An object recommendation method and apparatus, an electronic device, and a storage medium

Abstract

The present disclosure relates to an object recommendation method and device, electronic equipment and storage medium, the method comprising: obtaining a current timeliness parameter of a candidate object; the current timeliness parameter representing a probability that the candidate object has a recommendation value at a current time; determining a recommendation index of the candidate object according to an account feature of a to-be-recommended account, the current timeliness parameter of the candidate object and an object feature of the candidate object; the recommendation index of the candidate object being positively correlated with the current timeliness parameter of the candidate object; and performing object recommendation to the to-be-recommended account based on the recommendation index of the candidate object. The object recommendation method of the present disclosure can enhance the influence of the timeliness of the object itself on the recommendation result in the object recommendation process, improve the exposure probability of the candidate object with a high probability of having a recommendation value at the current time, ensure that the finally recommended object has a high recommendation value, and further improve the problem of unfair exposure of the object caused by the recommendation method biased by popularity or excessively dependent on user preference.

Description

Technical Field

[0001] This disclosure relates to the field of Internet technology, and in particular to an object recommendation method, apparatus, electronic device, and storage medium. Background Technology

[0002] In recent years, the bias problem in recommendation systems has attracted increasing attention and concern, because bias (such as popularity bias) often leads to a small number of objects receiving the vast majority of exposure resources in the system, causing phenomena such as the long tail and the Matthew effect. This can lead to problems such as a decline in consumer experience, loss of creators, and loss of diversity in the system ecosystem.

[0003] Therefore, it is urgent to develop an object-level fair exposure mechanism from the perspective of recommendation systems, and achieving reasonable allocation of recommendation resources is the key to improving recommendation performance. Summary of the Invention

[0004] This disclosure provides a method, apparatus, electronic device, and storage medium for object recommendation. The technical solution of this disclosure is as follows:

[0005] According to a first aspect of the present disclosure, an object recommendation method is provided, comprising:

[0006] Obtain the current timeliness parameter of the candidate object; the current timeliness parameter represents the probability that the candidate object has recommendation value at the current moment; the current timeliness parameter is determined based on the interaction data of the candidate object in the historical time period and the current basic failure parameter; the current basic failure parameter represents the basic probability that the candidate object has recommendation value at the current moment but loses recommendation value in the instant after the current moment;

[0007] Based on the account characteristics of the account to be recommended, the current timeliness parameters of the candidate object, and the object characteristics of the candidate object, the recommendation index of the candidate object is determined; the recommendation index represents the probability that the candidate object is successfully recommended to the account to be recommended; the recommendation index of the candidate object is positively correlated with the current timeliness parameters of the candidate object.

[0008] Based on the recommendation metrics of candidate objects, object recommendations are made to the accounts to be recommended.

[0009] In some possible embodiments, obtaining the current timeliness parameters of the candidate object includes:

[0010] In response to the recommendation request sent by the client corresponding to the account to be recommended, the current time is determined based on the timestamp information corresponding to the recommendation request;

[0011] Obtain the time-sensitive data of the candidate objects; the time-sensitive data includes multiple time-sensitive parameters, which correspond to multiple different times. Each time-sensitive parameter represents the probability that the candidate object has recommendation value at the time corresponding to each time-sensitive parameter.

[0012] Based on the current moment, the current timeliness parameter is determined from multiple timeliness parameters.

[0013] In some possible embodiments, the historical time period is initialized with the upload time of the candidate object; the method further includes:

[0014] Start timing from the initial moment;

[0015] When the timer reaches the first preset duration, the interaction data between the candidate object and the recommended account within the first preset duration is obtained; the interaction data includes the interaction feature value of the candidate object for each of the multiple interaction behaviors;

[0016] The model for determining timeliness parameters is obtained; the model for determining timeliness parameters includes a trained basic failure parameter determination model and weight coefficients corresponding to each interaction behavior; the basic failure parameter determination model is used to determine the basic probability that a candidate object has recommendation value at each time point in multiple different time points, but loses recommendation value in the instant after each time point;

[0017] Based on the trained basic failure parameters, determine the model, the weight coefficients corresponding to each interaction behavior, and the interaction feature values ​​of the candidate object with respect to each interaction behavior, and determine the timeliness parameters of the candidate object at each different time.

[0018] In some possible embodiments, the current timeliness parameter determines the model determination; the timeliness parameter determines the training method of the model, including:

[0019] Acquire historical interaction data for each of multiple historical objects; the historical interaction data for each historical object includes first interaction data within a first preset duration starting from the upload time of each historical object and second interaction data within a second preset duration starting from the middle time; the middle time is the time when the first preset duration ends.

[0020] The tag data for each historical object is determined based on the second interaction data of each historical object; the tag data for each historical object represents the historical moment when each historical object lost its recommendation value.

[0021] Obtain initial time-sensitivity parameters to determine the model; the initial time-sensitivity parameter determination model includes the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior;

[0022] Using the label data and first interaction data of each historical object, the basic failure parameters to be trained are determined to determine the model and the initial weight coefficients corresponding to each interaction behavior, and the initial weight coefficients are iteratively updated.

[0023] The training continues until the preset end-of-training conditions are met, resulting in a well-trained, time-determined model with the correct parameters.

[0024] In some possible embodiments, the second interaction data for each historical object includes the number of exposures of each historical object at each moment within a second preset duration and the interaction feature value of each historical object at each moment with respect to each interaction behavior;

[0025] The tag data for each historical object is determined based on the second interaction data of each historical object, including:

[0026] Based on the interaction feature values ​​of each historical object with respect to each interaction behavior at each time step, determine the recommendation ranking percentage of each historical object among multiple historical objects at each time step;

[0027] At each moment, if the number of exposures of each historical object at that moment is greater than or equal to the preset number, the recommendation degree value of each historical object at that moment is determined based on the recommendation ranking percentage of each historical object at that moment.

[0028] The recommendation score of each historical object at each time point is summed to obtain the cumulative recommendation score of each historical object at each time point;

[0029] For each historical object, when the cumulative recommendation value is less than or equal to the preset recommendation value, the moment corresponding to the cumulative recommendation value is determined as the historical moment when the historical object loses its recommendation value.

[0030] In some possible embodiments, the first interaction data for each historical object includes interaction feature values ​​for each historical object with respect to each interaction behavior;

[0031] Using the label data and first interaction data of each historical object, the basic failure parameters to be trained are determined, and the initial weight coefficients corresponding to each interaction behavior are iteratively updated, including:

[0032] Based on the failure parameters to be trained, determine the model, the initial weight coefficients corresponding to each interaction behavior, and the interaction feature values ​​of each historical object with respect to each interaction behavior, and determine the prediction timeliness parameter of each historical object at each time step; the prediction timeliness parameter of each historical object at each time step represents the predicted probability that each historical object has recommendation value at each time step.

[0033] Based on the historical moment when each historical object loses its recommendation value and the predicted timeliness parameters of each historical object at each moment, the maximum likelihood estimation method is used to iteratively update the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior, so as to obtain the trained basic failure parameter determination model and the weight coefficients corresponding to each interaction behavior.

[0034] In some possible embodiments, recommendation metrics for candidate objects are determined based on the account characteristics of the account to be recommended, the current validity parameters of the candidate objects, and the object characteristics of the candidate objects, including:

[0035] Based on the account characteristics of the account to be recommended and the object characteristics of the candidate object, determine the correlation index between the account to be recommended and the candidate object;

[0036] By using the current timeliness parameters of the candidate objects, the relevance indicators are adjusted to obtain the recommended indicators for the candidate objects.

[0037] According to a second aspect of the present disclosure, an object recommendation apparatus is provided, comprising:

[0038] The acquisition module is configured to acquire the current validity parameters of candidate objects. The current validity parameters represent the probability that a candidate object has recommendation value at the current moment. The current validity parameters are determined based on the interaction data of the candidate object in the historical time period and the current basic failure parameters. The current basic failure parameters represent the basic probability that a candidate object has recommendation value at the current moment but loses its recommendation value in the instant after the current moment.

[0039] The determination module is configured to determine the recommendation index of the candidate object based on the account characteristics of the account to be recommended, the current timeliness parameters of the candidate object, and the object characteristics of the candidate object. The recommendation index represents the probability that the candidate object is successfully recommended to the account to be recommended. The recommendation index of the candidate object is positively correlated with the current timeliness parameters of the candidate object.

[0040] The recommendation module is configured to perform object recommendations to accounts to be recommended based on recommendation metrics for candidate objects.

[0041] In some possible embodiments, the acquisition module is further configured to execute a recommendation request sent in response to the client corresponding to the account to be recommended, determine the current time based on the timestamp information corresponding to the recommendation request; acquire the timeliness data of the candidate object; the timeliness data includes multiple timeliness parameters, the multiple timeliness parameters correspond to multiple different times, and each timeliness parameter in the multiple timeliness parameters represents the probability that the candidate object has recommendation value at the time corresponding to each timeliness parameter; and determine the current timeliness parameter from the multiple timeliness parameters based on the current time.

[0042] In some possible embodiments, the historical time period begins with the upload time of the candidate object; the apparatus further includes:

[0043] The timeliness parameter determination module is configured to execute a timer starting from an initial moment; when the timer reaches a first preset duration, it acquires the interaction data between the candidate object and the recommended account within the first preset duration; the interaction data includes the interaction feature values ​​of the candidate object for each of multiple interaction behaviors; it acquires the timeliness parameter determination model; the timeliness parameter determination model includes a trained basic failure parameter determination model and the weight coefficients corresponding to each interaction behavior; the basic failure parameter determination model is used to determine the basic probability that the candidate object has recommendation value at each time point in multiple different moments and loses recommendation value within an instant after each time point; based on the trained basic failure parameter determination model, the weight coefficients corresponding to each interaction behavior, and the interaction feature values ​​of the candidate object for each interaction behavior, it determines the timeliness parameter of the candidate object at each different time point.

[0044] In some possible embodiments, the current timeliness parameter is determined based on the timeliness parameter determination model; the device further includes a training module for the timeliness parameter determination model, configured to perform the following: acquiring historical interaction data of each historical object among multiple historical objects; the historical interaction data of each historical object includes first interaction data within a first preset duration starting from the upload time of each historical object and second interaction data within a second preset duration starting from an intermediate moment; the intermediate moment is the moment when the first preset duration ends; determining the tag data of each historical object based on the second interaction data of each historical object; the tag data of each historical object represents the historical moment when each historical object lost its recommendation value; acquiring an initial timeliness parameter determination model; the initial timeliness parameter determination model includes a basic failure parameter determination model to be trained and initial weight coefficients corresponding to each interaction behavior; iteratively updating the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior using the tag data of each historical object and the first interaction data of each historical object; until a preset end-of-training condition is met, a trained timeliness parameter determination model is obtained.

[0045] In some possible embodiments, the second interaction data for each historical object includes the number of exposures of each historical object at each moment within a second preset duration and the interaction feature value of each historical object at each moment with respect to each interaction behavior;

[0046] The training module of the timeliness parameter determination model is also configured to perform the following actions: determine the recommendation ranking percentage of each historical object among multiple historical objects at each time step based on the interaction feature value of each historical object with respect to each interaction behavior at each time step; at each time step, if the number of exposures of each historical object at that time step is greater than or equal to a preset number, determine the recommendation degree value of each historical object at that time step based on the recommendation ranking percentage of each historical object at that time step; accumulate the recommendation degree values ​​of each historical object at each time step to obtain the cumulative recommendation degree value of each historical object at each time step; for each historical object, when the cumulative recommendation degree value is less than or equal to a preset recommendation degree value, determine the time step corresponding to the cumulative recommendation degree value as the historical time step in which the historical object loses its recommendation value.

[0047] In some possible embodiments, the first interaction data for each historical object includes interaction feature values ​​for each historical object with respect to each interaction behavior;

[0048] The training module of the timeliness parameter determination model is also configured to perform the following: based on the failure parameter determination model to be trained, the initial weight coefficients corresponding to each interaction behavior, and the interaction feature values ​​of each historical object with respect to each interaction behavior, determine the predicted timeliness parameter of each historical object at each time step; the predicted timeliness parameter of each historical object at each time step represents the predicted probability that each historical object has recommendation value at each time step; based on the historical time when each historical object loses recommendation value and the predicted timeliness parameter of each historical object at each time step, the maximum likelihood estimation method is used to iteratively update the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior to obtain the trained basic failure parameter determination model and the weight coefficients corresponding to each interaction behavior.

[0049] In some possible embodiments, the determining module is configured to perform the following actions: determining a correlation index between the account to be recommended and the candidate object based on the account characteristics of the account to be recommended and the object characteristics of the candidate object; adjusting the correlation index using the current timeliness parameters of the candidate object to obtain the recommendation index of the candidate object.

[0050] According to a third aspect of the present disclosure, an electronic device is provided, comprising:

[0051] processor;

[0052] Memory used to store processor-executable instructions;

[0053] The processor is configured to execute instructions to implement the object recommendation method of the first aspect of the present disclosure.

[0054] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided that, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the electronic device to perform the object recommendation method of the first aspect of the present disclosure.

[0055] According to a fifth aspect of the present disclosure, a computer program product is provided, the computer program product including a computer program stored in a readable storage medium, wherein at least one processor of a computer device reads from the readable storage medium and executes the computer program, causing the computer device to perform the object recommendation method of the first aspect of the present disclosure.

[0056] The technical solutions provided by the embodiments of this disclosure have at least the following beneficial effects:

[0057] During the object recommendation process, the server obtains the current timeliness parameters of candidate objects; it comprehensively considers the account characteristics of the account to be recommended, the current timeliness parameters of different candidate objects, and the object characteristics of different candidate objects to determine the recommendation index for each candidate object. Since the current timeliness parameters of different candidate objects reflect the probability that different candidate objects have recommendation value at the current moment, the higher the probability of having recommendation value at the current moment, the higher the recommendation index. In this way, the impact of the object's own timeliness on the recommendation results can be enhanced, the exposure probability of candidate objects with a higher probability of having recommendation value at the current moment can be increased, and the final recommended objects can be guaranteed to have high recommendation value. Furthermore, this can improve the problem of unfair exposure of objects caused by recommendation methods that are biased by popularity or overly dependent on user preferences.

[0058] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0059] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure, and are not intended to unduly limit this disclosure.

[0060] Figure 1 This is a schematic diagram illustrating an application environment according to an exemplary embodiment;

[0061] Figure 2 This is a flowchart illustrating an object recommendation method according to an exemplary embodiment;

[0062] Figure 3 This is a flowchart illustrating an exemplary embodiment for obtaining the current timeliness parameters of a candidate object;

[0063] Figure 4This is a flowchart illustrating the determination of timeliness parameters at different times, according to an exemplary embodiment;

[0064] Figure 5 This is a schematic diagram illustrating the changing trend of a timeliness parameter according to an exemplary embodiment;

[0065] Figure 6 This is a flowchart illustrating a method for determining model training based on a timeliness parameter according to an exemplary embodiment;

[0066] Figure 7 This is a schematic diagram illustrating a training process according to an exemplary embodiment;

[0067] Figure 8 This is a flowchart illustrating a method for determining tag data for each historical object according to an exemplary embodiment;

[0068] Figure 9 This is a schematic diagram illustrating a trend of cumulative recommendation level values ​​according to an exemplary embodiment;

[0069] Figure 10 This is a flowchart illustrating an iterative update according to an exemplary embodiment;

[0070] Figure 11 This is a schematic diagram illustrating a prediction process according to an exemplary embodiment;

[0071] Figure 12 This is a schematic diagram illustrating a recommendation process according to an exemplary embodiment;

[0072] Figure 13 This is a flowchart illustrating a recommendation metric for determining candidate objects according to an exemplary embodiment;

[0073] Figure 14 This is a block diagram illustrating an object recommendation device according to an exemplary embodiment;

[0074] Figure 15 This is a block diagram illustrating an electronic device for object recommendation according to an exemplary embodiment. Detailed Implementation

[0075] To enable those skilled in the art to better understand the technical solutions of this disclosure, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings.

[0076] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar first objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0077] It should be noted that the user information involved in this disclosure (including but not limited to user device information, user personal information, etc.) is all information authorized by the user or fully authorized by all parties.

[0078] An ideal recommender system should not only satisfy users' diverse preferences but also provide fair exposure opportunities for recommended items. Currently, most recommender systems focus primarily on modeling user preferences to generate satisfactory recommendations, neglecting the object exposure mechanism and ignoring the issue of object fairness.

[0079] A lack of a reasonable object exposure mechanism can lead to serious problems in real-world scenarios. For example, due to issues such as popularity bias, a small number of objects often receive the vast majority of exposure resources in the system, resulting in phenomena such as the long tail and the Matthew effect. Unfair exposure among different objects negatively impacts the dissemination of new objects in the recommendation system, further affecting the quality of recommendations and ultimately harming user satisfaction and creator enthusiasm.

[0080] Based on this, the present disclosure provides an object recommendation method that, based on the timeliness of objects, provides fair exposure opportunities for different objects, achieves reasonable allocation of recommendation resources, and thus improves recommendation performance.

[0081] Please see Figure 1 , Figure 1 A schematic diagram of an application environment is shown according to an exemplary embodiment, such as... Figure 1 As shown, it includes server 01 and terminal device 02. Optionally, server 01 and terminal device 02 can be connected via a wireless link or a wired link, which is not limited herein.

[0082] Reference Figure 1Server 01 can provide background services to terminal device 02, such as object recommendation services. When providing object recommendation services to terminal device 02, server 01 can obtain multiple candidate objects. For each candidate object, server 01 can comprehensively determine the recommendation index of the candidate object based on the account characteristics of the account to be recommended corresponding to terminal device 02, the object characteristics of the candidate object, and the current timeliness parameters of the candidate object. Then, server 01 can push candidate objects that meet certain requirements to terminal device 02 based on the recommendation index of each candidate object.

[0083] In some possible embodiments, server 01 may be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The operating system running on the server may include, but is not limited to, Android, iOS, Linux, Windows, Unix, etc.

[0084] In some possible embodiments, the terminal device 02 described above may include, but is not limited to, smartphones, desktop computers, tablets, laptops, smart speakers, digital assistants, augmented reality (AR) / virtual reality (VR) devices, smart wearable devices, and other types of clients. It may also be software running on the aforementioned client, such as applications or mini-programs. Optionally, the operating system running on the client may include, but is not limited to, Android, iOS, Linux, Windows, Unix, etc.

[0085] In addition, it should be noted that, Figure 1 The application environment shown is merely an example. In practical applications, the object recommendation method of this disclosure embodiment can be executed by a terminal device and a server in cooperation, or it can be executed independently by a terminal device or a server. This disclosure embodiment does not limit the specific application environment.

[0086] Figure 2 This is a flowchart illustrating an object recommendation method according to an exemplary embodiment, such as... Figure 2 As shown, the object recommendation method can be applied to the server and includes the following steps:

[0087] In step S201, the current timeliness parameter of the candidate object is obtained; the current timeliness parameter represents the probability that the candidate object has recommendation value at the current moment; the current timeliness parameter is determined based on the interaction data of the candidate object in the historical time period and the current basic failure parameter; the current basic failure parameter represents the basic probability that the candidate object has recommendation value at the current moment but loses recommendation value in the instant after the current moment.

[0088] In this embodiment, candidate objects refer to objects that the server intends to recommend to terminal devices. These objects are content that matches the actual application scenario. In some possible application scenarios, objects may include music, short videos, products, news, advertisements, etc. The server can respond to an object recommendation request from a terminal device, obtain multiple candidate objects, and then perform a series of operations such as filtering and sorting on these candidate objects before pushing the qualified objects to the terminal device. Optionally, the server can pre-collect multiple candidate objects, and when it receives an object recommendation request from a terminal device, it can immediately perform filtering and sorting operations on these multiple candidate objects. Alternatively, after receiving an object recommendation request from a terminal device, the server can use the currently logged-in account of the terminal device as the account to be recommended. The server determines multiple candidate objects based on the account to be recommended, and then performs filtering and sorting operations on these multiple candidate objects, so that the candidate objects corresponding to different terminal devices are not entirely the same. The former solution has a fast server response speed, while the latter solution allows the server to provide more personalized recommendation services to different terminal devices.

[0089] As mentioned earlier, in the process of recommending objects to terminal devices, servers generally need to perform operations such as filtering and sorting multiple candidate objects. In related technologies, the long-term reliance on the preferences of the accounts to be recommended for filtering and sorting leads to various bias problems in the recommendation results, causing phenomena such as the Matthew effect. Specifically, newly uploaded objects cannot obtain sufficient exposure opportunities, which is very detrimental to the timely dissemination of current events; popular objects are generally uploaded a long time ago and are frequently recommended due to their popularity, preventing the public from accessing new content in a timely manner. Therefore, the unfair exposure between different objects can easily lead to a decline in user experience, a decrease in creators' enthusiasm for creation, and a loss of recommendation diversity.

[0090] Based on this, in this embodiment, when the server recommends objects to the terminal device, it first obtains multiple candidate objects. Simultaneously, it also obtains the current timeliness parameter for each candidate object, which represents the probability that the candidate object has recommendation value at the current moment. The current moment can be the time of acquisition, the time of recommendation, or the time of responding to the object recommendation request from the terminal device. It should be noted that at any given time, an object either has recommendation value or does not. Objects with a high probability of having recommendation value have higher timeliness; for example, current events objects generally have higher timeliness and thus higher recommendation value. Conversely, objects with a low probability of having recommendation value or no recommendation value have lower timeliness and thus lower recommendation value. Therefore, in subsequent processing, the server can comprehensively consider the current timeliness parameter of each candidate object to increase the exposure probability of candidate objects with a high probability of having recommendation value at the current moment, ensuring that all objects ultimately pushed to the terminal device have high recommendation value. Here, recommendation value refers to value measured from the perspectives of user satisfaction, usability, etc.

[0091] Generally, among current affairs objects, those uploaded closer to the current moment are more timely and have a higher probability of being recommended at the current moment. Thus, newly uploaded current affairs objects can obtain a higher recommendation probability due to their higher current timeliness parameter, which can indirectly reduce the recommendation probability of popular but less timely objects. In this way, this disclosure starts from the timeliness of objects to ensure that highly timely objects can be recommended in a timely manner, solve the problem of unfair exposure between different objects, shorten the cold start time of current affairs objects, and avoid pushing expired current affairs objects to users.

[0092] The meaning of "timeliness parameters" will be explained in more detail below.

[0093] First, the current timeliness parameter of a candidate object refers to its timeliness parameter at the current moment. The timeliness parameter of each candidate object changes over time; it differs at different current moments or recommendation moments. In other words, the current timeliness parameter of each candidate object is dynamically changing. Generally, the trend of timeliness parameter changes over time differs for different types of objects, while the trend is basically the same for objects of the same type. For example, current affairs objects show a significant decrease in their display value over time, quickly losing their recommendation value after hot topics evolve or new information emerges. Therefore, the timeliness parameter of current affairs objects decreases as time increases. On the other hand, book-related topics, such as sharing content from classic works, do not significantly decrease in recommendation value with increasing upload time and are largely unaffected by time.

[0094] Considering that there are still significant individual differences among objects of the same type, in this embodiment of the disclosure, the server predetermines the timeliness parameters of each object at different times; thus, when receiving an object recommendation request from a terminal device, the server can directly call to obtain the timeliness parameters of the candidate objects at the current time.

[0095] Specifically, in some possible embodiments, the above step S201 of obtaining the current timeliness parameters of the candidate object may include, for example: Figure 3 The following steps are shown:

[0096] In step S301, in response to the recommendation request sent by the client corresponding to the account to be recommended, the current time is determined based on the timestamp information corresponding to the recommendation request.

[0097] In this context, the client corresponding to the account to be recommended is the terminal device. The timestamp information corresponding to the recommendation request can include the time when the server received the recommendation request or the time when the client sent the recommendation request, which is carried in the recommendation request. Therefore, in this step, after receiving the recommendation request, the server can determine the current time as the time when it received the recommendation request or the time when the client sent the recommendation request.

[0098] In step S303, the timeliness data of the candidate object is obtained; the timeliness data includes multiple timeliness parameters, which correspond to multiple different times. Each timeliness parameter represents the probability that the candidate object has recommendation value at the time corresponding to each timeliness parameter.

[0099] Here, there are multiple candidate objects; as mentioned above, the multiple candidate objects may be predetermined by the server before receiving the recommendation request, or the server may determine them based on the account to be recommended corresponding to the recommendation request; the specific determination method can refer to the prior art, and this disclosure does not limit it.

[0100] In this step, the server obtains the timeliness data for each candidate object. This timeliness data includes multiple timeliness parameters corresponding to multiple different times. Each timeliness parameter represents the probability that the candidate object has recommendation value at the corresponding time. The timeliness data is pre-calculated by the server; the specific calculation time and method will be explained in detail below and will not be elaborated here. The multiple different times are determined based on the calculation time and include every moment within a future time period after the calculation time. Specifically, the time granularity can be minutes, hours, days, weeks, etc. For example, taking an hourly time granularity, if the multiple different times include every hour within the next 7 days, then the timeliness data includes the timeliness parameters corresponding to each hour within 7*24 hours. Similarly, taking a daily time granularity, if the multiple different times include every day within the next 7 days, then the timeliness data includes the timeliness parameters corresponding to each day within those 7 days.

[0101] In step S305, the current timeliness parameter is determined from multiple timeliness parameters based on the current time.

[0102] In this step, the server determines the time that matches the current time from multiple different times based on the current time, and then sets the timeliness parameter corresponding to that time as the current timeliness parameter.

[0103] For example, if the server calculates the timeliness data for candidate A at 8:00 AM on March 1, 2022, including the timeliness parameters for each hour between 9:00 AM on March 1, 2022 and 9:00 AM on March 8, 2022; and the server receives a recommendation request at 12:30 PM on March 3, 2022, then the server can determine that the current time is 12:00 PM on March 3, 2022, and then find the timeliness parameter corresponding to 12:00 PM on March 3, 2022 from the pre-calculated timeliness data and determine it as the current timeliness parameter.

[0104] In the above embodiments, the server pre-calculates the timeliness parameters of each candidate object at different times. During the object recommendation process, the corresponding timeliness parameters can be matched according to the actual recommendation time. In this way, the calculation of a large number of parameters is processed offline, which can save the time of online real-time calculation, as well as save storage and computing resources, which is conducive to speeding up the processing of subsequent steps and improving the overall recommendation efficiency.

[0105] The following section will further explain the specific meaning of "timeliness parameters".

[0106] In this embodiment, based on the relevant knowledge of survival analysis, a current timeliness parameter of the candidate object is defined. The current timeliness parameter represents the probability that the candidate object has recommendation value at the current moment. Survival analysis is mainly used to analyze and infer the time of occurrence of a given event. When applied to the recommendation scenario, the given event can be understood as the object changing from having recommendation value to losing recommendation value. In this embodiment, the current timeliness parameter can also be understood as the probability that the object changes from having recommendation value to losing recommendation value at a certain moment. Therefore, the current timeliness parameter of the object can be expressed by formula (1):

[0107] GRV i (t)=P(T i >t) (1)

[0108] Among them, GRV i (t) represents the time-dependent parameter of object i at time t; T i This indicates the moment when object i changes from having a recommendation value to losing its recommendation value. The specific calculation method will be explained in detail below.

[0109] Therefore, combined with equation (1), the current validity parameter of an object can be understood as the probability that the object has not lost its recommendation value at the current moment.

[0110] Before explaining how to calculate the time-related parameters of an object, let's first introduce some basic concepts in survival analysis:

[0111] Hazard function: h i Let (t) represent the risk that object i survives at time t and dies instantaneously within an instant after time t in the time interval t+δ(t). It can be expressed by a limit, as shown in equation (2):

[0112]

[0113] When applied to the recommendation scenario disclosed herein, it can be understood as the probability that object i has recommendation value at time t and loses recommendation value within an instant t+δ(t) after time t.

[0114] Through equivalent derivation, the risk function h i (t) and the time-dependent parameter GRV i The relationship (t) is as follows: (3)

[0115]

[0116] Cumulative hazard function: H i (t), by h i (t) is derived from the integral, representing the cumulative mortality risk, which is related to the time-dependent parameter GRV. i The equivalence relations of (t) are as follows: (4) and (5):

[0117]

[0118] GRV i (t)=exp(-H i (t)) (5)

[0119] It can be seen that as long as the h of object i at each time step is determined... i (t) can be used to obtain the GRV of object i at future time t. i (t). Therefore, in this disclosure, the GRV of the computational object i can be calculated. i (t) is transformed into the computational object i at each time step h. i (t);

[0120] Furthermore, consider P(T) i>t) is influenced by many factors; in recommendation scenarios, the probability that an object has not lost its recommendation value at time t is related to various attributes of the object. Meanwhile, to determine the h of object i at each time step... i (t), in this embodiment of the disclosure, a Cox proportional hazards regression model is used to analyze the relationship between various attributes of the object and h. i The relationship between (t) is shown in the formula (6) of the Cox proportional hazards regression model.

[0121]

[0122] Among them, i m This represents the value of the m-th attribute of object i; Represents the average attribute value; ∑ i 1 indicates the number of attributes; α m Let represent the weight coefficient of the m-th attribute; h0(t) is the basic risk function for the object group, used to determine the basic failure parameter, i.e., the basic probability that an object has recommendation value at time t but loses recommendation value in the instant after time t; h0(t) is the same for any object, therefore the difference between different objects lies only in The difference. Among them, α m h0(t) needs to be obtained through training.

[0123] It should be noted that in equation (6)... The whole as a covariate, h i (t) is the dependent variable; in other embodiments, as shown in equation (7) below, i can be used. m to replace with i m h is a covariate i (t) is the dependent variable:

[0124] h i (t)=h0(t)*exp(∑ m α m (i m (7)

[0125] The advantage of equation (6) over equation (7) is that it performs mean-normalization on each attribute value of the object, that is, subtracts the average attribute value from each attribute value. This is beneficial for α during the parameter training process. m Fast fitting of h0(t).

[0126] Based on the above concept of survival analysis, it can be seen that the timeliness parameters of an object are affected by multiple attributes of the object. Therefore, the timeliness parameters of the object at different times can be predicted by using multiple attributes of the object. In some possible embodiments, the multiple attributes of the object can be determined based on the interaction data of the candidate object in the historical time period. The multiple attributes of the object can include the interaction feature value of at least one interaction behavior, such as completion rate, click rate, like rate, etc. Then, combined with the pre-trained basic risk function and the weight coefficient corresponding to each interaction behavior, the h of object i at multiple different times is determined according to the above formula (6). i (t), and then according to the above formulas (4) and (5), the time-dependent parameters GRV of object i at multiple different times are obtained. i (t).

[0127] The above embodiments describe the calculation method of the timeliness parameter. Since the timeliness parameter needs to be determined based on the interaction data of the object within a historical time period, the calculation time of the timeliness parameter must be after the historical time period following the upload of the object. Accordingly, in some possible embodiments, the historical time period takes the upload time of the candidate object as the initial time. The object recommendation method of this disclosure also includes, for example, Figure 4 The following steps are shown:

[0128] In step S401, timing begins from the initial moment.

[0129] In step S403, when the timer reaches the first preset duration, the interaction data between the candidate object and the recommended account within the first preset duration is obtained.

[0130] In this step, the server starts timing for each candidate object, taking the upload time of that candidate object as the initial time. When the timing reaches a first preset duration, it acquires the interaction data between the candidate object and the recommended account within that first preset duration. The first preset duration can be determined according to actual needs; for example, it could be half a day, one day, or two days. The interaction data includes the interaction feature values ​​of each of the candidate object's multiple interaction behaviors. These interaction behaviors, as described above, include, but are not limited to, at least one of playback, click, like, follow, and comment. Accordingly, the interaction feature value for playback may include completion rate; for click, click-through rate; for like, like rate; for follow, follow rate; and for comment, comment rate. The specific calculation method for the interaction feature values ​​of each behavior adopts commonly used methods in the art, which will not be elaborated upon here.

[0131] In step S405, the time-sensitivity parameter determination model is obtained; the time-sensitivity parameter determination model includes the trained basic failure parameter determination model and the weight coefficients corresponding to each interaction behavior.

[0132] In this step, the server obtains the timeliness parameter determination model. This model is designed to predict the object's timeliness parameters at different times by utilizing the object's interaction feature values ​​related to each interaction behavior and through a series of mathematical operations. The timeliness parameter determination model requires pre-training of the basic failure parameter determination model and the weight coefficient α corresponding to each interaction behavior. m The basic failure parameter determination model refers to the basic risk function h0(t) mentioned above. The basic risk function h0(t) is the same for different objects and is only related to time. It is used to determine the basic probability that an object has recommendation value at each time point in multiple different times and loses recommendation value in the instant after each time point.

[0133] In step S407, the model, the weight coefficients corresponding to each interaction behavior, and the interaction feature values ​​of the candidate object with respect to each interaction behavior are determined based on the trained basic failure parameters, and the timeliness parameters of the candidate object at each different time in multiple different times are determined.

[0134] During the recommendation process, the server uses steps S405 to S407 to input the interaction feature value of each candidate object with respect to each interaction behavior into the timeliness parameter determination model for each candidate object. The model automatically calls the trained basic failure parameter determination model and the weight coefficient corresponding to each interaction behavior, and performs calculations in sequence according to the above formulas (6), (4) and (5), and finally outputs the timeliness parameter of each candidate object at each different time in multiple different times.

[0135] In a specific example, the interaction data of candidate A (a current affairs category) within one day after its upload is obtained. This data can include candidate A's completion rate, click-through rate, and like rate within that day. The completion rate, click-through rate, and like rate of candidate A within one day are then input into the timeliness parameter model to predict the timeliness parameters of candidate A for each hour within the next seven days. (Refer to...) Figure 5 , Figure 5 The example demonstrates the changing trend of the timeliness parameter of candidate object A for each hour over the next 7 days.

[0136] In the above embodiments, a timeliness parameter determination model is established by combining survival analysis technology. The server collects the interaction data of candidate objects over a period of time, and then uses the interaction data to predict the timeliness parameters of candidate objects at different times in the future. The timeliness parameters of different objects can reflect the probability of different objects having recommendation value at the same time dimension. In this way, it can provide object-level data support for the selection of recommended objects in the subsequent recommendation process, enhance the impact of the object's own timeliness on the recommendation results, and improve the problem of unfair exposure of objects caused by biases such as popularity.

[0137] As mentioned above, the basic failure parameter determination model and the weight coefficients corresponding to each interaction behavior in the timeliness parameter determination model need to be trained in advance. The training method of the timeliness parameter determination model is introduced below.

[0138] In some possible embodiments, the timing parameter determines how the model is trained, and may include, for example... Figure 6 The following steps are shown:

[0139] In step S601, historical interaction data of each historical object among multiple historical objects is obtained.

[0140] During training, the server primarily utilizes historical interaction data corresponding to multiple historical objects to train parameters. Therefore, the server first acquires the historical interaction data for each historical object. This historical interaction data includes first interaction data within a first preset duration starting from the upload time of each historical object, and second interaction data within a second preset duration starting from the midpoint of the upload time; where the midpoint is the end time of the first preset duration. The first preset duration is as described in the embodiment of step S403 above, and the second preset duration can be determined according to actual needs, such as 5 days, 7 days, etc.

[0141] The following is combined with Figure 7 Each step is explained in detail. For example... Figure 7 As shown, Figure 7 The example shows three historical objects: Historical Object B, Historical Object C, and Historical Object D. Taking Historical Object B as an example, the specific content of its historical interaction data is explained. Overall, the historical interaction data of Historical Object B includes: first interaction data within a first preset duration (T1) starting from the upload time of Historical Object B (denoted by t0), and second interaction data within a second preset duration (T2) starting from an intermediate time (denoted by t1); where the intermediate time t1 is the end time of the first preset duration T1, and the end time of the second preset duration T2 is denoted by t2. In terms of specific dimensions, each column represents the interaction characteristic values ​​of an object regarding a certain interaction behavior at different times. For example... Figure 7The four columns can be click-through rate, completion rate, like rate, and follow rate, respectively; each row represents the interaction feature value of an object at a certain moment with respect to different interactive behaviors; thus, the historical interaction data of each historical object can be understood as a matrix of time dimension and interaction behavior dimension.

[0142] In step S603, the tag data of each historical object is determined based on the second interaction data of each historical object.

[0143] Before training the model parameters, it is necessary to determine the labeled data for training. In this step, the server determines the labeled data for each historical object based on the second interaction data of each historical object. The labeled data of each historical object represents the historical moment when each historical object lost its recommendation value. It should be noted that the historical moment when the historical object actually lost its recommendation value must have occurred within the aforementioned second preset time period; otherwise, effective labeled data cannot be obtained, which is not conducive to model training and model parameter convergence.

[0144] Label data can be determined manually or automatically using a specific strategy. The following describes an example of automatic determination using a corresponding strategy.

[0145] In some possible embodiments, the second interaction data for each historical object includes the number of exposures of each historical object at each moment within a second preset duration and the interaction feature value of each historical object at each moment with respect to each interaction behavior;

[0146] Accordingly, the tag data for each historical object determined based on the second interaction data of each historical object may include, for example: Figure 8 The following steps are shown:

[0147] In step S801, based on the interaction feature value of each historical object with respect to each interaction behavior at each time step, the recommendation ranking percentage of each historical object among multiple historical objects at each time step is determined.

[0148] In this step, the server considers the recommended ranking percentage of each historical object at different time dimensions, i.e. at different times. Specifically, for each time moment, the server determines the ranking position of each historical object among multiple historical objects based on the interaction feature value of each interaction behavior of each historical object at that time moment, and then converts it into a ranking percentage.

[0149] Combination Figure 7 This will be explained below. Taking historical object B as an example, the server performs the same steps for other historical objects. For each time point within T2 (between t1 and t2), the server determines the recommended ranking percentage of historical object B at that time, using R... B Let R be represented by (t), B(t)=Rank(F B (t),f I (t)); where F B (t) represents the interaction characteristic value of each interaction behavior of historical object B at time t; f I (t) represents the set of interaction feature values ​​for each interaction behavior of all historical objects at time t; here, it is assumed that there are a total of 8 historical objects. Therefore, the recommendation ranking percentage of historical object B at time t1 is R. B (t1) = 25%.

[0150] In step S803, at each time point, if the number of exposures of each historical object at time point is greater than or equal to a preset number, the recommendation degree value of each historical object at each time point is determined based on the recommendation ranking percentage of each historical object at time point.

[0151] In this step, the server compares the number of exposures of each historical object at a given time with a preset number. If the number of exposures is greater than or equal to the preset number, the recommendation level of each historical object at that time is determined based on its recommendation ranking percentage. The preset number can be set to 1 or set according to actual needs.

[0152] In a specific example, if the number of exposures is greater than or equal to 1 (i.e., the number of exposures is not 0), at each moment, the server can directly determine the recommendation degree value of each historical object at that moment by the percentage of its recommended ranking.

[0153] Alternatively, in another specific example, the recommendation level value can be determined according to the following formula (8):

[0154] v i (t)=1 (Expi(t)＞0) *(R i (t)-β E )+1 (Expi(t)＝0) *(-β nE (8)

[0155] Among them, v i (t) represents the recommendation level of object i at time t; R i (t) represents the percentage of the recommended ranking of object i at time t; Exp i (t) represents the number of exposures of object i at time t; satisfying Exp i When (t)>0, otherwise Similarly, satisfying Exp i When (t) = 0, otherwise βE β nE This is a hyperparameter that can be set according to actual needs. For example, β E and β nE All are set to 0.5;

[0156] In step S805, the recommendation level value of each historical object at each time point is accumulated to obtain the cumulative recommendation level value of each historical object at each time point.

[0157] In this step, the server sums up the recommendation level value of each historical object at each past moment to obtain the cumulative recommendation level value of each historical object at the current moment.

[0158] In step S807, for each historical object, when the cumulative recommendation level value is less than or equal to the preset recommendation level value, the time corresponding to the cumulative recommendation level value is determined as the historical time when the historical object loses its recommendation value.

[0159] In this step, the server monitors the cumulative recommendation level value of each historical object at each time point; optionally, the server can visualize the monitored data to form a representation such as... Figure 9 The cumulative recommendation value trend chart shows that the cumulative recommendation value of different historical objects generally decreases gradually over time. The horizontal dotted line in the chart represents the preset recommendation value, which is the minimum value at which an object can be confirmed as having recommendation value. If the value is lower than the preset recommendation value, the object loses its recommendation value. When the cumulative recommendation value of a historical object touches the dotted line, it means that the cumulative recommendation value equals the preset recommendation value. Subsequent cumulative recommendation values ​​may be lower than the preset recommendation value, meaning the object loses its recommendation value. Therefore, the moment when the cumulative recommendation value touches the dotted line is the historical moment when the historical object loses its recommendation value.

[0160] Specifically, the server can determine the historical moment when a historical object loses its recommendation value based on the following formula (9):

[0161]

[0162] Among them, T i This represents the historical moment when object i lost its recommendation value; Indicates t x The cumulative recommendation level of object i at time point i; β d This represents the preset recommendation level value. Equation (9) can be understood as solving for the minimum time when the cumulative recommendation level value is less than the preset recommendation level value.

[0163] In step S605, the initial time-sensitivity parameter determination model is obtained; the initial time-sensitivity parameter determination model includes the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior.

[0164] In this step, the server obtains the initial timeliness parameter determination model, that is, builds the corresponding mathematical model based on the above equations (4) to (6), and then determines the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior. For example, for Figure 7 The four interactive behaviors—click, complete playback, like, and follow—are each assigned a corresponding initial weight coefficient.

[0165] In step S607, the label data and first interaction data of each historical object are used to iteratively update the basic failure parameters of the model to be trained and the initial weight coefficients corresponding to each interaction behavior.

[0166] In step S609, the training continues until the preset end-of-training conditions are met, and the trained time-effect parameter determination model is obtained.

[0167] In this step, the server uses the initial timeliness parameters to determine the model, analyzes the first interaction data of each historical object, and predicts the timeliness parameters of each historical object at each moment within a second preset time period. Then, based on the label data of each historical object and the predicted timeliness parameters of each historical object, the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior are iteratively updated until the preset end-of-training conditions are met, and the trained timeliness parameter determination model is obtained. The preset end-of-training conditions include the fitting of the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior.

[0168] The iterative update process will be explained in detail below.

[0169] In some possible embodiments, the first interaction data for each historical object includes the interaction feature value of each historical object with respect to each interaction behavior; the above-mentioned iterative update of the model for determining the basic failure parameters to be trained and the initial weight coefficients corresponding to each interaction behavior using the label data and the first interaction data of each historical object may include, for example: Figure 10 The following steps are shown:

[0170] In step S1001, the model, the initial weight coefficients corresponding to each interaction behavior, and the interaction feature values ​​of each historical object with respect to each interaction behavior are determined based on the basic failure parameters to be trained, and the prediction timeliness parameters of each historical object at each time step are determined.

[0171] In this step, for each historical object, the server inputs the interaction feature values ​​of the historical object at each moment within the first preset time period into the initial time-sensitivity parameter to determine the model, determines the model and the initial weight coefficient corresponding to each interaction behavior based on the basic failure parameters to be trained, and predicts the probability that the historical object has recommendation value at each moment within the second preset time period, thus obtaining the predicted time-sensitivity parameter of the historical object at each moment within the second preset time period.

[0172] like Figure 11 As shown, the interaction feature values ​​of historical object B at each moment within the first preset time period T1, i.e., from t0 to t1-1, are input into the initial timeliness parameters to determine the model. Based on the basic failure parameters to be trained, the model and the initial weight coefficients corresponding to each interaction behavior are determined. The probability that historical object B has recommendation value at each moment within the second preset time period T2 is predicted. Here, taking each moment of the second preset time period as an example of each hour within 7 days, the predicted timeliness parameters of historical object B for each hour within the next 7 days after t1-1 can be obtained.

[0173] Accordingly, in the actual recommendation process, the server uses pre-trained basic failure parameters to determine the model and the weight coefficients corresponding to each interaction behavior, and performs the following operations on the interaction feature values ​​of candidate object 1, candidate object 2, and candidate object 3 within a first preset time period. Figure 11 The corresponding process ultimately yields the timeliness parameters of candidate object 1, candidate object 2, and candidate object 3 at each moment within the second preset time period; the server can present the prediction results in a visual form, for example, Figure 11 The top right corner shows the trend of the timeliness parameters for each hour over the next 7 days for candidate object 1, candidate object 2, and candidate object 3.

[0174] In step S1003, based on the historical moment when each historical object lost its recommendation value and the predicted timeliness parameters of each historical object at each moment, the maximum likelihood estimation method is used to iteratively update the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior, so as to obtain the trained basic failure parameter determination model and the weight coefficients corresponding to each interaction behavior.

[0175] In this step, the server uses the maximum likelihood estimation method to iteratively update the basic risk function to be trained and the initial weight coefficients corresponding to each interaction behavior based on the label data (i.e., the historical time when each historical object lost its recommendation value) and the prediction data (i.e., the predicted timeliness parameters of each historical object at each time point), until the fitted basic risk function and the weight coefficients corresponding to each interaction behavior are obtained. The specific iterative process is described in accordance with commonly used iterative methods in this field, and will not be elaborated further in this disclosure.

[0176] The above embodiments describe the training method of the timeliness parameter determination model, the label data required for training, and the data flow process during training. The timeliness parameter determination model is built based on survival analysis and Cox proportional hazards regression model. The server uses historical interaction data of historical objects to fit the unknown parameters in the timeliness parameter determination model. After parameter fitting, a trained timeliness parameter determination model is obtained, which is subsequently used to determine the timeliness parameters of candidate objects. Unlike the definition of object timeliness in related technologies, the embodiments of this disclosure innovatively introduce survival analysis technology and Cox proportional hazards regression model, which can simultaneously analyze the influence of multiple different factors on the timeliness parameters of candidate objects, and can effectively improve the accuracy of timeliness parameters.

[0177] In step S203, the recommendation index of the candidate object is determined based on the account characteristics of the account to be recommended, the current timeliness parameters of the candidate object, and the object characteristics of the candidate object.

[0178] In step S205, based on the recommendation metrics of the candidate objects, object recommendations are made to the accounts to be recommended.

[0179] In this embodiment of the disclosure, the server can comprehensively determine the recommendation index of the candidate object based on the account characteristics of the account to be recommended, the current validity parameters of the candidate object, and the object characteristics of the candidate object. The recommendation index represents the probability that the candidate object is successfully recommended to the account to be recommended, and the recommendation index of the candidate object is positively correlated with the current validity parameters of the candidate object. Finally, based on the recommendation index of each candidate object among multiple candidate objects, the object is recommended to the account to be recommended.

[0180] like Figure 12 As shown, in a specific application scenario, the server responds to the recommendation request of the account to be recommended, obtains multiple candidate objects, including candidate object 1, candidate object 2, candidate object 3, etc., and also obtains the time-sensitive data corresponding to each candidate object; then, based on the current time, it determines the current time-sensitive parameters of each candidate object from their respective time-sensitive data; then, the server performs fusion processing on the account characteristics of the account to be recommended, the current time-sensitive parameters of the candidate objects, and the object characteristics of the candidate objects to determine the recommendation index of each candidate object.

[0181] Optionally, the server sorts multiple candidate objects in descending order of recommendation metrics, and pushes the top N candidate objects to the client corresponding to the account to be recommended; where N is an integer greater than or equal to 1. The recommendation metrics of a candidate object are positively correlated with its current timeliness parameter; therefore, the higher the current timeliness parameter of a candidate object, the higher its corresponding recommendation metric, thus increasing the probability of time-sensitive objects being recommended.

[0182] The account features may include the behavioral features and / or attribute features of the account to be recommended. The behavioral features include, but are not limited to, the type of object the user prefers and the interactive behavior the user prefers. The attribute features include, but are not limited to, the user's gender, age, and region. The object features of the candidate object are determined based on the content actually displayed by the object and can be determined using commonly used object feature extraction algorithms in this field. This disclosure will not elaborate further.

[0183] In some possible embodiments, the above-described determination of recommendation metrics for candidate objects based on the account characteristics of the account to be recommended, the current validity parameters of the candidate objects, and the object characteristics of the candidate objects may include, for example: Figure 13 The following steps are shown:

[0184] In step S1301, the correlation index between the account to be recommended and the candidate object is determined based on the account characteristics of the account to be recommended and the object characteristics of the candidate object.

[0185] In this step, the server can determine a relevance index between the account to be recommended and the candidate object based on the account characteristics of the account to be recommended and the object characteristics of the candidate object. The relevance index represents the degree of interest of the account to be recommended in the candidate object. Optionally, the server can use a feature fusion algorithm commonly used in the field, such as dot product operation, to perform feature fusion processing on the account characteristics of the account to be recommended and the object characteristics of the candidate object to obtain the relevance index between the account to be recommended and the candidate object.

[0186] In step S1303, the relevance index is adjusted using the current timeliness parameters of the candidate object to obtain the recommended index of the candidate object.

[0187] In this step, the server uses the current timeliness parameters of the candidate object to adjust the relevance index and obtain the recommended index of the candidate object; optionally, the current timeliness parameters of the candidate object are directly multiplied by the relevance index, and the result of the multiplication is the recommended index of the candidate object.

[0188] In the above embodiments, the server comprehensively considers the account characteristics of the account to be recommended, the current timeliness parameters of the candidate object, and the object characteristics of the candidate object to determine the recommendation index of the candidate object. Compared with related technologies, which directly use the correlation between the account to be recommended and the candidate object to guide the final ranking of multiple candidate objects, this embodiment introduces the current timeliness parameter of the candidate object to adjust the correlation index between the account to be recommended and the candidate object, increasing the impact of the object's own timeliness on the final recommendation result, which can improve the problem of the single recommendation effect caused by over-reliance on user preferences.

[0189] In summary, the object recommendation method of this disclosure, considering the differences in timeliness among different candidate objects, comprehensively considers the account characteristics of the account to be recommended, the current timeliness parameters of different candidate objects, and the object characteristics of different candidate objects to determine the recommendation index of each candidate object. The current timeliness parameters of different candidate objects reflect the probability that different candidate objects have recommendation value at the current moment. Those with a higher probability of having recommendation value at the current moment have a higher recommendation index. Thus, in the object recommendation process, the impact of the object's own timeliness on the recommendation result is enhanced, increasing the exposure probability of candidate objects with a higher probability of having recommendation value at the current moment. This ensures that all objects ultimately pushed to the terminal device have high recommendation value, and can improve the problem of unfair object exposure caused by recommendation methods that are biased by popularity or overly reliant on user preferences.

[0190] Figure 14 This is a block diagram illustrating an object recommendation device according to an exemplary embodiment. (Refer to...) Figure 14 The device includes an acquisition module 1401, a determination module 1402, and a recommendation module 1403;

[0191] The acquisition module 1401 is configured to acquire the current timeliness parameters of the candidate object; the current timeliness parameters represent the probability that the candidate object has recommendation value at the current moment; the current timeliness parameters are determined based on the interaction data of the candidate object in the historical time period and the current basic failure parameters; the current basic failure parameters represent the basic probability that the candidate object has recommendation value at the current moment but loses recommendation value in the instant after the current moment.

[0192] The determination module 1402 is configured to determine the recommendation index of the candidate object based on the account characteristics of the account to be recommended, the current timeliness parameters of the candidate object, and the object characteristics of the candidate object; the recommendation index represents the probability that the candidate object is successfully recommended to the account to be recommended; the recommendation index of the candidate object is positively correlated with the current timeliness parameters of the candidate object;

[0193] Recommendation module 1403 is configured to perform object recommendations to accounts to be recommended based on recommendation metrics for candidate objects.

[0194] In some possible embodiments, the acquisition module 1401 is further configured to execute a recommendation request sent in response to the client corresponding to the account to be recommended, determine the current time based on the timestamp information corresponding to the recommendation request; acquire the timeliness data of the candidate object; the timeliness data includes multiple timeliness parameters, the multiple timeliness parameters correspond to multiple different times, and each timeliness parameter in the multiple timeliness parameters represents the probability that the candidate object has recommendation value at the time corresponding to each timeliness parameter; and determine the current timeliness parameter from the multiple timeliness parameters based on the current time.

[0195] In some possible embodiments, the historical time period begins with the upload time of the candidate object; the apparatus further includes:

[0196] The timeliness parameter determination module is configured to execute a timer starting from an initial moment; when the timer reaches a first preset duration, it acquires the interaction data between the candidate object and the recommended account within the first preset duration; the interaction data includes the interaction feature values ​​of the candidate object for each of multiple interaction behaviors; it acquires the timeliness parameter determination model; the timeliness parameter determination model includes a trained basic failure parameter determination model and the weight coefficients corresponding to each interaction behavior; the basic failure parameter determination model is used to determine the basic probability that the candidate object has recommendation value at each time point in multiple different moments and loses recommendation value within an instant after each time point; based on the trained basic failure parameter determination model, the weight coefficients corresponding to each interaction behavior, and the interaction feature values ​​of the candidate object for each interaction behavior, it determines the timeliness parameter of the candidate object at each different time point.

[0197] In some possible embodiments, the current timeliness parameter is determined based on the timeliness parameter determination model; the device further includes a training module for the timeliness parameter determination model, configured to perform the following: acquiring historical interaction data of each historical object among multiple historical objects; the historical interaction data of each historical object includes first interaction data within a first preset duration starting from the upload time of each historical object and second interaction data within a second preset duration starting from an intermediate moment; the intermediate moment is the moment when the first preset duration ends; determining the tag data of each historical object based on the second interaction data of each historical object; the tag data of each historical object represents the historical moment when each historical object lost its recommendation value; acquiring an initial timeliness parameter determination model; the initial timeliness parameter determination model includes a basic failure parameter determination model to be trained and initial weight coefficients corresponding to each interaction behavior; iteratively updating the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior using the tag data of each historical object and the first interaction data of each historical object; until a preset end-of-training condition is met, a trained timeliness parameter determination model is obtained.

[0198] In some possible embodiments, the second interaction data for each historical object includes the number of exposures of each historical object at each moment within a second preset duration and the interaction feature value of each historical object at each moment with respect to each interaction behavior;

[0199] The training module of the timeliness parameter determination model is also configured to perform the following actions: determine the recommendation ranking percentage of each historical object among multiple historical objects at each time step based on the interaction feature value of each historical object with respect to each interaction behavior at each time step; at each time step, if the number of exposures of each historical object at that time step is greater than or equal to a preset number, determine the recommendation degree value of each historical object at that time step based on the recommendation ranking percentage of each historical object at that time step; accumulate the recommendation degree values ​​of each historical object at each time step to obtain the cumulative recommendation degree value of each historical object at each time step; for each historical object, when the cumulative recommendation degree value is less than or equal to a preset recommendation degree value, determine the time step corresponding to the cumulative recommendation degree value as the historical time step in which the historical object loses its recommendation value.

[0200] In some possible embodiments, the first interaction data for each historical object includes interaction feature values ​​for each historical object with respect to each interaction behavior;

[0201] The training module of the timeliness parameter determination model is also configured to perform the following: based on the failure parameter determination model to be trained, the initial weight coefficients corresponding to each interaction behavior, and the interaction feature values ​​of each historical object with respect to each interaction behavior, determine the predicted timeliness parameter of each historical object at each time step; the predicted timeliness parameter of each historical object at each time step represents the predicted probability that each historical object has recommendation value at each time step; based on the historical time when each historical object loses recommendation value and the predicted timeliness parameter of each historical object at each time step, the maximum likelihood estimation method is used to iteratively update the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior to obtain the trained basic failure parameter determination model and the weight coefficients corresponding to each interaction behavior.

[0202] In some possible embodiments, the determining module 1402 is configured to perform the following: determine the correlation index between the account to be recommended and the candidate object based on the account characteristics of the account to be recommended and the object characteristics of the candidate object; adjust the correlation index using the current timeliness parameters of the candidate object to obtain the recommendation index of the candidate object.

[0203] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0204] Figure 15 This is a block diagram illustrating an electronic device for object recommendation according to an exemplary embodiment. The electronic device may be a terminal, and its internal structure diagram may be as follows: Figure 15As shown, the electronic device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements an object recommendation method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0205] Those skilled in the art will understand that Figure 15 The structure shown is merely a block diagram of a portion of the structure related to the present disclosure and does not constitute a limitation on the electronic device to which the present disclosure is applied. A specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0206] In an exemplary embodiment, an electronic device is also provided, including: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the instructions to implement the object recommendation method as described in the embodiments of this disclosure.

[0207] In an exemplary embodiment, a computer-readable storage medium is also provided, wherein when the instructions in the storage medium are executed by a processor of an electronic device, the electronic device is enabled to perform the object recommendation method of the present disclosure embodiments.

[0208] In an exemplary embodiment, a computer program product including instructions is also provided. The computer program product includes a computer program stored in a readable storage medium. At least one processor of a computer device reads from the readable storage medium and executes the computer program, causing the computer device to perform the object recommendation method of the present disclosure embodiments.

[0209] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0210] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the following claims.

[0211] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. An object recommendation method, characterized in that, include: Get the current expiration parameters of the candidate object; The current timeliness parameter indicates that the candidate object exists at the current moment. The probability of recommendation value; The current timeliness parameter is determined based on the interaction data of the candidate object within a historical time period and the current basic failure parameter; the current basic failure parameter represents the basic probability that the candidate object has recommendation value at the current moment but loses recommendation value within an instant after the current moment. The current timeliness parameter is determined based on the timeliness parameter determination model. The training method of the timeliness parameter determination model includes: acquiring historical interaction data of each historical object from multiple historical objects; the historical interaction data of each historical object includes first interaction data within a first preset duration starting from the upload time of each historical object and second interaction data within a second preset duration starting from an intermediate time; the intermediate time is the time when the first preset duration ends; determining the tag data of each historical object based on the second interaction data of each historical object; the tag data of each historical object represents the historical time when each historical object lost its recommendation value; acquiring an initial timeliness parameter determination model; the initial timeliness parameter determination model includes a basic failure parameter determination model to be trained and an initial weight coefficient corresponding to each interaction behavior; the basic failure parameter determination model is used to determine the basic probability that the candidate object has recommendation value at each time in multiple different times and loses recommendation value in the instant after each time; using the tag data of each historical object and the first interaction data of each historical object, iteratively updating the basic failure parameter determination model to be trained and the initial weight coefficient corresponding to each interaction behavior; until the preset end training condition is met, the trained timeliness parameter determination model is obtained. Based on the account characteristics of the account to be recommended, the current validity parameters of the candidate object, and the object characteristics of the candidate object, a recommendation index for the candidate object is determined; the recommendation index represents the probability that the candidate object is successfully recommended to the account to be recommended; the recommendation index of the candidate object is positively correlated with the current validity parameters of the candidate object. Based on the recommendation metrics of the candidate objects, object recommendations are made to the accounts to be recommended.

2. The object recommendation method according to claim 1, characterized in that, The step of obtaining the current timeliness parameters of the candidate object includes: In response to a recommendation request sent by the client corresponding to the account to be recommended, the current time is determined based on the timestamp information corresponding to the recommendation request; Obtain the timeliness data of the candidate object; the timeliness data includes multiple timeliness parameters, the multiple timeliness parameters correspond to multiple different times, and each of the multiple timeliness parameters represents the probability that the candidate object has recommendation value at the time corresponding to each timeliness parameter; Based on the current time, the current timeliness parameter is determined from the plurality of timeliness parameters.

3. The object recommendation method according to claim 2, characterized in that, The historical time period is initialized with the upload time of the candidate object; the method further includes: The timing begins from the initial moment; When the timer reaches the first preset duration, the interaction data between the candidate object and the recommended account within the first preset duration is acquired; the interaction data includes the interaction feature value of the candidate object for each of the multiple interaction behaviors; A timeliness parameter determination model is obtained; the timeliness parameter determination model includes a trained basic failure parameter determination model and weight coefficients corresponding to each interaction behavior; The model is determined based on the trained basic failure parameters, the weight coefficients corresponding to each interaction behavior, and the... The interaction feature values ​​of the candidate object with respect to each interaction behavior are used to determine the timeliness parameters of the candidate object at each different time.

4. The object recommendation method according to claim 1, characterized in that, The second interaction data for each historical object includes the number of exposures of each historical object at each moment within the second preset duration and the interaction feature value of each historical object with respect to each interaction behavior at each moment. The step of determining the tag data for each historical object based on the second interaction data of each historical object includes: Based on the interaction feature value of each historical object with respect to each interaction behavior at each time, determine the recommendation ranking percentage of each historical object among the multiple historical objects at each time. At each time point, if the number of exposures of each historical object at that time point is greater than or equal to a preset number, the recommendation degree value of each historical object at each time point is determined based on the recommendation ranking percentage of each historical object at that time point; The recommendation score of each historical object at each time point is summed to obtain the cumulative recommendation score of each historical object at each time point. For each historical object, when the cumulative recommendation level value is less than or equal to the preset recommendation level value, the time corresponding to the cumulative recommendation level value is determined as the historical time when the historical object loses its recommendation value.

5. The object recommendation method according to claim 1, characterized in that, The first interaction data for each historical object includes the interaction feature value of each historical object with respect to each interaction behavior; The step of iteratively updating the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior using the label data and the first interaction data of each historical object includes: Based on the basic failure parameters to be trained, the model, the initial weight coefficients corresponding to each interaction behavior, and the interaction feature values ​​of each historical object with respect to each interaction behavior are determined. Then, the prediction timeliness parameter of each historical object at each time step is determined. The prediction timeliness parameter of each historical object at each time step characterizes the prediction timeliness parameter of each historical object at each time step. The predicted probability that the object has recommendation value at each time point; Based on the historical moment when each historical object loses its recommendation value and the predicted timeliness parameters of each historical object at each moment, the maximum likelihood estimation method is used to iteratively update the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior, so as to obtain the trained basic failure parameter determination model and the weight coefficients corresponding to each interaction behavior.

6. The object recommendation method according to any one of claims 1-5, characterized in that, The step of determining the recommendation index of the candidate object based on the account characteristics of the account to be recommended, the current timeliness parameters of the candidate object, and the object characteristics of the candidate object includes: Based on the account characteristics of the account to be recommended and the object characteristics of the candidate object, the account to be recommended and the candidate object are determined. The correlation index between the candidate objects; By adjusting the relevance index using the current timeliness parameters of the candidate objects, the candidate objects are obtained. Recommended indicators.

7. An object recommendation device, characterized in that, include: The acquisition module is configured to retrieve the current validity parameters of the candidate objects; The current timeliness parameter characterizes the The probability that a candidate object has recommendation value at the current moment; the current timeliness parameter is based on the candidate object's historical time... The interaction data within the interval and the current basic failure parameters are determined; the current basic failure parameters characterize the candidate object in The basic probability that a recommendation has value at the current moment but loses its recommendation value within an instant after the current moment; The current timeliness parameter is determined based on the timeliness parameter determination model; the training module of the timeliness parameter determination model is configured to acquire historical interaction data of each historical object among multiple historical objects; the historical interaction data of each historical object includes first interaction data within a first preset duration starting from the upload time of each historical object and second interaction data within a second preset duration starting from an intermediate moment; the intermediate moment is the moment when the first preset duration ends; the tag data of each historical object is determined based on the second interaction data of each historical object; the tag data of each historical object represents the historical moment when each historical object lost its recommendation value; an initial timeliness parameter determination model is acquired; the initial timeliness parameter determination model includes a basic failure parameter determination model to be trained and initial weight coefficients corresponding to each interaction behavior; The basic failure parameter determination model is used to determine the basic probability that the candidate object has recommendation value at each time point in multiple different times and loses recommendation value in the instant after each time point; Using the label data and first interaction data of each historical object, the basic failure parameter determination model to be trained and the initial weight coefficients corresponding to each interaction behavior are iteratively updated until the preset end training condition is met, and the trained time-failure parameter determination model is obtained. The determination module is configured to perform the following actions: determining the recommendation index of the candidate object based on the account characteristics of the account to be recommended, the current validity parameters of the candidate object, and the object characteristics of the candidate object; the recommendation index characterizes the candidate object. The probability that an object is successfully recommended to the account to be recommended; the recommendation metric of the candidate object is positively correlated with the candidate. The object's current expiration parameters; The recommendation module is configured to perform recommendation based on the candidate objects and recommend them to the accounts to be recommended. recommend.

8. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the instructions to implement the object recommendation method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, When the instructions in the computer-readable storage medium are... When the processor of the electronic device executes, it enables the electronic device to perform the object recommendation method as described in any one of claims 1-6.

Images