Information processing method, model training method, electronic device, computer program product and storage medium

Abstract

The invention provides an information processing method, a model training method, electronic equipment, a computer program product and a storage medium. The method comprises the steps of performing feature decoupling on an object feature of a first object in a first time period to obtain a first willingness feature and a first idle feature of the first object in the first time period, and performing feature decoupling on an information feature of first information to obtain a second willingness feature and a second idle feature of the first information in the first time period; based on the first willingness feature and the second willingness feature, determining a first probability that the first object gets through a call corresponding to the first information in the first time period, and based on the first idle feature and the second idle feature, determining a second probability that the first object gets through the call in the first time period; and determining a calling strategy corresponding to the first information on the first object based on the first probability and the second probability.

Description

Technical Field

[0001] The present application relates to machine learning technology, and in particular to an information processing method, a model training method, an electronic device, a computer program product, and a storage medium. Background Art

[0002] When recommending information to users, information is often recommended to users through telephone calls. In related technologies, the method of recommending information to users through telephone calls often relies on experience to select the time of outbound calls, such as avoiding working hours, choosing fixed time periods such as evenings or weekends to make calls. Although this method can ensure that users have time to make phone calls to a certain extent, it does not take into account the differences between individual users, resulting in phone calls to individual users being unable to be connected due to the users being busy, thereby reducing the effectiveness of information recommendations. Summary of the Invention

[0003] The embodiments of the present application provide an information processing method, a model training method, an electronic device, a computer program product, and a computer-readable storage medium, which can improve the effectiveness of information recommendation.

[0004] The technical solution of the embodiment of the present application is implemented as follows:

[0005] This embodiment of the present application provides an information processing method, the method comprising:

[0006] Performing feature decoupling on the object features of the first object in the first time period to obtain a first willingness feature and a first availability feature of the first object in the first time period, and performing feature decoupling on the information features of the first information to obtain a second willingness feature and a second availability feature of the first information in the first time period;

[0007] Determining, based on the first willingness feature and the second willingness feature, a first probability that the first subject answers the call corresponding to the first information during the first time period, and determining, based on the first idle feature and the second idle feature, a second probability that the first subject answers the call during the first time period;

[0008] A calling strategy corresponding to the first information performed by the first object is determined based on the first probability and the second probability.

[0009] The present invention provides a method for training a model, which includes:

[0010] By using the model, feature decoupling is performed on the first sample feature of the object sample in the second period to obtain a third willingness feature and a third idle feature of the object sample in the second period, and feature decoupling is performed on the second sample feature to obtain a fourth willingness feature and a fourth idle feature of the information sample in the second period;

[0011] Determining, based on the third willingness feature and the fourth willingness feature, a third probability that the subject sample connects the call of the information sample during the second time period, and determining, based on the third idle feature and the fourth idle feature, a fourth probability that the subject sample connects the call during the second time period;

[0012] A first loss function is constructed based on the third probability, a second loss function is constructed based on the fourth probability, and parameters of the model are updated based on the first loss function and the second loss function.

[0013] An embodiment of the present application provides an electronic device, comprising:

[0014] a memory for storing computer-executable instructions or computer programs;

[0015] The processor is used to implement the information processing method provided in the embodiment of the present application or the training method of the model provided in the embodiment of the present application when executing the computer-executable instructions or computer programs stored in the memory.

[0016] An embodiment of the present application provides a computer-readable storage medium storing a computer program or computer-executable instructions, which, when executed by a processor, implements the information processing method provided in the embodiment of the present application or the training method of the model provided in the embodiment of the present application.

[0017] An embodiment of the present application provides a computer program product, including a computer program or computer-executable instructions. When the computer program or computer-executable instructions are executed by a processor, the information processing method provided by the embodiment of the present application or the training method of the model provided by the embodiment of the present application is implemented.

[0018] The embodiments of the present application have the following beneficial effects:

[0019] The object features of the first object in the first time period are feature decoupled to obtain a first willingness feature and a first availability feature of the first object in the first time period, and the information features of the first information are feature decoupled to obtain a second willingness feature and a second availability feature of the first information in the first time period. By decoupling the features of the first object and the features of the first information, the features of the first object and the features of the first information are accurately divided into features of the willingness dimension and features of the availability dimension for measuring the probability of the first object connecting, thereby improving the accuracy of the subsequently determined comprehensive probability. Based on the first willingness feature and the second willingness feature, a first probability of the first object connecting to the call corresponding to the first information in the first time period is determined, and based on the first availability feature and the second availability feature, a second probability of the first object connecting to the call corresponding to the first information in the first time period is determined. Based on the first probability and the second probability, a comprehensive probability of the first object connecting to the call corresponding to the first information in the first time period is determined. By combining the features of the first object and the features of the first information, a first probability of the first object connecting to the call in the willingness dimension and a second probability of the first object connecting to the call in the availability dimension are determined. The first probability and the second probability are combined to determine a call strategy corresponding to the first information for the first object, thereby improving the probability of the user connecting the call and thereby improving the effectiveness of information recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 1 is a schematic diagram of the architecture of the information processing system 100 provided in an embodiment of the present application;

[0021] Figure 2A 1 is a first structural diagram of an electronic device 500 provided in an embodiment of the present application;

[0022] Figure 2B This is a second structural diagram of the electronic device 500 provided in an embodiment of the present application.

[0023] Figure 3A This is a first flow chart of the information processing method provided by an embodiment of the present application;

[0024] Figure 3B This is a first flow chart of the model training method provided in the embodiment of the present application;

[0025] Figure 4 It is a schematic diagram of the implementation of the information processing method provided in the embodiment of the present application.

[0026] It should be pointed out that the above-mentioned "first" and "second" are only used to distinguish different solutions, and do not represent the degree of distinction between the advantages and disadvantages of the solutions or the priority in the implementation process. DETAILED DESCRIPTION

[0027] In order to make the purpose, 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 limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0028] 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.

[0029] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0030] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0031] Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meanings as those commonly understood by those skilled in the art. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0032] 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.

[0033] In intelligent outbound calling systems, traditional methods often rely on simple rules or empirical guidelines to select outbound calling times, such as avoiding work hours or choosing fixed time periods such as evenings or weekends. However, this approach ignores individual differences and dynamically changing behavior patterns, potentially leading to a large number of meaningless or even offensive phone calls, especially when users are disturbed at inappropriate times.

[0034] Based on this, the embodiments of the present application provide an information processing method, a model training method, an apparatus, an electronic device, a computer program product, and a storage medium, which can improve the effectiveness of information recommendation. The following describes an exemplary application of the information processing device provided by the embodiments of the present application. The device provided by the embodiments of the present application can be implemented as various types of terminals such as laptops, tablets, desktop computers, set-top boxes, smart phones, smart speakers, smart watches, smart TVs, and in-vehicle terminals, and can also be implemented as a server. The following describes an exemplary application when the device is implemented as a server.

[0035] See also Figure 1 , Figure 1 This is a schematic diagram of the architecture of the information processing system 100 provided in an embodiment of the present application. To support an information processing application, the terminal 400 is connected to the server 200 via the network 300. The network 300 can be a wide area network or a local area network, or a combination of the two.

[0036] The terminal 400 transmits the user's information in multiple time periods and the first information to be recommended to the server 200, determines the time period with the highest comprehensive probability through the information processing method provided in the embodiment of the present application, and sends the determined time period to the terminal 400, which then makes a call to the user corresponding to the first information in the time period with the highest comprehensive probability.

[0037] The information processing provided in the embodiments of the present application can be applied to scenarios where it is necessary to occupy the user's time to recommend information to the user, such as telemarketing scenarios and after-sales feedback scenarios. The following examples illustrate:

[0038] 1) In a telemarketing scenario, the terminal transmits the user's information and the telemarketing information to the server 200. The information processing method provided in the embodiment of the present application determines the time period with the highest comprehensive probability of a call to the user being connected, and the determined time period is sent to the terminal 400. The terminal 400 then conducts telemarketing corresponding to the first information to the user during the time period with the highest comprehensive probability.

[0039] 2) In the after-sales feedback scenario, the terminal transmits the information of the user who has undergone after-sales service and the information generated during the after-sales process to the server 200. Through the information processing method provided in the embodiment of the present application, the time period with the highest comprehensive probability of calling the user and the user being connected is determined, and the determined time period is sent to the terminal 400. The terminal 400 makes a telephone call to the user during the time period with the highest comprehensive probability to collect the user's after-sales feedback.

[0040] In some embodiments, the server 200 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It may also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal and the server may be connected directly or indirectly via wired or wireless communication, which is not limited in the embodiments of the present application.

[0041] See also Figures 2A-2B , Figures 2A-2B This is a structural diagram of an electronic device 500 provided in an embodiment of the present application, with the electronic device 500 as an example. Figure 1 Take server 200 in the example as an example, Figures 2A-2B The electronic device 500 shown 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 achieve connection and communication between these components. In addition to including a data bus, the bus system 540 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 540 is not shown in FIG. Figures 2A-2B Various buses are labeled as bus system 540 .

[0042] 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., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0043] The user interface 530 includes one or more output devices 531 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The 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.

[0044] The memory 550 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 550 may optionally include one or more storage devices that are physically remote from the processor 510.

[0045] The memory 550 includes volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 550 described in the embodiments of the present application is intended to include any suitable type of memory.

[0046] In some embodiments, the memory 550 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.

[0047] Operating system 551, including system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic services and processing hardware-based tasks;

[0048] A network communication module 552 for reaching other electronic devices via one or more (wired or wireless) network interfaces 520 , exemplary network interfaces 520 including Bluetooth, Wi-Fi, and Universal Serial Bus (USB);

[0049] a presentation module 553 for enabling presentation of information via one or more output devices 531 (e.g., a display screen, a speaker, etc.) associated with the user interface 530 (e.g., a user interface for operating peripheral devices and displaying content and information);

[0050] The input processing module 554 is configured to detect one or more user inputs or interactions from one of the one or more input devices 532 and to translate the detected inputs or interactions.

[0051] In some embodiments, the apparatus provided in the embodiments of the present application may be implemented in software. Figure 2A Information processing device 555A stored in memory 550 is shown. This device can be software in the form of a program or plug-in, and includes the following software modules: a decoupling module 5551, a determination module 5552, and a call module 5553. These modules are logical and can be arbitrarily combined or further separated according to the functions they implement. The functions of each module will be described below.

[0052] In some embodiments, the training device of the model provided in the embodiments of the present application can be implemented in software. Figure 2BA training device 555B for a model stored in memory 550 is shown. This device can be software in the form of a program or plug-in, and includes the following software modules: a decoupling module 5554, a determination module 5555, and a training module 5556. These modules are logical and can be arbitrarily combined or further separated according to the functions they implement. The functions of each module will be described below.

[0053] In other embodiments, the apparatus provided in the embodiments of the present application may be implemented in hardware. As an example, the apparatus provided in the embodiments of the present application may be a processor in the form of a hardware decoding processor, which is programmed to execute the information processing method provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor may be one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0054] The following describes the information processing method provided by the embodiment of the present application. As mentioned above, the electronic device that implements the information processing method of the embodiment of the present application can be a terminal, a server, or a combination of the two. Therefore, the execution entity of each step will not be repeated below.

[0055] See also Figure 3A , Figure 3A This is a first flow chart of the information processing method provided by the embodiment of the present application, which will be combined with Figure 3A The steps shown are explained.

[0056] In step 101, feature decoupling is performed on the object features of the first object in the first time period to obtain the first intention feature and the first idle feature of the first object in the first time period, and feature decoupling is performed on the information features of the first information to obtain the second intention feature and the second idle feature of the first information in the first time period.

[0057] In some embodiments, before executing step 101, the following technical solutions can also be implemented: feature extraction is performed on the behavior data of the first object in the first time period to obtain the behavior characteristics of the first object in the first time period; feature mapping is performed on the sparse features in the behavior characteristics to obtain the first mapping features; feature fusion is performed on the dense features in the behavior characteristics and the first mapping features to obtain the object characteristics of the first object in the first time period.

[0058] In actual implementation, behavioral data generated when interacting with the first subject can be collected first, such as the first subject's gender, age, occupation, call time, call duration, and call frequency. The data can also be collected regarding the first subject's communication methods during different time periods, such as calls, text messages, and internet usage. The collected data can then be divided by time period. For example, the data can be divided hourly, with a day divided into 24 time periods, and the data can be categorized into each time period. For example, if the first subject answered a call at 10:30 a.m., this data would be classified into the time period corresponding to 10:00 a.m. to 11:00 a.m. For another example, if the first subject is male, this data cannot be divided by time period, so the data can be replicated to include the data for each time period.

[0059] In actual implementation, multiple methods can be used to extract the user's behavioral characteristics in the first time period, where the first time period can be any time period obtained by dividing the time period:

[0060] For example, time series analysis tools are used to extract time series-related features from the user's historical data, such as trends, seasonality, and periodicity of user behavior; for another example, statistics of the user's behavior data in the first period are calculated, such as the mean click-through rate, the variance of the click-through rate, the maximum click-through rate, the minimum click-through rate, etc. as behavioral features; for another example, machine learning algorithms (such as decision trees, random forests, support vector machines, etc.) are used to automatically extract the behavioral features of the first object; for another example, deep learning models (such as convolutional neural networks, recurrent neural networks, etc.) can be used to automatically learn the behavioral features of the first object.

[0061] In actual implementation, feature mapping of sparse features in behavioral features usually refers to mapping these sparse features from the original high-dimensional sparse space to a low-dimensional dense space, so as to facilitate better subsequent data processing and analysis. Specifically, it is first necessary to identify which features in the behavioral features are sparse features and which are dense features. The so-called sparse features usually refer to those features whose values ​​are mostly zero or empty, such as the click history and purchase history of the first object; the so-called dense features refer to those features whose values ​​in the data set are mostly non-zero or non-empty. These dense features have values ​​in every row of the data, so their data density is high. Dense features usually contain continuous data, such as age, height, temperature, etc. The values ​​of these dense features change continuously within a certain range, rather than being discrete.

[0062] In practical applications, sparse features can be mapped to low-dimensional dense space through various feature mapping methods to obtain dense features. For example, the main components of the behavioral data, such as the time when the first object's click is generated, can be found, and then the main components can be reduced in dimension to obtain dense features. For example, assuming that the data is generated by several potential factors, such as the screen-off time of the terminal device used by the first object, these potential factors can be reduced in dimension to obtain dense features. For example, sparse features can be subjected to matrix decomposition to obtain dense features. For example, sparse features can be subjected to dimensionality reduction through neural networks to obtain dense features, etc.

[0063] In actual implementation, the dense features and the first mapping features (i.e., the features after the sparse features are mapped to the low-dimensional dense space) in the behavioral features of the first object are fused. The purpose is to integrate the information of these two types of features to obtain a more comprehensive and richer object feature representation. Specifically, the dense features and the first mapping features can be standardized first so that they are at the same scale to facilitate fusion. Before fusion, it may be necessary to perform feature selection based on the importance or contribution of the features to remove some unimportant features. After that, the dense features and the first mapping features can be directly spliced ​​into a large feature vector to serve as the object features of the first object in the first period.

[0064] In the above way, by fusing different types of data features, the diversity of the feature set can be increased, which helps the model capture more data features and patterns. Different features may contain different types of information. Combining dense features with low-dimensional features after sparse feature mapping can provide more comprehensive information, which can usually improve the performance of the model in prediction tasks.

[0065] In some embodiments, before executing step 101, the following technical solutions can also be implemented: feature extraction is performed on the attribute data of the first information to obtain the attribute characteristics of the first information; feature mapping is performed on the sparse features in the attribute characteristics to obtain the second mapping characteristics; feature fusion is performed on the dense features in the attribute characteristics and the second mapping features to obtain the information characteristics of the first information.

[0066] In actual implementation, attribute data of the first information can also be collected. It should be noted that the first information here can be a product or a service. For example, the first information can be toothpaste or a resource borrowing service. The data of the first information can include the first subject's rating or evaluation of the first information, the purchase history of the product corresponding to the first information or the receipt of the service of the first information, the frequency of clicks on the first information, etc. Similarly, the collected data of the first information can be divided into time periods. For example, if the first information is clicked by the first subject at 8:20 am, the data will be divided into the time period from 8:00 am to 9:00 am.

[0067] In actual implementation, a variety of attribute features for extracting the first information may be used:

[0068] For example, time series analysis tools can be used to extract time series-related features from the first information, such as trends, seasonality, and periodicity in product sales. Another example is the calculation of statistics of the data corresponding to the first information in each time period, such as the mean of product sales, the variance of product click-through rate, the maximum number of product views, the minimum number of product views, etc. Another example is the use of machine learning algorithms (such as decision trees, random forests, support vector machines, etc.) to automatically extract attribute features of the first information. Another example is the use of deep learning models (such as convolutional neural networks, recurrent neural networks, etc.) to automatically learn the attribute features of the first information.

[0069] In actual implementation, in order to perform feature mapping on sparse features in attribute features, it is first necessary to identify which features in the attribute features are sparse features, and then the coefficient features can be mapped to a low-dimensional dense space through feature mapping. For example, the main components of the attribute data can be found, such as the time when the product is clicked, and then the main components can be reduced in dimension to obtain dense features; for example, assuming that the attribute data is generated by several potential factors, such as the raw material supply channel of the product, these potential factors can be reduced in dimension to obtain dense features; for example, sparse features can be obtained by matrix decomposition; for example, sparse features can be mapped to a low-dimensional dense space through a neural network to obtain dense features.

[0070] In actual implementation, feature fusion is performed on the dense features and the second mapping features (i.e., the features after the sparse features are mapped to the low-dimensional dense space) in the attribute features. The dense features and the second mapping features can be standardized first so that they are at the same scale for easy fusion. Before fusion, feature selection may be required based on the importance or contribution of the features to remove some unimportant features. After that, the dense features and the first mapping features can be directly spliced ​​into a large feature vector as the information feature of the first information.

[0071] Among them, the first willingness feature represents the willingness of the first object to answer the call corresponding to the first information in the corresponding time period, and the first idle feature represents the idleness of the first object in the corresponding time period; the second willingness feature represents the attractiveness of the first information to the first object in the corresponding time period, and the second idle feature represents the probability that the call corresponding to the first information will be answered by the first object in the corresponding time period.

[0072] Willingness refers to the first person's willingness or inclination to answer the call corresponding to the first message during the corresponding time period. This can be quantified using a willingness score ranging from 0 to 1, where 0 indicates no willingness at all and 1 indicates strong willingness. The following factors can be considered when calculating the willingness score: 1) Historical behavior analysis: Analyzing the frequency and pattern of the first person's past calls answered during the same or similar time periods; 2) Schedule: Examining the first person's schedule to determine whether they have scheduled events or meetings during a specific time period; 3) Preferences: Considering the first person's personal preferences, such as whether they prefer to answer calls during work hours or during breaks; and 4) Emotional state: Analyzing the first person's emotional state, for example, by inferring it through social media activity or physiological data (such as heart rate). Combining these factors, the willingness score can be calculated using the following formula: Willingness score = (Historical call frequency weight * Historical call frequency) + (Schedule weight * Schedule availability) + (Preference weight * Preference match) + (Emotional state weight * Emotional positivity). Each weight can be adjusted based on its impact on the willingness score to ensure the total is 1.

[0073] Availability refers to the degree of availability of the first participant during a specific time period. It can be quantified using an availability score ranging from 0 to 1, where 0 indicates being extremely busy with no free time, and 1 indicates being completely idle. The availability score can be calculated by considering the following factors: 1) Schedule: Check the first participant's schedule to determine whether there are any scheduled events or meetings during a specific time period; 2) Workload: Assess the first participant's workload during a specific time period, for example, by inferring it through task management tools or project progress; 3) Physiological state: Analyze the first participant's physiological state, for example, by using an activity detector to determine their activity level; and 4) Environmental factors: Consider the first participant's environment, such as whether they are in a noisy environment, which may affect their willingness to answer calls. Combining these factors, the availability score can be calculated using the following formula: Availability score = (Schedule weight * Schedule availability) + (Workload weight * Workload severity) + (Physiological state weight * Physiological activity level) + (Environmental factor weight * Environmental suitability). Similarly, each weight can be adjusted based on its impact on availability to ensure the total is 1.

[0074] Attractiveness refers to the degree of appeal of a first piece of information to a first recipient during a specific time period. This can be quantified using an attractiveness score ranging from 0 to 1, with 0 indicating no appeal at all and 1 indicating very attractive. The following factors are considered when calculating the attractiveness score: 1) Content relevance: This assesses the relevance of the first piece of information to the first recipient's interests and needs; 2) Information urgency: This determines the urgency of the first piece of information, such as whether it requires immediate attention or a response; 3) Source credibility: This assesses the credibility and authority of the source providing the first piece of information; and 4) Historical engagement: This considers the first recipient's past engagement with similar information, such as whether they frequently clicked on or responded to it. Combining these factors, the attractiveness score can be calculated using the following formula: Attractiveness score = (Content relevance weight * Content match) + (Information urgency weight * Urgency) + (Source credibility weight * Source credibility) + (Historical engagement weight * Interaction frequency). Each weight can be adjusted based on its impact on attractiveness to ensure the total sum is 1.

[0075] It's important to note that the calculation of the willingness score, availability score, and attractiveness score is based on a comprehensive evaluation of multiple factors, each with its own weight. These factors and weights can be adjusted based on actual circumstances to ensure that the scores accurately reflect the likelihood that the first contact will answer the call within a specific time period. This comprehensive evaluation can optimize calling strategies and improve call success rates and efficiency. For example, if the first contact's willingness and availability are both high during a specific time period, and the first message's attractiveness is also high, then the success rate of calling the first contact during that time period is likely to be high. Conversely, if any of these characteristics are low, the call success rate is likely to be low.

[0076] In actual implementation, the first willingness feature reflects the first subject's preference for different types of information, emotional state, points of interest, etc. These factors determine whether the first subject is willing to interact with information in a specific time period. The first idle feature describes the first subject's schedule, living habits and work mode. These factors together determine the first subject's degree of idleness in different time periods.

[0077] In actual implementation, the second willingness feature corresponding to the first information can represent the attractiveness or relevance of the first information itself to the user, that is, the possibility of the first information arousing the user's interest in different time periods. The second willingness feature can be determined based on factors such as the content, type, urgency, and matching degree with the user's historical behavior of the information. For example, a news item about a topic of interest to the user may have a second willingness feature representing a higher willingness during the time period when the user is usually active.

[0078] In some embodiments, the feature decoupling of the object features in step 101 to obtain the first willingness feature and the first idle feature of the first object in each time period can be achieved by the following technical solution: determining a first influencing factor and a second influencing factor, the first influencing factor representing the willingness of the first object to connect the call corresponding to the first information in the first time period, and the second influencing factor representing the idleness of the first object in the first time period; based on the first influencing factor, extracting the first willingness feature of the first object in the first time period from the object features; based on the second influencing factor, extracting the first idle feature of the first object in the first time period from the object features.

[0079] In actual implementation, the first influencing factor may refer to a factor that affects the first subject's willingness to answer the call corresponding to the first information in each time period. The first influencing factor is usually related to factors such as the first subject's psychological state, personal preferences, current activities or emotions, and may include the following: the first subject's personal interests: for example, the degree of interest in a certain topic or activity. The first subject's emotional state: for example, whether the first subject feels happy, anxious or tired. The first subject's time sensitivity: for example, whether the first subject tends to answer calls during specific time periods. The second influencing factor may refer to a factor that affects the first subject's availability in each time period. The second influencing factor is usually related to factors such as the first subject's living habits, work arrangements, social activities, and may include the following: the first subject's work schedule: for example, the first subject's working hours and rest time; the first subject's social activities: for example, whether the first subject has parties or other social activities; the first subject's daily habits: for example, the first subject's sleep time, exercise time, etc.

[0080] In actual implementation, the object feature is a multidimensional feature, so the object feature can be regarded as an object sub-feature including each dimension. The feature related to the first influencing factor can be selected from the object feature in a variety of ways. For example, the correlation between each object sub-feature and the first influencing factor is determined, and the object sub-feature with a correlation higher than a first correlation threshold is used as the feature related to the first influencing factor; for another example, the transformation amount of the object sub-feature after the first influencing factor is combined with each sub-feature is determined, and the object sub-feature with a transformation amount higher than the first transformation threshold is used as the feature related to the first influencing factor; for another example, the object features are clustered through visualization and statistical methods, and then the object sub-features in the object features that have a correlation with the features representing willingness are found as the features related to the first influencing factor; for another example, the correlation coefficient between each object sub-feature and the first influencing factor is determined, and the object sub-feature with a correlation coefficient higher than the first correlation coefficient threshold is determined as the feature related to the first influencing factor.

[0081] Similarly, features related to the second influencing factor can be selected from the object features in a variety of ways. For example, the correlation between each object sub-feature and the second influencing factor is determined, and the sub-features with a correlation higher than a second correlation threshold are used as features related to the second influencing factor. For another example, the transformation amount of the object sub-feature after the second influencing factor is combined with each object sub-feature is determined, and the sub-features with a transformation amount higher than a second transformation threshold are used as features related to the second influencing factor. For another example, object features can be clustered through visualization and statistical methods to find object sub-features in each object feature that have a correlation relationship with the feature representing idleness, and use them as features related to the second influencing factor. For another example, the correlation coefficient between each object sub-feature and the second influencing factor is determined, and the features with a correlation coefficient higher than the second correlation coefficient threshold are determined as features related to the second influencing factor.

[0082] The first willingness feature is reconstructed based on the selected features related to the first influencing factor, and the first idle feature is reconstructed based on the selected features related to the second influencing factor. Finally, cross-validation and other methods are used to verify the performance of the decoupled first willingness feature and first idle feature in the prediction model, and the first willingness feature and first idle feature that meet the performance standards are used for subsequent applications.

[0083] In some embodiments, the feature decoupling of the information features in step 101 to obtain the second willingness feature and the second idle feature of the first information in each time period can be achieved by the following technical solution: determining a third influencing factor and a fourth influencing factor, the third influencing factor representing the attractiveness of the first information to the first object in the first time period, and the fourth influencing factor representing the probability that the call corresponding to the first information is answered by the first object in the first time period; based on the third influencing factor, extracting the second willingness feature of the first information in the first time period from the information feature; based on the fourth influencing factor, extracting the second idle feature of the first information in the first time period from the information feature.

[0084] In practice, the third influencing factor describes the appeal of the first message to the first recipient at each time period. This factor can include the following: information content relevance (how closely the message matches the first recipient's interests and needs); information source importance (the sender's fame, authority, or closeness to the first recipient); information immediacy (whether the message addresses an urgent or real-time event); and information format appeal (the form of the message, such as video, audio, or text). The fourth influencing factor describes the probability that the first recipient will answer the call corresponding to the first message at each time period. This factor can include the first recipient's availability (the first recipient's availability at each time period); the first recipient's historical response rate (the frequency with which the first recipient has responded to similar messages in the past); and call timing (whether the call is convenient for the first recipient to answer at that time).

[0085] In actual implementation, the information feature is a multidimensional feature, so the information feature can be regarded as including information sub-features of various dimensions. Features related to the third influencing factor can be selected from the information features in a variety of ways. For example, the correlation between each information sub-feature and the third influencing factor is determined, and the information sub-feature with a correlation higher than a third correlation threshold is regarded as a feature related to the third influencing factor; for another example, the transformation amount of the information sub-feature after the third influencing factor is combined with each information sub-feature is determined, and the information sub-feature with a transformation amount higher than the third transformation threshold is regarded as a feature related to the third influencing factor; for another example, the information features are clustered through visualization and statistical methods, and then the information sub-features in the information features that have a correlation with the features representing willingness are found as features related to the third influencing factor; for another example, the correlation coefficient between each information sub-feature and the third influencing factor is determined, and the information sub-feature with a correlation coefficient higher than the third correlation coefficient threshold is determined as a feature related to the first influencing factor.

[0086] Similarly, features related to the fourth influencing factor can be selected from the information features in a variety of ways. For example, the correlation between each information sub-feature and the fourth influencing factor is determined, and the information sub-feature with a correlation higher than a fourth correlation threshold is used as the feature related to the fourth influencing factor. For another example, the transformation amount of the information sub-feature after the fourth influencing factor is combined with each information sub-feature is determined, and the information sub-feature with a transformation amount higher than a fourth transformation threshold is used as the feature related to the fourth influencing factor. For another example, the information features are clustered through visualization and statistical methods, and then the information sub-features that have a correlation relationship with the features representing the idleness are found in each information feature as the features related to the fourth influencing factor. For another example, the correlation coefficient between each information sub-feature and the fourth influencing factor is determined, and the features with a correlation coefficient higher than the fourth correlation coefficient threshold are determined as the features related to the fourth influencing factor.

[0087] The second willingness feature is reconstructed based on the selected features related to the third influencing factor, and the second idle feature is reconstructed based on the selected features related to the fourth influencing factor. Finally, cross-validation and other methods are used to verify the performance of the decoupled second willingness feature and second idle feature in the prediction model, and the second willingness feature and second idle feature that meet the performance standards are used for subsequent applications.

[0088] In the above manner, by decoupling the independent features related to willingness and idleness, more accurate prediction models can be constructed. These models can better predict the response behavior and idle status of the first object in different time periods, thereby optimizing the information sending strategy and helping to provide a personalized first object experience. For example, based on the willingness characteristics of the first object, the system can recommend information that the first object is more willing to receive, and send information when the first object is more idle, thereby improving the first object's satisfaction. By identifying the time period when the first object is most likely to respond, resource allocation can be optimized, which helps to reduce invalid attempts and save costs.

[0089] In step 102, based on the first willingness feature and the second willingness feature, a first probability that the first object connects the call corresponding to the first information in the first time period is determined, and based on the first idle feature and the second idle feature, a second probability that the first object connects the call in the first time period is determined.

[0090] In some embodiments, determining the first probability that the first object connects the call corresponding to the first information at each moment based on the first willingness feature and the second willingness feature in step 102 can be achieved by the following technical solution: performing feature fusion on the first willingness feature and the second willingness feature to obtain a first fused feature; performing feature mapping on the first fused feature based on the weight matrix and the bias vector to obtain a first probability that the first object connects the call corresponding to the first information in the first time period.

[0091] In actual implementation, the method of fusing the first intention feature and the second intention feature can be to directly splice the first intention feature and the second intention feature, and then use the spliced ​​feature as the first fusion feature.

[0092] In actual implementation, feature mapping can be performed on the first fusion feature based on the following formula to obtain the first probability:

[0093] y=Sigmoid(Wx+b) (1)

[0094] In formula (1), y is the first probability of the output, x is the first fused feature, W is the weight matrix, b is the bias vector, and Sigmoid() is a mathematical function that can map any real value to a value between 0 and 1, which is used to convert the result of the linear combination into a probability. Among them, the weight matrix is ​​a two-dimensional matrix whose elements represent the strength of the association between the input feature and the output, and the bias vector is a one-dimensional vector whose elements represent the inherent bias or threshold of each output node.

[0095] Through the above formula (1), the first fusion feature can be substituted into formula (1), and after adjusting the first fusion feature through the weight matrix and the bias vector, it is mapped through the sigmoid function to obtain the first probability.

[0096] By performing prediction in combination with the features of the first object's willingness dimension, the accuracy of the predicted first probability is improved.

[0097] In actual implementation, based on the first idle feature and the second idle feature, a general method for determining the second probability that the first object connects to the call corresponding to the first information in each time period can be to first concatenate the first idle feature and the second idle feature to obtain a second fused feature, and then map the second fused feature through the function in the above formula (1) to obtain a second probability.

[0098] In step 103, a calling strategy for corresponding the first information to the first object is determined based on the first probability and the second probability.

[0099] In some embodiments, the calling strategy corresponding to the first information for the first object determined based on the first probability and the second probability in step 103 can be implemented by the following technical solution: based on the first probability and the second probability, determine the first comprehensive probability that the first object will connect to the call corresponding to the first information in the first time period; in the first time period with the highest first comprehensive probability, make a call corresponding to the first information to the first object.

[0100] In actual implementation, the comprehensive probability represents the probability of the first object answering the call corresponding to the first information in each time period. The comprehensive probability can be obtained by performing a weighted summation of the first probability and the second probability, as shown in the following formula (2):

[0101] S=r*E_will+(1-r)*E_avail (2)

[0102] In formula (2), S is the comprehensive probability, r is the adjustment parameter, E_will is the first probability, and E_avail is the second probability. Substitute the values ​​of the first probability and the second probability into the above formula (2), adjust the proportion of the first probability and the second probability by adjusting the parameter, and then add them together to obtain the comprehensive probability.

[0103] In actual implementation, in order to increase the connection rate of the first object as much as possible and avoid calling the first object when the first object does not want to be disturbed, after determining the comprehensive probability corresponding to each time period, the first object can be called in the time period with the highest comprehensive probability. For example, the comprehensive probability of time period A is 30%, the comprehensive probability of time period B is 50%, and the comprehensive probability of time period C is 70%, then the first object is called in time period C.

[0104] In some embodiments, the determination of the call strategy corresponding to the first information to the user based on the first probability and the second probability in step 103 can be implemented by the following technical solution: based on the first probability and the second probability, determine the first comprehensive probability that the first object connects the call corresponding to the first information in the first time period; in response to the second comprehensive probability corresponding to the second information in the first time period being the same as the first comprehensive probability, and the second comprehensive probability and the first comprehensive probability being higher than the probability threshold, determine the priority of the first information and the second information; in order of priority, determine the target information from the first information and the second information, and make a call corresponding to the target information to the first object in the first time period.

[0105] In actual implementation, due to the individual differences of the first object, there may be a relatively short time when the first object is willing to accept the call, resulting in the highest comprehensive probability values ​​obtained from many information predictions being concentrated in one time period. For example, the comprehensive probability of information A in time period A is 30%, the comprehensive probability of time period B is 50%, and the comprehensive probability of time period C is 70%. The comprehensive probability of information B in time period A is 40%, the comprehensive probability of time period B is 50%, and the comprehensive probability of time period C is 80%. The comprehensive probability of information C in time period A is 20%, the comprehensive probability of time period B is 60%, and the comprehensive probability of time period C is 77%. It can be seen that the time period with the highest comprehensive probability corresponding to information A, information B and information C is all time period C; and excessive concentration of information calls may make The first object is disgusted with the calling information. Therefore, when there are multiple second information in the time period with the highest comprehensive probability, the multiple information can be screened. First, the information with a comprehensive probability not higher than the probability threshold can be eliminated, and then the priority of each information included in the time period can be determined. For example, the time period with the highest comprehensive probability of each information is time period A, and time period A includes information A, information B and information C, among which information A has a higher priority than information C, and information C has a higher priority than information B. The result of sorting according to priority is information A, information C and information B. If the first number is 2 at this time, information A and information C are selected to call the first object, and information A is called to the first object first, and then information B is called to the first object.

[0106] Through the above-mentioned method, it is possible to avoid excessive concentration of information calls, thereby preventing the first party from feeling disgusted with the information in the calls, and ensuring the first party's experience.

[0107] Below, the training method of the model provided in the embodiment of the present application is described. As mentioned above, the electronic device that implements the training method of the model in the embodiment of the present application can be a terminal, a server, or a combination of the two. Therefore, the execution subject of each step will not be repeated below.

[0108] See also Figure 3B , Figure 3BThis is a first flow chart of the training method of the model provided in the embodiment of the present application, which will be combined with Figure 3B The steps shown are explained.

[0109] In step 201, the first sample feature of the object sample in the second period is feature decoupled through the model to obtain the third intention feature and the third idle feature of the object sample in the second period, and the second sample feature is feature decoupled to obtain the fourth intention feature and the fourth idle feature of the information sample in the second period.

[0110] In actual implementation, the sample pair can be composed of an object sample and an information sample. The label value of the sample pair represents the probability of the object sample being connected when facing a call from the information sample. For example, sample pair A includes object sample A and information sample B. The representation value of sample pair A is the probability that object sample A is connected when it receives a call from information sample B.

[0111] In actual implementation, the first sample feature and the second sample feature can be decoupled by first analyzing the first sample feature and the second sample feature to identify which features may affect willingness and availability. Specifically, the following methods can be used to determine which features affect willingness and availability. For example, by clustering object features through visualization and statistical methods, the features in each first sample feature are found to be correlated with the features representing willingness, and the features in each first sample feature are found to be correlated with the features representing availability. Similarly, the second sample features can be clustered to find the features in each second sample feature that are correlated with the features representing willingness, and the features in each first sample feature are found to be correlated with the features representing availability. The characteristics of the idleness have a correlation relationship; for example, the correlation coefficient between the first sample feature and the feature representing the willingness, and the correlation coefficient between the first sample feature and the feature representing the idleness can be determined. Similarly, the correlation coefficient between the second sample feature and the feature representing the willingness, and the correlation coefficient between the second sample feature and the feature representing the idleness can be determined, and the characteristics with correlation coefficients higher than the correlation coefficient threshold are determined to have a correlation relationship; then, based on the results of the feature analysis, features related to the willingness or features related to the idleness can be selected, wherein there can be multiple feature selection methods, for example, evaluating the correlation between a single first sample feature and the feature representing the willingness, and evaluating a single first sample feature. The correlation between a single sample feature and the feature representing idleness can be evaluated, and the correlation between a single second sample feature and the feature representing willingness can be evaluated, as well as the correlation between a single second sample feature and the feature representing idleness. For example, the correlation between the combination of multiple first sample features and the feature representing willingness can be evaluated, as well as the correlation between the combination of multiple first sample features and the feature representing idleness can be evaluated. At the same time, the correlation between the combination of multiple second sample features and the feature representing willingness can be evaluated, as well as the correlation between the combination of multiple second sample features and the feature representing idleness can be evaluated; then, the features are transformed using dimensionality reduction techniques such as PCA to find the feature that best represents the original data (i.e. The features related to willingness or idleness in the first sample features and the second sample features are selected. The features related to willingness or idleness can capture the key changes in the data. The recursive feature elimination technique is used to select the most important features. Features are selected by recursively reducing the size of the feature set. The least important features are removed in each iteration. According to the selected features related to willingness or idleness, the third willingness feature, the third idleness feature, the fourth willingness feature and the fourth idleness feature are reconstructed to ensure that they are as independent as possible. The performance of the decoupled features in the prediction model is verified by cross-validation and other methods, and it is determined that the features that meet the performance standards are the third willingness feature, the third idleness feature, the fourth willingness feature and the fourth idleness feature.

[0112] In step 202, based on the third willingness feature and the fourth willingness feature, a third probability of the object sample connecting to the call corresponding to the information sample at each moment is determined, and based on the third idle feature and the fourth idle feature, a fourth probability of the object sample connecting to the call corresponding to the information sample in the second time period is determined.

[0113] In actual implementation, the third probability and the fourth probability can be determined by concatenating the third intention feature and the fourth intention feature to obtain a third fusion feature, concatenating the third idle feature and the fourth idle feature to obtain a fourth fusion feature, and then inputting the third fusion feature and the fourth fusion feature into the following formula (3) to obtain the third probability and the fourth probability respectively.

[0114] y=Sigmoid(Wx+b) (3)

[0115] In formula (3), y is the third probability or the fourth probability of the output (if the input is the third fusion feature, the output is the third probability, and if the input is the fourth fusion feature, the output is the third probability), x is the third fusion feature or the fourth fusion feature, W is the weight matrix, and b is the bias vector.

[0116] In step 203, a first loss function is constructed based on the third probability, a second loss function is constructed based on the fourth probability, and the parameters of the model are updated based on the first loss function and the second loss function.

[0117] In some embodiments, constructing the first loss function based on the third probability in step 203 can be implemented by the following technical solution: when there are multiple sample pairs, sort the multiple sample pairs in descending order according to the third probability to obtain a sample pair sequence; select a second number of first sample pairs from the head of the sample pair sequence, and select a second number of second sample pairs from the tail of the sample pair sequence; determine the fifth probability that the object sample in the first sample pair actually connects the call corresponding to the information sample, and determine the sixth probability that the object sample in the second sample pair actually connects the call corresponding to the information sample; determine the difference between the fifth probability and the sixth probability, and construct a first loss function based on the difference, and the first loss function is negatively correlated with the difference.

[0118] In actual implementation, after obtaining the third probability of each sample pair, the sample pairs can be sorted according to the third probability value. Then, a second number of first sample pairs can be selected from the beginning, and a second number of second sample pairs can be selected from the end. For example, if the third probability of sample pair A is 0.5, the third probability of sample pair B is 0.6, the third probability of sample pair C is 0.7, and the third probability of sample pair D is 0.8, then the sample pair sequence obtained by sorting is sample pair D, sample pair C, sample pair B, and sample pair A. If the second number is 2, then the first sample pairs selected are sample pair D and sample pair C, and the second sample pairs are sample pair B and sample pair A.

[0119] In actual implementation, the fifth probability representing that the object sample in the first sample pair actually connects the call corresponding to the information sample is the label value of the first sample pair. Similarly, the sixth probability representing that the object sample in the second sample pair actually connects the call corresponding to the information sample is the label value of the second sample pair.

[0120] In actual implementation, the first loss function constructed based on the fifth probability and the sixth probability can be as follows:

[0121] L_will=-exp(will1-will2) (4)

[0122] In formula (4), L_will is the first loss function, will1 is the fifth probability, and will2 is the sixth probability. Based on the first loss function, we can first determine the difference between the fifth and sixth probabilities, then use this difference as the independent variable of an exponential function, and use the inverse of the dependent variable of the exponential function as the first loss function.

[0123] By the above-mentioned method, the deviation of the homogeneous scores of the first subject in all time periods is eliminated, so that the scores between different time periods have differentiated performance close to the real level.

[0124] In some embodiments, constructing the second loss function based on the fourth probability in step 203 can be implemented by the following technical solution: grouping the sample pairs according to the second time period to obtain the sample pair group corresponding to the second time period; for the sample pair group corresponding to the second time period, averaging the fourth probabilities corresponding to each sample pair in the sample pair group to obtain the seventh probability that the call for the information sample of the object sample in the second time period is predicted to be connected; determining the eighth probability that the call for the information sample of the object sample in the second time period is actually connected, and determining the mean square error between the seventh probability and the eighth probability as the second loss function.

[0125] In actual implementation, sample pairs can be grouped by time period. For example, the time period corresponding to sample pair A is time period A, the time period corresponding to sample pair B is time period B, the time period corresponding to sample pair C is time period B, and the time period corresponding to sample pair D is time period A. The grouping result is that sample pair A and sample pair B are grouped together, and sample pair C and sample pair B are grouped together. The fourth probabilities of the sample pairs included in each group are averaged. For example, if sample pair A and sample pair D are grouped together, the fourth probability of sample pair A is 0.6, and the fourth probability of sample pair D is 0.8, the average seventh probability is 0.7.

[0126] In actual implementation, the second loss function can be determined as follows:

[0127] L_avail=MSE(h avail , h rate ) (5)

[0128] In formula (5), L_avail is the second loss function, MSE() is the mean square error function, h avail is the seventh probability, h rate is the eighth probability, and the seventh and eighth probabilities are substituted into the mean square error function to obtain the output of the mean square error function, and the output is used as the second loss function. The mean square error loss is a loss function commonly used in regression tasks, which calculates the average of the squares of the differences between the predicted values ​​and the true values.

[0129] In some embodiments, updating the parameters of the prediction model based on the first loss function and the second loss function in step 203 can be achieved by the following technical solution: performing probability fusion on the third probability and the fourth probability to obtain a ninth probability, and determining the tenth probability that the call of the information sample of the object sample in the sample pair is actually connected; constructing the cross entropy loss based on the ninth probability and the tenth probability to obtain the third loss function; fusing the first loss function, the second loss function and the third loss function to obtain the total loss function, and updating the parameters of the model based on the total loss function.

[0130] In actual implementation, the third probability and the fourth probability may be fused by weighted summing the third probability and the fourth probability. For details, see the following formula:

[0131] S=r*E_will+(1-r)*E_avail (6)

[0132] In formula (6), S is the ninth probability, r is the adjustment parameter, E_will is the third probability, and E_avail is the fourth probability. Substituting the values ​​of the third and fourth probabilities into the above formula (6), adjusting the proportions of the third and fourth probabilities by adjusting the parameters and then adding them together, the ninth probability can be obtained.

[0133] In actual implementation, the cross entropy loss constructed based on the ninth probability and the tenth probability can be:

[0134]

[0135] In formula (7), L_main is the cross entropy loss (i.e., the third loss function), N is the number of sample pairs, and y i is the tenth probability, is the ninth probability. By constructing the cross-entropy loss function with the ninth and tenth probabilities, the loss value of the cross-entropy loss function is used as the third loss. Cross-entropy loss is a commonly used loss function in machine learning. It is used to evaluate the difference between the probability distribution predicted by the classification model and the probability distribution of the actual label. The cross-entropy loss function is used to guide the model to adjust weights during the optimization process to reduce the difference between the predicted distribution and the true distribution.

[0136] In implementation, the first loss function, the second loss function and the third loss function are fused, and the total loss can be obtained by weighted summing the first loss function, the second loss function and the third loss function. For details, please refer to the following formula:

[0137] Loss=a*L_main+b*L_avail+c*L_will (8)

[0138] In formula (8), Loss is the total loss function, a, b, and c are weight parameters, L_main is the third loss function, L_avail is the second loss function, and L_will is the first loss function. The weight parameters are used to adjust the proportion of each loss in the total loss, thereby adjusting the training effect of the model.

[0139] Through the above method, the trained model can help design more attractive interaction methods by understanding the willingness characteristics and idle characteristics of the first object, improve the participation and loyalty of the first object, reduce the sending of information during the time period when the first object is less idle or does not want to be disturbed, reduce interference with the first object, avoid causing dissatisfaction of the first object, and be able to identify the time period when the first object is most likely to respond, optimize resource allocation, help reduce invalid attempts, and save costs.

[0140] The following describes an exemplary application of the embodiments of the present application in a practical application scenario.

[0141] See also Figure 4 , Figure 4 It is a schematic diagram of the implementation of the information processing method provided in the embodiment of the present application.

[0142] exist Figure 4In the method, the user side features (i.e., the object features mentioned above) are first divided into the first willingness features and the first idle features, and the project side features (i.e., the information features of the first information mentioned above) are divided into the second willingness features and the second idle features. Then, the first probability is determined based on the first willingness features and the second willingness features, and the second probability is determined based on the first idle features and the second idle features. The first probability and the second probability are integrated to obtain the comprehensive probability, and finally, an information call is made based on the comprehensive probability.

[0143] First, the prediction period is divided into t∈T periods, and the user's portrait features, historical features, and statistical features in different periods are collected to form the corresponding user-side features in the input sample (i.e., the above-mentioned behavioral features). Information such as time, date, holidays, weeks, and marketing attributes are collected to form the item-side features corresponding to the input sample (i.e., the attribute features of the first information mentioned above).

[0144] The above user-side features and item-side features, a total of f features, are encoded. If the feature is sparse, feature mapping (embedding) is performed. Otherwise, it is retained as the source data to obtain where g i is the length of the encoding query list of the i-th feature. Finally, we get X={x1,x2…,x f}, where i∈f represents the embedded features of f features (i.e., the user features mentioned above).

[0145] The user features derived from the user-side features are decoupled and learned through different multilayer perceptron (MLP) modules to obtain the user-dimensional willingness and availability representation vectors, which are respectively denoted as E(w,u) (i.e., the first willingness feature mentioned above) and E(a,u) (i.e., the first availability feature mentioned above);

[0146] Similarly, the information features obtained from the item-side features are decoupled through different MLP modules to obtain the item-dimensional willingness and availability representation vectors, which are respectively recorded as E(w,i) (i.e., the second willingness feature mentioned above) and E(a,i) (i.e., the second availability feature mentioned above);

[0147] Pass E(w,u) and E(w,i) through the willingness conversion layer to obtain the user's willingness score E_will for the current hour (i.e., the first probability mentioned above). Similarly, pass E(a,u) and E(a,i) through the availability conversion layer to obtain the user's willingness score E_avail for the current hour (i.e., the second probability mentioned above). The above conversion layer is a fully connected layer. Assuming that the input vector is x and the output vector y has the following mapping relationship:

[0148] y=Sigmoid(Wx+b) (9)

[0149] In formula (9), W is the learnable weight matrix and b is the learnable bias vector;

[0150] The higher the user's willingness, the higher the corresponding connection rate. The samples in the training sample are sorted in reverse order by the willingness score E_will (i.e., the third probability mentioned above). The actual connection rate of the first k samples with the willingness score at the top is calculated and recorded as will1 (i.e., the fifth probability mentioned above). The actual connection rate of the last k samples with the willingness score at the bottom is calculated and recorded as will2 (i.e., the sixth probability mentioned above). Therefore, the willingness loss is designed as follows:

[0151] L_will=-exp(will1-will2) (10)

[0152] In formula (10), L_will is the first loss, will1 is the fifth probability, and will2 is the sixth probability. Based on the first loss function, we can first determine the difference between the fifth and sixth probabilities, then use this difference as the independent variable of an exponential function, and use the inverse of the dependent variable of the exponential function as the first loss.

[0153] A higher hourly availability indicates a higher connection rate. The training samples are grouped by the hour field. The E_avail availability score corresponding to each hour (i.e., the fourth probability mentioned above) is calculated and the average is calculated as the predicted availability for each hour, recorded as h_avail (i.e., the seventh probability mentioned above). The actual connection rate for each hour in the training sample is determined and recorded as h_rate (i.e., the eighth probability mentioned above). The second loss constructed based on the seventh and eighth probabilities is as follows:

[0154] L_avail=MSE(h avail , h rate ) (11)

[0155] In formula (11), L_avail is the second loss, MSE() is the mean square error function, and h avail is the seventh probability, h rate is the eighth probability, and the seventh probability and the eighth probability are substituted into the mean square error function to obtain the output of the mean square error function, and the output is used as the second loss.

[0156] The present invention uses the collected samples as the training label whether the user is connected

[0157] The MLP structure involved in the present invention uses a 2-layer network structure, the number of nodes in each hidden layer is [256,128], the learning rate is 0.003, and the Adam learner is used for training

[0158] The output prediction is divided into Score (that is, the above-mentioned comprehensive probability) and label for cross entropy loss parameter transfer. The calculation formula is as follows:

[0159]

[0160] In formula (12), L_main is the cross entropy loss (i.e., the third loss), N is the number of sample pairs, and y i is the tenth probability, is the ninth probability. By constructing a cross entropy loss function with the ninth and tenth probabilities, the loss value of the cross entropy loss function is used as the third loss.

[0161] Among them, the obtained willingness score E_will and availability score E_avail are fused:

[0162] S=r*E_will+(1-r)*E_avail (13)

[0163] In formula (13), S is the ninth probability, r is the adjustment parameter, E_will is the third probability, and E_avail is the fourth probability. Substituting the values ​​of the third and fourth probabilities into the above formula (13), adjusting the proportions of the third and fourth probabilities by adjusting the parameters and then adding them together can obtain the ninth probability.

[0164] The loss is calculated as the overall connection rate in the user dimension, where r is a hyperparameter and Score is the final predicted connection rate probability;

[0165] Loss=a*L_main+b*L_avail+c*L_will (14)

[0166] In formula (14), Loss is the total loss, a, b, and c are weight parameters, L_main is the third loss, L_avail is the second loss, and L_will is the first loss. The weight parameters are used to adjust the proportion of each loss in the total loss, thereby adjusting the training effect of the model.

[0167] In the intelligent outbound calling system, by decoupling the factors that influence user call completion into two dimensions: "willingness" and "availability," and further designing four features with different meanings from the user side and the time side, this approach can produce the following beneficial effects:

[0168] Willingness reflects the user's interest in or acceptance of a specific marketing campaign. By analyzing historical behavioral data (such as past purchase records, website browsing history, etc.), a model can be built to predict the user's attitude towards new marketing campaigns. This means that companies can prioritize contacting target customers who are most likely to be interested. Availability focuses on whether the user has enough time and willingness to answer the phone at different times of the day. Taking into account people's daily living habits, for example, weekday evenings are more likely to be free than daytime, or weekends are more suitable for receiving calls than weekdays. Therefore, by analyzing the user's daily routine, the optimal call window can be determined, thereby increasing the possibility of getting through.

[0169] By modeling these two dimensions separately and utilizing comparative learning of high-intention user-low-intention user pairs (i.e., the first loss function mentioned above) and hourly availability-hourly connection rate pairs (i.e., the second loss function mentioned above), we can identify the user groups most likely to respond and the optimal communication time points. The designed comparative learning loss can eliminate the deviation of user connection rate homogeneity in different time periods, ensuring that the score in each time period reflects the actual user status.

[0170] It is understandable that in the embodiments of the present application, when user information and other related data are involved, when the embodiments of the present application are applied to specific products or technologies, user permission or consent must be obtained, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards.

[0171] The following continues to describe the exemplary structure of the information processing device 555A provided in the embodiment of the present application implemented as a software module. In some embodiments, such as Figure 2A As shown, the software modules stored in the information processing device 555A of the memory 550 may include:

[0172] Decoupling module 5551 is configured to perform feature decoupling on the object features of the first object in the first time period to obtain a first willingness feature and a first availability feature of the first object in the first time period, and to perform feature decoupling on the information features of the first information to obtain a second willingness feature and a second availability feature of the first information in the first time period;

[0173] The determining module 5552 is further configured to determine, based on the first willingness characteristic and the second willingness characteristic, a first probability that the first subject answers the call corresponding to the first information during the first time period, and based on the first idle characteristic and the second idle characteristic, determine a second probability that the first subject answers the call corresponding to the first information during the first time period;

[0174] The calling module 5553 is configured to determine a calling strategy for the first object corresponding to the first information based on the first probability and the second probability.

[0175] In some embodiments, the above-mentioned determination module 5552 is also used to extract features from the behavioral data of the first object in the first time period to obtain the behavioral features of the first object in the first time period; perform feature mapping on the sparse features in the behavioral features to obtain first mapping features; perform feature fusion on the dense features in the behavioral features and the first mapping features to obtain the object features of the first object in the first time period.

[0176] In some embodiments, the above-mentioned determination module 5552 is also used to extract features from the attribute data of the first information to obtain attribute features of the first information; perform feature mapping on the sparse features in the attribute features to obtain second mapping features; perform feature fusion on the dense features in the attribute features and the second mapping features to obtain information features of the first information.

[0177] In some embodiments, the above-mentioned decoupling module 5551 is also used to determine a first influencing factor and a second influencing factor, wherein the first influencing factor represents the willingness of the first object to connect the call corresponding to the first information in the first time period, and the second influencing factor represents the idleness of the first object in the first time period; based on the first influencing factor, the first willingness feature of the first object in the first time period is extracted from the object features; based on the second influencing factor, the first idle feature of the first object in the first time period is extracted from the object features.

[0178] In some embodiments, the above-mentioned decoupling module 5551 is also used to determine a third influencing factor and a fourth influencing factor, wherein the third influencing factor represents the attractiveness of the first information to the first object in the first time period, and the fourth influencing factor represents the probability that the call corresponding to the first information is answered by the first object in the first time period; based on the third influencing factor, the second willingness feature of the first information in the first time period is extracted from the information feature; based on the fourth influencing factor, the second idle feature of the first information in the first time period is extracted from the information feature.

[0179] In some embodiments, the above-mentioned determination module 5552 is also used to perform feature fusion on the first intention feature and the second intention feature to obtain a first fused feature; based on the weight matrix and the bias vector, perform feature mapping on the first fused feature to obtain a first probability that the first object connects the call corresponding to the first information in the first time period.

[0180] In some embodiments, the above-mentioned call module 5553 is also used to determine the first comprehensive probability that the first object will connect to the call corresponding to the first information in the first time period based on the first probability and the second probability; and in the first time period with the highest first comprehensive probability, make a call corresponding to the first information to the first object.

[0181] In some embodiments, the above-mentioned call module 5553 is also used to determine the first comprehensive probability that the first object connects the call corresponding to the first information in the first time period based on the first probability and the second probability; in response to the second comprehensive probability corresponding to the second information in the first time period being the same as the first comprehensive probability, and the second comprehensive probability and the first comprehensive probability being higher than the probability threshold, determine the priority of the first information and the second information; determine the target information from the first information and the second information in the order of the priorities, and make a call corresponding to the target information to the first object in the first time period.

[0182] The following continues to describe the exemplary structure of the training device 555B of the model provided in the embodiment of the present application implemented as a software module. In some embodiments, such as Figure 2B As shown, the software modules stored in the model training device 555B of the memory 550 may include:

[0183] a decoupling module 5554 configured to perform feature decoupling on the first sample feature of the object sample in the second time period using a model to obtain a third willingness feature and a third availability feature of the object sample in the second time period, and to perform feature decoupling on the second sample feature to obtain a fourth willingness feature and a fourth availability feature of the information sample in the second time period;

[0184] The determining module 5555 is further configured to determine, based on the third willingness characteristic and the fourth willingness characteristic, a third probability that the subject sample answers the call of the information sample during the second time period, and to determine, based on the third idle characteristic and the fourth idle characteristic, a fourth probability that the subject sample answers the call during the second time period;

[0185] The training module 5556 is used to construct a first loss function based on the third probability, construct a second loss function based on the fourth probability, and update the parameters of the model based on the first loss function and the second loss function.

[0186] In some embodiments, the above-mentioned training module 5556 is also used to, when there are multiple sample pairs, sort the multiple sample pairs in descending order according to the third probability to obtain a sample pair sequence; select a second number of first sample pairs from the head of the sample pair sequence, and select a second number of second sample pairs from the tail of the sample pair sequence; determine the fifth probability that the object sample in the first sample pair actually connects the call corresponding to the information sample, and determine the sixth probability that the object sample in the second sample pair actually connects the call corresponding to the information sample; determine the difference between the fifth probability and the sixth probability, and construct the first loss function based on the difference, and the first loss function is negatively correlated with the difference.

[0187] In some embodiments, the above-mentioned training module 5556 is also used to group the sample pairs according to the second time period to obtain sample pair groups corresponding to each second time period; for the sample pair groups corresponding to each second time period, the fourth probabilities corresponding to each sample pair in the sample pair group are averaged to obtain the seventh probability that the call for the information sample of the object sample in the second time period is predicted to be connected; determine the eighth probability that the call for the information sample of the object sample in the second time period is actually connected, and determine the mean square error between the seventh probability and the eighth probability as the second loss function.

[0188] In some embodiments, the above-mentioned training module 5556 is also used to perform probability fusion on the third probability and the fourth probability to obtain a ninth probability, and determine the tenth probability that the call of the information sample of the object sample in the sample pair is actually connected; construct a cross-entropy loss based on the ninth probability and the tenth probability to obtain a third loss function; fuse the first loss function, the second loss function and the third loss function to obtain a total loss function, and update the parameters of the model based on the total loss function.

[0189] The present invention provides a computer program product comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the electronic device to perform the information processing method or model training method described in the present invention.

[0190] The embodiment of the present application provides a computer-readable storage medium in which computer-executable instructions or computer programs are stored. When the computer-executable instructions or computer programs are executed by a processor, the processor will execute the information processing method or model training method provided in the embodiment of the present application, for example, Figure 3AThe information processing method shown.

[0191] In some embodiments, the computer-readable storage medium may be a memory such as RAM, ROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or may be various devices including one or any combination of the above memories.

[0192] In some embodiments, computer-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.

[0193] As an example, computer-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).

[0194] By way of example, computer-executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed across multiple sites and interconnected by a communication network.

[0195] In summary, the following beneficial effects can be achieved through the embodiments of the present application:

[0196] The object features of the first object in the first time period are feature decoupled to obtain a first willingness feature and a first availability feature of the first object in the first time period, and the information features of the first information are feature decoupled to obtain a second willingness feature and a second availability feature of the first information in the first time period. By decoupling the features of the first object and the features of the first information, the features of the first object and the features of the first information are accurately divided into features of the willingness dimension and features of the availability dimension for measuring the probability of the first object connecting, thereby improving the accuracy of the subsequently determined comprehensive probability. Based on the first willingness feature and the second willingness feature, a first probability of the first object connecting to the call corresponding to the first information in the first time period is determined, and based on the first availability feature and the second availability feature, a second probability of the first object connecting to the call corresponding to the first information in the first time period is determined. Based on the first probability and the second probability, a comprehensive probability of the first object connecting to the call corresponding to the first information in the first time period is determined. By combining the features of the first object and the features of the first information, a first probability of the first object connecting to the call in the willingness dimension and a second probability of the first object connecting to the call in the availability dimension are determined. The first probability and the second probability are combined to determine a call strategy corresponding to the first information for the first object, thereby improving the probability of the user connecting the call and thereby improving the effectiveness of information recommendation.

[0197] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the scope of protection of the present application.

Claims

1. An information processing method, characterized in that: The method comprises: Performing feature decoupling on the object features of the first object in the first time period to obtain a first willingness feature and a first availability feature of the first object in the first time period, and performing feature decoupling on the information features of the first information to obtain a second willingness feature and a second availability feature of the first information in the first time period; Determining, based on the first willingness feature and the second willingness feature, a first probability that the first subject answers the call corresponding to the first information during the first time period, and determining, based on the first idle feature and the second idle feature, a second probability that the first subject answers the call during the first time period; A calling strategy for making a call corresponding to the first information to the first object is determined based on the first probability and the second probability.

2. The method according to claim 1, characterized in that Before performing feature decoupling on the object features of the first object in the first time period, the method further includes: performing feature extraction on the behavior data of the first object in the first time period to obtain a behavior feature of the first object in the first time period; Performing feature mapping on the sparse features in the behavior features to obtain first mapping features; Feature fusion is performed on the dense features in the behavior features and the first mapping features to obtain object features of the first object in the first time period.

3. The method according to claim 1, characterized in that Before performing feature decoupling on the information features of the first information, the method further includes: performing feature extraction on the attribute data of the first information to obtain attribute features of the first information; Performing feature mapping on the sparse features in the attribute features to obtain second mapping features; Feature fusion is performed on the dense feature in the attribute feature and the second mapping feature to obtain information features of the first information.

4. The method according to claim 1, wherein Decoupling the object features of the first object in the first time period to obtain a first willingness feature and a first availability feature of the first object in the first time period includes: Determining a first influencing factor and a second influencing factor, wherein the first influencing factor represents a willingness of the first subject to answer the call corresponding to the first information in the first time period, and the second influencing factor represents a degree of availability of the first subject in the first time period; extracting, based on the first influencing factor, a first willingness feature of the first subject in the first time period from the subject features; Based on the second influencing factor, a first idle feature of the first object in the first time period is extracted from the object features.

5. The method according to claim 1, wherein Decoupling the information feature of the first information to obtain the second willingness feature and the second availability feature of the first information in the first time period includes: Determining a third impact factor and a fourth impact factor, wherein the third impact factor represents the attractiveness of the first information to the first target in the first time period, and the fourth impact factor represents the probability that a call corresponding to the first information is answered by the first target in the first time period; extracting, based on the third influencing factor, a second intention feature of the first information in the first time period from the information feature; Based on the fourth impact factor, a second idle feature of the first information in the first time period is extracted from the information feature.

6. The method according to claim 1, wherein The determining, based on the first willingness feature and the second willingness feature, a first probability that the first subject answers the call corresponding to the first information at the first moment includes: Performing feature fusion on the first intention feature and the second intention feature to obtain a first fused feature; Based on the weight matrix and the bias vector, feature mapping is performed on the first fused feature to obtain a first probability that the first object answers the call corresponding to the first information during the first time period.

7. The method according to claim 1, characterized in that The determining, based on the first probability and the second probability, a calling strategy for the first object corresponding to the first information includes: Determining, based on the first probability and the second probability, a first comprehensive probability that the first subject connects the call corresponding to the first information in the first time period; During the first time period when the first comprehensive probability is the highest, a call corresponding to the first information is made to the first object.

8. The method according to claim 1, characterized in that The determining, based on the first probability and the second probability, a calling strategy corresponding to the first information for the user includes: Determining, based on the first probability and the second probability, a first comprehensive probability that the first subject connects the call corresponding to the first information in the first time period; In response to a second comprehensive probability corresponding to the second information within the first time period being the same as the first comprehensive probability, and the second comprehensive probability and the first comprehensive probability being higher than a probability threshold, determining the priority of the first information and the second information; A first number of third information is determined from the first information and the second information in the order of priority, and calls corresponding to the first number of third information are made to the first object in the first time period.

9. A model training method, characterized in that: The method comprises: By using the model, feature decoupling is performed on the first sample feature of the object sample in the second period to obtain a third willingness feature and a third idle feature of the object sample in the second period, and feature decoupling is performed on the second sample feature to obtain a fourth willingness feature and a fourth idle feature of the information sample in the second period; Determining, based on the third willingness feature and the fourth willingness feature, a third probability that the subject sample connects the call of the information sample during the second time period, and determining, based on the third idle feature and the fourth idle feature, a fourth probability that the subject sample connects the call during the second time period; A first loss function is constructed based on the third probability, a second loss function is constructed based on the fourth probability, and parameters of the model are updated based on the first loss function and the second loss function.

10. The method according to claim 9, characterized in that The constructing a first loss function based on the third probability includes: In the case where there are multiple sample pairs, sorting the multiple sample pairs in descending order of the third probability to obtain a sample pair sequence; Selecting a second number of first sample pairs from the head of the sample pair sequence, and selecting the second number of second sample pairs from the tail of the sample pair sequence; Determining a fifth probability that the subject sample in the first sample pair actually connects the call corresponding to the information sample, and determining a sixth probability that the subject sample in the second sample pair actually connects the call corresponding to the information sample; A difference between the fifth probability and the sixth probability is determined, and a first loss function is constructed based on the difference, wherein the first loss function is negatively correlated with the difference.

11. The method according to claim 9, characterized in that The constructing a second loss function based on the fourth probability includes: Grouping the sample pairs according to each of the second time periods to obtain sample pair groups corresponding to each of the second time periods; For each sample pair group corresponding to the second time period, averaging the fourth probabilities corresponding to each sample pair in the sample pair group to obtain a seventh probability that the call corresponding to the information sample of the object sample in the second time period is predicted to be connected; An eighth probability that the call for the information sample to the object sample in the second time period is actually connected is determined, and a mean square error between the seventh probability and the eighth probability is determined as the second loss function.

12. The method according to claim 9, characterized in that The updating of the parameters of the model based on the first loss function and the second loss function includes: Performing probability fusion on the third probability and the fourth probability to obtain a ninth probability, and determining a tenth probability that the call corresponding to the information sample of the object sample in the sample pair is actually connected; Performing a cross entropy loss based on the ninth probability and the tenth probability to obtain a third loss function; The first loss function, the second loss function and the third loss function are fused to obtain a total loss function, and the parameters of the model are updated based on the total loss function.

13. An electronic device, characterized in that: The electronic device comprises: a memory for storing computer-executable instructions or computer programs; A processor, for implementing the information processing method described in any one of claims 1 to 8 or the model training method described in any one of claims 9 to 12 when executing the computer-executable instructions or computer programs stored in the memory.

14. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that: When the computer-executable instructions or computer programs are executed by a processor, the information processing method described in any one of claims 1 to 8 or the model training method described in any one of claims 9 to 12 is implemented.

15. A computer program product comprising computer executable instructions or a computer program, characterized in that When the computer-executable instructions or computer programs are executed by a processor, the information processing method described in any one of claims 1 to 8 or the model training method described in any one of claims 9 to 12 is implemented.

Images