Method and device for identifying strategy of cheating order based on deep learning
By using a deep learning autoencoder model, the timeliness and scalability issues of ride-hailing fraud order identification are solved, enabling more accurate fraud order prediction and risk control, and adapting to changes in driver fraud methods.
Patent Information
- Application Number
- CN202411193105.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-11-26
AI Technical Summary
Existing technologies suffer from insufficient timeliness, insufficient learning depth, and poor scalability when identifying fraudulent ride-hailing orders. Traditional methods are difficult to cope with drivers' constantly changing cheating methods and cannot be extended to new scenarios on their own.
An autoencoder model based on deep learning is adopted. Through feature engineering and iterative optimization, an autoencoder model is generated. The autoencoder model is used to extract features from order data and predict the probability of fraud, thereby realizing the identification of fraudulent orders.
It improves the accuracy and stability of fraudulent order identification, can learn the internal evolution patterns of data, has strong generalization ability, can cover more scenarios, and adapt to changes in driver fraud methods.
Smart Images

Figure CN119027143A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for identifying cheating orders, in particular to a method for identifying cheating orders based on deep learning, and belongs to the technical field of information identification. Background Art
[0002] With the rapid development of online ride-hailing, various ride-hailing platforms have emerged one after another. Each platform has also launched various marketing activities to reward drivers, thereby retaining drivers and maintaining an active and highly dependent source of vehicles. However, there are some drivers who use various means, including but not limited to false passenger information, simulated GPS, etc., to obtain improper benefits, thereby damaging the interests of the platform and possibly committing a crime.
[0003] The traditional method of identifying cheating in online ride-hailing is to derive some strategies through data anomalies or classification algorithms (random forest) in machine learning for identification. For example, some drivers frequently match with the same passenger: because according to the existing ecosystem of online ride-hailing, the drivers that passengers can match for travel are random, and the probability of drivers and passengers matching multiple times at the same time is very small; some drivers have abnormally stable and high turnover: normal drivers will have a stable income range, but if some drivers take advantage of a large amount of wool, their income flow will be different from the performance of drivers in the market; the trip fee confirmed by the driver each time is higher than the estimated price: there is a high probability that the driver uses simulated GPS software, through additional routes, the platform's pricing is higher, thereby achieving the purpose of collecting more money, etc.; Random forest inputs a series of driver characteristics, such as order acceptance rate, order completion rate, hourly wage rate, vehicle dispatch time, price / distance estimated / actual ratio, order qualification rate and other aspects, and classifies them through the random forest model, and outputs a probability value between 0 and 1 to indicate the authenticity of the driver.
[0004] However, there are still the following problems in practical application:
[0005] 1. Lack of timeliness. Drivers are constantly cheating in various ways. Relying on certain specific strategies and random forest models can only play a certain role in a short period of time. Drivers will inevitably develop new cheating methods in the future, and it will be impossible to identify the new cheating methods by then.
[0006] 2. Insufficient learning depth. Usually through strategies or machine learning, the ability to handle exceptions is poor. It is only a reasoning demonstration of the surface of the data. It cannot learn the internal laws of data evolution, cannot make correct judgments on sudden situations, and cannot give exact outputs for some ambiguous data.
[0007] 3. Poor scalability. It is a historical experience that cannot cover all scenarios and cannot be extended in batches. It is only used to discover some existing scenarios and cannot be expanded to new scenarios by itself.
[0008] To this end, a method and device for identifying a cheating order strategy based on deep learning are proposed. Summary of the invention
[0009] In view of this, the present invention provides a method and device for identifying a fraudulent order strategy based on deep learning to solve or alleviate the technical problems existing in the prior art and at least provide a beneficial option.
[0010] The technical solution of the embodiment of the present invention is implemented as follows: A method for identifying a cheating order strategy based on deep learning, comprising the following steps:
[0011] S1. Historical data acquisition;
[0012] S2, feature engineering processing - generating training sets, test sets and validation sets;
[0013] S3, repeat S1, and generate data of order dimension according to the data;
[0014] S4. Use the autoencoder to load the data set and generate the autoencoder model.
[0015] S5, iterator model effect verification;
[0016] S6, repeat S4 and S5 to optimize the model, and determine the final optimized model based on the evaluation indicators;
[0017] S7, obtain real-time data and perform feature engineering on the data;
[0018] S8. Import model - import data into the autoencoder model to generate cheating probability prediction results based on each order.
[0019] Further preferably, in S1, real historical data is randomly selected from an online database.
[0020] Further preferably, in S2, the training set accounts for 70-80% of the total data, the test set accounts for 15-20% of the total data, and the validation set accounts for 5-10% of the total data.
[0021] Further preferably, in S3, real historical data is randomly selected again from the online database, and data of the order dimension is generated based on the selected historical data.
[0022] Further preferably, the step S4 further includes the following steps:
[0023] S41, using the autoencoder to load the training set and the test set, generating a data iterator for gradually calling the iterative model;
[0024] S42, initialize the model and hyperparameters, including hidden layer, output layer, gradient propagation, additional gradient, loss function, number of training rounds and size of training data for each training;
[0025] S43. Model training is performed until the loss function converges and an autoencoder model is generated.
[0026] Further preferably, in S43, the autoencoder model can check the feature importance of each input dimension;
[0027] The basis for feature importance is the influence / contribution of each dimension to the result Y during model training.
[0028] Further preferably, in S5, the validation set data is read and the data is input into the autoencoder model to verify the effect of the model in the validation set.
[0029] Further preferably, in S6, the evaluation indicators of the model include precision and recall;
[0030] Among them, the precision rate indicates the proportion of samples predicted as positive samples that are actually positive samples, that is, how many samples predicted as positive samples by the model are actually positive samples, and the expression is:
[0031] Precision = (number of true positive samples) / (number of samples predicted to be positive samples);
[0032] The recall rate is expressed as the proportion of true samples predicted as positive samples by the model, that is, how many true positive samples the model can correctly predict, and the expression is:
[0033] Recall rate = (number of true positive samples) / (number of true positive samples + number of samples predicted to be negative samples but actually positive samples).
[0034] Further preferably, in S8, the evaluation of the cheating probability prediction result continues to use the evaluation index in S6.
[0035] A device for identifying a cheating order strategy based on deep learning includes a memory, a processor, and a computer program stored in the memory and running on the processor. The processor executes the computer program to implement the above method.
[0036] The embodiment of the present invention has the following advantages due to the adoption of the above technical solution:
[0037] 1. The present invention uses the autoencoder model in the online car-hailing marketing subsidy model to more accurately and realistically predict whether an order is a cheating order, thereby more conveniently and targetedly controlling risks, which is incomparable to traditional strategies and machine learning;
[0038] 2. The autoencoder model of the present invention is constructed based on a neural network of deep learning, which can learn more internal evolution laws, is more sensitive to abnormal values or emergencies, has higher stability, and deep learning has strong generalization ability, scalability, and extensibility, and can cover more scenarios through one model.
[0039] The above summary is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments and features described above, further aspects, embodiments and features of the present invention will be readily apparent by reference to the accompanying drawings and the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0041] Figure 1 It is a schematic diagram of the steps of the present invention;
[0042] Figure 2 This is a schematic diagram of the first part of the data after feature engineering processing of the present invention;
[0043] Figure 3 This is a schematic diagram of the second part of the data after feature engineering processing of the present invention;
[0044] Figure 4 This is a schematic diagram of the third part of the data after feature engineering processing of the present invention;
[0045] Figure 5 This is a schematic diagram of the fourth part of the data after feature engineering processing of the present invention;
[0046] Figure 6 This is a fifth schematic diagram of data after feature engineering processing of the present invention;
[0047] Figure 7 Schematic diagram of the process of the autoencoder of the present invention;
[0048] Figure 8 It is a schematic diagram of the influence / contribution of the dimension of the present invention on the result Y. DETAILED DESCRIPTION
[0049] In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the present invention. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.
[0050] It should be noted that the terms "first", "second", "symmetrical", "array", etc. are only used to distinguish between description and position description purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the definition of "first", "symmetrical", etc. can explicitly or implicitly include one or more of these features; similarly, when the number of certain features is not limited in the form of words such as "two" or "three", it should be noted that this feature also explicitly or implicitly includes one or more feature quantities.
[0051] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0052] Explanation of relevant terms:
[0053] Deep learning: A method of achieving advanced data representation and analysis by building multi-layer neural networks. Its core idea is to learn complex feature representations from a large amount of input data through nonlinear transformations at multiple neural network levels. It is a branch of machine learning.
[0054] Cheating Orders: A fraudulent or illegal online ride-hailing order generated by fraudulent means. Usually, this behavior is to obtain improper benefits and obtain additional discounts, commissions, rewards or other economic benefits by manipulating or violating platform rules and services. Examples include fake orders (users deliberately create fake online ride-hailing orders in order to obtain cancellation fees, subsidies, etc., without actually having real travel needs), GPS cheating (drivers use various means to change or forge their GPS location information to obtain additional mileage subsidies, peak-time fare increases or other rewards), and collusion fraud (drivers and passengers conspire to jointly operate cheating behaviors for personal gain).
[0055] Decision Tree: A machine learning algorithm based on a tree structure, commonly used for classification and regression tasks. It gradually divides the data set into smaller subsets and makes decisions or predictions based on the value of the features.
[0056] Random Forest: An ensemble learning method that constructs multiple decision trees to perform classification, regression, or feature importance assessment.
[0057] Autoencoder: An unsupervised learning algorithm used to learn features that effectively represent input data. It consists of an encoder and a decoder that can convert input data into low-dimensional encodings and reconstruct the original data from the encodings. The goal of the autoencoder is to minimize the reconstruction error, that is, to achieve accurate reconstruction of the input data to the output data. Autoencoders include standard autoencoders (using fully connected neural networks), sparse autoencoders (adding sparsity constraints), convolutional autoencoders (for data image data), and variational autoencoders (for generating models). Autoencoders have a wide range of applications in deep learning and artificial intelligence.
[0058] Embodiment 1
[0059] like Figure 1-8 As shown, an embodiment of the present invention provides a method for identifying a cheating order strategy based on deep learning, comprising the following steps:
[0060] S1. Historical data acquisition;
[0061] S2, feature engineering processing - generating training sets, test sets and validation sets;
[0062] S3, repeat S1, and generate data of order dimension according to the data;
[0063] S4. Use the autoencoder to load the data set and generate the autoencoder model.
[0064] S5, iterator model effect verification;
[0065] S6, repeat S4 and S5 to optimize the model, and determine the final optimized model based on the evaluation indicators;
[0066] S7, obtain real-time data and perform feature engineering on the data;
[0067] S8. Import model - import data into the autoencoder model to generate cheating probability prediction results based on each order.
[0068] In one embodiment, in S1, real historical data is randomly selected from an online database.
[0069] In one embodiment, in S2, the training set accounts for 70% of the total data, the test set accounts for 20% of the total data, and the validation set accounts for 10% of the total data;
[0070] By setting the ratio of training set, test set and validation set according to actual needs, the training, testing and validation of the model can be completed more efficiently;
[0071] The data after feature engineering is as follows Figure 2-7 Description and examples of features are shown below.
[0072] In one embodiment, in S3, real historical data is randomly selected again from the online database, and order dimension data is generated based on the selected historical data.
[0073] The order dimension considers and analyzes business data from the perspective of orders. Order dimension data involves various information related to orders. By analyzing order dimension data, companies can gain insights into customer preferences, optimize inventory management, improve logistics distribution, enhance customer service experience, and develop more effective marketing and sales strategies. In addition, order data analysis can also help identify potential problem areas, such as order delays, insufficient inventory, or decreased customer satisfaction, so that timely measures can be taken to make improvements.
[0074] In one embodiment, Figure 7 As shown, S4 also includes the following steps:
[0075] S41, using the autoencoder to load the training set and the test set, generating a data iterator for gradually calling the iterative model;
[0076] S42, initialize the model and hyperparameters, including hidden layer, output layer, gradient propagation, additional gradient, loss function, number of training rounds and size of training data for each training;
[0077] S43. Model training is performed until the loss function converges and an autoencoder model is generated.
[0078] In one embodiment, Figure 8 As shown, in S43, the autoencoder model can view the feature importance of each input dimension;
[0079] The basis for feature importance is the influence / contribution of each dimension to the result Y during model training.
[0080] In one embodiment, in S5, the validation set data is read and the data is input into the autoencoder model to verify the effect of the model in the validation set.
[0081] The effect of the autoencoder model is verified by using the data in the validation set.
[0082] In one embodiment, in S6, the evaluation index of the model includes precision and recall;
[0083] Among them, the precision rate indicates the proportion of samples predicted as positive samples that are actually positive samples, that is, how many samples predicted as positive samples by the model are actually positive samples, and the expression is:
[0084] Precision = (number of true positive samples) / (number of samples predicted to be positive samples);
[0085] The recall rate is expressed as the proportion of true samples predicted as positive samples by the model, that is, how many true positive samples the model can correctly predict, and the expression is:
[0086] Recall rate = (number of true positive samples) / (number of true positive samples + number of samples predicted to be negative samples but actually positive samples).
[0087] By repeating steps S4 and S5, the model is optimized to generate the optimal autoencoder model, and then the optimized model is evaluated using precision and recall.
[0088] In one embodiment, in S8, the evaluation of the cheating probability prediction result continues to use the evaluation index in S6.
[0089] The real-time data processed by feature engineering is input into the autoencoder model, and then the model is used to generate the cheating probability prediction results for each order. The main online evaluation indicators include precision and recall.
[0090] Embodiment 2
[0091] The embodiment of the present invention also provides a method for identifying a cheating order strategy based on deep learning, comprising the following steps:
[0092] S1. Historical data acquisition;
[0093] S2, feature engineering processing - generating training sets, test sets and validation sets;
[0094] S3, repeat S1, and generate data of order dimension according to the data;
[0095] S4. Use the autoencoder to load the data set and generate the autoencoder model.
[0096] S5, iterator model effect verification;
[0097] S6, repeat S4 and S5 to optimize the model, and determine the final optimized model based on the evaluation indicators;
[0098] S7, obtain real-time data and perform feature engineering on the data;
[0099] S8. Import model - import data into the autoencoder model to generate cheating probability prediction results based on each order.
[0100] In one embodiment, in S1, real historical data is randomly selected from an online database.
[0101] In one embodiment, in S2, the training set accounts for 80% of the total data, the test set accounts for 15% of the total data, and the validation set accounts for 5% of the total data;
[0102] By setting the ratio of training set, test set and validation set according to actual needs, the training, testing and validation of the model can be completed more efficiently.
[0103] In one embodiment, in S3, real historical data is randomly selected again from the online database, and order dimension data is generated based on the selected historical data.
[0104] The order dimension considers and analyzes business data from the perspective of orders. Order dimension data involves various information related to orders. By analyzing order dimension data, companies can gain insights into customer preferences, optimize inventory management, improve logistics distribution, enhance customer service experience, and develop more effective marketing and sales strategies. In addition, order data analysis can also help identify potential problem areas, such as order delays, insufficient inventory, or decreased customer satisfaction, so that timely measures can be taken to make improvements.
[0105] In one embodiment, S4 further includes the following steps:
[0106] S41, using the autoencoder to load the training set and the test set, generating a data iterator for gradually calling the iterative model;
[0107] S42, initialize the model and hyperparameters, including hidden layer, output layer, gradient propagation, additional gradient, loss function, number of training rounds and size of training data for each training;
[0108] S43. Model training is performed until the loss function converges and an autoencoder model is generated.
[0109] In one embodiment, in S43, the autoencoder model may check the feature importance of each input dimension;
[0110] The basis for feature importance is the influence / contribution of each dimension to the result Y during model training.
[0111] In one embodiment, in S5, the validation set data is read and the data is input into the autoencoder model to verify the effect of the model in the validation set.
[0112] The effect of the autoencoder model is verified by using the data in the validation set.
[0113] In one embodiment, in S6, the evaluation index of the model includes precision and recall;
[0114] Among them, the precision rate indicates the proportion of samples predicted as positive samples that are actually positive samples, that is, how many samples predicted as positive samples by the model are actually positive samples, and the expression is:
[0115] Precision = (number of true positive samples) / (number of samples predicted to be positive samples);
[0116] The recall rate is expressed as the proportion of true samples predicted as positive samples by the model, that is, how many true positive samples the model can correctly predict, and the expression is:
[0117] Recall rate = (number of true positive samples) / (number of true positive samples + number of samples predicted to be negative samples but actually positive samples).
[0118] By repeating steps S4 and S5, the model is optimized to generate the optimal autoencoder model, and then the optimized model is evaluated using precision and recall.
[0119] In one embodiment, in S8, the evaluation of the cheating probability prediction result continues to use the evaluation index in S6.
[0120] The real-time data processed by feature engineering is input into the autoencoder model, and then the model is used to generate the cheating probability prediction results for each order. The main online evaluation indicators include precision and recall.
[0121] Embodiment 3
[0122] The embodiment of the present invention also provides a method for identifying a fraudulent order strategy based on deep learning, which differs from the first and second embodiments in that the autoencoder model adopts a convolutional neural network algorithm;
[0123] Convolutional neural network (CNN) is a powerful type of neural network designed for processing image data. In short, convolutional neural networks can effectively process spatial information, while autoencoders are an unsupervised learning method that discovers features in data by learning the potential representation of the data. Compared with supervised learning methods that require annotation, autoencoders do not require label information and can be trained without labeled data. They have a certain degree of noise tolerance during the training process and can learn a robust representation of the data. This enables the autoencoder to better reconstruct and restore the original data in the case of noisy or missing data.
[0124] Embodiment 4
[0125] An embodiment of the present invention also provides a device for identifying a cheating order strategy based on deep learning, including a memory, a processor, and a computer program stored in the memory and running on the processor. The processor executes the computer program to implement the method of embodiment one, embodiment two, or embodiment three.
[0126] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of various changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.
Claims
1. A method for identifying cheating orders based on deep learning, characterized in that: The following steps are involved: S1. Historical data acquisition; S2, feature engineering processing - generating training sets, test sets and validation sets; S3, repeat S1, and generate data of order dimension according to the data; S4. Use the autoencoder to load the data set and generate the autoencoder model. S5, iterator model effect verification; S6, repeat S4 and S5 to optimize the model, and determine the final optimized model based on the evaluation indicators; S7, obtain real-time data and perform feature engineering on the data; S8. Import model - import data into the autoencoder model to generate cheating probability prediction results based on each order.
2. The method for identifying cheating orders based on deep learning according to claim 1, characterized in that: In S1, real historical data is randomly selected from the online database.
3. The method for identifying cheating orders based on deep learning according to claim 1, characterized in that: In S2, the training set accounts for 70-80% of the total data, the test set accounts for 15-20% of the total data, and the validation set accounts for 5-10% of the total data.
4. The method for identifying cheating orders based on deep learning according to claim 2, characterized in that: In S3, real historical data is randomly selected again from the online database, and data of the order dimension is generated based on the selected historical data.
5. The method for identifying cheating orders based on deep learning according to claim 1, characterized in that: The S4 further includes the following steps: S41, using the autoencoder to load the training set and the test set, generating a data iterator for gradually calling the iterative model; S42, initialize the model and hyperparameters, including hidden layer, output layer, gradient propagation, additional gradient, loss function, number of training rounds and size of training data for each training; S43. Model training is performed until the loss function converges and an autoencoder model is generated.
6. The method for identifying cheating orders based on deep learning according to claim 5, characterized in that: In S43, the autoencoder model can check the feature importance of each input dimension; The basis for feature importance is the influence / contribution of each dimension to the result Y during model training.
7. The method for identifying cheating orders based on deep learning according to claim 3, characterized in that: In S5, the validation set data is read and the data is input into the autoencoder model to verify the effect of the model in the validation set.
8. The method for identifying cheating orders based on deep learning according to claim 1, characterized in that: In S6, the evaluation indicators of the model include precision and recall; Among them, the precision rate indicates the proportion of samples predicted as positive samples that are actually positive samples, that is, how many samples predicted as positive samples by the model are actually positive samples, and the expression is: Precision = (number of true positive samples) / (number of samples predicted to be positive samples); The recall rate is expressed as the proportion of true samples predicted as positive samples by the model, that is, how many true positive samples the model can correctly predict, and the expression is: Recall rate = (number of true positive samples) / (number of true positive samples + number of samples predicted to be negative samples but actually positive samples).
9. The method for identifying cheating orders based on deep learning according to claim 8, characterized in that: In S8, the evaluation of the cheating probability prediction result continues to use the evaluation index in S6.
10. A device for identifying a fraudulent order based on deep learning, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: The processor executes the computer program to implement the method according to any one of claims 1 to 9.
