Feature processing method, device and electronic device for artificial intelligence recommendation model

Abstract

The present invention provides a feature processing method, device, electronic device and storage medium of an artificial intelligence recommendation model; the feature processing method of the artificial intelligence recommendation model includes: obtaining input data, converting the data structure of the input data into a unified sample structure; determining a feature group identifier and a feature extraction function corresponding to the artificial intelligence recommendation model; performing feature extraction processing on the converted input data according to the feature extraction function to obtain a feature group identifier, feature identifier and feature value in the input data; and constructing a sample of the artificial intelligence recommendation model according to the feature group identifier, feature identifier and feature value. Through the present invention, it is possible to perform unified feature extraction on different types of input data, improve the strong consistency of features, and improve the efficiency and flexibility of feature extraction.

Description

Technical Field

[0001] The present invention relates to artificial intelligence technology, and in particular to a feature processing method, device, electronic device and storage medium for an artificial intelligence recommendation model. Background Art

[0002] Artificial Intelligence (AI) is the theory, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. Machine Learning (ML), a multidisciplinary field, is at the core of AI and the fundamental way to make computers intelligent. Its applications span a wide range of fields.

[0003] For artificial intelligence recommendation models that rely on machine learning, such as advertising recommendation models, the use process involves two aspects: training and application. In the solutions provided by related technologies, features are usually extracted from the training data stored in the log in an offline state to train the model, and the trained model is pushed to the online model service. Then, features are extracted from the online test data and input into the model service to obtain the recommendation results. In the above solution, the offline and online processes are separated, and there may be potential differences between the offline feature extraction process and the online feature extraction process. As a result, the extracted features of the same data in the offline and online states may be inconsistent, and the strong consistency of the features is poor. Summary of the Invention

[0004] The embodiments of the present invention provide a feature processing method, device, electronic device and storage medium for an artificial intelligence recommendation model, which can ensure the strong consistency of features extracted under different situations and improve the flexibility and efficiency of feature extraction.

[0005] The technical solution of the embodiment of the present invention is achieved as follows:

[0006] An embodiment of the present invention provides a feature processing method for an artificial intelligence recommendation model, including:

[0007] Acquire input data, and convert the data structure of the input data into a unified sample structure;

[0008] Determine the feature group identifier and feature extraction function corresponding to the artificial intelligence recommendation model;

[0009] Performing feature extraction processing on the converted input data according to the feature extraction function to obtain a feature group identifier, a feature identifier, and a feature value in the input data;

[0010] A sample of the artificial intelligence recommendation model is constructed based on the feature group identifier, feature identifier and feature value.

[0011] An embodiment of the present invention provides a feature processing device for an artificial intelligence recommendation model, comprising:

[0012] A conversion module, configured to obtain input data and convert the data structure of the input data into a unified sample structure;

[0013] A function determination module is used to determine the feature group identifier and feature extraction function corresponding to the artificial intelligence recommendation model;

[0014] A feature extraction module is used to perform feature extraction processing on the converted input data according to the feature extraction function to obtain a feature group identifier, a feature identifier and a feature value in the input data;

[0015] A construction module is used to construct a sample of the artificial intelligence recommendation model based on the feature group identifier, feature identifier and feature value.

[0016] An embodiment of the present invention provides an electronic device, including:

[0017] a memory for storing executable instructions;

[0018] The processor is configured to implement the feature processing method of the artificial intelligence recommendation model provided by an embodiment of the present invention when executing the executable instructions stored in the memory.

[0019] An embodiment of the present invention provides a computer-readable storage medium storing executable instructions for causing a processor to execute and implement the feature processing method of the artificial intelligence recommendation model provided by the embodiment of the present invention.

[0020] The embodiments of the present invention have the following beneficial effects:

[0021] The embodiment of the present invention converts the data structure of the input data into a unified sample structure, performs feature extraction processing on the converted input data according to the feature extraction function, obtains a processing result including a feature group identifier, a feature identifier and a feature value, and then constructs a sample of the artificial intelligence recommendation model based on the processing result. The embodiment of the present invention overcomes the data structure differences of input data from different sources, and uniformly processes the input data through the feature extraction function, thereby improving the flexibility and efficiency of feature extraction and ensuring the strong consistency of features in the sample. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a structural diagram of the feature extraction solution provided by the related technology;

[0023] Figure 2This is an optional structural diagram of a feature processing system for an artificial intelligence recommendation model provided by an embodiment of the present invention;

[0024] Figure 3 This is an optional structural diagram of a server provided by an embodiment of the present invention;

[0025] Figure 4A This is an optional flowchart of a feature processing method of an artificial intelligence recommendation model provided by an embodiment of the present invention;

[0026] Figure 4B This is another optional flowchart of the feature processing method of the artificial intelligence recommendation model provided by an embodiment of the present invention;

[0027] Figure 4C 1 is a flow chart of determining a feature group identifier and a feature extraction function corresponding to an artificial intelligence recommendation model provided by an embodiment of the present invention;

[0028] Figure 4D This is a flow chart of performing feature extraction processing on converted input data according to a feature extraction function to obtain feature group identifiers, feature identifiers, and feature values ​​in the input data, provided by an embodiment of the present invention;

[0029] Figure 4E This is a schematic diagram of a sample flow of constructing an artificial intelligence recommendation model based on feature group identifiers, feature identifiers, and feature values, provided by an embodiment of the present invention;

[0030] Figure 5 is a comparative schematic diagram of the feature extraction architecture provided by an embodiment of the present invention;

[0031] Figure 6 It is an optional structural diagram of a feature extraction framework provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0032] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0033] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0034] In the following description, the terms "first\second\third" are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It is understandable that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0035] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention pertains. The terms used herein are for the purpose of describing embodiments of the present invention only and are not intended to limit the present invention.

[0036] The relevant data collection and processing in the embodiments of this application should be strictly in accordance with the requirements of relevant laws and regulations when applied in examples, and the informed consent or separate consent of the personal information subject should be obtained. Subsequent data use and processing should be carried out within the scope of authorization of laws and regulations and the personal information subject.

[0037] Before further explaining the embodiments of the present invention in detail, the nouns and terms involved in the embodiments of the present invention are explained. The nouns and terms involved in the embodiments of the present invention are subject to the following interpretations.

[0038] 1) Instance: Also known as an example, it is used as input to the AI ​​recommendation model to complete the corresponding task. Typically, a sample contains several feature groups, a feature group contains several features, and a feature contains a feature identifier and feature value.

[0039] 2) Feature Group: Also known as a feature domain, a feature group includes at least one feature, and the included features are usually related. In addition, each feature group corresponds to a feature group identifier, which can be abbreviated as GID.

[0040] 3) Feature ID: The feature ID is located at the next level of the feature group ID. The feature ID can be referred to as fid.<gid,fid> Only the combination of can represent a global feature, and the feature identifier cannot exist independently from the feature group identifier.

[0041] 4) Eigenvalue: This is used to represent the meaning of a feature. Depending on the actual application scenario, the eigenvalue can be binary (i.e., 0 or 1) or continuous. The eigenvalue is represented by a floating-point number and can be abbreviated as fvalue.

[0042] 5) Atomic feature: refers to the smallest, indivisible feature. The feature group identifier of an atomic feature is called the atomic feature group identifier, agid for short; the feature identifier of an atomic feature is called the atomic feature identifier, afid for short; and the feature value of an atomic feature is called the atomic feature value, afvalue for short. For example, a sample is "user age: 20:1", where the user age is the atomic feature group identifier, 20 is the atomic feature identifier, and 1 is the atomic feature value. Assuming that in the atomic feature value, the value 1 means correct and the value 0 means incorrect, then in this sample, the result that the user age is 20 is correct.

[0043] 6) Combined features: New features are obtained by combining any number of atomic features. The feature group identifier of a combined feature is called a combined feature group identifier, or cgid for short; the feature identifier of a combined feature is called a combined feature identifier, or cfid for short; and the feature value of a combined feature is called a combined feature value, or cfvalue for short. For example, a sample is "user-age_user-gender: 20_female: 1", where user-age_user-gender is the combined feature group identifier, 20_female is the combined feature identifier, and 1 is the atomic feature value. Assuming that in the atomic feature value, the value 1 means correct and the value 0 means wrong, then in this sample, the result that user-age is 20 and user-gender is female is correct.

[0044] 7) Data structure: The way computers store and organize data, such as a collection of data elements that have one or more specific relationships with each other.

[0045] The inventors found in the process of implementing the present invention that in advertising or recommendation services, artificial intelligence recommendation models are usually used to process relevant data. For algorithmic tasks such as recall and sorting involved in the business, the data flow is usually as follows: Figure 1 As shown, in the offline case, offline training data is extracted from the data log. The training data has an offline data structure. Then, the training data is processed by the offline feature framework to extract the features, so as to perform model training. For example, Figure 1 Model a, model b and model c in the training set, and then push the trained model to the online model service. For the online case, the online feature framework is used to test the data with online data structure, such as Figure 1The user profile, document index, and context data shown in the figure are used for feature extraction. The extracted features are then fed into the model service, the results returned by the model service after model processing are obtained, and the corresponding business operations are performed. In the feature framework, there may be potential inconsistencies between the offline feature framework and the online feature framework. For example, business personnel may update the extraction method of a feature in the offline feature framework based on business requirements, but fail to synchronize the new extraction method with the online feature framework. As a result, for data with the same content but different data structures, the features extracted by the offline and online feature frameworks may be inconsistent, which may adversely affect the application of AI recommendation models.

[0046] An embodiment of the present invention provides a feature processing method, device, electronic device and storage medium for an artificial intelligence recommendation model, which can improve the strong consistency of features and improve the efficiency and flexibility of feature extraction. The following describes an exemplary application of the electronic device provided by an embodiment of the present invention.

[0047] See also Figure 2 , Figure 2 This is an optional architectural diagram of a feature processing system 100 for an artificial intelligence recommendation model provided by an embodiment of the present invention. To implement a feature processing application that supports an artificial intelligence recommendation model, a terminal device 400 (terminal device 400-1 and terminal device 400-2 are shown as examples) is connected to a server 200 via a network 300. The network 300 can be a wide area network or a local area network, or a combination of the two. In addition, the server 200 is in communication with a database 500.

[0048] The terminal device 400 is used to display business applications on the graphical interface 410 (graphical interface 410-1 and graphical interface 410-2 are shown as examples); it is also used to respond to user operations on business applications, generate online test data, and send the test data to the server 200; the server 200 is used to obtain the test data of the terminal device 400; it is also used to convert the data structure of the test data into a unified sample structure; it is also used to determine the feature group identifier and feature extraction function corresponding to the artificial intelligence recommendation model; it is also used to perform feature extraction processing on the converted test data according to the feature extraction function to obtain the feature group identifier, feature identifier and feature value in the test data; it is also used to construct a test sample based on the feature group identifier, feature identifier and feature value in the test data; it is also used to perform prediction processing on the test sample through the artificial intelligence recommendation model to obtain a recommendation result, and send the recommendation result to the terminal device 400; the terminal device 400 is used to display the recommendation result in the business application of the graphical interface 410, and Figure 2 , advertisement A, advertisement B, and advertisement C are shown as examples.

[0049] In addition, the server 200 is also used to obtain training data from the database 500; it is also used to convert the data structure of the training data into a unified sample structure; it is also used to determine the feature group identifier and feature extraction function corresponding to the artificial intelligence recommendation model; it is also used to perform feature extraction processing on the converted training data according to the feature extraction function to obtain the feature group identifier, feature identifier and feature value in the training data; it is also used to construct training samples based on the feature group identifier, feature identifier and feature value in the training data; it is also used to train the artificial intelligence recommendation model based on the training samples.

[0050] The following describes exemplary applications of electronic devices provided by embodiments of the present invention. The electronic devices can be implemented as various types of terminal devices, such as laptop computers, tablet computers, desktop computers, set-top boxes, mobile devices (e.g., mobile phones, portable music players, personal digital assistants, dedicated messaging devices, portable gaming devices), and can also be implemented as servers. The following description uses the electronic device as a server as an example.

[0051] See also Figure 3 , Figure 3 The server 200 provided by the embodiment of the present invention (for example, it can be Figure 2 The server 200 shown in FIG. Figure 3 The server 200 shown includes: at least one processor 210, a memory 250, at least one network interface 220, and a user interface 230. The various components in the server 200 are coupled together via a bus system 240. It is understood that the bus system 240 is used to achieve connection and communication between these components. In addition to including a data bus, the bus system 240 also includes a power bus, a control bus, and a status signal bus. However, for the sake of clarity, the bus system 240 is not described in detail. Figure 3 Various buses are labeled as bus system 240 .

[0052] The processor 210 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., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0053] The user interface 230 includes one or more output devices 231 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 230 also includes one or more input devices 232, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0054] The memory 250 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 250 may optionally include one or more storage devices that are physically remote from the processor 210.

[0055] The memory 250 includes volatile memory or nonvolatile memory, or may include both volatile and nonvolatile memory. The nonvolatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 250 described in the embodiments of the present invention is intended to include any suitable type of memory.

[0056] In some embodiments, the memory 250 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplified below.

[0057] Operating system 251, including system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, and driver layer, which are used to implement various basic services and process hardware-based tasks;

[0058] A network communication module 252 for reaching other computing devices via one or more (wired or wireless) network interfaces 220 , exemplary network interfaces 220 including Bluetooth, WiFi, and USB;

[0059] a presentation module 253 for enabling presentation of information via one or more output devices 231 (e.g., a display screen, a speaker, etc.) associated with the user interface 230 (e.g., a user interface for operating peripheral devices and displaying content and information);

[0060] The input processing module 254 is configured to detect one or more user inputs or interactions from one of the one or more input devices 232 and to translate the detected inputs or interactions.

[0061] In some embodiments, the feature processing device of the artificial intelligence recommendation model provided by the embodiments of the present invention can be implemented in software. Figure 3A feature processing device 255 of an artificial intelligence recommendation model stored in a memory 250 is shown, which can be software in the form of programs and plug-ins, including the following software modules: a conversion module 2551, a function determination module 2552, a feature extraction module 2553 and a construction module 2554. These modules are logical and can therefore be arbitrarily combined or further split according to the functions implemented.

[0062] The functions of each module will be described below.

[0063] In other embodiments, the feature processing device of the artificial intelligence recommendation model provided by the embodiment of the present invention can be implemented in hardware. As an example, the feature processing device of the artificial intelligence recommendation model provided by the embodiment of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the feature processing method of the artificial intelligence recommendation model provided by the embodiment of the present invention. For example, the processor in the form of a hardware decoding processor can adopt 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.

[0064] The feature processing method of the artificial intelligence recommendation model provided in the embodiment of the present invention can be executed by the above-mentioned server, or by a terminal device (for example, Figure 2 The process is executed by the terminal device 400-1 and the terminal device 400-2 shown in the figure, or is executed jointly by the server and the terminal device.

[0065] The following will combine the exemplary applications and structures of the electronic device recorded above to illustrate the process of implementing the feature processing method of the artificial intelligence recommendation model through the feature processing device of the embedded artificial intelligence recommendation model in the electronic device.

[0066] See also Figure 4A , Figure 4A This is an optional flow chart of the feature processing method of the artificial intelligence recommendation model provided by the embodiment of the present invention, which will be combined with Figure 4A The steps shown are explained.

[0067] In step 101, input data is acquired, and the data structure of the input data is converted into a unified sample structure.

[0068] Here, the input data can be the data to be tested or the training data. For example, in the case where the artificial intelligence recommendation model is an advertising recommendation model, the data to be tested can be user portrait data, such as user age and gender, etc. The training data can include user portrait data and user click data (such as click or not click) for the advertisement. Since the data to be tested is usually obtained online so that it can be quickly predicted and fed back, the training data is usually stored in a log. In actual application scenarios, in order to facilitate storage, the data structure of the data in the log is usually different from the data structure online. Therefore, in an embodiment of the present invention, the data structure of the input data is converted into a unified sample structure, thereby overcoming the differences between the data structures and ensuring the consistency of the meaning of the fields in different types of input data. Among them, the unified sample structure is a cross-language data format, including but not limited to protobuf format, flatbuffers format, thrift format, avro format and msgpack format. When converting, the cross-language data format adopted is first determined, and then the toolkit of the corresponding format is called to convert the data structure of the input data into the cross-language data format. After the conversion is completed, a file of the corresponding format is generated, which contains the input data described using the cross-language data format.

[0069] In step 102, a feature group identifier and a feature extraction function corresponding to the artificial intelligence recommendation model are determined.

[0070] The embodiments of the present invention do not limit the type of artificial intelligence recommendation model. For example, the artificial intelligence recommendation model may be a recommendation model that utilizes linear regression, support vector machines, or neural networks. In this step, the feature group identifier corresponding to the artificial intelligence recommendation model configuration is determined. This feature group identifier includes an atomic feature group identifier and / or a combined feature group identifier, determined based on the actual application scenario. For artificial intelligence recommendation models applicable to the same business type, there is a certain commonality in the required data. Therefore, corresponding feature group identifiers can be pre-configured for different business types. In this step, the corresponding configured feature group identifier is determined based on the business type of the artificial intelligence recommendation model. For example, business types include game recommendations, cosmetics recommendations, and TV series recommendations. A corresponding feature group identifier is configured for each business type. When determining that the artificial intelligence recommendation model is applicable to game recommendations, the feature group identifier corresponding to the game recommendation type is determined. Simultaneously with determining the feature group identifier corresponding to the artificial intelligence recommendation model, a feature extraction function corresponding to the artificial intelligence recommendation model is determined. These feature extraction functions are pre-configured, and each feature extraction function corresponds to an atomic feature group identifier. The feature extraction function is used to parse the converted input data to obtain the atomic feature group identifier, as well as the atomic feature identifiers and atomic feature values ​​corresponding to the atomic feature group identifier. When configuring the feature extraction function, for example, if the atomic feature group identifier is "user age", then the "user age" field is configured in the atomic extraction function. In this way, by executing the atomic extraction function, the atomic feature group identifier of "user age" and the atomic feature identifiers and atomic feature values ​​related to the atomic feature group identifier can be identified in the converted input data.

[0071] In step 103, feature extraction processing is performed on the converted input data according to the feature extraction function to obtain a feature group identifier, a feature identifier, and a feature value in the input data.

[0072] After feature extraction processing is performed by the feature extraction function, the atomic feature group identifier, atomic feature identifier, and atomic feature value in the converted input data are obtained. Of course, if the converted input data does not contain a certain atomic feature group identifier, then after the input data is processed by the feature extraction function corresponding to the atomic feature group identifier, the result obtained is empty. It is worth noting that when at least two feature extraction functions are determined, feature extraction processing is performed on the converted input data according to each feature extraction function.

[0073] In step 104, a sample of the artificial intelligence recommendation model is constructed based on the feature group identifier, feature identifier and feature value.

[0074] Based on the triple of <feature group identifier: feature identifier: feature value>, samples for the AI ​​recommendation model are constructed for input into the AI ​​recommendation model. Each feature group identifier corresponds to a separate sample. For example, for the atomic feature "user age," the constructed sample is "user age: 20: 1." It is worth noting that since only the atomic feature group identifier, atomic feature identifier, and atomic feature value are obtained in step 103, if the AI ​​recommendation model also corresponds to combined features, the atomic feature identifiers and atomic feature values ​​must be further combined. The details are explained later.

[0075] Through the embodiment of the invention Figure 4A It can be seen from the above exemplary implementation that the embodiment of the present invention, based on a unified sample structure, utilizes the feature extraction function corresponding to the artificial intelligence recommendation model to perform feature extraction on the input data to obtain the required samples. The embodiment of the present invention shields the data structure differences between the predicted data and the training data, ensures the strong consistency of the features, and improves the flexibility and efficiency of feature extraction.

[0076] In some embodiments, see Figure 4B , Figure 4B This is an optional flow chart of a feature processing method of an artificial intelligence recommendation model provided by an embodiment of the present invention, based on Figure 4A After step 104, in step 105, when the input data is training data, the artificial intelligence recommendation model can be trained according to the sample.

[0077] For example, the artificial intelligence recommendation model is an advertising recommendation model, and the sample constructed based on the input data is "User Age_User Gender_Advertisement: 20_Female_Advertisement A: 1", where the input data is training data, and the eigenvalue "1" indicates that the user clicked on Advertisement A. During training, the user age "20" and the user gender "Female" are input as input parameters to the artificial intelligence recommendation model. Based on the difference between the output parameters of the artificial intelligence recommendation model and the eigenvalue "1" in the sample, the parameters in the artificial intelligence recommendation model are adjusted so that when training is subsequently performed based on new samples, the difference between the output parameters of the artificial intelligence recommendation model and the eigenvalues ​​in the sample is reduced, thereby improving the accuracy of the prediction made by the artificial intelligence recommendation model. For example, when the artificial intelligence recommendation model is a recommendation model that applies a neural network, backpropagation is performed based on the difference between the output parameters and the eigenvalue "1" in the sample, thereby adjusting the weight parameters of each layer of the neural network.

[0078] In some embodiments, the above-mentioned training of the artificial intelligence recommendation model based on the sample can also be achieved in the following manner: creating a copy of the artificial intelligence recommendation model; and training the copy of the artificial intelligence recommendation model based on the sample.

[0079] Here, the model training and model prediction processes can be separated. Specifically, a copy of the AI ​​recommendation model is created and trained on the samples corresponding to the training data. This avoids the confusion caused by the AI ​​recommendation model processing two types of samples at the same time.

[0080] In some embodiments, before training the copy of the artificial intelligence recommendation model based on the sample, it also includes: determining a first accuracy rate of the copy of the artificial intelligence recommendation model based on the sample.

[0081] Here, an accuracy mechanism can be set to determine the timing of updating the artificial intelligence recommendation model. First, the first accuracy of the copy of the artificial intelligence recommendation model is determined based on the sample. For example, the sample includes 100 subsamples, where the subsample refers to a triple in the form of <feature group identifier: feature identifier: feature value>. The input parameters in each subsample are input to the copy of the artificial intelligence recommendation model. For 70 subsamples in the sample, the output parameters of the copy of the artificial intelligence recommendation model match the feature values ​​in the subsample; for the other 30 subsamples, the output parameters do not match the feature values. In this case, the first accuracy can be determined to be 70%.

[0082] After training the copy of the artificial intelligence recommendation model according to the sample, it also includes: determining the second accuracy of the trained copy of the artificial intelligence recommendation model according to the sample corresponding to the new training data; when the second accuracy exceeds the first accuracy, updating the artificial intelligence recommendation model according to the trained copy of the artificial intelligence recommendation model.

[0083] After the training of the copy of the artificial intelligence recommendation model is completed, new training data is obtained and samples are constructed. Similarly, based on the newly constructed samples, the second accuracy of the copy of the trained artificial intelligence recommendation model is determined. When the second accuracy exceeds the first accuracy, the artificial intelligence recommendation model is updated based on the copy of the trained artificial intelligence recommendation model (the artificial intelligence recommendation model here is used to process the samples corresponding to the test data); when the second accuracy does not exceed the first accuracy, the copy of the artificial intelligence recommendation model can continue to be trained based on the newly constructed samples. In this way, the accuracy of the artificial intelligence recommendation model used for prediction is guaranteed, and the reduction in accuracy due to the low quality of training data is avoided.

[0084] In some embodiments, the above-mentioned updating of the artificial intelligence recommendation model based on the trained copy of the artificial intelligence recommendation model can be achieved in the following manner: obtaining the generation rate of the training data; when the generation rate is lower than the generation rate threshold, updating the artificial intelligence recommendation model based on the trained copy of the artificial intelligence recommendation model.

[0085] For example, at specific intervals, the number of newly acquired training data items is counted. When the generation rate falls below a set threshold, such as 10 items per minute and 11 items per minute, the AI ​​recommendation model is updated based on a copy of the trained model. This approach reduces service delays caused by model updates.

[0086] In some embodiments, before training the artificial intelligence recommendation model based on the sample, it also includes: obtaining a feature adjustment identifier and a corresponding adjustment value; in the sample, determining each group of feature group identifiers, feature identifiers and feature values ​​that meet the feature adjustment identifier as a sub-sample to be adjusted; updating the weight value of the sub-sample to be adjusted to the adjustment value; wherein the feature adjustment identifier is one of the following: feature group identifier, feature group identifier and feature identifier.

[0087] For example, based on the business's priorities, a feature adjustment flag and adjustment value are set. For example, if the business prioritizes users aged 20, the feature adjustment flag, including the feature group flag and feature flag, is set to <user age: 20>, with a corresponding adjustment value of 2. A sample contains a triplet of <feature group flag:feature flag:feature value>. For ease of understanding, this triplet is named a subsample. The feature values ​​in a subsample can range from 0 to 1, where 0 represents the corresponding user clicking on ad A and 1 represents the corresponding user clicking on ad B. The subsample's weight (sample_weight) defaults to 1. Based on the feature adjustment flag, subsamples containing <user age: 20> are identified as subsamples to be adjusted (regardless of their feature value), and their weight is updated to 2 to increase their proportion in the total sample. This approach enhances the flexibility of model training, allowing the weight of various subsamples to be adjusted based on actual application scenarios.

[0088] In step 106, when the input data is test data, the artificial intelligence recommendation model is used to perform prediction processing on the sample to obtain a recommendation result.

[0089] Similarly, using the example above, after feature extraction, the sample constructed from the input data is "User Age _ User Gender: 20 _ Female: 1." The input data is the test data, and a feature value of "1" indicates that the sample data is correct. During prediction, the user age "20" and user gender "Female" in the sample are input parameters to the AI ​​recommendation model. The output parameters of the AI ​​recommendation model are used as the probability of the user clicking Ad A. Business logic then determines whether to recommend Ad A to the user, resulting in a corresponding recommendation result. For example, if the probability exceeds 70%, Ad A is recommended; if the probability does not exceed 70%, Ad A is not recommended.

[0090] Through the embodiment of the invention Figure 4B From the above exemplary implementation, it can be seen that the embodiment of the present invention uses samples corresponding to training data as training samples to train the artificial intelligence recommendation model; uses samples corresponding to the test data as test samples, which are input into the artificial intelligence recommendation model to obtain recommendation results, thereby improving the pertinence to different types of input data.

[0091] In some embodiments, see Figure 4C , Figure 4C This is a flow chart of determining the feature group identifier and feature extraction function corresponding to the artificial intelligence recommendation model provided by the embodiment of the present invention, based on Figure 4A , step 102 can be implemented through steps 201 to 206, which will be explained in conjunction with each step.

[0092] In step 201, atomic feature registration information and combined feature registration information are obtained, wherein the atomic feature registration information includes a registered atomic feature group identifier and a corresponding feature extraction function, and the combined feature registration information includes a registered combined feature group identifier and corresponding at least two atomic feature group identifiers.

[0093] In this step, the atomic feature registration information and combined feature registration information are not limited to the AI ​​recommendation model, but are global registration information. Specifically, the atomic feature registration information includes the registered atomic feature group identifier and the corresponding feature extraction function, and the combined feature registration information includes the registered combined feature group identifier and at least two corresponding atomic feature group identifiers. For example, the combined feature registration information includes the combined feature group identifier "user age_user gender" and the corresponding two atomic feature group identifiers "user age" and "user gender."

[0094] In step 202, a feature group identifier set corresponding to the artificial intelligence recommendation model is obtained, and the feature group identifier set is traversed.

[0095] Unlike the registration information in step 201, the acquired feature group identifier set is not global. Instead, it includes only feature group identifiers that serve as building blocks for samples, where samples correspond to the AI ​​recommendation model. The feature group identifier set includes at least one feature group identifier. After acquiring the feature group identifier set, a traversal operation is performed on it.

[0096] In step 203, when the traversed feature group identifier matches the atomic feature registration information, the feature extraction function corresponding to the feature group identifier in the atomic feature registration information is added to the feature extraction function chain.

[0097] Each time a feature group identifier is traversed, the feature group identifier is matched one by one with the atomic feature group identifier in the atomic feature registration information to determine whether the feature group identifier and the atomic feature group identifier in the atomic feature registration information are the same. If they are the same, the feature group identifier and the atomic feature registration information are determined to match, and the feature extraction function corresponding to the feature group identifier in the atomic feature registration information is added to the set feature extraction function chain. It is worth noting that the feature extraction function chain here refers to a collection of feature extraction functions, not a specific storage structure.

[0098] In step 204, when the traversed feature group identifier does not match the atomic feature registration information but matches the combined feature registration information, the atomic feature group identifier corresponding to the feature group identifier in the combined feature registration information is determined.

[0099] If the traversed feature group identifier does not match any of the atomic feature group identifiers in the atomic feature registration information, the feature group identifier is matched one-to-one with each combined feature group identifier in the combined feature registration information to determine whether the feature group identifier is the same as the combined feature group identifier in the combined feature registration information. If they are the same, all atomic feature group identifiers corresponding to the feature group identifier in the combined feature registration information are determined. If the traversed feature group identifier does not match the atomic feature registration information and does not match the combined feature registration information, no action is taken.

[0100] In step 205, a feature extraction function corresponding to the atomic feature group identifier is determined in the atomic feature registration information.

[0101] For each determined atomic feature group identifier, a corresponding feature extraction function is determined in the atomic feature registration information.

[0102] In step 206 , when the feature extraction function does not exist in the feature extraction function chain, the feature extraction function is added to the feature extraction function chain.

[0103] When the determined feature extraction function exists in the feature extraction function chain, no processing is performed; when the determined feature extraction function does not exist in the feature extraction function chain, the feature extraction function is added to the feature extraction function chain.

[0104] For ease of understanding, let's take an example. Let's denote the feature extraction function as fa, and the feature extraction function chain as fa_chain. Suppose the atomic feature registration information obtained includes "agid1-fa1" and "agid2-fa2," where "agid1-fa1" represents the atomic feature group identifier agid1 and the corresponding feature extraction function fa1. The combined feature registration information obtained includes "cgid1-(agid1, agid2)," and the feature group identifier set includes the feature group identifiers agid1 and cgid1. The process of constructing fa_chain is described in numbered form:

[0105] (1) Traverse the feature group identifier agid1 in the feature group identifier set, determine that agid1 matches the atomic feature registration information, and add fa1 corresponding to agid1 in the atomic feature registration information to fa_chain.

[0106] (2) Traverse the feature group identifier set to feature group identifier cgid1, determine that cgid1 does not match the atomic feature registration information, and then determine that cgid1 matches the combined feature registration information, obtaining the two atomic feature group identifiers agid1 and agid2 corresponding to cgid1 in the combined feature registration information. For agid1, its corresponding fa1 in the atomic feature registration information has been added to fa_chain, so it is not processed; for agid2, its corresponding fa2 in the atomic feature registration information does not exist in fa_chain, so fa2 is added to fa_chain.

[0107] (3) All feature group identifiers in the feature group identifier set have been traversed, and fa_chain is output.

[0108] In some embodiments, see Figure 4D , Figure 4D The embodiment of the present invention provides a flow chart of performing feature extraction processing on the converted input data according to the feature extraction function to obtain the feature group identifier, feature identifier and feature value in the input data. Figure 4A and Figure 4C , step 103 can be implemented through steps 207 to 208, which will be explained in conjunction with each step.

[0109] In step 207, the feature extraction function chain is traversed.

[0110] The converted input data is subjected to feature extraction processing according to the feature extraction function chain. First, the feature extraction function chain is traversed.

[0111] In step 208, feature extraction processing is performed on the converted input data according to the traversed feature extraction function to obtain the atomic feature group identifier, atomic feature identifier and atomic feature value in the input data.

[0112] According to each feature extraction function traversed, feature extraction processing is performed on the converted input data until the traversal is completed. Among them, the feature extraction function only identifies the atomic feature group identifier corresponding to the function in the input data, as well as the atomic feature identifiers and atomic feature values ​​related to the atomic feature group identifier.

[0113] In some embodiments, see Figure 4E , Figure 4E This is a flow chart of a sample of constructing the artificial intelligence recommendation model according to the feature group identifier, feature identifier and feature value provided by an embodiment of the present invention, based on Figure 4A 、 Figure 4C and Figure 4D , step 104 can be implemented through steps 209 to 214, which will be explained in conjunction with each step.

[0114] In step 209, the feature group identifier set is traversed.

[0115] When constructing samples of the artificial intelligence recommendation model, the feature group identifier set is also traversed.

[0116] In step 210 , when the traversed feature group identifier matches the atomic feature registration information, the feature group identifier is determined as the atomic feature group identifier, and the atomic feature group identifier, the corresponding atomic feature identifier, and the atomic feature value are added to the sample.

[0117] When the traversed feature group identifier is the same as an atomic feature group identifier in the atomic feature registration information, the feature group identifier is determined as the atomic feature group identifier, and the atomic feature group identifier, the corresponding atomic feature identifier and the atomic feature value (determined by step 208) are added to the sample.

[0118] In step 211 , when the traversed feature group identifier does not match the atomic feature registration information but matches the combined feature registration information, the feature group identifier is determined as a combined feature group identifier.

[0119] When the traversed feature group identifier is different from all atomic feature group identifiers in the atomic feature registration information and is the same as a combined feature group identifier in the combined feature registration information, the feature group identifier is determined as the combined feature group identifier.

[0120] In step 212, at least two atomic feature group identifiers corresponding to the combined feature group identifier in the combined feature registration information are determined, and the atomic feature identifiers and atomic feature values ​​corresponding to the atomic feature group identifiers are determined.

[0121] For the sake of convenience, for example, the traversed feature group identifier is cgid1, and the atomic feature group identifiers corresponding to cgid1 in the combined feature registration information include agid1 and agid2. Then, according to the feature extraction result of step 208, the atomic feature identifier afid1 and the atomic feature value afvalue1 corresponding to agid1, as well as the atomic feature identifier afid2 and the atomic feature value afvalue2 corresponding to agid2 are determined.

[0122] In step 213, at least two of the atomic feature identifiers are combined to obtain a combined feature identifier, and at least two of the atomic feature values ​​are combined to obtain a combined feature value.

[0123] For example, afid1 and afid2 are combined to obtain a combined feature identifier cfid1, and afvalue1 and afvalue2 are combined to obtain a combined feature value cfvalue1.

[0124] In some embodiments, the above-mentioned combined processing of at least two of the atomic feature identifiers to obtain a combined feature identifier can also be achieved in the following way: a first combination operation is performed on at least two of the atomic feature identifiers to obtain a combined feature identifier; wherein, the first combination operation includes at least one of the following: hash operation, bitwise OR operation, bitwise AND operation, bitwise XOR operation.

[0125] Depending on the actual application scenarios, different first combination operations can be applied to combine at least two atomic feature identifiers (such as the above-mentioned afid1 and afid2) to obtain a combined feature identifier (such as the above-mentioned cfid1). The first combination operation includes but is not limited to hash operation, bitwise OR operation, bitwise AND operation and bitwise XOR operation.

[0126] In some embodiments, the above-mentioned combination processing of at least two of the atomic eigenvalues ​​to obtain a combined eigenvalue can also be achieved in the following way: a second combination operation is performed on at least two of the atomic eigenvalues ​​to obtain a combined eigenvalue; wherein the second combination operation includes at least one of the following: a product operation, an addition operation.

[0127] Depending on the actual application scenario, different second combination operations can be applied to combine at least two atomic eigenvalues ​​(such as the above-mentioned afvalue1 and afvalue2) to obtain a combined eigenvalue (such as the above-mentioned cfvalue1). The second combination operation includes but is not limited to multiplication operations and addition operations.

[0128] In step 214 , the combined feature group identifier, the combined feature identifier, and the combined feature value are added to the sample.

[0129] After the combination processing is performed to obtain the combined feature identifier and combined feature value, they are added to the sample in the form of a triplet of <combined feature group identifier: combined feature identifier: combined feature value>, for example<cgid1:cfid1:cfval ue1> .

[0130] Through the embodiment of the invention Figure 4C It can be seen from the above exemplary implementation that the embodiment of the present invention realizes sample generation related to the artificial intelligence recommendation model by obtaining registration information and determining the corresponding feature extraction function according to the feature group identification set of the artificial intelligence recommendation model, thereby improving the flexibility and scalability of feature extraction, that is, determining the relevant features in the input data according to different feature group identification sets.

[0131] See also Figure 5 , Figure 5 3 is a comparative diagram of the feature extraction architecture provided by an embodiment of the present invention. Figure 5 The left figure is the feature extraction architecture provided by the related technology. For offline training data, the offline feature framework is used to extract features, and the model is trained based on the extracted features; for online test data, the online feature framework is used to extract features, and the extracted features are processed according to the model to obtain the results. Figure 5 The right figure is a feature extraction architecture obtained by applying the feature processing method of an embodiment of the present invention, wherein both offline training data and online test data are extracted using a unified feature framework, avoiding inconsistency in the extracted features caused by differences between frameworks.

[0132] See also Figure 6 , Figure 6 This is a structural diagram of the feature extraction framework provided by an embodiment of the present invention, including a data structure adaptation layer, an atomic feature management layer, a combined feature management layer, a model configuration layer, and a feature extraction layer. For ease of understanding, the following text uses a bottom-up approach and combines pseudocode to explain the functions of each layer one by one.

[0133] The first layer of the feature framework, the data structure adaptation layer, adapts the data structures of different input data types, converting them into a unified sample structure to ensure consistent field meanings. This unified sample structure can be defined as a universal cross-language data format, including but not limited to protobuf, flatbuffers, thrift, avro, and msgpack. Specifically, the data structure conversion is completed by invoking toolkits corresponding to the cross-language data format to process the input data.

[0134] The atomic feature management layer stores atomic feature registration information. For ease of explanation, this information is represented by a fa_rule, which includes several atomic feature group identifiers (agids) and a feature extraction function (fa) mapped to each agid. fa processes data using a unified sample structure, outputting the corresponding agid and a mapping from afid to afvalue. fa can be pre-configured based on the actual application scenario and mapped to the corresponding agid.

[0135] The combined feature management layer is used to store combined feature registration information. For ease of explanation, combined feature registration information is represented by fc_rule, which includes the combined feature group identifier cgid and at least two agids mapped to cgid. It is worth noting that the data structure adaptation layer, atomic feature management layer, and combined feature management layer all belong to the "registration" stage. In other words, the feature extraction framework applied to different types of artificial intelligence recommendation models can all use the same data structure adaptation layer, atomic feature management layer, and combined feature management layer.

[0136] As for the "configuration" stage, it is the configuration related to the artificial intelligence recommendation model itself.

[0137] In the model configuration layer, based on the pre-registered fa_rule and fc_rule, several atomic features and / or several combined features are selected and added to the feature framework as model configuration. The selection is based on the feature group identifier set corresponding to the AI ​​recommendation model, referred to as model_rule. In the feature framework, the registered atomic features and combined features are combined to generate a set of fa chains, namely fa_chain, which are used for subsequent feature extraction. The pseudo code flow of the generation process is as follows:

[0138]

[0139] The feature extraction layer is the core layer of the feature extraction framework, which is used to output a series of features present in the input data in the form of samples. The pseudo code is as follows:

[0140]

[0141]

[0142] Among them, the pseudo code of the feature combination function fc is as follows:

[0143]

[0144]

[0145] Through the above-mentioned unified feature framework, features can be extracted from the converted input data according to the feature group identification set corresponding to the artificial intelligence recommendation model to form samples, thereby performing model training or model prediction tasks, improving the strong consistency of features and overcoming the differences in extracted features caused by the application of two frameworks.

[0146] Next, the exemplary structure of the feature processing device 255 of the artificial intelligence recommendation model provided by the embodiment of the present invention is implemented as a software module. In some embodiments, Figure 3 As shown, the software modules stored in the feature processing device 255 of the artificial intelligence recommendation model in the memory 250 may include:

[0147] The conversion module 2551 is used to obtain input data and convert the data structure of the input data into a unified sample structure;

[0148] Function determination module 2552, used to determine the feature group identifier and feature extraction function corresponding to the artificial intelligence recommendation model;

[0149] A feature extraction module 2553 is configured to perform feature extraction processing on the converted input data according to the feature extraction function to obtain a feature group identifier, a feature identifier, and a feature value in the input data;

[0150] The construction module 2554 is used to construct a sample of the artificial intelligence recommendation model based on the feature group identifier, feature identifier and feature value.

[0151] In some embodiments, the function determination module 2552 is also used to: obtain atomic feature registration information and combined feature registration information, wherein the atomic feature registration information includes a registered atomic feature group identifier and a corresponding feature extraction function, and the combined feature registration information includes a registered combined feature group identifier and corresponding at least two atomic feature group identifiers; obtain a feature group identifier set corresponding to the artificial intelligence recommendation model, and traverse the feature group identifier set; when the traversed feature group identifier matches the atomic feature registration information, add the feature extraction function corresponding to the feature group identifier in the atomic feature registration information to the feature extraction function chain.

[0152] In some embodiments, the function determination module 2552 is also used to: when the traversed feature group identifier does not match the atomic feature registration information, but matches the combined feature registration information, determine the atomic feature group identifier corresponding to the feature group identifier in the combined feature registration information; in the atomic feature registration information, determine the feature extraction function corresponding to the atomic feature group identifier; when the feature extraction function does not exist in the feature extraction function chain, add the feature extraction function to the feature extraction function chain.

[0153] In some embodiments, the feature extraction module 2553 is further used to: traverse the feature extraction function chain; perform feature extraction processing on the converted input data according to the traversed feature extraction function to obtain the atomic feature group identifier, atomic feature identifier and atomic feature value in the input data.

[0154] In some embodiments, the construction module 2554 is further used to: traverse the feature group identifier set; when the traversed feature group identifier matches the atomic feature registration information, determine the feature group identifier as the atomic feature group identifier, and add the atomic feature group identifier, the corresponding atomic feature identifier and the atomic feature value to the sample.

[0155] In some embodiments, the construction module 2554 is further used to: when the traversed feature group identifier does not match the atomic feature registration information and matches the combined feature registration information, determine the feature group identifier as a combined feature group identifier; determine at least two atomic feature group identifiers corresponding to the combined feature group identifier in the combined feature registration information, and determine the atomic feature identifier and atomic feature value corresponding to the atomic feature group identifier; combine the at least two atomic feature identifiers to obtain a combined feature identifier, and combine the at least two atomic feature values ​​to obtain a combined feature value; and add the combined feature group identifier, the combined feature identifier and the combined feature value to the sample.

[0156] In some embodiments, the combining processing of at least two of the atomic feature identifiers to obtain a combined feature identifier includes: performing a first combining operation on the at least two atomic feature identifiers to obtain a combined feature identifier; wherein the first combining operation includes at least one of the following: a hash operation, a bitwise OR operation, a bitwise AND operation, and a bitwise XOR operation;

[0157] The combining process of at least two of the atomic eigenvalues ​​to obtain a combined eigenvalue includes: performing a second combining operation on at least two of the atomic eigenvalues ​​to obtain a combined eigenvalue; wherein the second combining operation includes at least one of the following: a product operation, an addition operation.

[0158] In some embodiments, the feature processing device 255 of the artificial intelligence recommendation model also includes: a training module, which is used to train the artificial intelligence recommendation model according to the sample when the input data is training data; and a prediction module, which is used to predict the sample through the artificial intelligence recommendation model to obtain a recommendation result when the input data is test data.

[0159] An embodiment of the present invention provides a storage medium storing executable instructions, wherein the executable instructions are stored. When the executable instructions are executed by a processor, the processor will execute the feature processing method of the artificial intelligence recommendation model provided by the embodiment of the present invention, for example, Figure 4A and 4B The feature processing method of the artificial intelligence recommendation model shown.

[0160] In some embodiments, the storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EE PROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or various devices including one or any combination of the above memories.

[0161] In some embodiments, executable instructions may be in 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 stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0162] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0163] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0164] To sum up, the embodiments of the present invention overcome the data structure differences of input data from different sources, and uniformly process the input data through feature extraction functions, thereby improving the flexibility and efficiency of feature extraction, ensuring the strong consistency of features in the sample, and having a good sample generation effect for large-scale and sparse advertising or recommendation business scenarios, where "large-scale" refers to a large number of features, which may reach a scale of more than 100 million, and "sparse" refers to each sample containing only part of the features of a part of the feature group, which may be hundreds or thousands.

[0165] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present invention are included in the scope of protection of the present invention.

Claims

1. A feature processing method for an artificial intelligence recommendation model, characterized in that: include: Acquire input data, and convert the data structure of the input data into a unified sample structure, wherein the input data includes user portrait data and user click data on information; Determine a feature group identifier set corresponding to the artificial intelligence recommendation model, wherein the feature group identifier set includes at least one feature group identifier; For each of the feature group identifiers, when the atomic feature registration information matches the feature group identifier, adding the feature extraction function corresponding to the feature group identifier in the atomic feature registration information to the feature extraction function chain; Performing feature extraction processing on the converted input data according to the feature extraction function chain to obtain a feature group identifier, a feature identifier and a feature value in the input data; Constructing a sample of the artificial intelligence recommendation model according to the feature group identifier, feature identifier and feature value; When the input data is the data to be tested, the sample is predicted and processed by the artificial intelligence recommendation model to obtain the probability of the user clicking on the information.

2. The feature processing method according to claim 1, characterized in that: For each of the feature group identifiers, when the atomic feature registration information matches the feature group identifier, adding the feature extraction function corresponding to the feature group identifier in the atomic feature registration information to the feature extraction function chain includes: Acquire atomic feature registration information and combined feature registration information, wherein the atomic feature registration information includes a registered atomic feature group identifier and a corresponding feature extraction function, and the combined feature registration information includes a registered combined feature group identifier and corresponding at least two atomic feature group identifiers; Traversing each of the feature group identifiers in the feature group identifier set; When the traversed feature group identifier matches the atomic feature registration information, the feature extraction function corresponding to the feature group identifier in the atomic feature registration information is added to the feature extraction function chain.

3. The feature processing method according to claim 2, characterized in that: Also includes: When the traversed feature group identifier does not match the atomic feature registration information but matches the combined feature registration information, determining the atomic feature group identifier in the combined feature registration information corresponding to the feature group identifier; Determining, in the atomic feature registration information, a feature extraction function corresponding to the atomic feature group identifier; When the feature extraction function does not exist in the feature extraction function chain, the feature extraction function is added to the feature extraction function chain.

4. The feature processing method according to claim 2, characterized in that: The step of performing feature extraction processing on the converted input data according to the feature extraction function chain to obtain a feature group identifier, a feature identifier and a feature value in the input data includes: Traversing the feature extraction function chain; The converted input data is subjected to feature extraction processing according to the traversed feature extraction function to obtain the atomic feature group identifier, atomic feature identifier and atomic feature value in the input data.

5. The feature processing method according to claim 4, characterized in that: The step of constructing a sample of the artificial intelligence recommendation model according to the feature group identifier, feature identifier and feature value includes: Traversing the feature group identifier set; When the traversed feature group identifier matches the atomic feature registration information, the feature group identifier is determined as the atomic feature group identifier, and the atomic feature group identifier, the corresponding atomic feature identifier and the atomic feature value are added to the sample.

6. The feature processing method according to claim 5, characterized in that: Also includes: When the traversed feature group identifier does not match the atomic feature registration information but matches the combined feature registration information, determining the feature group identifier as a combined feature group identifier; Determine at least two atomic feature group identifiers corresponding to the combined feature group identifier in the combined feature registration information, and determine the atomic feature identifiers and atomic feature values ​​corresponding to the atomic feature group identifiers; Combining at least two of the atomic feature identifiers to obtain a combined feature identifier, and combining at least two of the atomic feature values ​​to obtain a combined feature value; The combined feature group identifier, the combined feature identifier and the combined feature value are added to the sample.

7. The feature processing method according to claim 6, characterized in that: The combining process of at least two of the atomic feature identifiers to obtain a combined feature identifier includes: Performing a first combination operation on at least two of the atomic feature identifiers to obtain a combined feature identifier; wherein the first combination operation includes at least one of the following: a hash operation, a bitwise OR operation, a bitwise AND operation, and a bitwise XOR operation; The combining process of at least two of the atomic eigenvalues ​​to obtain a combined eigenvalue comprises: A second combination operation is performed on at least two of the atomic eigenvalues ​​to obtain a combined eigenvalue; wherein the second combination operation includes at least one of the following: a product operation, an addition operation.

8. The feature processing method according to any one of claims 1 to 7, characterized in that: Also includes: When the input data is training data, the artificial intelligence recommendation model is trained according to the samples.

9. The feature processing method according to claim 8, characterized in that: The training of the artificial intelligence recommendation model according to the sample includes: Creating a copy of the artificial intelligence recommendation model; A copy of the artificial intelligence recommendation model is trained based on the sample.

10. The feature processing method according to claim 9, characterized in that: Before training the copy of the artificial intelligence recommendation model according to the sample, the method further includes: determining a first accuracy of a copy of the artificial intelligence recommendation model based on the sample; After training the copy of the artificial intelligence recommendation model according to the sample, the method further includes: Determining a second accuracy rate of a trained copy of the artificial intelligence recommendation model based on samples corresponding to the new training data; When the second accuracy rate exceeds the first accuracy rate, the artificial intelligence recommendation model is updated according to the trained copy of the artificial intelligence recommendation model.

11. The feature processing method according to claim 10, characterized in that: The updating of the artificial intelligence recommendation model according to the trained copy of the artificial intelligence recommendation model comprises: Get the generation rate of training data; When the generation rate is lower than a generation rate threshold, the artificial intelligence recommendation model is updated according to the trained copy of the artificial intelligence recommendation model.

12. The feature processing method according to claim 8, characterized in that: Before training the artificial intelligence recommendation model according to the sample, the method further includes: Get feature adjustment identifier and corresponding adjustment value; In the sample, each group of feature group identifiers, feature identifiers and feature values ​​that meet the feature adjustment identifier is determined as a sub-sample to be adjusted; Updating the weight value of the sub-sample to be adjusted to the adjustment value; The feature adjustment identifier is one of the following: a feature group identifier, a feature group identifier and a feature identifier.

13. A feature processing device for an artificial intelligence recommendation model, characterized in that: include: A conversion module, used to obtain input data and convert the data structure of the input data into a unified sample structure, wherein the input data includes user portrait data and user click data on information; A function determination module, used to determine a feature group identifier set corresponding to the artificial intelligence recommendation model, wherein the feature group identifier set includes at least one feature group identifier; for each feature group identifier, when the atomic feature registration information matches the feature group identifier, a feature extraction function corresponding to the feature group identifier in the atomic feature registration information is added to a feature extraction function chain; A feature extraction module, used for performing feature extraction processing on the converted input data according to the feature extraction function chain to obtain a feature group identifier, a feature identifier and a feature value in the input data; A construction module is used to construct a sample of the artificial intelligence recommendation model based on the feature group identifier, feature identifier and feature value; when the input data is the test data, the sample is predicted and processed by the artificial intelligence recommendation model to obtain the probability of the user clicking on the information.

14. An electronic device, characterized in that: include: A memory for storing executable instructions; A processor, configured to implement the feature processing method according to any one of claims 1 to 12 when executing the executable instructions stored in the memory.

15. A computer-readable storage medium, characterized in that: Executable instructions are stored, which are used to cause a processor to execute and implement the feature processing method described in any one of claims 1 to 12.

Images