derive the Q function: "Formula 2"
[0121]
[0122] Where λ represents the parameters of the hidden Markov model that need to be maximized. This represents the current estimate of the Hidden Markov Model.
[0123] Substituting Formula 1 into Formula 2, we obtain Formula 3.
[0124]
[0125] Step c: Maximize each parameter individually.
[0126] In Formula 3, the first term can also be expressed as:
[0127]
[0128] Where π i Constraints must be met Using the method of Lagrange multipliers, write out the Lagrange function:
[0129]
[0130] Taking its partial derivative and setting the result to 0, we obtain "Formula Four".
[0131]
[0132] Summing over i yields γ: Then, substituting it into formula "Formula Four" yields:
[0133] In Formula 3, the second term can also be expressed as:
[0134]
[0135] The application has constraints. The Lagrange multiplier method can be used to find:
[0136]
[0137] In Formula 3, the third term can also be expressed as:
[0138]
[0139] The application has constraints. The Lagrange multiplier method can be used to find:
[0140]
[0141] Among them, only in o t =v k time b j (o t ) for b j (k) is not 0, so I(o) t =v k )express.
[0142] The fourth step is to update the parameters of the hidden Markov model using a Bayesian filter based on the first output result, thereby obtaining the first-layer hidden Markov model after training.
[0143] In step four, some risk points may present revenue loss risks in the logical judgment. However, considering that handling such risks would have a significant impact on existing users, these risks are classified as risk-free. Therefore, based on experience, a Bayesian filter needs to be set to update the predicted state.
[0144] According to the forward and backward algorithm, after obtaining the current observation, the forward variable is updated. The current observation represents the posterior probability of the current state after processing by the Bayesian filter. Therefore, the updated result is the maximum posterior probability of the state.
[0145] The fifth step is to use the Viterbi algorithm to obtain the optimal risk state sequence corresponding to the observation after the first-layer Hidden Markov Model has been trained.
[0146] Step 6: Input the optimal risk state sequence into the next layer of the Hidden Markov Model, and repeat the above steps p times to obtain the trained p-layer Hidden Markov Model, where p is a positive integer.
[0147] In step six, the trained p-layer Hidden Markov Model is as follows:
[0148]
[0149] Among them, f iLet F represent the sequence of risk states predicted by the i-th layer of the Hidden Markov Model. The risk state sequences output by each layer are combined in chronological order to obtain a combined state feature F of the same length, where 1 ≤ i ≤ p, 1 ≤ j ≤ p. Let λ = (A, B, π) be the probability in the Hidden Markov Model, where the state at time t is q. i Under the given conditions, the observation sequence from time t+1 to time T is o t+1 ,o t+2 ,o t+3 ,...o T The probability of is the reverse probability, denoted as β. t (i)=P(o t+1 ,o t+2 ,o t+3 ,...o T |i t =q i ,λ); Given that the die used in the t-th round is q i ,β t (i) represents the number of observation points from the (t+1)th to the Tth time, which is 0. t+1 ,o t+2 ,o t+3 ,...o T The probability of.
[0150] In this embodiment, after the p-layer Hidden Markov Model is constructed, each layer of the Hidden Markov Model needs to be tested sequentially, and the model parameters adjusted based on the test results. Specifically, this includes randomly selecting users with risk labels and users without risk labels as the test set. In the first layer of the model, for users with existing risk labels, the prediction results are directly compared with the risk labels to calculate the accuracy of the model's predicted state, and the model parameters are adjusted based on the accuracy. After adjusting the parameters of the first layer, the parameters of the next layer are tested and updated using the same method.
[0151] The trained multilayer hidden Markov model (MLM) can decompose multiple states in a single hidden Markov model into multiple sub-MMMs, thus decomposing the high-dimensional features corresponding to multiple states. This addresses the efficiency decline problem caused by the increased feature dimensionality in a single hidden Markov model. Secondly, the trained MLM can predict not only multiple risk states simultaneously but also superimposed states of multiple risk states, thereby uncovering deeper risk information. Furthermore, the use of Bayesian filters can improve the accuracy of model predictions.
[0152] In another optional embodiment of the present invention, step 13 may include:
[0153] Step 131: Input the feature data into the first layer of the trained multilayer hidden Markov model for processing to obtain the test users of the first type of risk.
[0154] Step 132: Input the feature data of the user to be tested for the first type of risk into the next layer of the first layer of the multilayer hidden Markov model for processing, and obtain the risk state output by the next layer model, until the risk state output by each layer of the multilayer hidden Markov model is obtained.
[0155] Step 133: Obtain the business proportion of each business handled by the user under the risk state output by each layer of the multi-layer Hidden Markov Model.
[0156] Step 134: Based on the business proportion and the associated threshold, determine the risk prediction result of the risk combination status of the business handled by the user to be tested.
[0157] In this embodiment, the feature data is first input into the first layer of the trained multilayer hidden Markov model for processing, and risk users exceeding the risk probability threshold and test users with the first type of risk in their risk status are selected.
[0158] Next, the feature data of the user to be tested for the first type of risk is input into the next layer of the first layer of the multilayer hidden Markov model for processing to obtain the risk state output by the next layer model, until the risk state output by each layer of the multilayer hidden Markov model is obtained. In this way, users with multiple risk states superimposed due to risk can be identified.
[0159] After identifying users with the first type of risk, further risk analysis is conducted. Specifically, for users with the first type of risk identified by each sub-model in the first layer, the proportion of various business transactions processed under each risk state in each model over a month is calculated. The Apriori algorithm is then used to identify strongly correlated transactions exceeding the correlation threshold for each risk user group. This further identifies risky transactions and risky transaction combinations. For users with risk types identified in the second layer or higher, the risk combination states are selected using the same method as in the first layer analysis, and then the risk prediction results are obtained using the same method.
[0160] After obtaining the risk prediction results, the risk prediction results can be further output and uploaded to the audit platform. The risk prediction results include: the analyzed risk status, relevant information of risk users, and risk business or combinations thereof compiled into corresponding documents.
[0161] In yet another alternative embodiment of the invention, step 13 may be followed by:
[0162] Step 135: Update the multilayer hidden Markov model.
[0163] In this embodiment, as business processes are updated and iterated, new business processes will emerge, potentially generating new types of risks with novel characteristics. Therefore, for these new risks, it is necessary to extract their risk features, train a sub-Hidden Markov Model (NMM) for this type of risk, place it in the first layer of the model, and then update the model parameters of the other layers of the multi-layer NMM. Furthermore, before updating the multi-layer NMM, the corresponding feature data should also be updated.
[0164] Figure 3 A flowchart illustrating a specific communication service risk prediction method provided by an embodiment of the present invention is shown. Figure 3 As shown, the method includes the following steps:
[0165] Step 31: Obtain business processing, business-related data, and risk tags from the business system;
[0166] Step 32: Use the Apriori algorithm and trained word embedding model to extract user business transaction features, including extracting business attribute features, extracting business association features, extracting business text features, extracting business time features, extracting business combination features, and extracting user attribute features.
[0167] Step 33: After normalizing and vectorizing the features, input them into the multilayer hidden Markov model for layer-by-layer training to construct the multilayer hidden Markov model.
[0168] Step 34: Test each layer of the Hidden Markov Model sequentially, and update and adjust the model parameters based on the test results;
[0169] Step 35: Input the feature information of the user to be predicted into the trained multilayer hidden Markov model to obtain the risk prediction result. Then, analyze and organize the result and upload it to the risk audit platform for relevant departments to review.
[0170] Step 36: Update the model regularly as the business updates and iterates.
[0171] In the above embodiments of the present invention, communication service data of the user to be tested is acquired; feature data of the communication service data is obtained based on the communication service data of the user to be tested; the feature data is input into a preset service risk prediction model for processing to obtain the risk prediction result; the preset service risk prediction model is obtained after training a multilayer hidden Markov model based on a training set. This can make communication service risk prediction more time-saving, labor-saving, and secure. At the same time, the association features obtained by the association algorithm can make risk mining more time-saving and labor-saving; the word embedding model of the service description is trained by the Word2Vector algorithm, which can extract a series of semantically similar words based on a single word. This method can replace manual inspection of each service description by eye. In the early stage, it is only necessary to select the service descriptions of services under a few service types, then extract the risk keywords, and then use the algorithm to find all similar keywords and use these keywords to label the service features. In the later stage, services can be found only based on these labels, saving manpower and time.
[0172] Figure 4 A schematic diagram of the structure of a communication service risk prediction device 40 provided in an embodiment of the present invention is shown. Figure 4 As shown, the device includes:
[0173] Module 41 is used to acquire communication service data of the user under test;
[0174] Processing module 42 is used to obtain feature data of the communication service data based on the communication service data of the user under test;
[0175] Output module 43 is used to input the feature data into a preset business risk prediction model for processing to obtain the risk prediction result; the preset business risk prediction model is obtained by training a multilayer hidden Markov model based on a training set.
[0176] Optionally, the processing module 42 is further configured to obtain at least one of the text features, service association features, service time features, service attribute features, and user attribute features of the communication service data based on the communication service data of the user under test;
[0177] The feature data of the communication service data is obtained by combining at least two of the text features, service association features, service time features, service attribute features and user attribute features.
[0178] Optionally, the processing module 42 is further configured to obtain a service description of the communication service data based on the communication service data of the user under test;
[0179] The word vector model is used to extract keywords from the business description to obtain keywords related to communication business risk assessment.
[0180] The keywords are classified to obtain the text features of the communication service.
[0181] Optionally, the processing module 42 is further configured to obtain all services of the target user who has handled the associated services based on the communication service data of the user under test;
[0182] Obtain the support and credibility of all services for the target user under test;
[0183] Based on the support and credibility, a business association table is obtained;
[0184] Based on the association table and the pre-stored association relationships of associated services, the business association characteristics of the user under test are obtained.
[0185] Optionally, a multilayer Hidden Markov Model is trained through the following process:
[0186] Obtain feature data of the training business data in the training set;
[0187] The feature data of the training business data are normalized and vectorized to obtain the vectorized processing result;
[0188] The vectorization result is input into the first layer of the Hidden Markov Model for processing to obtain the first output result.
[0189] Based on the first output result, the parameters of the hidden Markov model are updated through a Bayesian filter to obtain the first layer of the trained hidden Markov model.
[0190] Based on the trained first-layer Hidden Markov Model, obtain the risk state sequence;
[0191] The optimal risk state sequence is input into the next layer of the Hidden Markov Model, and the above steps are repeated p times to obtain a trained p-layer Hidden Markov Model, where p is a positive integer.
[0192] Optionally, the trained p-layer Hidden Markov Model is:
[0193]
[0194] Among them, f i Let F represent the sequence of risk states predicted by the i-th layer of the Hidden Markov Model. The risk state sequences output by each layer are combined in chronological order to obtain a combined state feature F of the same length, where 1 ≤ i ≤ p, 1 ≤ j ≤ p. Let λ = (A, B, π) be the probability in the Hidden Markov Model, where the state at time t is q. i Under the given conditions, the observation sequence from time t+1 to time T is o t+1 ,o t+2 ,o t+3 ,...o T The probability of is the reverse probability, denoted as β. t (i)=P(o t+1 ,o t+2 ,o t+3 ,...o T |i t =q i ,λ); Given that the die used in the t-th round is q i ,β t (i) represents the number of observation points from the (t+1)th to the Tth time, which is 0. t+1 ,o t+2 ,o t+3 ,...o T The probability of.
[0195] Optionally, the output module 43 is further configured to input the feature data into the first layer of the trained multilayer hidden Markov model for processing to obtain the test user of the first type of risk.
[0196] The feature data of the user to be tested for the first type of risk is input into the next layer of the first layer of the multilayer hidden Markov model for processing to obtain the risk state output by the next layer model, until the risk state output by each layer of the multilayer hidden Markov model is obtained.
[0197] Under the risk state obtained from the output of each layer of the multi-layer Hidden Markov Model, the business proportion of each business handled by the user under test;
[0198] Based on the business proportion and associated thresholds, the risk prediction results for the risk combination status of the business handled by the user under test are determined.
[0199] Figure 5 This diagram illustrates a specific implementation structure of a communication service risk prediction device provided in an embodiment of the present invention. Figure 5 As shown, the device includes:
[0200] Data acquisition module 51 is used to acquire business processing and business-related data and risk tags from the business system;
[0201] Feature extraction module 52 is used to extract user transaction features, including extracting transaction attribute features, extracting transaction association features, extracting transaction text features, extracting transaction time features, extracting transaction combination features, and extracting user attribute features.
[0202] The multilayer hidden Markov model construction module 53 is used to normalize and vectorize the features and then input them into the multilayer hidden Markov model for layer-by-layer training to construct the multilayer hidden Markov model.
[0203] The model testing and parameter adjustment module 54 is used to test each layer of the Hidden Markov Model sequentially and update and adjust the model parameters based on the test results.
[0204] The prediction and result output module 55 is used to input the feature information of the user to be predicted into the trained multilayer hidden Markov model to obtain its risk prediction result. This result is then analyzed, organized, and uploaded to the risk audit platform for relevant departments to review.
[0205] The model update module 56 is used to periodically update the model as the business updates and iterates.
[0206] In the two communication service risk prediction devices, the acquisition module 41 includes a data acquisition module 51, the processing module 42 includes a feature extraction module 52, a multi-layer hidden Markov model construction module 53, and a model testing and parameter adjustment module 54, and the output module 43 includes a prediction and result output module 55 and a model update module 56.
[0207] It should be noted that this embodiment is a device embodiment corresponding to the above method embodiment. All implementation methods in the above method embodiment are applicable to this device embodiment and can achieve the same technical effect.
[0208] This invention provides a non-volatile computer storage medium storing at least one executable instruction that can execute the communication service risk prediction method in any of the above method embodiments.
[0209] Figure 6 The diagram shows a structural schematic of a computing device provided in an embodiment of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computing device.
[0210] like Figure 6 As shown, the computing device may include a processor, a communications interface, memory, and a communications bus.
[0211] The processor, communication interface, and memory communicate with each other via a communication bus. The communication interface is used to communicate with other network elements, such as clients or other servers. The processor executes programs, specifically the relevant steps described in the embodiment of the communication service risk prediction method for computing devices.
[0212] Specifically, the program may include program code, which includes computer operation instructions.
[0213] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computing device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.
[0214] Memory is used to store programs. Memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk drive.
[0215] Specifically, the program can be used to cause the processor to execute the communication service risk prediction method in any of the above method embodiments. The specific implementation of each step in the program can be found in the corresponding descriptions of the steps and units in the above communication service risk prediction method embodiments, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments, and will not be repeated here.
[0216] The algorithms or displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, the embodiments of the present invention are not directed to any particular programming language. It should be understood that the embodiments of the present invention described herein can be implemented using various programming languages, and the above description of specific languages is for the purpose of disclosing the best mode of implementation of the embodiments of the present invention.
[0217] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.
[0218] Similarly, it should be understood that, in order to streamline the embodiments of the invention and aid in understanding one or more of the various inventive aspects, features of the embodiments of the invention are sometimes grouped together in a single embodiment, figure, or description thereof in the above description of exemplary embodiments of the invention. However, this disclosure should not be construed as reflecting an intention that the claimed embodiments of the invention require more features than are expressly recited in each claim. Rather, as reflected in the following claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of the invention.
[0219] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.
[0220] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination.
[0221] The various component embodiments of the present invention can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some or all of the components according to the embodiments of the present invention. The embodiments of the present invention can also be implemented as device or apparatus programs (e.g., computer programs and computer program products) for performing part or all of the methods described herein. Such programs implementing the embodiments of the present invention can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.
[0222] It should be noted that the above embodiments are illustrative of the present invention and not restrictive of the invention, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. Embodiments of the present invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be construed as limiting the order of execution.
Claims
1. A method of communication service risk prediction, characterized by, The method includes: Obtain the communication service data of the user under test; Based on the communication service data of the user under test, obtain the feature data of the communication service data; The feature data is input into a preset business risk prediction model for processing to obtain the risk prediction result; the preset business risk prediction model is obtained by training a multilayer hidden Markov model based on a training set. The process involves inputting the feature data into a preset business risk prediction model for processing to obtain the risk prediction result, including: The feature data is input into the first layer of the trained multilayer hidden Markov model for processing to obtain the risk state sequence of the first type of risk, wherein the first layer model includes multiple sub-hidden Markov models. The risk state sequence of the first type of risk is input into the next layer of the first layer of the multilayer hidden Markov model for processing to obtain the risk state output by the next layer model, until the risk state output by each layer of the multilayer hidden Markov model is obtained. Under the risk state obtained from the output of each layer of the multi-layer Hidden Markov Model, the business proportion of each business handled by the user under test; Based on the business proportion and associated thresholds, the risk prediction results for the risk combination status of the business handled by the user under test are determined.
2. The method of claim 1, wherein, Based on the communication service data of the user under test, characteristic data of the communication service data is obtained, including: Based on the communication service data of the user under test, obtain at least one of the following: text features, service association features, service time features, service attribute features, and user attribute features of the communication service data. The feature data of the communication service data is obtained by combining at least two of the text features, service association features, service time features, service attribute features and user attribute features.
3. The method of claim 2, wherein, Based on the communication service data of the user under test, obtain the text features of the communication service data, including: Based on the communication service data of the user under test, obtain the service description of the communication service data; The word vector model is used to extract keywords from the business description to obtain keywords related to communication business risk assessment. The keywords are classified to obtain the text features of the communication service.
4. The method of claim 2, wherein, Based on the communication service data of the user under test, obtain the service association features of the communication service data, including: Based on the communication service data of the target user, obtain all services of the target user who has handled the associated services; Obtain the support and credibility of all services for the target user under test; Based on the support and credibility, a business association table is obtained; Based on the association table and the pre-stored association relationships of associated services, the business association characteristics of the user under test are obtained.
5. The method of claim 1, wherein, Multilayer Hidden Markov Models are trained through the following process: Obtain feature data of the training business data in the training set; The feature data of the training business data are normalized and vectorized to obtain the vectorized processing result; The vectorization result is input into the first layer of the Hidden Markov Model for processing to obtain the first output result. Based on the first output result, the parameters of the hidden Markov model are updated through a Bayesian filter to obtain the first layer of the trained hidden Markov model. Based on the trained first-layer Hidden Markov Model, obtain the risk state sequence; The sequence of risk states is input into the next layer of the Hidden Markov Model, and the above steps are repeated p times to obtain a trained p-layer Hidden Markov Model, where p is a positive integer.
6. The method of claim 5, wherein, The trained p-layer Hidden Markov Model is as follows: in, Let F represent the sequence of risk states predicted by the i-th layer of the Hidden Markov Model. The risk state sequences output by each layer are combined in chronological order to obtain a combined state feature F of the same length, where 1 ≤ i ≤ p, 1 ≤ j ≤ p. Hidden Markov Model Medium probability, where the state at time t is Under the given conditions, the observation sequence from time t+1 to time T is as follows: The probability of is the reverse probability, denoted as . It is known that the die used in the t-th round is... , The number of observation points from the (t+1)th to the Tth time is The probability of.
7. A communication service risk prediction device, characterized in that, include: The acquisition module is used to acquire the communication service data of the user under test. The processing module is used to obtain feature data of the communication service data based on the communication service data of the user under test; The output module is used to input the feature data into a preset business risk prediction model for processing to obtain the risk prediction result; the preset business risk prediction model is obtained by training a multilayer hidden Markov model based on a training set. The output module is further configured to: input the feature data into the first layer of the trained multilayer hidden Markov model for processing to obtain a risk state sequence of the first type of risk, wherein the first layer model includes multiple sub-hidden Markov models; The risk state sequence of the first type of risk is input into the next layer of the first layer of the multilayer hidden Markov model for processing to obtain the risk state output by the next layer model, until the risk state output by each layer of the multilayer hidden Markov model is obtained. Under the risk state obtained from the output of each layer of the multi-layer Hidden Markov Model, the business proportion of each business handled by the user under test; Based on the business proportion and associated thresholds, the risk prediction results for the risk combination status of the business handled by the user under test are determined.
8. A computing device, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the communication service risk prediction method as described in any one of claims 1-6.
9. A computer storage medium storing at least one executable instruction that causes a processor to perform an operation corresponding to the communication service risk prediction method as described in any one of claims 1-6.
Citation Information
Patent Citations
Risk identification method and device
CN106296195A
Event time sequence data processing method, device and system for recognizing user behaviors
CN107402921A
