Binning method and apparatus based on neural network model, and electronic device

Abstract

The application provides a binning method and device based on a neural network model, an electronic device, and a computer readable storage medium. The method comprises: obtaining a sample set of a neural network model; performing sample binning processing on a plurality of feature samples in the sample set to obtain a binning sequence comprising at least two sample bins; determining a weight of evidence (WOE) of each sample bin; extracting a longest monotonic sequence in a monotonic relationship from the binning sequence based on the WOE of each sample bin, the longest monotonic sequence comprising at least two sample bins; and merging sample bins not in the longest monotonic sequence among the at least two sample bins based on the longest monotonic sequence to obtain at least two target bins. Through the application, the robustness and interpretability of the neural network model can be significantly improved.

Description

Technical Field

[0001] This application relates to data processing technology, and more particularly to a binning method, apparatus, and electronic device based on a neural network model. Background Technology

[0002] Binning is a technique for grouping multiple data sets, with each group referred to as a bin. In the field of machine learning, by binning continuous features, the features can be discretized. This allows for faster model iteration and effectively enhances the robustness and interpretability of the model when training it based on these discretized features.

[0003] Related technologies employ binning methods such as equidistant binning and equal-frequency binning for data binning. However, the binning obtained using these methods cannot effectively improve the robustness and interpretability of the trained model. Summary of the Invention

[0004] This application provides a binning method, apparatus, electronic device, and storage medium based on a neural network model, which can achieve binning with significantly improved robustness and interpretability of the neural network model.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] This application provides a binning method based on a neural network model, including:

[0007] Obtain the sample set for the neural network model;

[0008] Multiple feature samples in the sample set are binned to obtain a binning sequence that includes at least two sample bins.

[0009] Determine the evidence weight for each of the aforementioned sample bins;

[0010] Based on the evidence weights of each sample bin, the longest monotonic sequence that is monotonic is extracted from the binning sequence, and the longest monotonic sequence includes at least two sample bins.

[0011] Based on the longest monotonic sequence, sample bins that are not in the longest monotonic sequence in the at least two sample bins are merged to obtain at least two target bins.

[0012] This application provides a binning device based on a neural network model, comprising:

[0013] The acquisition module is used to acquire the sample set of the neural network model;

[0014] The binning processing module is used to perform sample binning processing on multiple feature samples in the sample set to obtain a binning sequence including at least two sample bins.

[0015] The evidence weight determination module is used to determine the evidence weight of each of the aforementioned sample bins;

[0016] An extraction module is used to extract the longest monotonic sequence that is monotonic from the binning sequence based on the evidence weight of each of the sample bins, wherein the longest monotonic sequence includes at least two sample bins.

[0017] The merging module is used to merge sample bins that are not in the longest monotonic sequence from the at least two sample bins, based on the longest monotonic sequence, to obtain at least two target bins.

[0018] In the above scheme, the merging module is further configured to determine at least one subsequence in the binning sequence that is not in the longest monotonic sequence, the subsequence including at least one sample bin; merge the sample bins in the at least one subsequence respectively to obtain at least one merged bin; and merge the at least one merged bin based on the longest monotonic sequence to obtain at least two target bins.

[0019] In the above scheme, the merging module is further configured to perform the following processing on each merged bin: obtain the sample bin that is adjacent to the merged bin and has the smallest difference in evidence weight in the longest monotonic sequence; merge the merged bin with the adjacent sample bin that has the smallest difference in evidence weight to obtain the at least two target bins.

[0020] In the above scheme, the binning device based on the neural network model further includes: a longest monotonic sequence determination module, used to determine the evidence weight of the at least one merged bin; based on the evidence weight of each sample bin in the longest monotonic sequence and the evidence weight of each merged bin, to redetermine the longest monotonic sequence; correspondingly, the merging module is also used to merge the at least one merged bin based on the redetermined longest monotonic sequence to obtain at least two target bins.

[0021] In the above scheme, the merging module is further configured to merge the at least one merged bin with the sample bins in the longest monotonic sequence to obtain a merged bin sequence, the merged bin sequence including at least two merged sample bins; determine the evidence weight of the sample bins in the merged bin sequence respectively; based on the evidence weight of the sample bins in the merged bin sequence, if it is determined that the merged bin sequence is not monotonic, redetermine the longest monotonic sequence; based on the redetermined longest monotonic sequence, merge the sample bins that are not in the redetermined longest monotonic sequence to obtain the at least two target bins.

[0022] In the above scheme, the merging module is further used to merge the sample bins that are not in the re-determined longest monotonic sequence with the sample bins that are in the re-determined longest monotonic sequence to obtain a new merged bin sequence; for the new merged bin sequence, the step of determining the evidence weights of the sample bins in the merged bin sequence is returned until the merged bin sequence is monotonic to obtain the at least two target bins.

[0023] In the above scheme, the merging module is further used to determine, based on the evidence weight of the sample bins in the merged binning sequence, that the merged binning sequence is monotonic, and then use the sample bins in the merged binning sequence as target bins to obtain the at least two target bins.

[0024] In the above scheme, the feature samples in the sample set include sample features; the binning device based on the neural network model further includes: a sorting module, used to sort the multiple feature samples based on the sample features included in each feature sample in the sample set to obtain a feature sample sequence; correspondingly, the binning processing module is also used to perform sample binning processing on multiple feature samples in the feature sample sequence.

[0025] This application provides an electronic device, including:

[0026] Memory, used to store executable instructions;

[0027] The processor, when executing executable instructions stored in the memory, implements the binning method based on a neural network model provided in the embodiments of this application.

[0028] This application provides a computer-readable storage medium storing executable instructions for inducing a processor to execute and implement the binning method based on a neural network model provided in this application.

[0029] The embodiments of this application have the following beneficial effects:

[0030] Compared to related technologies that use equidistant or equal-frequency binning for binning, this embodiment extracts the longest monotonic sequence from the binning sequence obtained by binning the sample set, and merges the sample bins that are not in the longest monotonic sequence to obtain multiple target bins with monotonic relationships. This allows the bin merging operation to revolve around the longest monotonic sequence, thus preserving more bins in the longest monotonic sequence and reducing the number of bin merging operations. As a result, a greater number of monotonic bins are preserved, significantly improving the robustness and interpretability of the neural network model trained based on the target bins obtained in this embodiment. This overcomes the limitation of limited improvement in model performance caused by binning methods used in related technologies. Attached Figure Description

[0031] Figure 1 This is an optional structural diagram of the binning system based on a neural network model provided in an embodiment of this application;

[0032] Figure 2 This is an optional structural schematic diagram of the electronic device provided in an embodiment of this application;

[0033] Figure 3 This is an optional flowchart illustrating the binning method based on a neural network model provided in this application embodiment;

[0034] Figure 4A This is an optional schematic diagram of the binning sequence provided in the embodiments of this application;

[0035] Figure 4B This is an optional schematic diagram of the binning sequence after merging sub-sequences provided in an embodiment of this application;

[0036] Figure 4C This is an optional schematic diagram of the bin sequence after merging the merged bins provided in the embodiments of this application;

[0037] Figure 4D This is an optional schematic diagram of the binning sequence after merging the merged bins based on the redefined longest monotonic sequence provided in this application embodiment;

[0038] Figure 5 This is an optional flowchart illustrating the binning method based on a neural network model provided in this application embodiment;

[0039] Figure 6 This is an optional flowchart illustrating the binning method based on a neural network model provided in this application embodiment;

[0040] Figure 7This is an optional flowchart illustrating the binning method based on a neural network model provided in this application embodiment;

[0041] Figure 8 This is an optional flowchart illustrating the binning method based on a neural network model provided in this application embodiment;

[0042] Figure 9 This is an optional flowchart illustrating the binning method based on a neural network model provided in this application embodiment;

[0043] Figure 10 This is a schematic diagram of an optional component structure of the binning device based on a neural network model provided in this application embodiment. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0045] In the following description, references are made to “some embodiments”, which describe sample binning of all possible embodiments. However, it is understood that “some embodiments” may be the same sample binning or different sample binning of all possible embodiments and may be combined with each other without conflict.

[0046] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0047] In the following description, the terms "first", "second", etc. are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first", "second", etc., may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0048] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0049] 1) Weight of Evidence (WOE), used to measure the difference between the normal sample distribution and the default sample distribution. WOE = ln(normal sample distribution / default sample distribution).

[0050] In this embodiment, normal samples are those that have passed the credit assessment, and default samples are those that have failed the credit assessment. For each sample bin, if the ratio of the number of samples that have passed the credit assessment to the number of samples that have failed the credit assessment in that sample bin is a first ratio, and the ratio of the number of samples that have passed the credit assessment to the number of samples that have failed the credit assessment in the total sample set is a second ratio, then the WOE of the sample bin is ln(first ratio / second ratio).

[0051] 2) A risk control model is a classification model used to output a risk control assessment result for a target user based on input user characteristic data. It can be implemented using neural network models, such as Long Short-Term Memory (LSTM) networks.

[0052] 3) Feature segmentation data, which contains feature segmentation information of bins, used to indicate the feature intervals of bins.

[0053] For example, if the sample set is divided into five age groups based on age: under 10 years old, 10-20 years old, 20-40 years old, 40-60 years old, and over 60 years old, then the feature segmentation data can be feature intervals such as [0,10], (10,20], (20,40], (40,60], and (60,+∞], with the unit being years. Furthermore, the feature segmentation data can also be binning intervals, such as 10, 20, 40, and 60, with the unit being years.

[0054] In implementing the embodiments of this application, the inventors discovered that when the evidence weights of bins are monotonic, the robustness and interpretability of the model can be significantly improved. However, in related technologies, binning methods such as equidistant binning or equal-frequency binning do not guarantee monotonic evidence weights. To ensure monotonic evidence weights, the inventors, during binning, continuously merge the tails of monotonically increasing subsequences and the heads of monotonically decreasing subsequences, or merge the tails of monotonically decreasing subsequences and the heads of monotonically increasing subsequences, based on the discretized binning sequence, until the evidence weights of all bins are monotonic. However, this approach often leaves only a few bins, and its contribution to improving model performance is still insufficient.

[0055] Based on this, embodiments of this application provide a binning method, apparatus, electronic device, and computer-readable storage medium based on a neural network model, which can achieve binning with significantly improved robustness and interpretability of the neural network model.

[0056] First, the binning system based on a neural network model provided in the embodiments of this application will be described, see [link to relevant documentation]. Figure 1 , Figure 1This is an optional architecture diagram of the binning system 100 based on a neural network model provided in this application embodiment. Terminal 400 connects to server 200 via network 300, which can be a wide area network (WAN), a local area network (LAN), or a combination of both, using a wireless link for data transmission. In some embodiments, terminal 400 can be a laptop, tablet, desktop computer, smartphone, dedicated messaging device, portable gaming device, smart speaker, smartwatch, etc., but is not limited to these. Server 200 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Network 300 can be a wide area network (WAN), a local area network (LAN), or a combination of both. Terminal 400 and server 200 can be directly or indirectly connected via wired or wireless communication, which is not limited in this application embodiment.

[0057] Terminal 400 is used to collect sample sets of neural network models and send the sample sets to server 200.

[0058] Server 200 is used to acquire a sample set of neural network models; perform sample binning on multiple feature samples in the sample set to obtain a binning sequence including at least two sample bins; determine the evidence weight of each sample bin; based on the evidence weight of each sample bin, extract the longest monotonic sequence that is monotonic from the binning sequence, the longest monotonic sequence including at least two sample bins; based on the longest monotonic sequence, merge the sample bins that are not in the longest monotonic sequence among the at least two sample bins to obtain at least two target bins; and return at least two target bins to terminal 400.

[0059] Terminal 400 is also used to receive a message from server 200 indicating that binning processing is complete. When the neural network model is a risk control model, in response to a model training trigger operation for the risk control model, it sends a model training instruction to server 200. In response to the model training instruction, server 200 trains the risk control model based on at least two target bins and generates a message indicating that model training is complete, which is then returned to terminal 400.

[0060] Terminal 400 is also used to receive a message from server 200 indicating that model training is complete, and in response to a risk prediction trigger operation for risk prediction of a target user, send a risk prediction instruction for the target user to server 200, so that server 200 responds to the risk prediction instruction, performs risk prediction for the target user, obtains a risk control assessment result, and returns it to terminal 400, so that terminal 400 outputs the risk control assessment result.

[0061] See Figure 2 , Figure 2 This is a schematic diagram of the structure of the electronic device 500 provided in the embodiments of this application. See also... Figure 2 , Figure 2 This is an optional schematic diagram of the structure of the electronic device 500 provided in this application embodiment. In practical applications, the electronic device 500 can be implemented as follows: Figure 1 Terminal 400 or server 200 in the middle, with electronic devices as Figure 1 Taking server 200 as an example, an electronic device implementing the binning method based on a neural network model according to the embodiments of this application will be described. Figure 2 The illustrated electronic device 500 includes at least one processor 510, a memory 550, at least one network interface 520, and a user interface 530. The various components in the electronic device 500 are coupled together via a bus system 540. It is understood that the bus system 540 is used to implement communication between these components. In addition to a data bus, the bus system 540 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 2 The general labeled all buses as Bus System 540.

[0062] The processor 510 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0063] User interface 530 includes one or more output devices 531 that enable the presentation of media content, including one or more speakers and / or one or more visual displays. User interface 530 also includes one or more input devices 532, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0064] The memory 550 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 550 may optionally include one or more storage devices physically located away from the processor 510.

[0065] The memory 550 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 550 described in this application embodiment is intended to include any suitable type of memory.

[0066] In some embodiments, memory 550 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.

[0067] Operating system 551 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks;

[0068] The network communication module 552 is used to reach other computing devices via one or more (wired or wireless) network interfaces 520, exemplary network interfaces 520 including: Bluetooth, WiFi, and Universal Serial Bus (USB), etc.

[0069] Presentation module 553 is configured to enable the presentation of information (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices 531 (e.g., a display screen, a speaker, etc.) associated with user interface 530;

[0070] The input processing module 554 is used to detect and translate one or more user inputs or interactions from one or more input devices 532.

[0071] In some embodiments, the bin-sorting device based on a neural network model provided in this application can be implemented in software. Figure 2 A binning device 555 based on a neural network model, stored in memory 550, is shown. This device can be software in the form of programs and plug-ins, and includes the following software modules: an acquisition module 5551, a binning processing module 5552, an evidence weight determination module 5553, an extraction module 5554, and a merging module 5555. These modules are logically connected and can therefore be arbitrarily combined or further split according to their implemented functions. The functions of each module will be described below.

[0072] In other embodiments, the binning device based on the neural network model provided in this application can be implemented in hardware. As an example, the binning device based on the neural network model provided in this application can be a processor in the form of a hardware decoding processor, which is programmed to execute the binning method based on the neural network model provided in this application. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0073] The binning method based on a neural network model provided in this application will be described in conjunction with exemplary applications and implementations of the server provided in the embodiments of this application.

[0074] See Figure 3 , Figure 3 This is an optional flowchart illustrating a binning method based on a neural network model provided in an embodiment of this application, which will be combined with... Figure 3 The steps shown are explained.

[0075] Step 101: The server obtains the sample set of the neural network model.

[0076] Here, the neural network model can be a binary classification model or a multi-class classification model. This embodiment uses a neural network model as a binary classification model for illustration. Accordingly, each feature sample in the sample set has at least one sample feature, and each feature sample is labeled with an event result. This event result includes two types of events: first event and second event. For ease of subsequent explanation, this embodiment uses "event" and "non-event" to represent the first event and the second event, respectively. In actual implementation, the server can collect the sample set from a sample library containing a large number of feature samples, crawl the required data from web pages as feature samples to obtain the sample set, or generate the required feature samples through a large number of testing processes, etc. This embodiment does not limit the specific method of obtaining the sample set.

[0077] For example, if the neural network model is a risk control model, then each feature sample in the sample set of the risk control model has sample features, and the event results labeled on each feature sample are the target risk control assessment results. The risk control assessment results include two event types: a first event of passing the credit assessment and a second event of failing the credit assessment. Feature samples can be user feature samples, which include at least one user feature data. User feature data may include, but is not limited to, user account data, user business data, user device information, and user trajectory. Account data may include, but is not limited to, account name, mobile phone number, bank card number, name, age, gender, ID card number, and address. Business data refers to data within a business scenario; for example, if the business scenario is loan business, business data includes loan data, repayment data, and loan consumption data. Device information includes the model of the user's commonly used devices and the environmental information of the devices. User trajectory includes the user's access sequence, operation frequency, and key press information.

[0078] In practice, the server can obtain the sample set of the neural network model based on user uploads, or it can obtain the sample set by collecting data from relevant platforms. For example, it can obtain the sample set of the risk control model by collecting user characteristic data and corresponding risk control assessment results from a credit platform. Here, the credit platform can be a financial application platform, such as an online banking platform or a shopping platform.

[0079] Step 102: Perform sample binning on multiple feature samples in the sample set to obtain a binning sequence that includes at least two sample bins.

[0080] Here, the server bins multiple feature samples according to their sample characteristics, resulting in a binning sequence containing at least two sample bins. Specifically, the server can bin multiple feature samples according to their sample feature values. For example, for the sample feature of age, the server bins multiple feature samples based on age size. For instance, multiple feature samples can be divided into sample bins corresponding to several age groups, such as under 10 years old, 10-20 years old, 20-30 years old, 30-40 years old, 40-50 years old, 50-60 years old, and over 60 years old. The sample bins in the binning sequence are then sorted according to the size of the age group they correspond to. The server can also bin multiple feature samples according to their classification. For example, for the feature of address, multiple feature samples can be binned based on the geographical range to which the address belongs, such as by city, resulting in multiple sample bins corresponding to multiple cities. The sample bins in the binning sequence can be sorted based on the city's Gross Domestic Product (GDP) or by the city's geographical latitude and longitude, etc. In actual implementation, a feature sample includes multiple sample features, and the server bins multiple feature samples based on each sample feature to obtain at least two target bins corresponding to each sample feature.

[0081] In some embodiments, based on Figure 3 Following step 101, the following can also be performed: based on the sample features included in each feature sample in the sample set, sort the multiple feature samples to obtain a feature sample sequence. Correspondingly, step 102 includes: performing sample binning processing on the multiple feature samples in the feature sample sequence.

[0082] Here, after obtaining the sample set, the server first sorts the multiple feature samples based on the sample features included in each feature sample in the sample set, and then performs binning processing based on the sorted feature samples. In actual implementation, the server performs binning processing on multiple feature samples based on preset rules. Here, the preset rules can be, for example, based on the sample features of the feature samples, according to a certain feature value interval, dividing multiple feature samples into a binning sequence that includes at least two sample bins divided by the preset feature value interval. The preset rules can also be to divide multiple feature samples into a certain number of sample bins, with each sample bin based on an equal feature value interval.

[0083] Step 103: Determine the evidence weight for each sample bin.

[0084] In practice, the server determines the evidence weight of each sample bin based on the event results of different classifications of each feature sample in each sample bin. Specifically, the server determines the number of feature samples corresponding to the first event and the number of feature samples corresponding to the second event in the sample set, and performs the following processing for each sample bin: determines the number of feature samples corresponding to the first event and the number of feature samples corresponding to the second event in the sample bin, and determines the evidence weight of the sample bin based on the number of feature samples corresponding to the first event and the number of feature samples corresponding to the second event in the sample set, and the number of feature samples corresponding to the first event and the number of feature samples corresponding to the second event in the sample bin. For example, the server uses formula (1) to determine the evidence weight of the sample bin:

[0085]

[0086] Among them, WOE i The evidence weight for sample binning, N event N represents the number of feature samples corresponding to the first event in the sample set. non-event The number of feature samples corresponding to the second event in the sample set. This represents the number of feature samples corresponding to the first event in the sample binning. The number of feature samples corresponding to the second event in the sample binning.

[0087] For a fixed sample set, the evidence weights for any sample binning are defined as follows: N event / N non-event The value remains constant. This can be understood as the sample binning... The larger the ratio, that is, the greater the ratio of the number of feature samples corresponding to the first event to the number of feature samples corresponding to the second event in the sample bin, the greater the evidentiary weight of that sample bin, and vice versa. This is understandable. At that time, the evidence weight of the sample bin is zero; At that time, the evidence weight of the sample bin is less than zero; At that time, the evidence weight of the sample bin is greater than zero.

[0088] Step 104: Based on the evidence weight of each sample bin, extract the longest monotonic sequence that is monotonic from the binning sequence. The longest monotonic sequence includes at least two sample bins.

[0089] In practice, the server uses the evidence weight of each sample bin to represent the value of each sample bin. For example, refer to... Figure 4A , Figure 4AThis is an optional schematic diagram of the binning sequence provided in the embodiments of this application. The sample set is divided into a binning sequence consisting of 10 sample bins, numbered 1-10, with evidence weights of WOE1, WOE2, ..., WOE10, respectively. <WOE9<WO E6<WOE5<WOE2<WOE3 <woe10>WOE4 <WOE7<WOE8。

[0090] In practice, the server determines the longest monotonic sequence in which the evidence weights of each sample bin exhibit a monotonic relationship with the binning order, based on the evidence weights of each bin. Here, the server employs a dynamic programming algorithm to extract the longest monotonic sequence in which the evidence weights exhibit a monotonic relationship from the binning sequence, based on the evidence weights of each sample bin. For example, Figure 4A In the binning sequence shown, the longest monotonic sequence determined by the server is "1-2-3-4-7-8".

[0091] Step 105: Based on the longest monotonic sequence, merge the sample bins that are not in the longest monotonic sequence in at least two sample bins to obtain at least two target bins.

[0092] Here, based on the above formula (1), it can be seen that since N event / N non-event The value of is constant, therefore the weight of evidence varies with . The value of is monotonic. Based on this monotonicity, it can be known that for any sample bins i and i+1 to be merged, regardless of WOE i >WOE i+1 Or WOE i <WOE i+1 The final WOE value corresponding to the binning of the sample after merging will definitely be the WOE. i >WOE>WOE i+1 or WOE i <WOE<WOE i+1 , and WOE i =WOE i+1 At that time, the WOE value of the sample bins obtained after merging remains unchanged, while the number of sample bins in the binning sequence decreases by one.

[0093] Based on this, in actual implementation, the server, based on the longest monotonic sequence, merges the remaining sample bins that are not in the longest monotonic sequence into the longest monotonic sequence, so as to obtain at least two target bins that constitute the monotonic sequence.

[0094] In the above embodiments, the server extracts the longest monotonic sequence that is monotonic from the binning sequence obtained by binning the sample set, and merges the sample bins that are not in the longest monotonic sequence to obtain multiple target bins that are monotonic. This makes the bin merging operation revolve around the longest monotonic sequence, which can retain the bins in the longest monotonic sequence to a greater extent and reduce the bin merging operation, thereby retaining more bins to a greater extent. This significantly improves the performance of the neural network model trained based on at least two target bins.

[0095] In some embodiments, see Figure 5 , Figure 5 This is an optional flowchart illustrating a binning method based on a neural network model provided in an embodiment of this application. Figure 3 Step 105 can be achieved in the following way:

[0096] Step 201: The server determines at least one subsequence in the binning sequence that is not in the longest monotonic sequence, and the subsequence includes at least one sample bin.

[0097] Here, a subsequence consists of at least two adjacent sample bins in the binning sequence, and none of the sample bins in the subsequence are in the longest monotonic sequence. In practice, the server obtains all sample bins in the binning sequence except for the longest monotonic sequence, and determines adjacent sorted sample bins to form a subsequence based on the sorting of the sample bins in the binning sequence excluding the longest monotonic sequence. For example, Figure 4A In the binning sequence shown, the longest monotonic sequence is "1-2-3-4-7-8". The server then obtains sample bins "5", "6", "9", and "10" from the binning sequence excluding the longest monotonic sequence. Based on the sorting of the obtained sample bins, the server determines the adjacent sorted sample bins to form subsequences. Here, sample bins "5" and "6" are adjacent sorted, and sample bins "9" and "10" are adjacent sorted. Therefore, the server takes the sequence "5-6" formed by sample bins "5" and "6" as a subsequence, and the sequence "9-10" formed by sample bins "9" and "10" as a subsequence. It should be understood that in some embodiments, sample bins not in the longest monotonic sequence do not constitute subsequences, i.e., the number of subsequences not in the longest monotonic sequence in the binning sequence is zero. In this case, the server will not execute step 202, but will instead use the sample bins not in the longest monotonic sequence in the binning sequence as the merge sequence and execute step 203. Furthermore, if the binning sequence itself is a monotonic sequence, then each sample bin in that binning sequence will be directly used as the target bin, without performing bin merging operations.

[0098] Step 202: Merge the sample bins in at least one subsequence to obtain at least one merged bin.

[0099] In practice, the server merges the sample bins in each subsequence separately, thus combining each subsequence into a single merged bin. Specifically, the server merges the feature samples from each sample bin in each subsequence into a single sample bin, which is then used as the merged bin. It can be understood that the number of merged bins corresponds to the number of subsequences. For example, see [link to example]. Figure 4B , Figure 4B This is an optional schematic diagram of the binning sequence after merging the subsequences provided in the embodiments of this application. Here, the server merges the subsequences "5-6" and "9-10" respectively to obtain the merged bin "11" after merging the sample bins "5" and "6", and the merged bin "12" after merging the sample bins "9" and "10".

[0100] Step 203: Based on the longest monotonic sequence, merge at least one merged bin to obtain at least two target bins.

[0101] In practice, the server incorporates at least one merged bin into the longest monotonic sequence to obtain at least two target bins constituting the monotonic sequence. Here, the server identifies the sample bins adjacent to the merged bin in the longest sequence, merges the merged bin with its adjacent sample bins to obtain a merged bin sequence, and then determines the monotonicity of the merged bin sequence. If the merged bin sequence is monotonic, each sample bin in the merged bin sequence is used as a target bin, completing the binning process for the sample set. If the merged bin sequence is not monotonic, sample bins not in the merged bin sequence are used as merged bins, and the above merging process is repeated until the merged bin sequence becomes monotonic.

[0102] In some embodiments, based on Figure 5 Step 203 can also be implemented as follows: The server performs the following processing for each merged bin: obtain the sample bin that is adjacent to the merged bin and has the smallest difference in evidence weight in the longest monotonic sequence; merge the merged bin with the adjacent sample bin that has the smallest difference in evidence weight to obtain at least two target bins.

[0103] In practice, after the server merges the subsequences to obtain merged bins, since each merged bin is formed by combining at least two sample bins from the subsequences, the evidence weight of the merged bins is unknown. Therefore, in order to merge each merged bin based on its evidence weight, it is necessary to first determine the evidence weight of the merged bins. Here, the server merges each merged bin based on its evidence weight. The evidence weight of each merge bin is determined. In this embodiment, the server calculates the evidence weight of each merge bin using the formula (1) above, based on the ratio of the number of feature samples that passed the credit assessment to the number of feature samples that failed the credit assessment in each merge bin.

[0104] Next, after determining the evidence weights of each merged bin, the server identifies the adjacent bins in the longest monotonic sequence that are adjacent to the merged bins. Then, it compares the evidence weights of the adjacent bins with the evidence weights of the merged bins, obtains the adjacent bin with the smallest difference between the evidence weights of the adjacent bins and the merged bins, and merges the adjacent bin with the smallest difference in evidence weights with the merged bin to obtain at least two target bins after merging.

[0105] For example, see Figure 4C , Figure 4C This is an optional schematic diagram of the bin sequence after merging the merged bins provided in the embodiments of this application. Here, the server performs the following processing for merged bin "11": The server determines the evidence weight WOE11 of merged bin "11", and determines the sample bins "4" and "7" adjacent to merged bin "11" in the longest monotonic sequence. Then, based on the evidence weights WOE11, WOE4, and WOE7 of "11", "4", and "7", the server determines the difference between WOE11 and WOE4 and WOE7 respectively, and compares the absolute value of the difference between WOE11 and WOE4 |WOE11-WOE4| with the absolute value of the difference between WOE11 and WOE7 |WOE11-WOE7|. Based on the relationship between the magnitudes of |WOE11-WOE4| and |WOE11-WOE7|, the server determines the sample bin "4" with the smallest difference in evidence weight with merged bin "11". After that, the server merges merged bin "11" and sample bin "4" to obtain the merged sample bin "13". Similarly, the server performs the above processing on merged bin "12", determining that it has a sample bin "8" that is adjacent to merged bin "12" and has the smallest difference in evidence weight. Merged bin "12" and sample bin "8" are then merged to obtain the merged sample bin "14". At this point, the server obtains the merged bin sequence "1-2-3-13-7-14". In actual implementation, the server continues to redetermine the longest monotonic sequence and merges sample bins that are not in the longest monotonic sequence until all sample bin sequences are in the longest monotonic sequence. The sample bin in the longest monotonic sequence is then used as the target bin.

[0106] In the above embodiments, by merging the merged bins with the adjacent sample bins with the smallest difference in evidence weights, the loss of the number of bins during the bin merging process can be minimized, thereby retaining more target bins and significantly improving the robustness and interpretability of the neural network model trained based on the target bins.

[0107] In some embodiments, based on Figure 5 Before step 203, the following steps can also be performed: the server determines the evidence weight of at least one merged bin; based on the evidence weight of each sample bin and the evidence weight of each merged bin in the longest monotonic sequence, the longest monotonic sequence is re-determined. Correspondingly, step 203 can also be implemented as follows: the server merges at least one merged bin based on the re-determined longest monotonic sequence to obtain at least two target bins.

[0108] In practice, the server first determines the evidence weights of each merged bin. After determining the evidence weights of each merged bin, it redetermines the longest monotonic sequence based on the evidence weights of each merged bin and the evidence weights of each sample bin in the longest monotonic sequence. Here, the server uses a dynamic programming algorithm to redetermine the longest monotonic sequence. It can be understood that the redetermined longest monotonic sequence may include merged bins obtained by merging subsequences. After obtaining the redetermined longest monotonic sequence, the server continues to merge bins that are not in the redetermined longest monotonic sequence until all merged bins are in the longest monotonic sequence. Each sample bin in the longest monotonic sequence is then used as the target bin, completing the binning process for the sample set.

[0109] For example, refer to Figure 4D , Figure 4D This is an optional schematic diagram of the binning sequence after merging the merged bins based on the redefined longest monotonic sequence provided in this application embodiment. Here, after merging bins "11" and "12" to obtain the binning sequence "1-2-3-13-7-14" based on the longest monotonic sequence "1-2-3-4-7-8", the server redetermines the longest monotonic sequence based on this binning sequence. Here, the redefined longest monotonic sequence is "1-2-3-13-7". The server continues to merge the sample bin "14" that is not in the longest monotonic sequence based on the redefined longest monotonic sequence "1-2-3-13-7". Here, the server can determine the sample bin "7" that is adjacent to sample bin "14" and has the smallest difference in evidence weight in the longest monotonic sequence. Then, sample bin "14" is merged with the adjacent sample bin "7" to obtain the merged sample bin "15", so as to obtain the merged bin sequence "1-2-3-13-15". When the merged bin sequence is monotonic, at least two target bins that constitute the monotonic bin sequence are obtained.

[0110] In the above embodiments, by redetermining the longest monotonic sequence, the merged bins that can form monotonic sequences with the previous longest monotonic sequence are avoided, thus ensuring that the final target bins retain more bins.

[0111] In some embodiments, based on Figure 5 Step 203 can also be implemented as follows: The server merges at least one merged bin with the sample bins in the longest monotonic sequence to obtain a merged bin sequence, which includes at least two merged sample bins; the evidence weights of the sample bins in the merged bin sequence are determined respectively; based on the evidence weights of the sample bins in the merged bin sequence, if it is determined that the merged bin sequence is not monotonic, the longest monotonic sequence is re-determined; based on the re-determined longest monotonic sequence, the sample bins that are not in the re-determined longest monotonic sequence are merged to obtain at least two target bins.

[0112] In actual implementation, after the server merges each merged bin based on the longest monotonic sequence to obtain the merged bin sequence, it continues to determine the evidence weight of each sample bin in the merged bin sequence. Based on the evidence weight of each sample bin in the merged bin sequence, it determines whether the merged bin sequence is monotonic based on the evidence weight. If it is not monotonic, it continues to redetermine the longest monotonic sequence. Based on the longest monotonic sequence, it merges the sample bins that are not in the longest monotonic sequence until the merged bin sequence is monotonic. Then, it obtains each target bin of the merged bin sequence that constitutes the monotonic relationship.

[0113] In some embodiments, based on the redefined longest monotonic sequence, sample bins not in the redefined longest monotonic sequence are merged to obtain at least two target bins. This can be achieved as follows: the server merges sample bins not in the redefined longest monotonic sequence with sample bins in the redefined longest monotonic sequence to obtain a new merged bin sequence; for the new merged bin sequence, the steps of determining the evidence weights of the sample bins in the merged bin sequence are returned until the merged bin sequence is monotonic, thus obtaining at least two target bins.

[0114] For example, see Figure 4C and Figure 4D , Figure 4C The binning sequence shown is the merged binning sequence obtained by merging the merged bins based on the longest monotonic sequence. The server is based on... Figure 4C The merged binning sequence shown redetermines the longest monotonic sequence as "1-2-3-13-7". Then, based on the redetermined longest monotonic sequence, bin "14" (which is not in this longest monotonic sequence) is merged to obtain... Figure 4D The new merged binning sequence "1-2-3-13-15" is shown. Here, for the new merged binning sequence, the server returns the step of determining the evidence weight of each sample bin in the merged binning sequence to continue determining the evidence weight of each sample bin in the merged binning sequence "1-2-3-13-15", and determines whether the merged binning sequence is monotonic. If it is not monotonic, the server further redetermines the longest monotonic sequence and merges the sample bins that are not in the longest monotonic sequence.

[0115] In some embodiments, the server also performs the following processing: when the server determines that the merged binning sequence is monotonic based on the evidence weights of the sample bins in the merged binning sequence, it takes the sample bins in the merged binning sequence as target bins to obtain at least two target bins.

[0116] In practice, if the server determines that the merged binning sequence after binning is monotonic, the binning process ends, and the sample bins in the currently obtained merged binning sequence are used as target bins, resulting in at least two target bins, thus completing the binning process for the sample set. For example, Figure 4D The binning sequence "1-2-3-13-15" shown is monotonic, so the server ends the binning process and... Figure 4D Each sample bin in the shown merged binning sequence is used as the target bin, thus obtaining at least two target bins "1", "2", "3", "13" and "15".

[0117] In the above embodiments, after each bin merging, the evidence weight of each sample bin in the current binning sequence is continuously re-determined, and the current binning sequence is judged to be monotonic based on the evidence weight of each sample bin. If it is not monotonic, the current longest monotonic sequence is re-determined, and then the sample bins that are not in the longest monotonic sequence are merged until the merged binning sequence is monotonic, so that the target bin with the largest number of bins and a monotonic relationship can be obtained.

[0118] In some embodiments, see Figure 6 , Figure 6 This is an optional flowchart illustrating the binning method based on a neural network model provided in this application embodiment. When the neural network model is a risk control model, the sample set is the sample set of the risk control model, and each feature sample in the sample set is labeled with the target risk control assessment result, based on... Figure 3 After step 105, the following can also be executed:

[0119] Step 301: The server determines the feature segment data and target evidence weights for each target bin.

[0120] In practice, the server trains the risk control model based on at least two target bins obtained after binning the sample set. Here, the feature segmentation data of the target bin is the feature range formed by the sample features corresponding to all feature samples within the target bin. For example, if the sample set is binned based on age as a sample feature, the final feature segment data corresponding to each target bin can be five feature intervals: less than 10 years old, 10-20 years old, 20-40 years old, 40-60 years old, and greater than 60 years old. These can be represented as [0,10], (10,20], (20,40], (40,60], and (60,+∞], with the unit being years. The server can also use the interval points of the feature segments as feature segment data, so the feature segment data would be 10 years old, 20 years old, 40 years old, and 60 years old. In addition, the server also determines the target WOE value of each target bin based on the ratio of the number of feature samples that pass the credit assessment to the number of feature samples that fail the credit assessment.

[0121] In some embodiments, each feature sample in the sample set includes at least two sample features. The server performs feature binning on the sample set based on each sample feature to obtain at least two target bins corresponding to each sample feature. The binning sequence consisting of at least two target bins is used as the target binning sequence, thus obtaining the target binning sequence corresponding to each sample feature. For example, the server bins the sample set based on age to obtain the target binning sequence consisting of the above-mentioned multiple target bins. The server also bins the sample set based on education level to obtain the target binning sequence consisting of multiple target bins, such as the target binning sequence consisting of several target bins such as below high school, junior college, bachelor's degree, master's degree, and doctoral degree.

[0122] Step 302: Input each feature sample of the sample set into the risk control model.

[0123] Step 303: Using the risk control model, based on the feature segment data and target evidence weights of each target bin, perform risk prediction on each feature sample in the sample set to obtain the predicted risk control assessment results for each feature sample.

[0124] In practice, the risk control model is a binary classification neural network model, such as a Long Short-Term Memory (LSTM) network. Here, the server takes each feature sample from the sample set as input to the risk control model. Then, the risk control model predicts the risk of each feature sample based on the feature segment data of the target bin and the target evidence weight, thus obtaining the predicted risk control assessment result for each feature sample.

[0125] In some embodiments, based on Figure 6 Step 303 can also be implemented in the following way: The server encodes each feature sample based on the feature segment data of each target bin using the risk control model to obtain the feature code of each feature sample; and performs risk prediction on the feature code of each feature sample based on the target evidence weight of the target bin to which each feature sample belongs using the risk control model to obtain the predicted risk control assessment result of each feature sample.

[0126] Here, the server uses a risk control model to determine the target bin to which a feature sample belongs based on the feature segment data of each target bin and the sample features of the feature sample. It then encodes the feature sample by combining the feature segment data of each target bin and the target bin to which the feature sample belongs.

[0127] For example, if each target bin is obtained by binning based on the sample feature of age, the resulting feature segment data are [0,10], (10,20], (20,40], (40,60], and (60,+∞], in years. For a feature sample with the feature of being 25 years old, it can be encoded as [0 0 1 0 0]. T In this embodiment, the feature sample includes at least two sample features. The server encodes the feature sample based on the feature segment data of the target bin corresponding to each sample feature, obtaining the feature code corresponding to each sample feature. Then, the feature codes corresponding to each sample feature are concatenated to obtain the feature code of the feature sample. For example, if the feature sample includes two features, age and education level, and the feature segment data corresponding to the age sample feature is as described above, and the feature segment data corresponding to the education level sample feature is high school or below, junior college, bachelor's degree, master's degree, and doctoral degree, then for the feature sample with the feature feature of being 25 years old and having a bachelor's degree, the server encodes it as the feature code shown below:

[0128]

[0129] The number of bins corresponding to different sample features may vary. Here, the server takes the number of bins of the target bin for the sample feature with the most bins as the row number for feature encoding. For other sample features with fewer bins than the target bin, after encoding the row number, the remaining rows can be assigned a value of zero. In actual implementation, sample features can also be encoded according to specific needs. This application embodiment does not specifically limit the encoding form of sample features.

[0130] In addition, the server can obtain the feature weights corresponding to each sample feature, and perform feature concatenation based on the feature weights of each sample feature, or perform a weighted sum of the feature codes corresponding to each sample feature to obtain the feature code of the feature sample. It should be noted that in actual implementation, the number of sample features can be selected based on actual needs, and the server needs to encode the feature samples based on all sample features to obtain the feature code of the feature sample.

[0131] Next, the server performs probability prediction on the feature codes of each feature sample to determine the risk control assessment result. Based on the predicted probability value, the server obtains the predicted risk control assessment result for each feature sample. Here, when the predicted probability value, representing the probability of passing the credit assessment, is greater than the probability threshold, the predicted risk control assessment result is "credit assessment passed"; otherwise, it is "credit assessment failed."

[0132] Step 304: Update the model parameters of the risk control model based on the difference between the target risk control assessment results and the predicted risk control assessment results labeled on each feature sample.

[0133] In practical implementation, the server determines the data distribution of the target risk control assessment results for the sample set based on the target risk control assessment results labeled on each feature sample. Then, based on the predicted risk control assessment results for each feature sample, it determines the data distribution of the predicted risk control assessment results for the sample set. Finally, based on the difference between the data distribution of the target risk control assessment results and the data distribution of the predicted risk control assessment results, the server updates the model parameters of the risk control model. Here, the server uses the cross-entropy loss function to determine the difference between the data distribution of the target risk control assessment results and the data distribution of the predicted risk control assessment results. In some embodiments, the server may also use other methods to determine the difference between the two data distributions; this application embodiment does not specifically limit this. Here, through continuous iterative training, when the difference between the two data distributions reaches a difference threshold, the model training is completed, and the trained risk control model is obtained.

[0134] In the above embodiments, the risk control model is trained by using the target bins obtained from the binning process in the embodiments of this application. The trained risk control model has good robustness and interpretability.

[0135] In some embodiments, see Figure 7 , Figure 7 This is an optional flowchart illustrating the binning method based on a neural network model provided in this application embodiment. Each feature sample in the sample set is labeled with the target risk control assessment result, based on... Figure 6 After step 304, the following can also be executed:

[0136] Step 401: The server obtains the user characteristic data of the target user;

[0137] Here, when risk control assessment of a target user is required, the server collects the target user's user characteristic data. Specifically, the server obtains the target user's identity information and, based on this identity information and financial big data from the internet, collects the target user's user characteristic data. In practice, the server also filters and extracts the collected user characteristic data, extracting all user characteristic data corresponding to the input of the risk control model.

[0138] Step 402: Input user characteristic data into the risk control model;

[0139] Step 403: Using a risk control model and based on user characteristic data, perform risk prediction for the target user to obtain the risk control assessment result of the target user.

[0140] In some embodiments, based on Figure 7 Step 403 can also be implemented in the following way: The server uses the risk control model to determine the target bin to which the target user belongs based on user feature data and feature segment data of each target bin; and uses the risk control model to predict the risk of the target user based on the target evidence weight of the target bin to which the target user belongs, and obtains the risk control assessment result of the target user.

[0141] In actual implementation, the server inputs the obtained user feature data into the risk control model. Through the risk control model, the target bins corresponding to the target user based on the user features of each dimension are determined. The feature segment data of the target bins corresponding to the user features of each dimension, as well as the target bins to which the target user belongs based on each dimension, are used to encode the target user to obtain the target user's feature code. Then, the risk control model uses this feature code to predict the risk of the target user and obtains the predicted probability of the risk control assessment result of the target user. When the predicted probability, which represents the probability of passing the credit assessment, is greater than the probability threshold, the risk control assessment result of passing the credit assessment is obtained; otherwise, the risk control assessment result of failing the credit assessment is obtained.

[0142] In the risk control assessment methods described above, the risk control assessment results obtained by using a risk control model trained based on the target bins obtained in the embodiments of this application to predict the risk of target users are more accurate, providing a better reference for risk prediction of credit platforms.

[0143] The binning method based on a neural network model provided in the embodiments of this application will be described next. Figure 8 This is an optional flowchart illustrating the binning method based on a neural network model provided in this application embodiment. See [link / reference]. Figure 8 The binning method based on a neural network model provided in this application embodiment is implemented collaboratively by the terminal and the server.

[0144] Step 501: The terminal responds to the sample set acquisition operation and acquires the sample set of the neural network model.

[0145] Here, the acquisition operation can be an upload operation for a sample set, where the terminal directly receives the sample set uploaded by the user. Alternatively, the acquisition operation can be a trigger operation initiated by the user through the terminal's input device to collect a sample set. Based on this trigger operation, the terminal obtains the collection instruction for the sample set, parses the Uniform Resource Locator (URL) carried in the collection instruction, and accesses the target web server corresponding to that URL to collect the sample set.

[0146] Step 502: The terminal sends the sample set to the server.

[0147] Step 503: The server performs sample binning on multiple feature samples in the received sample set to obtain a binning sequence that includes at least two sample bins.

[0148] Step 504: The server determines the evidence weight of each sample bin.

[0149] Step 505: Based on the evidence weight of each sample bin, the server extracts the longest monotonic sequence that is monotonic from the binning sequence. The longest monotonic sequence includes at least two sample bins.

[0150] Step 506: Based on the longest monotonic sequence, the server merges the sample bins that are not in the longest monotonic sequence in at least two sample bins to obtain at least two target bins.

[0151] Step 507: The server sends a message to the terminal indicating that the binning process is complete.

[0152] Step 508: The terminal receives a message indicating that the binning process is complete.

[0153] In some embodiments, the server also sends at least two target bins to the terminal, and the terminal also receives at least two target bins sent by the server.

[0154] Step 509: In response to the model training trigger operation for the risk control model, the terminal sends a model training instruction to the server.

[0155] Here, the aforementioned neural network model is a risk control model, and its sample set is the sample set of the risk control model. Each feature sample in the sample set is labeled with the target risk control assessment result.

[0156] Here, the model training trigger operation is triggered by the user's input device based on the terminal.

[0157] Step 510: The server responds to the model training instruction and trains the risk control model.

[0158] Here, the specific process of the server training the risk control model is as follows: The server determines the feature segment data and target evidence weights for each target bin; each feature sample in the sample set is input into the risk control model; the risk control model performs risk prediction on each feature sample in the sample set based on the feature segment data and target evidence weights of each target bin, and obtains the predicted risk control assessment result for each feature sample; the model parameters of the risk control model are updated based on the difference between the target risk control assessment result and the predicted risk control assessment result marked on each feature sample.

[0159] Step 511: The server sends a message to the terminal indicating that the model training is complete.

[0160] Step 512: The terminal receives a message indicating that the model training is complete.

[0161] In some embodiments, the server also sends the trained risk control model to the terminal, and the terminal also receives the risk control model sent by the server.

[0162] Step 513: In response to the risk prediction trigger operation for the target user, the terminal sends a risk prediction instruction for the target user to the server.

[0163] Here, the risk prediction instruction can be generated by the terminal in response to a user's risk prediction trigger operation for a target user; or it can be issued by other devices and received by the terminal.

[0164] Step 514: In response to the risk prediction command, the server performs risk prediction for the target user and obtains the risk control assessment result.

[0165] Here, the specific process by which the server performs risk prediction on the target user to obtain the risk control assessment result is as follows: the server obtains the user characteristic data of the target user; the user characteristic data is input into the risk control model; and the risk control model, based on the user characteristic data, performs risk prediction on the target user to obtain the risk control assessment result of the target user.

[0166] Step 515: The server sends the risk control assessment results to the terminal.

[0167] Step 516: The terminal outputs the received risk control assessment results.

[0168] In the above embodiments, the server performs binning on the sample set sent by the terminal, extracts the longest monotonic sequence from the binning sequence, and merges the sample bins that are not in the longest monotonic sequence based on the longest monotonic sequence, resulting in at least two target bins that are monotonic. This allows the bin merging operation to revolve around the longest monotonic sequence, which can retain more bins in the longest monotonic sequence and reduce the number of bin merging operations, thereby retaining a larger number of bins. This significantly improves the performance of the risk control model trained on at least two target bins, and the risk control assessment results obtained by predicting the risk of target users based on this risk control model are more accurate.

[0169] The following will describe an exemplary application of the embodiments of this application in a real-world application scenario.

[0170] Reference Figure 9 , Figure 9 This is an optional flowchart of the binning method based on a neural network model provided in the embodiments of this application.

[0171] Step 601: The server obtains the sample set of the neural network model and performs sample binning on multiple feature samples in the sample set to obtain the binning sequence.

[0172] Here, the sample set includes multiple samples, each with features across multiple dimensions, such as user age, education level, occupation, spending history, loan records, repayment records, and social data. The server performs binning on each feature dimension, resulting in a binning sequence corresponding to each feature. The server bins the sample set based on certain rules, such as equal-distance binning or equal-frequency binning. The resulting binning sequence includes multiple sample bins. For example, see [link to example]. Figure 4A , Figure 4A A binning sequence consisting of 10 sample bins, numbered 1-10, is shown.

[0173] Step 602: Determine whether the evidence weights of the binning sequence are monotonic. If the evidence weights of the binning sequence are not monotonic, proceed to step 603. If the evidence weights of the binning sequence are monotonic, proceed to step 608.

[0174] Here, the method for determining the weight of evidence is the same as in the above embodiment, and will not be repeated here. Next, the server determines whether the weight of evidence in the binning sequence is monotonic based on the WOE value corresponding to each sample in the binning sequence. For example, Figure 4A The binning sequence shown is not monotonic.

[0175] Step 603: Determine the longest monotonic sequence in the binning sequence.

[0176] Here, the server uses a dynamic programming algorithm to determine the longest monotonic sequence based on the evidence weights of each sample bin in the binning sequence. For example, Figure 4A In the binning sequence shown, the longest monotonic sequence consists of sample bins "1", "2", "3", "4", "7" and "8".

[0177] Step 604: Determine if there is a subsequence that is not in the longest monotonic sequence. If it exists, proceed to step 605; otherwise, proceed to step 606.

[0178] In practice, a subsequence consists of at least two adjacent sample bins in the binning sequence, and none of the sample bins in the subsequence are in the longest monotonic sequence. See also Figure 4A , Figure 4A In the binning sequence shown, the sample bins other than the longest monotonic sequence are "5", "6", "9", and "10". Here, "5" and "6" are adjacent consecutive bins, and "9" and "10" are adjacent consecutive bins, thus forming the subsequences "5-6" and "9-10" respectively. Figure 4A The binning sequence shown contains a subsequence that is not in the longest monotonic sequence, and the server executes step 605.

[0179] Step 605: Perform a first-class merging on the longest monotonic sequence and the subsequences not in the longest monotonic sequence to obtain a new binning sequence, and return to step 602.

[0180] Here, the first type of merging specifically refers to merging subsequences that are not in the longest monotonic sequence. For example, see [link to example]. Figure 4A and Figure 4B The subsequences are "5-6" and "9-10". After merging these two subsequences, new sample bins "11" and "12" are obtained. Subsequently, the server performs step 602 on the new bin sequence to determine whether the new bin sequence is monotonic based on the evidence weights.

[0181] Step 606: Determine if there is a single sample bin that is not in the longest monotonic sequence. If it exists, proceed to step 607; otherwise, proceed to step 608.

[0182] In practice, if there is no subsequence that is not in the longest monotonic sequence, then it is necessary to continue to determine whether there is a single sample bin that is not in the longest monotonic sequence. If none of them exist, it means that the longest monotonic sequence has encompassed all the samples in the sample set, and the longest monotonic sequence is the required final binning sequence.

[0183] For example, see Figure 4B In the binning sequence shown in Figure 4 after the first type of merging, the longest monotonic sequence is "1-2-3-4-7-8". The sample bins "11" and "12" are not in the longest monotonic sequence. They are not adjacent and do not constitute a subsequence, that is, there is no subsequence not in the longest monotonic sequence. However, both of them are single sample bins not in the longest monotonic sequence. Therefore, the server executes step 607.

[0184] Step 607: Perform a second-class merging on the bins of the longest monotonic sequence and the individual samples not in the longest monotonic sequence to obtain a new bin sequence, and return to step 602.

[0185] Here, the second type of merging specifically involves: identifying the sample bin in the longest monotonic sequence that is adjacent to a single sample bin and has the smallest difference in evidence weights, and merging the single sample bin with the adjacent sample bin that has the smallest difference in evidence weights. For example, see [link to example]. Figure 4B and 4C After the server determines that individual sample bins "11" and "12" are not in the longest monotonic sequence, it determines that the sample bins adjacent to sample bins "11" and "12" with the smallest difference in evidence weight are "4" and "8". Therefore, the server merges sample bins "11" and "4" to obtain sample bin "13", and merges sample bins "12" and "8" to obtain sample bin "14". At this point, a new bin sequence "1-2-3-13-7-14" is obtained, and then the server returns to step 602 to continue to determine whether the new bin sequence is monotonic.

[0186] Step 608: Output the binning sequence to complete the binning process of the sample set.

[0187] In practice, when the binning sequence is monotonic based on the evidence weights, the server will no longer perform binning merging and will use the monotonic binning sequence as the final binning result.

[0188] The following description continues to illustrate the exemplary structure of the binning device 555 based on the neural network model provided in this application embodiment as a software module. In some embodiments, see [link to relevant documentation]. Figure 10 , Figure 10 This is an optional structural diagram of the binning device based on a neural network model provided in this application embodiment. The software modules stored in the binning device 555 based on the neural network model in the memory 550 may include:

[0189] The acquisition module 5551 is used to acquire the sample set of the neural network model;

[0190] The binning processing module 5552 is used to perform sample binning processing on multiple feature samples in the sample set to obtain a binning sequence including at least two sample bins.

[0191] The evidence weight determination module 5553 is used to determine the evidence weight of each of the sample bins respectively;

[0192] The extraction module 5554 is used to extract the longest monotonic sequence that is monotonic from the binning sequence based on the evidence weight of each of the sample bins, wherein the longest monotonic sequence includes at least two sample bins.

[0193] The merging module 5555 is used to merge sample bins that are not in the longest monotonic sequence in the at least two sample bins based on the longest monotonic sequence, so as to obtain at least two target bins.

[0194] In some embodiments, the merging module 5555 is further configured to determine at least one subsequence in the binning sequence that is not in the longest monotonic sequence, the subsequence including at least one sample bin; merge the sample bins in the at least one subsequence respectively to obtain at least one merged bin; and merge the at least one merged bin based on the longest monotonic sequence to obtain at least two target bins.

[0195] In some embodiments, the merging module 5555 is further configured to perform the following processing on each of the merged bins: obtain the sample bin that is adjacent to the merged bin and has the smallest difference in evidence weights in the longest monotonic sequence; merge the merged bin with the adjacent sample bin that has the smallest difference in evidence weights to obtain the at least two target bins.

[0196] In some embodiments, the software module stored in the binning device 555 based on the neural network model in the memory 550 may further include: a longest monotonic sequence determination module, configured to determine the evidence weights of the at least one merged bin; based on the evidence weights of each sample bin in the longest monotonic sequence and the evidence weights of each merged bin, redetermine the longest monotonic sequence; correspondingly, the merging module 5555 is further configured to merge the at least one merged bin based on the redetermined longest monotonic sequence to obtain at least two target bins.

[0197] In some embodiments, the merging module 5555 is further configured to merge the at least one merged bin with the sample bins in the longest monotonic sequence to obtain a merged bin sequence, the merged bin sequence including at least two merged sample bins; determine the evidence weight of the sample bins in the merged bin sequence; based on the evidence weight of the sample bins in the merged bin sequence, if it is determined that the merged bin sequence is not monotonic, redetermine the longest monotonic sequence; based on the redetermined longest monotonic sequence, merge the sample bins that are not in the redetermined longest monotonic sequence to obtain the at least two target bins.

[0198] In some embodiments, the merging module 5555 is further configured to merge sample bins that are not in the redefined longest monotonic sequence with sample bins that are in the redefined longest monotonic sequence to obtain a new merged bin sequence; for the new merged bin sequence, return to the step of determining the evidence weights of the sample bins in the merged bin sequence respectively, until the merged bin sequence is monotonic, to obtain the at least two target bins.

[0199] In some embodiments, the merging module 5555 is further configured to determine, based on the evidence weights of the sample bins in the merged binning sequence, that the merged binning sequence is monotonic, and then use the sample bins in the merged binning sequence as target bins to obtain the at least two target bins.

[0200] In some embodiments, the feature samples in the sample set include sample features; the software module stored in the binning device 555 based on the neural network model in the memory 550 may further include: a sorting module, used to sort the plurality of feature samples based on the sample features included in each feature sample in the sample set to obtain a feature sample sequence; correspondingly, the binning processing module is also used to perform sample binning processing on the plurality of feature samples in the feature sample sequence.

[0201] It should be noted that the description of the apparatus in this application embodiment is similar to the description of the method embodiment above, and has similar beneficial effects as the method embodiment, so it will not be repeated.

[0202] This application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the binning method based on a neural network model described in this application.

[0203] This application provides a computer-readable storage medium storing executable instructions. When these executable instructions are executed by a processor, they cause the processor to execute the binning method based on a neural network model provided in this application. For example, ... Figure 3 The binning method based on a neural network model is shown.

[0204] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0205] In some embodiments, executable instructions may take the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a standalone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0206] As an example, executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple collaborating files (e.g., a file that stores one or more modules, subroutines, or code sections).

[0207] As an example, executable instructions can be deployed to execute on a single computing device, or on multiple computing devices located in one location, or on multiple computing devices distributed across multiple locations and interconnected via a communication network.

[0208] In summary, the binning method described in this application can significantly improve the robustness and interpretability of neural network models.

[0209] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A binning method based on a neural network model, characterized in that, include: Obtain the sample set for the neural network model; Multiple feature samples in the sample set are binned to obtain a binning sequence comprising at least two sample bins. The feature samples are user feature samples, each including at least one user feature data point. This user feature data includes at least the user's account data, business data, device information, and user trajectory. The account data includes account name, mobile phone number, bank card number, name, age, gender, ID number, and address. The business data is data specific to a business scenario; when the business scenario is a loan business, the business data includes loan data, repayment data, and loan consumption data. The device information includes the device model and the environment in which the device is located. The user trajectory includes the user's access sequence, operation frequency, and keystroke information. Determine the evidence weight for each of the aforementioned sample bins; Based on the evidence weights of each sample bin, the longest monotonic sequence that exhibits a monotonic relationship is extracted from the binning sequence, and the longest monotonic sequence includes at least two sample bins; Determine at least one subsequence in the binning sequence that is not in the longest monotonic sequence, the subsequence comprising at least one sample bin; The samples in the at least one subsequence are merged into at least one merged bin; Based on the longest monotonic sequence, the at least one merged bin is merged to obtain at least two target bins; wherein, the step of merging the at least one merged bin based on the longest monotonic sequence to obtain at least two target bins includes: merging the at least one merged bin with sample bins in the longest monotonic sequence to obtain a merged bin sequence, the merged bin sequence including at least two merged sample bins; determining the evidence weights of the sample bins in the merged bin sequence; based on the evidence weights of the sample bins in the merged bin sequence, if it is determined that the merged bin sequence is not monotonic, re-determining the longest monotonic sequence; merging the sample bins not in the re-determined longest monotonic sequence with the sample bins in the re-determined longest monotonic sequence to obtain a new merged bin sequence; for the new merged bin sequence, returning to the step of determining the evidence weights of the sample bins in the merged bin sequence, until the merged bin sequence is monotonic, thus obtaining the at least two target bins.

2. The method according to claim 1, characterized in that, The process of merging at least one merged bin based on the longest monotonic sequence to obtain at least two target bins includes: The following processing is performed on each of the aforementioned merged bins: Obtain the sample bin that is adjacent to the merged bin and has the smallest difference in evidence weights in the longest monotonic sequence; The merged bin is merged with the adjacent sample bin with the smallest difference in evidence weight to obtain the at least two target bins.

3. The method according to claim 1, characterized in that, Before merging the at least one merged bin based on the longest monotonic sequence to obtain at least two target bins, the method further includes: Determine the evidence weight of each of the at least one merged bin; Based on the evidence weights of each sample bin in the longest monotonic sequence and the evidence weights of each merged bin, the longest monotonic sequence is re-determined. Accordingly, the merging of the at least one merged bin based on the longest monotonic sequence to obtain at least two target bins includes: Based on the newly determined longest monotonic sequence, the at least one merged bin is merged to obtain at least two target bins.

4. The method according to claim 1, characterized in that, The method further includes: Based on the evidence weights of the sample bins in the merged binning sequence, when it is determined that the merged binning sequence is monotonic, the sample bins in the merged binning sequence are taken as target bins to obtain the at least two target bins.

5. The method according to claim 1, characterized in that, The feature samples in the sample set include sample features; after obtaining the sample set of the neural network model, the method further includes: Based on the sample features included in each feature sample in the sample set, the multiple feature samples are sorted to obtain a feature sample sequence; Accordingly, the binning process for multiple feature samples in the sample set includes: Sample binning is performed on multiple feature samples in the feature sample sequence.

6. The method according to any one of claims 1-5, characterized in that, When the neural network model is a risk control model, the sample set is the sample set of the risk control model, and each feature sample in the sample set is labeled with the target risk control assessment result. The method further includes: Determine the feature segment data and target evidence weights for each of the target bins; Each feature sample of the sample set is input into the risk control model; Using the risk control model, based on the feature segmentation data and target evidence weights of each target bin, risk prediction is performed on each feature sample in the sample set to obtain the predicted risk control assessment result for each feature sample. Based on the difference between the target risk control assessment result and the predicted risk control assessment result marked on each of the aforementioned feature samples, the model parameters of the risk control model are updated.

7. The method according to claim 6, characterized in that, The risk control model is used to predict the risk of each feature sample in the sample set based on the feature segmentation data and target evidence weights of each target bin, respectively, to obtain the predicted risk control assessment result for each feature sample, including: Each feature sample is encoded based on the feature segmentation data of each target bin using the risk control model to obtain the feature code of each feature sample. The risk control model is used to predict the risk of each feature sample based on the target evidence weight of the target bin to which each feature sample belongs, thereby obtaining the predicted risk control assessment result of each feature sample.

8. A bin-splitting device based on a neural network model, characterized in that, include: The acquisition module is used to acquire the sample set of the neural network model; The binning module is used to bin multiple feature samples in the sample set to obtain a binning sequence including at least two sample bins. The feature samples are user feature samples, each including at least one user feature data point. This user feature data includes at least the user's account data, business data, device information, and user trajectory. The account data includes account name, mobile phone number, bank card number, name, age, gender, ID number, and address. The business data is data specific to a business scenario; when the business scenario is a loan business, the business data includes loan data, repayment data, and loan consumption data. The device information includes the device model and the environment in which the device is located. The user trajectory includes the user's access sequence, operation frequency, and key press information. The evidence weight determination module is used to determine the evidence weight of each of the aforementioned sample bins; An extraction module is used to extract the longest monotonic sequence that is monotonic from the binning sequence based on the evidence weight of each of the sample bins, wherein the longest monotonic sequence includes at least two sample bins. A merging module is used to determine at least one subsequence in the binning sequence that is not in the longest monotonic sequence, the subsequence including at least one sample bin; The sample bins in the at least one subsequence are merged to obtain at least one merged bin; based on the longest monotonic sequence, the at least one merged bin is merged to obtain at least two target bins; wherein, the step of merging the at least one merged bin based on the longest monotonic sequence to obtain at least two target bins includes: merging the at least one merged bin with the sample bins in the longest monotonic sequence to obtain a merged bin sequence, the merged bin sequence including at least two merged sample bins; determining the evidence weights of the sample bins in the merged bin sequence; based on the evidence weights of the sample bins in the merged bin sequence, if it is determined that the merged bin sequence is not monotonic, re-determining the longest monotonic sequence; merging the sample bins not in the re-determined longest monotonic sequence with the sample bins in the re-determined longest monotonic sequence to obtain a new merged bin sequence; for the new merged bin sequence, returning to the step of determining the evidence weights of the sample bins in the merged bin sequence, until the merged bin sequence is monotonic, to obtain the at least two target bins.

9. An electronic device, characterized in that, include: Memory, used to store executable instructions; A processor, when executing executable instructions stored in the memory, implements the binning method based on a neural network model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions for use by a processor to implement the binning method based on a neural network model as described in any one of claims 1 to 7.

Images