Training method of prediction model, storage medium and electronic device

Abstract

The application discloses a training method of a prediction model, a storage medium and an electronic device. The method comprises the following steps: obtaining first media resource information of a published first multimedia resource, a first historical actual publishing result sequence of the first multimedia resource, and a first future actual publishing result of the first multimedia resource at a target time; converting the first historical actual publishing result sequence into a first target abnormal publishing result sequence when a first target conversion probability expression is used to convert the first historical actual publishing result sequence into the first target abnormal publishing result sequence; and training a publishing result prediction model to be trained by using the first media resource information, the first target abnormal publishing result sequence and the first future actual publishing result. The application can be applied to an artificial intelligence scene, and the application also relates to a neural network model and other technologies. The application solves the technical problem of low prediction accuracy of a prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computers, and in particular, to a training method of a prediction model, a storage medium and an electronic device. BACKGROUND

[0002] In recent years, the application of share rate prediction is more and more widespread. For example, through share rate prediction, the development or popular trend of content can be better grasped, so as to complete the targeted optimization of the content.

[0003] For share rate prediction, the related technology is often to train a prediction model based on exposed content information. However, if the exposure of the content information itself is insufficient, the reliability will be greatly discounted. The training of the prediction model in the prior art often ignores the content information with insufficient exposure, and thus the trained prediction model loses the prediction ability for the content information with insufficient exposure. Based on this, when facing the content information with insufficient exposure, the trained prediction model often cannot guarantee high prediction accuracy.

[0004] In summary, in the scenario of using a prediction model to predict a share rate, the prediction model in the related technology does not consider the content information with insufficient exposure in the training process, which leads to the technical problem that the prediction accuracy of the trained prediction model is low.

[0005] At present, no effective solution has been proposed for the above problems. SUMMARY

[0006] The embodiments of the present application provide a training method of a prediction model, a storage medium and an electronic device to at least solve the technical problem that the prediction accuracy of the prediction model is low.

[0007] According to an aspect of some embodiments of the present application, a method for training a prediction model is provided. The method comprises: obtaining first media resource information of a published first multimedia resource, a first historical actual publishing result sequence of the first multimedia resource, and a first future actual publishing result of the first multimedia resource at a target time, wherein the first historical actual publishing result sequence comprises a set of actual publishing results of the first multimedia resource within a first time period, and the target time is later than the first time period; obtaining a first target conversion probability randomly generated by a sequence conversion module, wherein the first target conversion probability is used to indicate whether the first historical actual publishing result sequence is converted into a first target abnormal publishing result sequence, and at least part of the publishing results in the first target abnormal publishing result sequence are abnormal publishing results; when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, converting the first historical actual publishing result sequence into the first target abnormal publishing result sequence, and training a to-be-trained publishing result prediction model using the first media resource information, the first target abnormal publishing result sequence, and the first future actual publishing result, wherein the to-be-trained publishing result prediction model is used to generate a first future predicted publishing result of the first multimedia resource at the target time; when the first target conversion probability indicates that the first historical actual publishing result sequence is not converted into the first target abnormal publishing result sequence, training the to-be-trained publishing result prediction model using the first media resource information, the first historical actual publishing result sequence, and the first future actual publishing result.

[0008] According to an aspect of some embodiments of the present application, there is provided another method for training a prediction model, comprising: obtaining target media resource information of a published target multimedia resource, and a target historical actual publishing result sequence of the target multimedia resource, wherein the target historical actual publishing result sequence comprises a set of actual publishing results of the target multimedia resource within a target time period; inputting the target media resource information and the target historical actual publishing result sequence into a target publishing result prediction model to obtain a future predicted publishing result of the target multimedia resource at a target future time; wherein the target publishing result prediction model is a model obtained by training a to-be-trained publishing result prediction model using a target sample set, each sample in the target sample set comprising: media resource information of a published multimedia resource, a future actual publishing result of the multimedia resource at a target time, and one of a historical actual publishing result sequence, a first abnormal publishing result sequence, and a second abnormal publishing result sequence of the multimedia resource; wherein the historical actual publishing result sequence comprises a set of actual publishing results of the multimedia resource within a time period, the target time being later than the time period; the first abnormal publishing result sequence is a sequence converted from the historical actual publishing result sequence, and each publishing result in the first abnormal publishing result sequence is an abnormal publishing result; and the second abnormal publishing result sequence is a sequence converted from the historical actual publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results.

[0009] According to another aspect of the embodiments of the present application, a training device of a prediction model is also provided, comprising: a first obtaining unit, configured to obtain first media resource information of a published first multimedia resource, a first historical actual publishing result sequence of the first multimedia resource, and a first future actual publishing result of the first multimedia resource at a target time, wherein the first historical actual publishing result sequence comprises a set of actual publishing results of the first multimedia resource within a first time period, and the target time is later than the first time period; a second obtaining unit, configured to obtain a first target conversion probability randomly generated by a sequence conversion module, wherein the first target conversion probability is used to indicate whether the first historical actual publishing result sequence is converted into a first target abnormal publishing result sequence, and at least part of the publishing results in the first target abnormal publishing result sequence are abnormal publishing results; a first training unit, configured to convert the first historical actual publishing result sequence into the first target abnormal publishing result sequence when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, and train a publishing result prediction model to be trained by using the first media resource information, the first target abnormal publishing result sequence and the first future actual publishing result, wherein the publishing result prediction model to be trained is used to generate a first future predicted publishing result of the first multimedia resource at the target time; and a second training unit, configured to train the publishing result prediction model to be trained by using the first media resource information, the first historical actual publishing result sequence and the first future actual publishing result when the first target conversion probability indicates that the first historical actual publishing result sequence is not converted into the first target abnormal publishing result sequence.

[0010] As an optional solution, the first training unit comprises: a first conversion module, configured to convert the first historical actual publishing result sequence into the first abnormal publishing result sequence when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence, wherein the first target abnormal publishing result sequence comprises the first abnormal publishing result sequence, and the publishing results in the first abnormal publishing result sequence are all abnormal publishing results; or a second conversion module, configured to convert the first historical actual publishing result sequence into a second abnormal publishing result sequence when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence, wherein the first target abnormal publishing result sequence comprises the second abnormal publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results.

[0011] As an optional solution, the apparatus further comprises: a first determining module, configured to determine that the first target conversion probability indicates not to convert the first historical actual publishing result sequence into the first target abnormal publishing result sequence when the value of the first target conversion probability is located in a first value interval; and a second determining module, configured to determine that the first target conversion probability indicates to convert the first historical actual publishing result sequence into the first abnormal publishing result sequence or to convert the first historical actual publishing result sequence into the second abnormal publishing result sequence when the value of the first target conversion probability is located in a second value interval.

[0012] As an optional solution, the first training unit comprises: a third conversion module, configured to convert the first historical actual publishing result sequence into the first abnormal publishing result sequence when the first target conversion probability indicates to convert the first historical actual publishing result sequence into the first abnormal publishing result sequence, wherein the first target abnormal publishing result sequence comprises the first abnormal publishing result sequence, and each publishing result in the first abnormal publishing result sequence is an abnormal publishing result; and a fourth conversion module, configured to convert the first historical actual publishing result sequence into the second abnormal publishing result sequence when the first target conversion probability indicates to convert the first historical actual publishing result sequence into the second abnormal publishing result sequence, wherein the first target abnormal publishing result sequence comprises the second abnormal publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results.

[0013] As an optional solution, the apparatus further comprises: a third determining module, configured to determine that the first target conversion probability indicates not to convert the first historical actual publishing result sequence into the first target abnormal publishing result sequence when the value of the first target conversion probability is located in a third value interval; a fourth determining module, configured to determine that the first target conversion probability indicates to convert the first historical actual publishing result sequence into the first abnormal publishing result sequence when the value of the first target conversion probability is located in a fourth value interval; and a fifth determining module, configured to determine that the first target conversion probability indicates to convert the first historical actual publishing result sequence into the second abnormal publishing result sequence when the value of the first target conversion probability is located in a fifth value interval.

[0014] As an optional solution, the first conversion module comprises: a first obtaining sub-module, configured to determine a same value for each historical actual publishing result in the first historical actual publishing result sequence, and obtain a first value sequence; and a second obtaining sub-module, configured to multiply the first historical actual publishing result sequence and the data at the same position in the first value sequence, and obtain the first abnormal publishing result sequence.

[0015] As an optional solution, the second conversion module comprises: a third acquisition submodule for randomly determining a value for each historical actual publishing result in the first historical actual publishing result sequence, to obtain a second value sequence; a fourth acquisition submodule for determining a value of a corresponding flag for each historical actual publishing result in the first historical actual publishing result sequence according to each value in the second value sequence and a preset target threshold, to obtain a third value sequence, wherein the value of the flag is used to indicate whether the corresponding publishing result is an abnormal publishing result; and a fifth acquisition submodule for determining the second abnormal publishing result sequence according to the first historical actual publishing result sequence and the third value sequence.

[0016] As an optional solution, the fourth acquisition submodule comprises: a first setting subunit for setting the value of the flag corresponding to the i th historical actual publishing result in the first historical actual publishing result sequence to a first value when the i th value in the second value sequence is greater than the target threshold, wherein the second value sequence comprises N values, 1≤i≤N, and N is a natural number greater than 1; and a second setting subunit for setting the value of the flag corresponding to the i th historical actual publishing result in the first historical actual publishing result sequence to a second value when the i th value in the second value sequence is less than or equal to the target threshold; wherein the first value is used to indicate that the corresponding publishing result is an abnormal publishing result, and the second value is used to indicate that the corresponding publishing result is not an abnormal publishing result, or the first value is used to indicate that the corresponding publishing result is not an abnormal publishing result, and the second value is used to indicate that the corresponding publishing result is an abnormal publishing result.

[0017] As an optional solution, the fifth acquisition submodule comprises a calculation subunit for multiplying the data at the same position in the first historical actual publishing result sequence and the third value sequence to obtain the second abnormal publishing result sequence.

[0018] As an optional solution, the second training unit comprises: a sixth determination module for determining the features of the first multimedia resource in multiple dimensions according to the first media resource information and the first target abnormal publishing result sequence, to obtain a first feature set; a seventh determination module for determining the first future predicted publishing result of the first multimedia resource at the target time according to the first feature set; and an adjustment module for adjusting the parameters in the publishing result prediction model to obtain a first publishing result prediction model to be trained when the value of the target loss function determined according to the first future predicted publishing result and the first future actual publishing result does not satisfy a preset condition.

[0019] As an optional solution, the sixth determining module includes at least two of the following steps: a first determining submodule is configured to determine a text feature of the first multimedia resource according to text information of the first multimedia resource, wherein the first media resource information includes the text information of the first multimedia resource, and the first feature set includes the text feature of the first multimedia resource; a second determining submodule is configured to determine a visual feature of the first multimedia resource according to picture information of the first multimedia resource, wherein the first media resource information includes the picture information of the first multimedia resource, and the first feature set includes the visual feature of the first multimedia resource; a third determining submodule is configured to determine a sequence feature of the first multimedia resource according to the first target abnormal publishing result sequence, wherein the first feature set includes the sequence feature of the first multimedia resource; and a fourth determining submodule is configured to determine an attribute feature of the first multimedia resource according to attribute information of the first multimedia resource, wherein the first media resource information includes the attribute information of the first multimedia resource, and the first feature set includes the attribute feature of the first multimedia resource.

[0020] As an optional solution, the device further comprises: a first obtaining module, configured to obtain second media resource information of a published second multimedia resource, a second historical actual publishing result sequence of the second multimedia resource, and a second future actual publishing result of the second multimedia resource at the target time, wherein the second historical actual publishing result sequence comprises a set of actual publishing results of the second multimedia resource within the first time period; a second obtaining module, configured to obtain a second target conversion probability randomly generated by the sequence conversion module, wherein the second target conversion probability is used to indicate whether the second historical actual publishing result sequence is converted into a second target abnormal publishing result sequence, and at least part of the publishing results in the second target abnormal publishing result sequence are abnormal publishing results; a first training module, configured to, when the second target conversion probability indicates that the second historical actual publishing result sequence is converted into the second target abnormal publishing result sequence, convert the second historical actual publishing result sequence into the second target abnormal publishing result sequence, and train the first publishing result prediction model by using the second media resource information, the second target abnormal publishing result sequence, and the second future actual publishing result, wherein the first publishing result prediction model is used to generate a second future predicted publishing result of the second multimedia resource at the target time; and a second training module, configured to, when the second target conversion probability indicates that the second historical actual publishing result sequence is not converted into the second target abnormal publishing result sequence, train the first publishing result prediction model by using the second media resource information, the second historical actual publishing result sequence, and the second future actual publishing result.

[0021] As an optional solution, the second training unit comprises: a third training module, configured to train a sharing rate prediction model to be trained by using the first media resource information, the first historical actual publishing result sequence, and the first future actual publishing result, wherein the publishing result prediction model is the sharing rate prediction model, the sharing rate prediction model is used to generate a first future predicted sharing rate of the first multimedia resource at the target time according to the first media resource information and the first historical actual publishing result sequence, the first future predicted sharing rate indicates a predicted number of times of sharing of the first multimedia resource divided by a number of times of exposure, and the first future actual publishing result is a first future actual sharing rate of the first multimedia resource at the target time, which indicates an actual number of times of sharing of the first multimedia resource divided by a number of times of exposure.

[0022] According to another aspect of the embodiments of the present application, there is further provided another device for training a prediction model, comprising: a third obtaining unit configured to obtain target media resource information of a published target multimedia resource and a target historical actual publishing result sequence of the target multimedia resource, wherein the target historical actual publishing result sequence comprises a set of actual publishing results of the target multimedia resource within a target time period; and an input unit configured to input the target media resource information and the target historical actual publishing result sequence into a target publishing result prediction model to obtain a future predicted publishing result of the target multimedia resource at a target future time; wherein the target publishing result prediction model is a model obtained by training the publishing result prediction model to be trained using a target sample set, each sample in the target sample set comprising: media resource information of a published multimedia resource, a future actual publishing result of the multimedia resource at a target time, and one of a historical actual publishing result sequence of the multimedia resource, a first abnormal publishing result sequence, and a second abnormal publishing result sequence; wherein the historical actual publishing result sequence comprises a set of actual publishing results of the multimedia resource within a time period, and the target time is later than the time period; the first abnormal publishing result sequence is a sequence converted from the historical actual publishing result sequence, and each publishing result in the first abnormal publishing result sequence is an abnormal publishing result; and the second abnormal publishing result sequence is a sequence converted from the historical actual publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results.

[0023] As an optional solution, the third obtaining unit comprises a third obtaining module configured to obtain target media resource information of a published target multimedia resource and a target historical actual sharing rate sequence of the target multimedia resource, wherein the target historical actual publishing result sequence comprises a set of actual sharing rates of the target multimedia resource within the target time period, and each actual sharing rate represents a number of times that the target multimedia resource is actually shared divided by a number of times that the target multimedia resource is exposed; and the input unit comprises an input module configured to input the target media resource information and the target historical actual sharing rate sequence into the target publishing result prediction model to obtain a future predicted sharing rate of the target multimedia resource at the target future time, wherein the future predicted sharing rate represents a number of times that the target multimedia resource is predicted to be shared divided by a number of times that the target multimedia resource is exposed.

[0024] According to still another aspect of the embodiments of the present application, there is further provided a computer readable storage medium having a computer program stored therein, wherein the computer program is configured to perform the above-mentioned method for training a prediction model when executed.

[0025] According to a further aspect of the embodiments of the present application, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the training method of the prediction model by the computer program.

[0026] In the embodiments of the present application, the result publishing prediction model is trained by using the actually published result sequence, so that the result publishing prediction model has the ability to predict the sharing rate of normal state data; the actually published result sequence is converted into an abnormal result publishing sequence by the target conversion probability randomly generated by the sequence conversion module, and the result publishing prediction model is further trained by using the abnormal result publishing sequence, so that the result publishing prediction model has the ability to predict the sharing rate of abnormal data; based on this, the trained result publishing prediction model has the ability to predict the sharing rate of normal data and the ability to predict the sharing rate of abnormal state data, and thus the result publishing prediction model can maintain high sharing rate prediction accuracy when facing normal data or abnormal data, thereby solving the technical problem of low prediction accuracy of the prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0027] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0028] Figure 1 is a schematic diagram of an application environment of an optional training method of a prediction model according to an embodiment of the present application;

[0029] Figure 2 is a schematic diagram of a flow of an optional training method of a prediction model according to an embodiment of the present application;

[0030] Figure 3 is a schematic diagram of an optional training method of a prediction model according to an embodiment of the present application;

[0031] Figure 4 is a schematic diagram of another optional training method of a prediction model according to an embodiment of the present application;

[0032] Figure 5 is a schematic diagram of another optional training method of a prediction model according to an embodiment of the present application;

[0033] Figure 6 is a schematic diagram of another optional training method of a prediction model according to an embodiment of the present application;

[0034] Figure 7 is a schematic diagram of another optional training method of a prediction model according to an embodiment of the present application;

[0035] Figure 8 is a schematic diagram of another optional training method of a prediction model according to an embodiment of the present application;

[0036] Figure 9 is a schematic diagram of another optional training method of a prediction model according to an embodiment of the present application;

[0037] Figure 10 is a schematic diagram of another optional training method of a prediction model according to an embodiment of the present application;

[0038] Figure 11 is a schematic diagram of another optional training method of a prediction model according to an embodiment of the present application;

[0039] Figure 12 is a schematic diagram of a flow of an optional prediction method of a publishing result according to an embodiment of the present application;

[0040] Figure 13 is a schematic diagram of an optional training device of a prediction model according to an embodiment of the present application;

[0041] Figure 14 is a schematic diagram of an optional prediction device of a publishing result according to an embodiment of the present application;

[0042] Figure 15 is a structural schematic diagram of an optional electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0043] In order to make the personnel in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by the person of ordinary skill in the art without making creative labor should belong to the scope of protection of the present application.

[0044] It is to be understood that the terminology "first", "second" and the like used in the specification and the claims of the application as well as the preceding description of the drawings is merely used to distinguish between similar objects and does not necessarily imply a specific order or chronology. It is to be understood that the data thus distinguished can be interchanged, where appropriate, so that the embodiments of the application described herein can be carried out in a different order than the one illustrated or described herein. Furthermore, the terms "comprise" and "have" and any variations thereof are intended to cover non-exclusive inclusions, for example, processes, methods, systems, products or devices that comprise a list of steps or units do not necessarily have to include only those steps or units expressly listed, but can include other steps or units that are not expressly listed or inherent to such processes, methods, products or devices.

[0045] According to an aspect of the embodiments of the present application, there is provided a method for training a prediction model. Optionally, the method for training a prediction model can be applied to, but is not limited to, the environment as shown in Figure 1 The environment can include, but is not limited to, a user device 102, a network 110 and a server 112. The user device 102 can include, but is not limited to, a display 108, a processor 106 and a memory 104.

[0046] The specific process can include the following steps:

[0047] In step S102, the user device 102 obtains a prediction instruction triggered by a click operation triggered on a sharing rate prediction website. The prediction instruction is used to request a prediction of a sharing rate of an article currently displayed on the sharing rate prediction website in a future period of time (e.g., one month later). The prediction instruction also carries sharing rate data of the article in a historical period of time and basic information of the article (e.g., article name, author name, reading times, like times, image information, text information, etc.).

[0048] In steps S104-S106, the user device 102 sends the prediction instruction to the server 112 through the network 110.

[0049] In step S108, the server 112 responds to the prediction instruction and processes the sharing rate data of the article in the historical period of time and the basic information of the article through a processing engine 116. For example, the sharing rate data and the basic information are input into a pre-trained publication result prediction model to generate a prediction result.

[0050] In steps S110-S112, the server 112 sends the prediction result to the user device 102 through the network 110. The processor 106 in the user device 102 displays the prediction result on the display 108 and stores the prediction result in the memory 104.

[0051] In addition,Figure 1 In addition to the examples shown, the above steps can be completed independently by the user device 102, i.e., the user device 102 performs the steps of processing the sharing rate data of the article in a historical period of time, and processing the basic information of the article, etc., thereby reducing the processing pressure of the server. The user device 102 includes but is not limited to handheld devices (such as mobile phones), notebook computers, desktop computers, vehicle-mounted devices, etc., and the present application does not limit the specific implementation manner of the user device 102.

[0052] Optionally, as an optional implementation manner, as shown in Figure 2 The training method of the prediction model includes:

[0053] S202, obtaining first media resource information of a published first multimedia resource, a first historical actual publishing result sequence of the first multimedia resource, and a first future actual publishing result of the first multimedia resource at a target time, wherein the first historical actual publishing result sequence includes a group of actual publishing results of the first multimedia resource in a first time period, and the target time is later than the first time period;

[0054] S204, obtaining a first target conversion probability randomly generated by a sequence conversion module, wherein the first target conversion probability is used to indicate whether to convert the first historical actual publishing result sequence into a first target abnormal publishing result sequence, and at least part of the publishing results in the first target abnormal publishing result sequence are abnormal publishing results;

[0055] S206, when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, converting the first historical actual publishing result sequence into the first target abnormal publishing result sequence, and using the first media resource information, the first target abnormal publishing result sequence and the first future actual publishing result to train a to-be-trained publishing result prediction model, wherein the to-be-trained publishing result prediction model is used to generate a first future predicted publishing result of the first multimedia resource at the target time;

[0056] S208, when the first target conversion probability indicates that the first historical actual publishing result sequence is not converted into the first target abnormal publishing result sequence, using the first media resource information, the first historical actual publishing result sequence and the first future actual publishing result to train the to-be-trained publishing result prediction model.

[0057] Optionally, in the present embodiment, the above training method of the prediction model can be but is not limited to applied in the scenario of predicting the sharing rate of content information in a future period of time, for example, in order to better grasp the development trend of video content, the staff triggers a prediction operation on the video D (target multimedia resource 302) in the background, for example Figure 3As shown in (a), video D can be understood as a video resource played on certain applications; further, the media resource information 304-1 of the target multimedia resource 302 (such as video title, video description, video publishing platform, video click rate, video likes, etc.) and the historical actual publishing result sequence 304-2 (such as the video's sharing rate, number of shares, and effective return visit rate after sharing within a historical preset time period, etc.) are input into the publishing result prediction model 306 trained by the above prediction model training method to obtain the future predicted publishing result 308 for video D, wherein the future predicted publishing result 308 can be used, but is not limited to, to represent the predicted sharing rate of video D in the future period of time; furthermore, the future predicted publishing result 308 is displayed on the background display screen for staff reference, such as Figure 3 As shown in (c);

[0058] In addition, but not limited to, targeted content optimization solutions can be provided to staff based on the future prediction release result 308. If the sharing rate indicated by the future prediction release result 308 does not reach the expected level, reference information can be generated based on the input media resource information 304-1 and the historical actual release result sequence 304-2. The reference information is used to indicate the factors that affect the sharing rate indicated by the future prediction release result 308 from not reaching the expected level, such as the existence of optimization items in video D itself, or the problem of low user retention rate on the platform on which video D is played.

[0059] Optionally, in this embodiment, media resource information may be used, but is not limited to, to represent basic information of multimedia resources, such as text information, image information, attribute information, etc., and the basic information is usually fixed or remains unchanged over a long period of time.

[0060] The historical actual release result sequence can be used, but is not limited to, to represent multiple sets of information that have changed or remained unchanged within a historical period. For example, if the historical period is divided into 10 time points, the historical actual release result sequence can be, but is not limited to, a set of information composed of the actual release results corresponding to each time point.

[0061] To further illustrate, the optional historical sequence of actual published results is as follows: Figure 4 As shown, assuming the horizontal axis represents time (T) and the vertical axis represents the sharing rate (S), the historical sharing rate sequence 402 can be used, but is not limited to, to represent the sum of the sharing rates corresponding to each time point within a historical time period. Specifically, assuming a historical time period includes n time points, namely T1, T2...Tn, and the sharing rates corresponding to the n time points are S1, S2...Sn respectively, then the historical sharing rate sequence 402 can be, but is not limited to, [S1, S2...Sn], or a numerical sequence can be obtained by calculation based on [S1, S2...Sn], without any limitation here.

[0062] Optionally, in the embodiment, the historical actual publishing result sequence may, but is not limited to, include the normal publishing result sequence and / or the abnormal publishing result sequence, and generally, the probability that the historical actual publishing result sequence includes the normal publishing result sequence is much higher than that the historical actual publishing result sequence includes the abnormal publishing result sequence, and then in the related art, the historical actual publishing result sequence is directly used to train the publishing result prediction model, but the trained publishing result prediction model depends on the normal publishing result sequence, and then when facing the abnormal publishing result sequence, the high prediction accuracy cannot be guaranteed.

[0063] Based on this, in the embodiment, multiple conversion modes are configured, the historical actual publishing result sequence is converted by using the target conversion probability randomly generated by the sequence conversion module, for example, the historical actual publishing result sequence is converted into the abnormal publishing result sequence, and the abnormal publishing result sequence is further used to train the publishing result prediction model, so as to reduce the dependence of the trained publishing result prediction model on the normal publishing result sequence, so as to guarantee the high prediction accuracy when facing the abnormal publishing result sequence.

[0064] Optionally, in the embodiment, when the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, the first media resource information of the first multimedia resource may, but is not limited to, be retained, and then the first media resource information, the first target abnormal publishing result sequence and the first future actual publishing result are used to train the publishing result prediction model to be trained, which can be understood as converting the first historical actual publishing result sequence into the first target abnormal publishing result sequence while retaining the first media resource information of the first multimedia resource, so that the publishing result prediction model pays more attention to the first media resource information than the first historical actual publishing result sequence in the training process, so as to reduce the dependence of the publishing result prediction model on the historical actual publishing result sequence.

[0065] Generally, in the training process of the above publishing result prediction model, the conversion may, but is not limited to, be only the historical actual publishing result sequence, and the media resource information of the multimedia resource and the future actual publishing result at the target time are not changed.

[0066] Optionally, in the embodiment, the relationship between the historical actual publishing result sequence and the future actual publishing result may, but is not limited to, be understood as follows: taking the published first multimedia resource as an example, the first time period and the target time are selected in the time period in which the first multimedia resource is published, and the target time is after the first time period.

[0067] Based on this, since both the first time period and the target time are the times when the first multimedia resource has been published, the actual publication results of the first multimedia resource in the first time period and the target time can be obtained directly, namely the historical actual publication result sequence and the future actual publication result. The time difference between the first time period and the target time can be, but is not limited to, fixed. Assuming that the time difference between the first time period and the target time is the target time difference, the target time difference can be, but is not limited to, used to determine the time of publication result after the publication result prediction model predicts the content information of the historical time. For example, assuming that the target time difference is 1 month, after the trained publication result prediction model predicts the content information published in the [T1~T2] time period, the output publication result is the predicted sharing rate of the content information 1 month after T2.

[0068] Further optionally, in this embodiment, for the training process of the release result prediction model, the historical actual release result sequence and the future actual release result can be, but are not limited to, a correspondence. The specific historical actual release result sequence or the transformed historical actual release result sequence (target abnormal release result sequence) is used as the input of the release result prediction model, while the future actual release result is used as the comparison object for the output of the release result prediction model. By comparing with the comparison object, the loss of the release result prediction model in this training is determined. If the convergence condition is met, the training ends. If the convergence condition is not met, the model parameters in the release result prediction model are adjusted according to the loss, and the next training begins based on the release result prediction model after adjusting the model parameters.

[0069] It should be noted that training the release result prediction model using historical release result sequences enables it to predict the sharing rate of normal data. By randomly generating target conversion probabilities through the sequence transformation module, the historical release result sequences are converted into abnormal release result sequences. Further training the release result prediction model using these abnormal release result sequences enables it to predict the sharing rate of abnormal data. Based on this, the trained release result prediction model possesses the ability to predict the sharing rate of both normal and abnormal data, thus maintaining high accuracy in sharing rate prediction regardless of whether the data is normal or abnormal.

[0070] To further illustrate, optional examples include... Figure 5 As shown, the media resource information 502-1 of the published multimedia resources, the historical actual publication result sequence 502-2 in T1 (first time period), and the actual sharing rate 502-3 (future actual publication result) of the multimedia resources in T2 (target time) are obtained.

[0071] Further, the acquisition sequence conversion module 504 randomly generates a target conversion probability (P), and in the case that P does not satisfy a target condition, converts the historical actual publishing result sequence into a target abnormal publishing result sequence 502-4, and takes the media resource information 502-1 and the target abnormal publishing result sequence 502-4 as current model input to train the publishing result prediction model 506 to be trained; or, in the case that P satisfies the target condition, takes the media resource information 502-1 and the historical actual publishing result sequence 502-2 as current model input to train the publishing result prediction model 506 to be trained.

[0072] Further, the publishing result prediction model 506 uses the media resource information 502-1 and the target abnormal publishing result sequence 502-4, and / or the media resource information 502-1 and the historical actual publishing result sequence 502-2 as current model input to start training, such as outputting a predicted sharing rate 508 corresponding to the model input, and comparing with the real sharing rate 502-3 to obtain a loss 510; in the case that the loss 510 reaches a convergence condition, the training of the publishing result prediction model 506 is ended; in the case that the loss 510 does not reach the convergence condition, the model parameters of the publishing result prediction model 506 are adjusted according to the loss 510, and the next training is continued based on the publishing result prediction model 506 after the model parameters are adjusted.

[0073] According to the embodiments provided in the present application, the first media resource information of the published first multimedia resource, the first historical actual publishing result sequence of the first multimedia resource, and the first future actual publishing result of the first multimedia resource at a target time are obtained, the first historical actual publishing result sequence includes a set of actual publishing results of the first multimedia resource within a first time period, and the target time is later than the first time period; a first target conversion probability randomly generated by a sequence conversion module is obtained, the first target conversion probability is used to indicate whether the first historical actual publishing result sequence is converted into a first target abnormal publishing result sequence, and at least part of the publishing results in the first target abnormal publishing result sequence are abnormal publishing results; when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, and the first media resource information, the first target abnormal publishing result sequence, and the first future actual publishing result are used to train a publishing result prediction model to be trained, wherein the publishing result prediction model to be trained is used to generate a first future predicted publishing result of the first multimedia resource at the target time; when the first target conversion probability indicates that the first historical actual publishing result sequence is not converted into the first target abnormal publishing result sequence, the first media resource information, the first historical actual publishing result sequence, and the first future actual publishing result are used to train the publishing result prediction model to be trained, so as to achieve the purpose that the publishing result prediction model can maintain high sharing rate prediction accuracy when facing normal data or abnormal data, and the effect of improving the prediction accuracy of the prediction model on the sharing rate is realized.

[0074] As an optional solution, when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, converting the first historical actual publishing result sequence into the first target abnormal publishing result sequence includes:

[0075] S1, when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence, wherein the first target abnormal publishing result sequence includes the first abnormal publishing result sequence, and the publishing results in the first abnormal publishing result sequence are all abnormal publishing results; or

[0076] S2, when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence, wherein the first target abnormal publishing result sequence includes the second abnormal publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results.

[0077] Optionally, in the embodiment, the target abnormal publishing result sequence can be but is not limited to divided into multiple types, such as the target abnormal publishing result sequence including a first abnormal publishing result sequence and a second abnormal publishing result sequence.

[0078] Optionally, in the embodiment, the first historical actual publishing result sequence includes a set of actual publishing results of the first multimedia resource in a first time period, and then in the case that the actual publishing result is an actual sharing rate, the actual sharing rate can be divided into multiple states according to uncertainty, and in the sharing rate prediction task, there are mainly two types of uncertainty, including historical sequence uncertainty and attribute uncertainty. The historical sequence uncertainty mainly occurs at the beginning of content publishing, which is generally caused by insufficient content exposure; in addition, the sharing rate of the content can have a large fluctuation during the entire life cycle of the content, which makes it difficult for the prediction model to obtain high credibility information from the historical sharing rate sequence; this uncertainty exists in the life cycle of each content, because each project has a cold start period, and most projects are long-tailed; in addition, the attribute uncertainty is derived from the noise or absence of content-related meta-information.

[0079] For further illustration, for example, the actual sharing rate is divided into a normal state (such as the state of the historical publishing result sequence) and an abnormal state (such as the state of the target abnormal publishing result sequence) according to uncertainty, and the abnormal state can be further divided into a cold start state (such as the state of the first abnormal publishing result sequence) and a noise state (such as the state of the second abnormal publishing result sequence), wherein the normal state can be but is not limited to used to represent that the historical sharing rate data sequence is reliable with sufficient exposure, and such historical sharing rate data can provide strong guidance for the sharing rate prediction model; the cold start state can be but is not limited to used to represent that the entire historical sharing rate data sequence is unreliable or even missing; the noise state can be but is not limited to used to represent that the historical sharing rate data sequence has partial uncertainty, which is usually caused by partial absence or noise, so that part of the data in the historical sequence data is unreliable, and in the actual scene, the multiple state data of the sequence is usually unbalanced, and the data of the normal state is much more than the data of the other two states (the cold start state data and the noise state data).

[0080] It should be noted that the conversion of the first historical actual publishing result sequence into the first abnormal publishing result sequence or the second abnormal publishing result sequence can be realized flexibly by pre-configuration, such as pre-configuring that the first target conversion probability corresponds to the conversion of the first abnormal publishing result sequence, and then when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence.

[0081] Further illustrated, optionally based on Figure 5 As shown in the scenario, continue for example Figure 6 As shown, assuming that the pre-configuration P (the first target conversion probability) corresponds to the conversion of the first abnormal publishing result sequence 702, the first historical actual publishing result sequence 502-2 is converted into the first abnormal publishing result sequence 602 when P meets the first condition.

[0082] Similarly, if the pre-configuration of the first target conversion probability corresponds to the conversion of the second abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence.

[0083] Further illustrated, optionally based on Figure 5 As shown in the scenario, continue for example Figure 7 As shown, assuming that the pre-configuration P (the first target conversion probability) corresponds to the conversion of the second abnormal publishing result sequence 702, the first historical actual publishing result sequence 502-2 is converted into the second abnormal publishing result sequence 702 when P meets the second condition.

[0084] Through the embodiments provided in the present application, when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence, wherein the first target abnormal publishing result sequence includes the first abnormal publishing result sequence, and all publishing results in the first abnormal publishing result sequence are abnormal publishing results; or when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence, wherein the first target abnormal publishing result sequence includes the second abnormal publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results. The purpose of flexibly converting the historical actual publishing result sequence into the first or second abnormal publishing result sequence is achieved, and the effect of improving the conversion flexibility of the historical actual publishing result sequence is achieved.

[0085] As an optional solution, the method further includes:

[0086] S1, when the value of the first target conversion probability is located in the first value interval, it is determined that the first target conversion probability indicates that the first historical actual publishing result sequence is not converted into the first target abnormal publishing result sequence;

[0087] S2, when the value of the first target conversion probability is in the second value interval, it is determined that the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence or the second abnormal publishing result sequence.

[0088] Optionally, in the embodiment, when the value of the first target conversion probability is in the first value interval, P can satisfy the target condition as shown in the following formula (1) but is not limited to the formula (1): Figure 6 Similarly, when the value of the first target conversion probability is in the second value interval, P can satisfy the first condition as shown in the following formula (2) or the second condition as shown in the following formula (3) but is not limited to the formula (2) or the formula (3): Figure 6 Figure 7

[0089] For further illustration, assuming that the first value interval is [1 / 2-1], when the value of the first target conversion probability is in the first value interval, it is determined that the first target conversion probability indicates that the first historical actual publishing result sequence is not converted into the first target abnormal publishing result sequence.

[0090] In addition, the second value interval and the first value interval can be but are not limited to average distribution, and then assuming that the second value interval is [0-1 / 2] on the basis of the first value interval being [1 / 2-1], when the value of the first target conversion probability is in the second value interval, it is determined that the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence or the second abnormal publishing result sequence.

[0091] As an optional solution, when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, including:

[0092] S1, when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence, wherein the first target abnormal publishing result sequence includes the first abnormal publishing result sequence, and the publishing results in the first abnormal publishing result sequence are all abnormal publishing results.

[0093] S2, when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence, wherein the first target abnormal publishing result sequence includes the second abnormal publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results. ​​

[0094] Optionally, in the embodiment, the first historical actual publishing result sequence includes a set of actual publishing results of the first multimedia resource in a first time period, and further in the case that the actual publishing result is actual sharing rate, the actual sharing rate can be divided into multiple states according to uncertainty, and in the sharing rate prediction task, there are mainly two types of uncertainty, including historical sequence uncertainty and attribute uncertainty. The historical sequence uncertainty mainly occurs at the beginning of content publishing, which is generally caused by insufficient content exposure; in addition, the sharing rate of the content may have a large fluctuation during the life cycle of the content, which makes it difficult for the prediction model to obtain high credibility information from the historical sharing rate sequence; this uncertainty exists in the life cycle of each content, because each project has a cold start period, and most projects are long-tailed; in addition, the attribute uncertainty is derived from the noise or absence of content-related meta information;

[0095] For further illustration, for example, the actual sharing rate is divided into normal state (such as the state of the historical publishing result sequence) and abnormal state (such as the state of the target abnormal publishing result sequence) according to uncertainty, and the abnormal state can be further divided into cold start state (such as the state of the first abnormal publishing result sequence) and noise state (such as the state of the second abnormal publishing result sequence), wherein the normal state can but is not limited to represent that the historical sharing rate data sequence is reliable with sufficient exposure, and such historical sharing rate data can provide strong guidance for the sharing rate prediction model; the cold start state can but is not limited to represent that the entire historical sharing rate data sequence is unreliable or even missing; the noise state can but is not limited to represent that the historical sharing rate data sequence has partial uncertainty, which is usually caused by partial absence or noise, so part of the data in the historical sequence data is unreliable, and in actual scenarios, the data of multiple states of the sequence is usually unbalanced, and the data of the normal state is much more than the data of the other two states (cold start state data and noise state data).

[0096] It should be noted that the three cases of retaining the historical actual publishing result sequence (i.e. not converting), converting the historical actual publishing result sequence into the first abnormal publishing result sequence, and converting the historical actual publishing result sequence into the second abnormal publishing result sequence can but are not limited to be regarded as three modes. In the face of the above three modes, one can but is not limited to be executed, two can but are not limited to be executed, or all three can but are not limited to be executed.

[0097] For further illustration by taking one of the three as an example, assuming that converting the historical actual publishing result sequence into the first abnormal publishing result sequence is mode one, retaining the historical actual publishing result sequence is mode two, and converting the historical actual publishing result sequence into the second abnormal publishing result sequence is mode three, optionally, based on the mode one, the sharing rate prediction model can but is not limited to be trained by using the historical actual publishing result sequence and the first abnormal publishing result sequence; based on the mode two, the sharing rate prediction model can but is not limited to be trained by using the historical actual publishing result sequence and the second abnormal publishing result sequence; and based on the mode three, the sharing rate prediction model can but is not limited to be trained by using the historical actual publishing result sequence, the first abnormal publishing result sequence and the second abnormal publishing result sequence. Figure 5The scenario shown, for example Figure 8 As shown, according to the selection basis of P being one of three, in the case that P meets the first condition, mode one is selected from the three modes, and the mode one is used to determine the related data (media resource information 502-1, first abnormal publishing result sequence 802-1) to be input into the publishing result prediction model 506, so as to complete the training of the publishing result prediction model 506 this time.

[0098] Similarly, in the case that P meets the target condition, mode one is selected from the three modes, and the mode two is used to determine the related data (media resource information 502-1, historical actual publishing result sequence 802-2) to be input into the publishing result prediction model 506, so as to complete the training of the publishing result prediction model 506 this time; or, in the case that P meets the second condition, mode three is selected from the three modes, and the mode three is used to determine the related data (media resource information 502-1, second abnormal publishing result sequence 802-2) to be input into the publishing result prediction model 506, so as to complete the training of the publishing result prediction model 506 this time.

[0099] As an optional solution, the method further comprises:

[0100] S1, when the value of the first target conversion probability is located in the third value interval, it is determined that the first target conversion probability represents not converting the first historical actual publishing result sequence into the first target abnormal publishing result sequence;

[0101] S2, when the value of the first target conversion probability is located in the fourth value interval, it is determined that the first target conversion probability represents converting the first historical actual publishing result sequence into the first abnormal publishing result sequence;

[0102] S3, when the value of the first target conversion probability is located in the fifth value interval, it is determined that the first target conversion probability represents converting the first historical actual publishing result sequence into the second abnormal publishing result sequence.

[0103] Optionally, in the embodiment, in the case that the value of the first target conversion probability is located in the third value interval, P can meet the target condition as shown in Figure 8 , but not limited to; similarly, in the case that the value of the first target conversion probability is located in the fourth value interval, P can meet the first condition as shown in Figure 8 , but not limited to; or in the case that the value of the first target conversion probability is located in the fifth value interval, P can meet the second condition as shown in Figure 8 , but not limited to.

[0104] Further, the third, fourth and fifth value intervals can be but are not limited to average distribution, and further, on the basis of the third value interval being [1 / 3≤P<2 / 3], it is assumed that the fourth value interval is [0≤P<1 / 3] and the fifth value interval is [2 / 3≤P<1], then when the value of the first target conversion probability is in the fourth value interval, it is determined that the first target conversion probability indicates converting the first historical actual publishing result sequence into the first abnormal publishing result sequence; or, when the value of the first target conversion probability is in the fifth value interval, it is determined that the first target conversion probability indicates converting the first historical actual publishing result sequence into the second abnormal publishing result sequence.

[0105] Further, the third, fourth and fifth value intervals can be but are not limited to average distribution, and further, on the basis of the third value interval being [1 / 3≤P<2 / 3], it is assumed that the fourth value interval is [0≤P<1 / 3] and the fifth value interval is [2 / 3≤P<1], then when the value of the first target conversion probability is in the fourth value interval, it is determined that the first target conversion probability indicates converting the first historical actual publishing result sequence into the first abnormal publishing result sequence; or, when the value of the first target conversion probability is in the fifth value interval, it is determined that the first target conversion probability indicates converting the first historical actual publishing result sequence into the second abnormal publishing result sequence.

[0106] As an optional solution, converting the first historical actual publishing result sequence into the first abnormal publishing result sequence comprises:

[0107] S1, determining a same value for each historical actual publishing result in the first historical actual publishing result sequence, and obtaining a first value sequence;

[0108] S2, multiplying the data at the same position in the first historical actual publishing result sequence and the first value sequence to obtain the first abnormal publishing result sequence.

[0109] Optionally, in the embodiment, converting the first historical actual publishing result sequence into the target abnormal publishing result sequence can be but is not limited to understanding as reducing the dependence or influence of the publishing result prediction model on the first historical actual publishing result sequence in the training process, and converting the first historical actual publishing result sequence into the first abnormal publishing result sequence can be but is not limited to understanding as playing a role of simulating the abnormal publishing result sequence in the cold start mode, thereby improving the adaptability of the publishing result prediction model to the abnormal publishing result sequence in the cold start mode in the training process, and indirectly reducing the dependence or influence of the publishing result prediction model on the first historical actual publishing result sequence in the training process.

[0110] Optionally, in the embodiment, converting the first historical actual publishing result sequence into the first abnormal publishing result sequence can be, but is not limited to, reducing the dependence or influence of the publishing result prediction model on the first historical actual publishing result sequence in the training process, and then determining a same numerical value (which can be, but is not limited to, less than 1 and greater than or equal to 0) for each historical actual publishing result in the first historical actual publishing result sequence, to obtain a first numerical value sequence; and multiplying the first historical actual publishing result sequence and the data at the same position in the first numerical value sequence to obtain the first abnormal publishing result sequence.

[0111] It should be noted that the abnormal publishing result sequence in the cold start mode has a high uncertainty in most or even all numerical values, and if the publishing result prediction model does not consider this uncertainty, the prediction accuracy of the publishing result prediction model will be inevitably affected when facing the abnormal publishing result sequence in the cold start mode. Therefore, the first historical actual publishing result sequence is converted into the abnormal publishing result sequence in the cold start mode by simulation, and the publishing result prediction model is trained accordingly.

[0112] Further, it is assumed that the first historical actual publishing result sequence is [S1, S2, …, Sn], and the first numerical value sequence is [0.5, 0.5, …, 0.5], and then [S1, S2, …, Sn] and [0.5, 0.5, …, 0.5] are multiplied to obtain the first abnormal publishing result sequence [S1×0.5, S2×0.5, …, Sn×0.5]. As can be seen, the numerical value amount of the obtained first abnormal publishing result sequence is reduced compared with the first historical actual publishing result sequence, so that the publishing result prediction model pays most attention to the media resource information in the training process, thereby achieving the effect of reducing the dependence or influence of the publishing result prediction model on the first historical actual publishing result sequence in the training process.

[0113] In addition, in another embodiment, it is assumed that the first historical actual publishing result sequence is [S1, S2, …, Sn], and the first numerical value sequence is [0, 0, …, 0], and then [S1, S2, …, Sn] and [0, 0, …, 0] are multiplied to obtain the first abnormal publishing result sequence [0, 0, …, 0]. As can be seen, the first historical actual publishing result sequence is completely removed from the first abnormal publishing result sequence, so that the publishing result prediction model pays full attention to the media resource information in the training process, thereby achieving the effect of eliminating the dependence or influence of the publishing result prediction model on the first historical actual publishing result sequence in the training process.

[0114] As an optional solution, converting the first historical actual publishing result sequence into the second abnormal publishing result sequence comprises:

[0115] S1, randomly determining a value for each historical actual publishing result in the first historical actual publishing result sequence, to obtain a second value sequence;

[0116] S2, determining a value of a corresponding flag for each historical actual publishing result in the first historical actual publishing result sequence according to each value in the second value sequence and a preset target threshold, to obtain a third value sequence, wherein the value of the flag is used to represent whether the corresponding publishing result is an abnormal publishing result;

[0117] S3, determining a second abnormal publishing result sequence according to the first historical actual publishing result sequence and the third value sequence.

[0118] Optionally, in the embodiment, converting the first historical actual publishing result sequence into the target abnormal publishing result sequence can be but is not limited to being understood as reducing the dependence or influence of the publishing result prediction model on the first historical actual publishing result sequence in the training process, and converting the first historical actual publishing result sequence into the second abnormal publishing result sequence can be but is not limited to being understood as playing a role of simulating the abnormal publishing result sequence in the noise mode, thereby improving the adaptability of the publishing result prediction model to the abnormal publishing result sequence in the noise mode in the training process, and indirectly reducing the dependence or influence of the publishing result prediction model on the first historical actual publishing result sequence in the training process.

[0119] It should be noted that the values of the abnormal publishing result sequence in the noise mode all have high uncertainty, and if the publishing result prediction model does not consider this uncertainty, the prediction accuracy of the publishing result prediction model will inevitably be greatly affected when facing the abnormal publishing result sequence in the cold start mode, and therefore the first historical actual publishing result sequence is converted into the abnormal publishing result sequence in the noise mode in a simulated manner, and the publishing result prediction model is trained in a targeted manner.

[0120] Further, it is assumed that the first historical actual publishing result sequence is [S1, S2, …, Sn], and the third value sequence is [0.5, 0.4, …, 0.1], and then [S1, S2, …, Sn] is multiplied by [0.5, 0.5, …, 0.5] to obtain the first abnormal publishing result sequence [S1x0.5, S2x0.4, …, Snx0.1]. As can be seen, the overall value amount of the obtained first abnormal publishing result sequence is reduced compared with the first historical actual publishing result sequence, so that the publishing result prediction model pays most attention to the media resource information in the training process, thereby achieving the effect of reducing the dependence or influence of the publishing result prediction model on the first historical actual publishing result sequence in the training process.

[0121] As an optional solution, according to each value in the second numerical sequence and a preset target threshold, the value of the corresponding flag of each historical actual publishing result in the first historical actual publishing result sequence is determined, and a third numerical sequence is obtained, including:

[0122] S1, when the ith value in the second numerical sequence is greater than the target threshold, the value of the flag corresponding to the ith historical actual publishing result in the first historical actual publishing result sequence is set to a first value, wherein the second numerical sequence includes N values, 1≤i≤N, and N is a natural number greater than 1;

[0123] S2, when the ith value in the second numerical sequence is less than or equal to the target threshold, the value of the flag corresponding to the ith historical actual publishing result in the first historical actual publishing result sequence is set to a second value;

[0124] S3, wherein the first value is used to indicate that the corresponding publishing result is an abnormal publishing result, and the second value is used to indicate that the corresponding publishing result is not an abnormal publishing result, or,

[0125] S4, the first value is used to indicate that the corresponding publishing result is not an abnormal publishing result, and the second value is used to indicate that the corresponding publishing result is an abnormal publishing result.

[0126] Optionally, in the embodiment, in order to construct a sequence noise, for each missing flag in the sequence, a value T is randomly selected from a uniform distribution U[0, 1], and at the same time, a threshold τ can be set, but not limited to; when T is greater than τ, the missing flag is set to 0, otherwise it is set to 1, wherein, τ can be but not limited to regarded as a missing rate, when τ is larger, the constructed sequence missing value is more;

[0127] Considering that the original sharing rate input sequence also contains missing data, the missing data flag in the original input sequence can be but not limited to kept unchanged. Therefore, the final missing flag sequence is m c =m c ∨m, wherein ∨ represents logical or. Therefore, the corresponding input sharing rate sequence y is reset by , wherein represents Hadamard product.

[0128] According to the embodiments provided in the present application, when the i th value in the second value sequence is greater than the target threshold value, the value of the i th historical actual publishing result in the first historical actual publishing result sequence is set to the first value, wherein the second value sequence includes N values, 1≤i≤N, and N is a natural number greater than 1; when the i th value in the second value sequence is less than or equal to the target threshold value, the value of the i th historical actual publishing result in the first historical actual publishing result sequence is set to the second value, thereby improving the degree of simulation of the publishing result sequence under the noise mode.

[0129] As an optional solution, the second abnormal publishing result sequence is determined according to the first historical actual publishing result sequence and the third value sequence, including:

[0130] The data at the same position in the first historical actual publishing result sequence and the third value sequence are multiplied to obtain the second abnormal publishing result sequence.

[0131] Further, optionally, it is assumed that the first historical actual publishing result sequence is [S1, S2, …, Sn], and the third value sequence is [0.5, 0.4, …, 0.1], then [S1, S2, …, Sn] and [0.5, 0.5, …, 0.5] are multiplied to obtain the first abnormal publishing result sequence [S1×0.5, S2×0.4, …, Sn×0.1]. As can be seen, the overall value of the obtained first abnormal publishing result sequence is reduced compared to the first historical actual publishing result sequence, so that the publishing result prediction model pays most attention to the media resource information in the training process, thereby reducing the dependence or influence of the publishing result prediction model on the first historical actual publishing result sequence in the training process.

[0132] As an optional solution, the first media resource information, the first target abnormal publishing result sequence, and the first future actual publishing result are used to train the publishing result prediction model, including:

[0133] S1, according to the first media resource information and the first target abnormal publishing result sequence, the features of the first multimedia resource in multiple dimensions are determined, and a first feature set is obtained;

[0134] S2, according to the first feature set, the first future predicted publishing result of the first multimedia resource at the target time is determined;

[0135] S3, when the value of the target loss function determined according to the first future predicted publishing result and the first future actual publishing result does not meet the preset condition, the parameters in the publishing result prediction model are adjusted to obtain the first publishing result prediction model to be trained.

[0136] Optionally, in the embodiment, for the training process of the publishing result prediction model, the historical actual publishing result sequence and the future actual publishing result can be but not limited to a corresponding relationship, specifically, the historical actual publishing result sequence or the converted historical actual publishing result sequence (target abnormal publishing result sequence) is taken as the input of the publishing result prediction model, and the future actual publishing result is taken as the comparison object of the output of the publishing result prediction model, and by comparison between the two, the loss of the publishing result prediction model in this training is determined, if the convergence condition is met, the training is ended, if the convergence condition is not met, the model parameters in the publishing result prediction model are adjusted according to the loss, and the next training is started based on the publishing result prediction model after adjusting the model parameters.

[0137] Further, as shown in Figure 9 , the publishing result prediction model can be trained by using the historical data 1-1 corresponding to T1-1 and the feedback data 2-1 corresponding to T2-1, and in the training process of the publishing result prediction model, the historical data 1-1 is processed as the first abnormal data (the first abnormal publishing result sequence); similarly, the publishing result prediction model can be trained by using the historical data 1-2 corresponding to T1-2 and the feedback data 2-2 corresponding to T2-2, and in the training process of the publishing result prediction model, the historical data 1-2 is processed as the second abnormal data (the second abnormal publishing result sequence); in addition, the publishing result prediction model can also be trained by using the historical data 1-3 corresponding to T1-3 and the feedback data 2-3 corresponding to T2-3, and in the training process of the publishing result prediction model, the historical data 1-2 is processed as normal data (the historical actual publishing result sequence), or the historical data 1-2 is directly regarded as normal data without processing.

[0138] In addition, in the embodiment, the first media resource information, the first historical actual publishing result sequence and the first future actual publishing result are used, including:

[0139] S1, determining the features of the first multimedia resource in multiple dimensions according to the first media resource information and the first historical actual publishing result sequence, and obtaining a second feature set;

[0140] S2, determining the second future predicted publishing result of the first multimedia resource at the target time according to the second feature set;

[0141] S3, when the value of the target loss function determined according to the second future predicted publishing result and the first future actual publishing result does not meet the preset condition, adjusting the parameters in the publishing result prediction model to obtain a second publishing result prediction model to be trained.

[0142] As an optional solution, according to the first media resource information and the first target abnormal publishing result sequence, a feature of the first multimedia resource in multiple dimensions is determined, and a first feature set is obtained, including at least two of the following steps:

[0143] S1, according to the text information of the first multimedia resource, the text feature of the first multimedia resource is determined, wherein the first media resource information includes the text information of the first multimedia resource, and the first feature set includes the text feature of the first multimedia resource;

[0144] S2, according to the picture information of the first multimedia resource, the visual feature of the first multimedia resource is determined, wherein the first media resource information includes the picture information of the first multimedia resource, and the first feature set includes the visual feature of the first multimedia resource;

[0145] S3, according to the first target abnormal publishing result sequence, the sequence feature of the first multimedia resource is determined, wherein the first feature set includes the sequence feature of the first multimedia resource;

[0146] S4, according to the attribute information of the first multimedia resource, the attribute feature of the first multimedia resource is determined, wherein the first media resource information includes the attribute information of the first multimedia resource, and the first feature set includes the attribute feature of the first multimedia resource.

[0147] Optionally, in the embodiment, the feature of the first multimedia resource in multiple dimensions can be but is not limited to a multi-modal feature, wherein the multi-modal feature can be but is not limited to a feature indicating a multi-state, such as a picture, a text, some discrete features and a historical sharing rate sequence, etc.

[0148] It should be noted that the multi-modal feature is introduced to alleviate the uncertainty of the feature. Since many contents are uploaded by users and lack professional guidance, it is easy to have a picture that does not match the text, which may cause a certain dimension feature to be invalid. Therefore, by introducing the multi-modal feature, the complementarity between multiple types of features can be generated.

[0149] Further, for example, it is assumed that the multi-modal feature includes the sequence feature, and at least one of the text feature, the picture feature and the attribute feature, then taking a video content as an example, the sharing rate information of the video content in a historical time period (a first time period) is extracted and used as the sequence feature; the cover of the video content is extracted and used as the picture feature; the text (such as video title, author name, publishing platform name, comment, dialogue, etc.) of the video content is extracted and used as the text feature; and the attribute (such as content category, content publishing duration, content publishing source, video click rate, video like number, etc.) information of the video content is extracted and used as the attribute feature.

[0150] Optionally, in the present embodiment, the extraction / treatment of each feature can be, but is not limited to, using the same or different network structure or model, such as Figure 10 As shown, the extraction / treatment of each feature in the multi-modal feature 1002 can be, but is not limited to, using different network structures or models, such as using a multi-state modeling module and a time convolution network to complete the extraction / treatment of the historical sharing rate sequence; using an xDeeoFM model to complete the extraction / treatment of meta-features (attribute features); using a RestNet-50 network structure to complete the extraction / treatment of visual (picture features); using a BERT model to extract / treat text features;

[0151] Among them, mining sufficient potential information from multi-modal features can introduce more complementary information between features for the sharing rate prediction task, thereby reducing the uncertainty of the features.

[0152] Specifically, the multi-modal feature modeling module is responsible for extracting the feature representation of text features, visual features, sequence features and meta-features, and balancing the correlation between different features through a gate fusion module;

[0153] The BERT model is a pre-training model that adopts a typical bidirectional encoding architecture. It has very outstanding results in many tasks. And, conventional NLP tasks, such as text classification, text clustering, language translation, question and answer systems, text feature extraction, etc., BERT can handle. Therefore, for the extraction of text features, the BERT model is directly used;

[0154] The ResNet network structure solves the "degeneration" problem of deep neural networks by introducing a residual module, that is, after stacking more layers to the network, the performance is rapidly reduced. It has strong representation ability and is applied in many computer vision tasks; In the present embodiment, the ResNet-50 network structure, that is, a convolutional neural network with a depth of 50 layers, is used to extract visual features in the cover picture;

[0155] Due to the dependence of the recurrent convolutional network on the sequence, it is relatively slow to operate. Therefore, the time convolution network (TCN) is selected to extract the time sequence feature, which can be parallel and has a flexible receptive field;

[0156] In order to learn the interaction between meta-features, the xdeepFM model is used for interaction learning between features. The core of xdeepFM is the architecture of CIN. It can learn controllable, automatic and explicit high-order feature interactions, that is, it can learn the highest order through the number of layers.

[0157] In this embodiment, different models are used to extract and process different types of features. In order to better learn the importance of different features for the final task, as shown in the release result prediction model 1004 in the middle, after the features are extracted / processed, attention fusion and multi-layer perception are combined to learn the importance of different features. Figure 10

[0158] In addition, in this embodiment, text feature extraction can not only select BERT but also select HAN[1 and other text processing methods, visual feature extraction can not only select ResNet-50, but also select many other deep models such as VGGNet, GoogLeNet, etc. Sequence feature extraction can not only select TCN, but also select LSTM, GRU and other recurrent neural networks as alternatives. In addition, the extraction of meta-feature interaction information can not only select xDeepFM, but also select other extractable interaction feature frameworks such as DeepFM] deep&cross and other models.

[0159] As an optional solution, the method further comprises:

[0160] S1, obtaining second media resource information of a published second multimedia resource, a second historical actual release result sequence of the second multimedia resource, and a second future actual release result of the second multimedia resource at a target time, wherein the second historical actual release result sequence comprises a group of actual release results of the second multimedia resource within a first time period;

[0161] S2, obtaining a second target conversion probability randomly generated by a sequence conversion module, wherein the second target conversion probability is used to indicate whether to convert the second historical actual release result sequence into a second target abnormal release result sequence, and at least part of the release results in the second target abnormal release result sequence are abnormal release results;

[0162] S3, when the second target conversion probability indicates that the second historical actual release result sequence is converted into the second target abnormal release result sequence, converting the second historical actual release result sequence into the second target abnormal release result sequence, and using the second media resource information, the second target abnormal release result sequence and the second future actual release result to train the first release result prediction model, wherein the first release result prediction model is used to generate a second future predicted release result of the second multimedia resource at the target time;

[0163] ​S4, when the second target conversion probability representation does not convert the second historical actual publishing result sequence into the second target abnormal publishing result sequence, training the first publishing result prediction model using the second media resource information, the second historical actual publishing result sequence, and the second future actual publishing result.

[0164] Optionally, in this embodiment, the training of the publishing result prediction model is an iterative training process. Without satisfying the convergence condition, the publishing result prediction model can be iteratively trained based on the historical actual publishing result sequence / target abnormal publishing result sequence corresponding to each multimedia resource (which can be but is not limited to a sample, and the first multimedia resource can be but is not limited to a current sample, and in this embodiment, the second multimedia resource is another current sample).

[0165] As an optional solution, the publishing result is regarded as a sharing rate, the corresponding historical actual publishing result is a historical actual sharing rate, and the future actual publishing result corresponds to a future actual sharing rate. Then, the publishing result prediction model to be trained is trained using the first media resource information, the first historical actual publishing result sequence, and the first future actual publishing result, including:

[0166] The publishing result prediction model to be trained is trained using the first media resource information, the first historical actual publishing result sequence, and the first future actual publishing result, wherein the publishing result prediction model is a sharing rate prediction model, the sharing rate prediction model is used to generate a first future predicted sharing rate of the first multimedia resource at a target time according to the first media resource information and the first historical actual publishing result sequence, the first future predicted sharing rate represents a predicted number of times of sharing of the first multimedia resource divided by a number of exposures, and the first future actual publishing result is a first future actual sharing rate of the first multimedia resource at the target time, the first future actual sharing rate represents an actual number of times of sharing of the first multimedia resource divided by a number of exposures.

[0167] Optionally, in this embodiment, in order to simulate the modal sequence of different modalities (such as normal modalities, cold start modalities, noise modalities, etc.), two perturbation blocks can be introduced, including a cold start perturbation block and a noise perturbation block, and in order to better represent the state of the time sequence from t-ω1 to t time window wherein m∈{0,1}. The missing flag m is 1 when the data is missing or uncertain, and the flag is 0 when the data is reliable or normal.

[0168] Further, the cold start perturbation block is responsible for simulating the cold start state data. The cold start perturbation block forces the model to learn other feature information by means of forced mask, so as to alleviate the excessive dependence of the model on the historical sequence information. Formally, we use a full one vector mask as the missing label vector, which means that the entire sharing rate data is missing or uncertain. Therefore, the corresponding sharing rate sequence is a zero vector erased. The purpose of this operation is to guide the model to pay more attention to other feature information rather than the historical sharing rate, so as to improve the content sharing rate prediction performance under the cold start state;

[0169] Similar to the cold start perturbation block, the noise perturbation block is used to simulate the uncertainty of part of the sequence data. In order to construct sequence noise randomly, for each missing label in the sequence, a value T is randomly selected from a uniform distribution U[0,1], and at the same time, a threshold τ is set. When T is greater than τ, the missing label is set to 0, otherwise it is set to 1. Note that τ can be regarded as a missing rate. When τ is large, the constructed sequence has more missing values. Considering that the original sharing rate input sequence also contains missing data, the missing data flag in the original input sequence should be kept unchanged. Therefore, the final missing flag sequence is m c = m c ∨m, where ∨ represents logical or. Therefore, the corresponding input sharing rate sequence y is reset as y , where represents the Hadamard product.

[0170] Further, with a probability of p, a certain state block is randomly entered, and p is randomly collected from a uniform distribution, as shown in the following formula (1): Figure 11 For example, in the case of obtaining the historical sharing rate sequence, mode selection is performed first, that is, with a probability of p, a certain state block is randomly entered, wherein mode 1 corresponds to the cold start mode (embedding feature), mode 2 corresponds to the normal mode (embedding feature), and mode 3 corresponds to the noise mode (embedding feature); finally, after processing by the corresponding state block, the processed sequence is input into the feature encoding layer to complete the extraction / processing of the feature.

[0171] The formal representation is shown in the following formula (1):

[0172] p ~ U[0,1]

[0173]

[0174] In addition, after the input sequence is processed by a certain state module, it is sent to the embedding layer emb(·) to obtain the final sharing rate representation sequence The formal representation is shown in the following formula (2):

[0175]

[0176] Optionally, as another optional implementation, as shown in Figure 12 the prediction method of the publishing result includes:

[0177] S1202, obtaining target media resource information of a published target multimedia resource, and a target historical actual publishing result sequence of the target multimedia resource, wherein the target historical actual publishing result sequence includes a set of actual publishing results of the target multimedia resource in a target time period;

[0178] S1204, inputting the target media resource information and the target historical actual publishing result sequence into a target publishing result prediction model to obtain a future predicted publishing result of the target multimedia resource at a target future time;

[0179] Wherein, the target publishing result prediction model is a model obtained by training a to-be-trained publishing result prediction model using a target sample set, each sample in the target sample set includes: media resource information of a published multimedia resource, a future actual publishing result of the multimedia resource at a target time; and one sequence of a historical actual publishing result sequence, a first abnormal publishing result sequence, and a second abnormal publishing result sequence of the multimedia resource;

[0180] Wherein, the historical actual publishing result sequence includes a set of actual publishing results of the multimedia resource in a time period, and the target time is later than the time period; the first abnormal publishing result sequence is a sequence converted from the historical actual publishing result sequence, and the publishing results in the first abnormal publishing result sequence are all abnormal publishing results; the second abnormal publishing result sequence is a sequence converted from the historical actual publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results.

[0181] Optionally, in the present embodiment, the above-mentioned prediction method of the publishing result can be but not limited to applied in the scene of predicting the sharing rate of the content information in a future period of time, for example, in order to better grasp the development trend of the video content, the staff triggers the prediction operation on the video D (target multimedia resource 302) in the background, for example Figure 3wherein the video D can be understood as a video resource played on some application; further, the media resource information 304-1 (such as a video title, a video introduction, a video publishing platform, a video click rate, a video like number, etc.) and the historical actual publishing result sequence 304-2 (such as a sharing rate, a sharing number, an effective revisit rate after sharing, etc. of the video D in a historical preset time period) of the target multimedia resource 302 are input into the publishing result prediction model 306 trained by the above-mentioned publishing result prediction method, so as to obtain a future prediction publishing result 308 of the video D, wherein the future prediction publishing result 308 can be but is not limited to used to represent a predicted sharing rate of the video D in a future time period; further, the future prediction publishing result 308 is displayed on a display screen in the background for reference by the staff, such as Figure 3 as shown in (c) of FIG. 4;

[0182] In addition, the staff can also be provided with a targeted content optimization scheme based on the future prediction publishing result 308, such as if the sharing rate indicated by the future prediction publishing result 308 does not reach an expectation, reference information is generated based on the input media resource information 304-1 and the historical actual publishing result sequence 304-2, wherein the reference information is used to represent factors affecting the sharing rate indicated by the future prediction publishing result 308 not reaching the expectation, such as the video D itself has an optimization item, the video D played platform has a problem of low user retention rate, etc.

[0183] Optionally, in the embodiment, the media resource information can be but is not limited to used to represent basic information of the multimedia resource, such as text information, image information, attribute information, etc., which is usually fixed or unchanged in a long period;

[0184] The historical actual publishing result sequence can be but is not limited to used to represent a plurality of information sets changed or unchanged in a historical time period, such as if the historical time period is divided into 10 time points, the historical actual publishing result sequence can be but is not limited to an information set composed of actual publishing results corresponding to each time point;

[0185] For further illustration, the historical actual publishing result sequence is for example Figure 4 as shown in FIG. 4, assuming that the horizontal axis is time (T) and the vertical axis is sharing rate (S), the historical sharing rate sequence 402 can be but is not limited to used to represent a sharing rate summary corresponding to each time point in a historical time period; specifically, assuming that the historical time period includes n time points, T1, T2, …, Tn, and the sharing rates corresponding to the n time points are S1, S2, …, Sn respectively, the historical sharing rate sequence 402 can be but is not limited to [S1, S2, …, Sn], or a numerical sequence calculated based on [S1, S2, …, Sn], which is not limited herein.​

[0186] Optionally, in the embodiment, the historical actual publishing result sequence may, but is not limited to, include the normal publishing result sequence and / or the abnormal publishing result sequence, and generally, the probability that the historical actual publishing result sequence includes the normal publishing result sequence is much higher than that the historical actual publishing result sequence includes the abnormal publishing result sequence, and then in the related art, the historical actual publishing result sequence is directly used to train the publishing result prediction model, but the trained publishing result prediction model depends on the normal publishing result sequence, and then when facing the abnormal publishing result sequence, the high prediction accuracy cannot be guaranteed.

[0187] Based on this, in the embodiment, multiple conversion modes are configured, the historical actual publishing result sequence is converted by using the target conversion probability randomly generated by the sequence conversion module, for example, the historical actual publishing result sequence is converted into the abnormal publishing result sequence, and the abnormal publishing result sequence is further used to train the publishing result prediction model, so as to reduce the dependence of the trained publishing result prediction model on the normal publishing result sequence, so as to guarantee the high prediction accuracy when facing the abnormal publishing result sequence.

[0188] Optionally, in the embodiment, when the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, the first media resource information of the first multimedia resource may, but is not limited to, be retained, and then the first media resource information, the first target abnormal publishing result sequence and the first future actual publishing result are used to train the publishing result prediction model to be trained, which can be understood as converting the first historical actual publishing result sequence into the first target abnormal publishing result sequence while retaining the first media resource information of the first multimedia resource, so that the publishing result prediction model pays more attention to the first media resource information than the first historical actual publishing result sequence in the training process, so as to reduce the dependence of the publishing result prediction model on the historical actual publishing result sequence.

[0189] Generally, in the training process of the above publishing result prediction model, the conversion may, but is not limited to, be only the historical actual publishing result sequence, and the media resource information of the multimedia resource and the future actual publishing result at the target time are not changed.

[0190] Optionally, in the embodiment, the relationship between the historical actual publishing result sequence and the future actual publishing result may, but is not limited to, be understood as follows: taking the published first multimedia resource as an example, the first time period and the target time are selected in the time period in which the first multimedia resource is published, and the target time is after the first time period.

[0191] Based on this, since both the first time period and the target time are the times when the first multimedia resource has been published, the actual publication results of the first multimedia resource in the first time period and the target time can be obtained directly, namely the historical actual publication result sequence and the future actual publication result. The time difference between the first time period and the target time can be, but is not limited to, fixed. Assuming that the time difference between the first time period and the target time is the target time difference, the target time difference can be, but is not limited to, used to determine the time of publication result after the publication result prediction model predicts the content information of the historical time. For example, assuming that the target time difference is 1 month, after the trained publication result prediction model predicts the content information published in the [T1~T2] time period, the output publication result is the predicted sharing rate of the content information 1 month after T2.

[0192] Further optionally, in this embodiment, for the training process of the release result prediction model, the historical actual release result sequence and the future actual release result can be, but are not limited to, a correspondence. The specific historical actual release result sequence or the transformed historical actual release result sequence (target abnormal release result sequence) is used as the input of the release result prediction model, while the future actual release result is used as the comparison object for the output of the release result prediction model. By comparing with the comparison object, the loss of the release result prediction model in this training is determined. If the convergence condition is met, the training ends. If the convergence condition is not met, the model parameters in the release result prediction model are adjusted according to the loss, and the next training begins based on the release result prediction model after adjusting the model parameters.

[0193] It should be noted that training the release result prediction model using historical release result sequences enables it to predict the sharing rate of normal data. By randomly generating target conversion probabilities through the sequence transformation module, the historical release result sequences are converted into abnormal release result sequences. Further training the release result prediction model using these abnormal release result sequences enables it to predict the sharing rate of abnormal data. Based on this, the trained release result prediction model possesses the ability to predict the sharing rate of both normal and abnormal data, thus maintaining high accuracy in sharing rate prediction regardless of whether the data is normal or abnormal.

[0194] To further illustrate, optional examples include... Figure 5 As shown, the media resource information 502-1 of the published multimedia resources, the historical actual publication result sequence 502-2 in T1 (first time period), and the actual sharing rate 502-3 (future actual publication result) of the multimedia resources in T2 (target time) are obtained.

[0195] Further, the acquisition sequence conversion module 504 randomly generates a target conversion probability (P), and in a case where P does not satisfy a target condition, converts the historical actual publishing result sequence into a target abnormal publishing result sequence 502-4, and takes the media resource information 502-1 and the target abnormal publishing result sequence 502-4 as current model input to train the publishing result prediction model 506 to be trained; or, in a case where P satisfies the target condition, takes the media resource information 502-1 and the historical actual publishing result sequence 502-2 as current model input to train the publishing result prediction model 506 to be trained.

[0196] Further, the publishing result prediction model 506 uses the media resource information 502-1 and the target abnormal publishing result sequence 502-4, and / or the media resource information 502-1 and the historical actual publishing result sequence 502-2 as current model input to start training, such as outputting a predicted sharing rate 508 corresponding to the model input, and comparing with a real sharing rate 502-3 to obtain a loss 510; in a case where the loss 510 reaches a convergence condition, ending the training of the publishing result prediction model 506; in a case where the loss 510 does not reach the convergence condition, adjusting model parameters of the publishing result prediction model 506 according to the loss 510, and continuing the next training based on the publishing result prediction model 506 after adjusting the model parameters.

[0197] Through the embodiments provided in the present application, target media resource information of a published target multimedia resource and a target historical actual publishing result sequence of the target multimedia resource are acquired, wherein the target historical actual publishing result sequence includes a group of actual publishing results of the target multimedia resource in a target time period; the target media resource information and the target historical actual publishing result sequence are input to a target publishing result prediction model to obtain a future predicted publishing result of the target multimedia resource at a target future time, achieving the purpose that the publishing result prediction model can maintain high sharing rate prediction accuracy regardless of facing normal data or abnormal data, and realizing the effect of improving the prediction accuracy of the prediction model on the sharing rate.

[0198] As an optional solution, the target media resource information of the published target multimedia resource and the target historical actual publishing result sequence of the target multimedia resource are acquired, including: acquiring the target media resource information of the published target multimedia resource and a target historical actual sharing rate sequence of the target multimedia resource, wherein the target historical actual publishing result sequence includes a group of actual sharing rates of the target multimedia resource in a target time period, and the actual sharing rate represents the number of actual target multimedia resources shared divided by the number of exposures;

[0199] The specific embodiments can refer to the examples shown in the training method of the prediction model, which will not be repeated here in this example.

[0200] As an optional solution, the target media resource information and the target historical actual publishing result sequence are input into the target publishing result prediction model to obtain a future predicted publishing result of the target multimedia resource at a target future time, including: inputting the target media resource information and the target historical actual sharing rate sequence into the target publishing result prediction model to obtain a future predicted sharing rate of the target multimedia resource at the target future time, wherein the future predicted sharing rate represents a predicted number of times of sharing of the target multimedia resource divided by a number of times of exposure.

[0201] The specific embodiments can refer to the examples shown in the training method of the prediction model, which will not be repeated here in this example.

[0202] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited to the action sequence described, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily necessary for the present application.

[0203] According to another aspect of the embodiments of the present application, a training device for training the prediction model of the above-mentioned training method of the prediction model is also provided. As shown in the Figure 13 The device comprises:

[0204] The first acquisition unit 1302 is configured to acquire first media resource information of a published first multimedia resource, a first historical actual publishing result sequence of the first multimedia resource, and a first future actual publishing result of the first multimedia resource at a target time, wherein the first historical actual publishing result sequence comprises a group of actual publishing results of the first multimedia resource within a first time period, and the target time is later than the first time period.

[0205] The second acquisition unit 1304 is configured to acquire a first target conversion probability randomly generated by the sequence conversion module, wherein the first target conversion probability is used to indicate whether to convert the first historical actual publishing result sequence into a first target abnormal publishing result sequence, and at least part of the publishing results in the first target abnormal publishing result sequence are abnormal publishing results.

[0206] The first training unit 1306 is configured to, when the first target conversion probability representation is used to convert the first historical actual publishing result sequence into the first target abnormal publishing result sequence, convert the first historical actual publishing result sequence into the first target abnormal publishing result sequence, and train the to-be-trained publishing result prediction model by using the first media resource information, the first target abnormal publishing result sequence, and the first future actual publishing result, wherein the to-be-trained publishing result prediction model is used to generate the first future predicted publishing result of the first multimedia resource at the target time.

[0207] The second training unit 1308 is configured to, when the first target conversion probability representation is not used to convert the first historical actual publishing result sequence into the first target abnormal publishing result sequence, train the to-be-trained publishing result prediction model by using the first media resource information, the first historical actual publishing result sequence, and the first future actual publishing result.

[0208] Optionally, in the embodiment, the media resource information can be, but is not limited to, used to represent the basic information of the multimedia resource, such as text information, image information, attribute information, etc., and generally the basic information is fixed or does not change in a long period.

[0209] The historical actual publishing result sequence can be, but is not limited to, used to represent a plurality of information sets that change or remain unchanged in a historical time period, for example, if the historical time period is divided into 10 time points, the historical actual publishing result sequence can be, but is not limited to, an information set composed of the actual publishing results corresponding to each time point.

[0210] Optionally, in the embodiment, the historical actual publishing result sequence can be, but is not limited to, composed of the normal publishing result sequence and / or the abnormal publishing result sequence, and generally the probability that the historical actual publishing result sequence includes the normal publishing result sequence is much higher than the probability that the historical actual publishing result sequence includes the abnormal publishing result sequence, and then in the related art, the historical actual publishing result sequence is directly used to train the publishing result prediction model, but the trained publishing result prediction model depends on the normal publishing result sequence, and then when facing the abnormal publishing result sequence, the high prediction accuracy cannot be guaranteed.

[0211] Based on this, in the embodiment, a plurality of conversion modes are configured, the historical actual publishing result sequence is converted by using the target conversion probability randomly generated by the sequence conversion module, for example, the historical actual publishing result sequence is converted into the abnormal publishing result sequence, and the publishing result prediction model is further trained by using the abnormal publishing result sequence, so as to reduce the dependence of the trained publishing result prediction model on the normal publishing result sequence, and to guarantee the high prediction accuracy when facing the abnormal publishing result sequence.

[0212] Optionally, in the embodiment, when the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, the first media resource information of the first multimedia resource can be but is not limited to be retained, and then the first media resource information, the first target abnormal publishing result sequence and the first future actual publishing result are used to train the to-be-trained publishing result prediction model. It can be understood that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, and the first media resource information of the first multimedia resource is retained, so that the publishing result prediction model pays more attention to the first media resource information than the first historical actual publishing result sequence in the training process, so as to reduce the dependence of the publishing result prediction model on the historical actual publishing result sequence.

[0213] Generally, in the training process of the above publishing result prediction model, the conversion can be but is not limited to only the historical actual publishing result sequence, the media resource information of the multimedia resource and the future actual publishing result at the target time without any change.

[0214] Optionally, in the embodiment, the relationship between the historical actual publishing result sequence and the future actual publishing result can be but is not limited to understood as follows: taking the published first multimedia resource as an example, a first time period and a target time are selected in the time period in which the first multimedia resource is published, and the target time is after the first time period.

[0215] Based on this, since the first time period and the target time are both the time in which the first multimedia resource is published, the actual publishing result of the first multimedia resource at the first time period and the target time, i.e. the historical actual publishing result sequence and the future actual publishing result, can be directly obtained, respectively. The time difference between the first time period and the target time can be but is not limited to fixed. Assuming that the time difference between the first time period and the target time is a target time difference, the target time difference can be but is not limited to used to determine the time of the publishing result predicted by the publishing result prediction model based on the content information of the historical time. For example, assuming that the target time difference is 1 month, after the trained publishing result prediction model predicts the content information published in the time period [T1-T2], the output publishing result is the sharing rate of the content information predicted 1 month after T2 as the time start point.

[0216] Further optionally, in the present embodiment, for the training process of the publishing result prediction model, the historical actual publishing result sequence and the future actual publishing result can be, but are not limited to, a corresponding relationship, specifically, the historical actual publishing result sequence or the converted historical actual publishing result sequence (target abnormal publishing result sequence) is taken as the input of the publishing result prediction model, and the future actual publishing result is taken as the comparison object of the output of the publishing result prediction model, by comparison between the two, the loss of the publishing result prediction model in the present training is determined, if the convergence condition is met, the training is ended, if the convergence condition is not met, the model parameters in the publishing result prediction model are adjusted according to the loss, and the next training is started based on the publishing result prediction model after the model parameters are adjusted.

[0217] It should be noted that the publishing result prediction model is trained by using the historical actual publishing result sequence, so that the publishing result prediction model has the ability to predict the share rate of normal state data; the historical actual publishing result sequence is converted into an abnormal publishing result sequence by the target conversion probability randomly generated by the sequence conversion module, and the publishing result prediction model is further trained by using the abnormal publishing result sequence, so that the publishing result prediction model has the ability to predict the share rate of abnormal data; based on this, the trained publishing result prediction model has the ability to predict the share rate of normal data and the ability to predict the share rate of abnormal state data, and thus the publishing result prediction model can maintain high share rate prediction accuracy when facing normal data or abnormal data.

[0218] The specific embodiments can refer to the examples shown in the training method of the prediction model described above, which will not be described herein in the present example.

[0219] According to the embodiments provided in the present application, the first media resource information of the published first multimedia resource, the first historical actual publishing result sequence of the first multimedia resource, and the first future actual publishing result of the first multimedia resource at a target time are obtained, the first historical actual publishing result sequence includes a set of actual publishing results of the first multimedia resource within a first time period, and the target time is later than the first time period; a first target conversion probability randomly generated by a sequence conversion module is obtained, the first target conversion probability is used to indicate whether the first historical actual publishing result sequence is converted into a first target abnormal publishing result sequence, and at least part of the publishing results in the first target abnormal publishing result sequence are abnormal publishing results; when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, and the first media resource information, the first target abnormal publishing result sequence, and the first future actual publishing result are used to train the to-be-trained publishing result prediction model, wherein the to-be-trained publishing result prediction model is used to generate a first future predicted publishing result of the first multimedia resource at the target time; when the first target conversion probability indicates that the first historical actual publishing result sequence is not converted into the first target abnormal publishing result sequence, the first media resource information, the first historical actual publishing result sequence, and the first future actual publishing result are used to train the to-be-trained publishing result prediction model, so as to achieve the purpose that the publishing result prediction model can maintain a high sharing rate prediction accuracy when facing normal data or abnormal data, and the effect of improving the prediction accuracy of the prediction model on the sharing rate is realized.

[0220] As an optional solution, the first training unit 1306 includes:

[0221] The first conversion module is configured to convert the first historical actual publishing result sequence into the first abnormal publishing result sequence when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence, wherein the first target abnormal publishing result sequence includes the first abnormal publishing result sequence, and the publishing results in the first abnormal publishing result sequence are all abnormal publishing results; or

[0222] The second conversion module is configured to convert the first historical actual publishing result sequence into the second abnormal publishing result sequence when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence, wherein the first target abnormal publishing result sequence includes the second abnormal publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results.

[0223] The specific embodiments can refer to the examples shown in the training method of the prediction model described above, which will not be described here in this example.

[0224] As an optional solution, the apparatus further comprises:

[0225] The first determining module is configured to determine that the first target conversion probability indicates not to convert the first historical actual publishing result sequence into the first target abnormal publishing result sequence when the value of the first target conversion probability is located in the first value interval.

[0226] The second determining module is configured to determine that the first target conversion probability indicates to convert the first historical actual publishing result sequence into the first abnormal publishing result sequence, or to convert the first historical actual publishing result sequence into the second abnormal publishing result sequence when the value of the first target conversion probability is located in the second value interval.

[0227] The specific embodiments can refer to the examples shown in the training method of the prediction model described above, which will not be repeated here in this example.

[0228] As an optional solution, the first training unit 1306 comprises:

[0229] The third conversion module is configured to convert the first historical actual publishing result sequence into the first abnormal publishing result sequence when the first target conversion probability indicates to convert the first historical actual publishing result sequence into the first abnormal publishing result sequence, wherein the first target abnormal publishing result sequence comprises the first abnormal publishing result sequence, and the publishing results in the first abnormal publishing result sequence are all abnormal publishing results.

[0230] The fourth conversion module is configured to convert the first historical actual publishing result sequence into the second abnormal publishing result sequence when the first target conversion probability indicates to convert the first historical actual publishing result sequence into the second abnormal publishing result sequence, wherein the first target abnormal publishing result sequence comprises the second abnormal publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results.

[0231] The specific embodiments can refer to the examples shown in the training method of the prediction model described above, which will not be repeated here in this example.

[0232] As an optional solution, the apparatus further comprises:

[0233] The third determining module is configured to determine that the first target conversion probability indicates not to convert the first historical actual publishing result sequence into the first target abnormal publishing result sequence when the value of the first target conversion probability is located in the third value interval.

[0234] The fourth determining module is configured to determine that the first target conversion probability indicates to convert the first historical actual publishing result sequence into the first abnormal publishing result sequence when the value of the first target conversion probability is located in the fourth value interval.

[0235] The fifth determining module is configured to determine that the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence when the value of the first target conversion probability is located in the fifth value interval.

[0236] The specific embodiments can refer to the examples shown in the training method of the prediction model, which will not be repeated here in this example.

[0237] As an optional solution, the first conversion module comprises:

[0238] The first obtaining sub-module is configured to determine a same numerical value for each historical actual publishing result in the first historical actual publishing result sequence, and obtain a first numerical value sequence;

[0239] The second obtaining sub-module is configured to multiply the data at the same position in the first historical actual publishing result sequence and the first numerical value sequence to obtain the first abnormal publishing result sequence.

[0240] The specific embodiments can refer to the examples shown in the training method of the prediction model, which will not be repeated here in this example.

[0241] As an optional solution, the second conversion module comprises:

[0242] The third obtaining sub-module is configured to randomly determine a numerical value for each historical actual publishing result in the first historical actual publishing result sequence to obtain a second numerical value sequence;

[0243] The fourth obtaining sub-module is configured to determine the value of the corresponding flag for each historical actual publishing result in the first historical actual publishing result sequence according to each numerical value in the second numerical value sequence and a preset target threshold value, to obtain a third numerical value sequence, wherein the value of the flag is used to indicate whether the corresponding publishing result is an abnormal publishing result.

[0244] The fifth obtaining sub-module is configured to determine the second abnormal publishing result sequence according to the first historical actual publishing result sequence and the third numerical value sequence.

[0245] The specific embodiments can refer to the examples shown in the training method of the prediction model, which will not be repeated here in this example.

[0246] As an optional solution, the fourth obtaining sub-module comprises:

[0247] The first setting subunit is configured to set a value of a flag corresponding to an i-th historical actual publishing result in the first historical actual publishing result sequence to a first value when an i-th value in the second value sequence is greater than the target threshold, where the second value sequence includes N values, 1≤i≤N, and N is a natural number greater than 1.

[0248] The second setting subunit is configured to set a value of a flag corresponding to an i-th historical actual publishing result in the first historical actual publishing result sequence to a second value when an i-th value in the second value sequence is less than or equal to the target threshold.

[0249] The first value is used to indicate that the corresponding publishing result is an abnormal publishing result, and the second value is used to indicate that the corresponding publishing result is not an abnormal publishing result, or

[0250] The first value is used to indicate that the corresponding publishing result is not an abnormal publishing result, and the second value is used to indicate that the corresponding publishing result is an abnormal publishing result.

[0251] The specific embodiments can refer to the examples shown in the training method of the prediction model described above, which will not be repeated here in this example.

[0252] As an optional solution, the fifth acquisition sub-module includes:

[0253] The calculation subunit is configured to multiply the first historical actual publishing result sequence and data at the same position in the third value sequence to obtain a second abnormal publishing result sequence.

[0254] The specific embodiments can refer to the examples shown in the training method of the prediction model described above, which will not be repeated here in this example.

[0255] As an optional solution, the second training unit includes:

[0256] The sixth determination module is configured to determine features of the first multimedia resource in multiple dimensions to obtain a first feature set according to the first media resource information and the first target abnormal publishing result sequence.

[0257] The seventh determination module is configured to determine a first future predicted publishing result of the first multimedia resource at the target time according to the first feature set.

[0258] The adjustment module is configured to adjust parameters in the publishing result prediction model to obtain a first publishing result prediction model to be trained when a value of a target loss function determined according to the first future predicted publishing result and the first future actual publishing result does not satisfy a preset condition.

[0259] The specific embodiments can refer to the examples shown in the training method of the prediction model described above, which will not be repeated here in this example.

[0260] As an optional solution, the sixth determining module comprises at least two of the following steps:

[0261] The first determining sub-module is configured to determine the text feature of the first multimedia resource according to the text information of the first multimedia resource, wherein the first media resource information comprises the text information of the first multimedia resource, and the first feature set comprises the text feature of the first multimedia resource.

[0262] The second determining sub-module is configured to determine the visual feature of the first multimedia resource according to the picture information of the first multimedia resource, wherein the first media resource information comprises the picture information of the first multimedia resource, and the first feature set comprises the visual feature of the first multimedia resource.

[0263] The third determining sub-module is configured to determine the sequence feature of the first multimedia resource according to the first target abnormal publishing result sequence, wherein the first feature set comprises the sequence feature of the first multimedia resource.

[0264] The fourth determining sub-module is configured to determine the attribute feature of the first multimedia resource according to the attribute information of the first multimedia resource, wherein the first media resource information comprises the attribute information of the first multimedia resource, and the first feature set comprises the attribute feature of the first multimedia resource.

[0265] The specific embodiments can refer to the examples shown in the training method of the prediction model, which will not be described here in this example.

[0266] As an optional solution, the device further comprises:

[0267] The first obtaining module is configured to obtain the second media resource information of the published second multimedia resource, the second historical actual publishing result sequence of the second multimedia resource, and the second future actual publishing result of the second multimedia resource at the target time, wherein the second historical actual publishing result sequence comprises a group of actual publishing results of the second multimedia resource within a first time period.

[0268] The second obtaining module is configured to obtain the second target conversion probability randomly generated by the sequence conversion module, wherein the second target conversion probability is used to indicate whether to convert the second historical actual publishing result sequence into the second target abnormal publishing result sequence, and at least part of the publishing results in the second target abnormal publishing result sequence are abnormal publishing results.

[0269] The first training module is used to convert the second historical actual release result sequence into the second target abnormal release result sequence when the second target conversion probability representation converts the second historical actual release result sequence into the second target abnormal release result sequence, and uses the second media resource information, the second target abnormal release result sequence and the second future actual release result to train the first release result prediction model, wherein the first release result prediction model is used to generate the second future predicted release result of the second multimedia resource at the target time;

[0270] The second training module is used to train the first release result prediction model using the second media resource information, the second historical actual release result sequence, and the second future actual release result when the second target conversion probability representation does not convert the second historical actual release result sequence into the second target abnormal release result sequence.

[0271] For specific implementation examples, please refer to the examples shown in the training method of the prediction model above, which will not be repeated here.

[0272] As an optional solution, the second training unit 1308 includes:

[0273] The third training module is used to train the sharing rate prediction model to be trained using the first media resource information, the first historical actual release result sequence, and the first future actual release result. The release result prediction model is a sharing rate prediction model, which is used to generate the first future predicted sharing rate of the first multimedia resource at the target time based on the first media resource information and the first historical actual release result sequence. The first future predicted sharing rate represents the predicted number of times the first multimedia resource is shared divided by the number of times it is exposed. The first future actual release result is the first future actual sharing rate of the first multimedia resource at the target time, which represents the actual number of times the first multimedia resource is shared divided by the number of times it is exposed.

[0274] For specific implementation examples, please refer to the examples shown in the training method of the prediction model above, which will not be repeated here.

[0275] According to another aspect of the present invention, an apparatus for predicting the publication results of implementing the above-described method for predicting publication results is also provided. For example... Figure 14 As shown, the device includes:

[0276] The third acquisition unit 1402 is used to acquire the target media resource information of the published target multimedia resource and the target historical actual release result sequence of the target multimedia resource, wherein the target historical actual release result sequence includes a set of actual release results of the target multimedia resource within the target time period.

[0277] Input unit 1404 is used to input target media resource information and the target's historical actual release result sequence into the target release result prediction model to obtain the target multimedia resource's future predicted release result in the target's future time.

[0278] The target release result prediction model is a model obtained by training the release result prediction model to be trained using the target sample set. Each sample in the target sample set includes: media resource information of the released multimedia resources, the future actual release result of the multimedia resources at the target time; and one of the following sequences: the historical actual release result sequence of the multimedia resources, the first abnormal release result sequence, and the second abnormal release result sequence.

[0279] The historical actual release result sequence includes a set of actual release results of multimedia resources within a time period, with the target time being later than a time period; the first abnormal release result sequence is a sequence converted from the historical actual release result sequence, and all release results in the first abnormal release result sequence are abnormal release results; the second abnormal release result sequence is a sequence converted from the historical actual release result sequence, and some release results in the second abnormal release result sequence are abnormal release results.

[0280] Optionally, in this embodiment, media resource information may be used, but is not limited to, to represent basic information of multimedia resources, such as text information, image information, attribute information, etc., and the basic information is usually fixed or remains unchanged over a long period of time.

[0281] The historical actual release result sequence can be used, but is not limited to, to represent multiple sets of information that have changed or remained unchanged within a historical period. For example, if the historical period is divided into 10 time points, the historical actual release result sequence can be, but is not limited to, a set of information composed of the actual release results corresponding to each time point.

[0282] Optionally, in this embodiment, the historical actual release result sequence may include, but is not limited to, normal release result sequences and / or abnormal release result sequences. Generally, the probability of the historical actual release result sequence including a normal release result sequence is much higher than that of the historical actual release result sequence including an abnormal release result sequence. Therefore, in related technologies, the historical actual release result sequence is directly used to train the release result prediction model. However, the trained release result prediction model depends on the normal release result sequence, and thus cannot guarantee high prediction accuracy when facing abnormal release result sequences.

[0283] Based on this, this embodiment is configured with multiple conversion modes. The target conversion probability is randomly generated by the sequence conversion module to convert the historical actual release result sequence. For example, the historical actual release result sequence is converted into an abnormal release result sequence. Furthermore, the abnormal release result sequence is used to train the release result prediction model, thereby reducing the dependence of the trained release result prediction model on the normal release result sequence, so as to ensure high prediction accuracy when facing abnormal release result sequences.

[0284] Optionally, in this embodiment, when converting the first historical actual release result sequence into the first target abnormal release result sequence, the first media resource information of the first multimedia resource may be retained, but is not limited to. Then, the first media resource information, the first target abnormal release result sequence, and the first future actual release result are used to train the release result prediction model to be trained. This can be understood as converting the first historical actual release result sequence into the first target abnormal release result sequence while retaining the first media resource information of the first multimedia resource, so that the release result prediction model pays more attention to the first media resource information rather than the first historical actual release result sequence during the training process, thereby reducing the dependence of the release result prediction model on the historical actual release result sequence.

[0285] In summary, during the training process of the above-mentioned release result prediction model, the transformation may be limited to the historical actual release result sequence, while the media resource information of multimedia resources and the future actual release results at the target time remain unchanged.

[0286] Optionally, in this embodiment, the relationship between the historical actual release result sequence and the future actual release result can be understood, but is not limited to, as follows: taking the first published multimedia resource as an example, a first time period and a target time are selected during the time period in which the first multimedia resource has been published, and the target time is after the first time period;

[0287] Based on this, since both the first time period and the target time are the times when the first multimedia resource has been published, the actual publication results of the first multimedia resource in the first time period and the target time can be obtained directly, namely the historical actual publication result sequence and the future actual publication result. The time difference between the first time period and the target time can be, but is not limited to, fixed. Assuming that the time difference between the first time period and the target time is the target time difference, the target time difference can be, but is not limited to, used to determine the time of publication result after the publication result prediction model predicts the content information of the historical time. For example, assuming that the target time difference is 1 month, after the trained publication result prediction model predicts the content information published in the [T1~T2] time period, the output publication result is the predicted sharing rate of the content information 1 month after T2.

[0288] Further optionally, in this embodiment, for the training process of the release result prediction model, the historical actual release result sequence and the future actual release result can be, but are not limited to, a correspondence. The specific historical actual release result sequence or the transformed historical actual release result sequence (target abnormal release result sequence) is used as the input of the release result prediction model, while the future actual release result is used as the comparison object for the output of the release result prediction model. By comparing with the comparison object, the loss of the release result prediction model in this training is determined. If the convergence condition is met, the training ends. If the convergence condition is not met, the model parameters in the release result prediction model are adjusted according to the loss, and the next training begins based on the release result prediction model after adjusting the model parameters.

[0289] It should be noted that training the release result prediction model using historical release result sequences enables it to predict the sharing rate of normal data. By randomly generating target conversion probabilities through the sequence transformation module, the historical release result sequences are converted into abnormal release result sequences. Further training the release result prediction model using these abnormal release result sequences enables it to predict the sharing rate of abnormal data. Based on this, the trained release result prediction model possesses the ability to predict the sharing rate of both normal and abnormal data, thus maintaining high accuracy in sharing rate prediction regardless of whether the data is normal or abnormal.

[0290] For specific implementation examples, please refer to the examples shown in the training method of the prediction model above, which will not be repeated here.

[0291] The embodiments provided in this application obtain target media resource information of published target multimedia resources and a target historical actual release result sequence of target multimedia resources. The target historical actual release result sequence includes a set of actual release results of the target multimedia resources within a target time period. The target media resource information and the target historical actual release result sequence are input into the target release result prediction model to obtain the future predicted release result of the target multimedia resources at the target future time. This achieves the goal of maintaining high sharing rate prediction accuracy of the release result prediction model regardless of whether it is facing normal or abnormal data, and realizes the effect of improving the prediction accuracy of the prediction model for sharing rate.

[0292] As an optional solution, the third acquisition unit 1402 includes: a third acquisition module, used to acquire the target media resource information of the published target multimedia resource and the target historical actual sharing rate sequence of the target multimedia resource, wherein the target historical actual publishing result sequence includes a set of actual sharing rates of the target multimedia resource within the target time period, and the actual sharing rate represents the actual number of times the target multimedia resource is shared divided by the number of times it is exposed.

[0293] For specific implementation examples, please refer to the examples shown in the training method of the prediction model above, which will not be repeated here.

[0294] As an optional solution, the input unit 1404 includes: an input module, used to input the target media resource information and the target historical actual sharing rate sequence into the target release result prediction model to obtain the future predicted sharing rate of the target multimedia resource in the target future time, wherein the future predicted sharing rate represents the predicted number of times the target multimedia resource is shared divided by the number of times it is exposed.

[0295] For specific implementation examples, please refer to the examples shown in the training method of the prediction model above, which will not be repeated here.

[0296] According to another aspect of the present invention, an electronic device for implementing the training method of the above-described prediction model is also provided, such as... Figure 15 As shown, the electronic device includes a memory 1502 and a processor 1504. The memory 1502 stores a computer program, and the processor 1504 is configured to execute the steps of any of the above method embodiments via the computer program.

[0297] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0298] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0299] S1, obtain the first media resource information of the published first multimedia resource, the first historical actual release result sequence of the first multimedia resource, and the first future actual release result of the first multimedia resource at the target time, wherein the first historical actual release result sequence includes a set of actual release results of the first multimedia resource within a first time period, and the target time is later than the first time period;

[0300] S2, obtain the first target conversion probability randomly generated by the sequence conversion module, wherein the first target conversion probability is used to indicate whether to convert the first historical actual release result sequence into the first target abnormal release result sequence, and at least some of the release results in the first target abnormal release result sequence are abnormal release results;

[0301] S3, when the first target conversion probability representation transforms the first historical actual release result sequence into the first target abnormal release result sequence, the first historical actual release result sequence is transformed into the first target abnormal release result sequence, and the first media resource information, the first target abnormal release result sequence and the first future actual release result are used to train the release result prediction model to be trained, wherein the release result prediction model to be trained is used to generate the first future predicted release result of the first multimedia resource at the target time;

[0302] S4, when the first target conversion probability indicates that the first historical actual release result sequence will not be converted into the first target abnormal release result sequence, the first media resource information, the first historical actual release result sequence, and the first future actual release result are used to train the release result prediction model to be trained; or,

[0303] S1, obtain the target media resource information of the published target multimedia resource, and the target historical actual release result sequence of the target multimedia resource, wherein the target historical actual release result sequence includes a set of actual release results of the target multimedia resource within the target time period;

[0304] S2, input the target media resource information and the target's historical actual release result sequence into the target release result prediction model to obtain the target multimedia resource's future predicted release result in the target's future time;

[0305] The target release result prediction model is a model obtained by training the release result prediction model to be trained using the target sample set. Each sample in the target sample set includes: media resource information of the released multimedia resources, the future actual release result of the multimedia resources at the target time; and one of the following sequences: the historical actual release result sequence of the multimedia resources, the first abnormal release result sequence, and the second abnormal release result sequence.

[0306] The historical actual release result sequence includes a set of actual release results of multimedia resources within a time period, with the target time being later than a time period; the first abnormal release result sequence is a sequence converted from the historical actual release result sequence, and all release results in the first abnormal release result sequence are abnormal release results; the second abnormal release result sequence is a sequence converted from the historical actual release result sequence, and some release results in the second abnormal release result sequence are abnormal release results.

[0307] Alternatively, as those skilled in the art will understand, Figure 15 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones (such as Android phones, iOS phones, etc.), tablets, PDAs, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 15 This does not limit the structure of the aforementioned electronic devices. For example, the electronic device may also include components that are more... Figure 15 The more or fewer components shown (such as network interfaces, etc.), or having the same Figure 15 The different configurations shown.

[0308] The memory 1502 can be used to store software programs and modules, such as the program instructions / modules corresponding to the training method and apparatus for the prediction model in this embodiment of the invention. The processor 1504 executes various functional applications and data processing by running the software programs and modules stored in the memory 1502, thereby implementing the above-mentioned training method for the prediction model. The memory 1502 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1502 may further include memory remotely located relative to the processor 1504, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Specifically, the memory 1502 may be used, but is not limited to, to store information such as first media resource information, first historical actual release result sequence, first future actual release result, and first target abnormal release result sequence. As an example, such as Figure 15 As shown, the memory 1502 may include, but is not limited to, the first acquisition unit 1302, the second acquisition unit 1304, the first training unit 1306, and the second training unit 1308, or the third acquisition unit 1402 (not shown) and the input unit 1404 (not shown) of the training device for the prediction model. Furthermore, it may include, but is not limited to, other module units in the training device for the prediction model, which will not be elaborated upon in this example.

[0309] Optionally, the transmission device 1506 described above is used to receive or send data via a network. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 1506 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 1506 is a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0310] In addition, the aforementioned electronic device also includes: a display 1508 for displaying information such as the first media resource information, the first historical actual release result sequence, the first future actual release result, and the first target abnormal release result sequence; and a connection bus 1510 for connecting the various module components in the aforementioned electronic device.

[0311] In other embodiments, the aforementioned terminal device or server can be a node in a distributed system, wherein the distributed system can be a blockchain system, which is a distributed system formed by connecting multiple nodes through network communication. The nodes can form a peer-to-peer (P2P) network, and any form of computing device, such as a server, terminal, or other electronic device, can become a node in the blockchain system by joining this peer-to-peer network.

[0312] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the training method for the prediction model and the prediction method for publishing results described above. The computer program is configured to execute the steps in any of the above method embodiments at runtime.

[0313] Optionally, in this embodiment, the computer-readable storage medium described above may be configured to store a computer program for performing the following steps:

[0314] S1, obtain the first media resource information of the published first multimedia resource, the first historical actual release result sequence of the first multimedia resource, and the first future actual release result of the first multimedia resource at the target time, wherein the first historical actual release result sequence includes a set of actual release results of the first multimedia resource within a first time period, and the target time is later than the first time period;

[0315] S2, obtain the first target conversion probability randomly generated by the sequence conversion module, wherein the first target conversion probability is used to indicate whether to convert the first historical actual release result sequence into the first target abnormal release result sequence, and at least some of the release results in the first target abnormal release result sequence are abnormal release results;

[0316] S3, when the first target conversion probability representation transforms the first historical actual release result sequence into the first target abnormal release result sequence, the first historical actual release result sequence is transformed into the first target abnormal release result sequence, and the first media resource information, the first target abnormal release result sequence and the first future actual release result are used to train the release result prediction model to be trained, wherein the release result prediction model to be trained is used to generate the first future predicted release result of the first multimedia resource at the target time;

[0317] S4, when the first target conversion probability indicates that the first historical actual release result sequence will not be converted into the first target abnormal release result sequence, the first media resource information, the first historical actual release result sequence, and the first future actual release result are used to train the release result prediction model to be trained; or,

[0318] S1, obtain the target media resource information of the published target multimedia resource, and the target historical actual release result sequence of the target multimedia resource, wherein the target historical actual release result sequence includes a set of actual release results of the target multimedia resource within the target time period;

[0319] S2, input the target media resource information and the target's historical actual release result sequence into the target release result prediction model to obtain the target multimedia resource's future predicted release result in the target's future time;

[0320] The target release result prediction model is a model obtained by training the release result prediction model to be trained using the target sample set. Each sample in the target sample set includes: media resource information of the released multimedia resources, the future actual release result of the multimedia resources at the target time; and one of the following sequences: the historical actual release result sequence of the multimedia resources, the first abnormal release result sequence, and the second abnormal release result sequence.

[0321] The historical actual release result sequence includes a set of actual release results of multimedia resources within a time period, with the target time being later than a time period; the first abnormal release result sequence is a sequence converted from the historical actual release result sequence, and all release results in the first abnormal release result sequence are abnormal release results; the second abnormal release result sequence is a sequence converted from the historical actual release result sequence, and some release results in the second abnormal release result sequence are abnormal release results.

[0322] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0323] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0324] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more computer devices (which may be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention.

[0325] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0326] In the several embodiments provided in this application, it should be understood that the disclosed client can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0327] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0328] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0329] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for training a prediction model, characterized in that, The method comprises: obtaining first media resource information of a published first multimedia resource, a first historical actual publishing result sequence of the first multimedia resource, and a first future actual publishing result of the first multimedia resource at a target time, wherein the first historical actual publishing result sequence comprises a set of actual publishing results of the first multimedia resource within a first time period, and the target time is later than the first time period; obtaining a first target conversion probability randomly generated by a sequence conversion module, wherein the first target conversion probability is used to indicate whether the first historical actual publishing result sequence is converted into a first target abnormal publishing result sequence, and at least part of publishing results in the first target abnormal publishing result sequence are abnormal publishing results; when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence, converting the first historical actual publishing result sequence into the first target abnormal publishing result sequence, and using the first media resource information, the first target abnormal publishing result sequence, and the first future actual publishing result to train a to-be-trained publishing result prediction model, wherein the to-be-trained publishing result prediction model is used to generate a first future predicted publishing result of the first multimedia resource at the target time; when the first target conversion probability indicates that the first historical actual publishing result sequence is not converted into the first target abnormal publishing result sequence, using the first media resource information, the first historical actual publishing result sequence, and the first future actual publishing result to train the to-be-trained publishing result prediction model.

2. The method of claim 1, wherein, The converting the first historical actual publishing result sequence into the first target abnormal publishing result sequence when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence comprises: when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into a first abnormal publishing result sequence, converting the first historical actual publishing result sequence into the first abnormal publishing result sequence, wherein the first target abnormal publishing result sequence comprises the first abnormal publishing result sequence, and all publishing results in the first abnormal publishing result sequence are abnormal publishing results; or when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into a second abnormal publishing result sequence, converting the first historical actual publishing result sequence into the second abnormal publishing result sequence, wherein the first target abnormal publishing result sequence comprises the second abnormal publishing result sequence, and part of publishing results in the second abnormal publishing result sequence are abnormal publishing results.

3. The method of claim 2, wherein, The method further comprises: when a value of the first target conversion probability is located in a first value interval, determining that the first target conversion probability indicates that the first historical actual publishing result sequence is not converted into the first target abnormal publishing result sequence; When the first target conversion probability is in the second value interval, it is determined that the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence or the second abnormal publishing result sequence.

4. The method of claim 1, wherein, The converting the first historical actual publishing result sequence into the first target abnormal publishing result sequence when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first target abnormal publishing result sequence comprises: The converting the first historical actual publishing result sequence into the first abnormal publishing result sequence when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence, wherein the first target abnormal publishing result sequence comprises the first abnormal publishing result sequence, and each publishing result in the first abnormal publishing result sequence is an abnormal publishing result; The converting the first historical actual publishing result sequence into the second abnormal publishing result sequence when the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence, wherein the first target abnormal publishing result sequence comprises the second abnormal publishing result sequence, and part of the publishing results in the second abnormal publishing result sequence are abnormal publishing results.

5. The method of claim 2, wherein, The method further comprises: When the first target conversion probability is in the third value interval, it is determined that the first target conversion probability indicates that the first historical actual publishing result sequence is not converted into the first target abnormal publishing result sequence; When the first target conversion probability is in the fourth value interval, it is determined that the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the first abnormal publishing result sequence; When the first target conversion probability is in the fifth value interval, it is determined that the first target conversion probability indicates that the first historical actual publishing result sequence is converted into the second abnormal publishing result sequence.

6. The method according to any one of claims 2 to 5, characterized in that, The converting the first historical actual publishing result sequence into the first abnormal publishing result sequence comprises: determining a same value for each historical actual publishing result in the first historical actual publishing result sequence, and obtaining a first value sequence; multiplying the first historical actual publishing result sequence and data at the same position in the first value sequence to obtain the first abnormal publishing result sequence.

7. The method according to any one of claims 2 to 5, characterized in that, The converting the first historical actual publishing result sequence into the second abnormal publishing result sequence comprises: randomly determining a value for each historical actual publishing result in the first historical actual publishing result sequence to obtain a second value sequence; determining a value of a corresponding mark for each historical actual publishing result in the first historical actual publishing result sequence according to each value in the second value sequence and a preset target threshold to obtain a third value sequence, wherein the value of the mark is used to indicate whether the corresponding publishing result is an abnormal publishing result; and determining a value of a corresponding mark for each historical actual publishing result in the first historical actual publishing result sequence according to each value in the second value sequence and a preset target threshold to obtain a third value sequence, wherein the value of the mark is used to indicate whether the corresponding publishing result is an abnormal publishing result. According to the first historical actual publishing result sequence and the third numerical sequence, the second abnormal publishing result sequence is determined.

8. The method of claim 7, wherein, According to each numerical value in the second numerical sequence and a preset target threshold, a value of a corresponding mark of each historical actual publishing result in the first historical actual publishing result sequence is determined, and a third numerical sequence is obtained, including: When the i th numerical value in the second numerical sequence is greater than the target threshold, a value of a mark corresponding to the i th historical actual publishing result in the first historical actual publishing result sequence is set to a first value, wherein the second numerical sequence includes N numerical values, 1≤i≤N, and N is a natural number greater than 1; When the i th numerical value in the second numerical sequence is less than or equal to the target threshold, a value of a mark corresponding to the i th historical actual publishing result in the first historical actual publishing result sequence is set to a second value; The first value is used to indicate that the corresponding publishing result is an abnormal publishing result, and the second value is used to indicate that the corresponding publishing result is not an abnormal publishing result, or The first value is used to indicate that the corresponding publishing result is not an abnormal publishing result, and the second value is used to indicate that the corresponding publishing result is an abnormal publishing result.

9. The method of claim 7, wherein, According to the first historical actual publishing result sequence and the third numerical sequence, the second abnormal publishing result sequence is determined, including: The data at the same position in the first historical actual publishing result sequence and the third numerical sequence are multiplied to obtain the second abnormal publishing result sequence.

10. The method according to any one of claims 1 to 5, characterized in that, The first media resource information, the first target abnormal publishing result sequence, and the first future actual publishing result are used to train the publishing result prediction model, including: According to the first media resource information and the first target abnormal publishing result sequence, features of the first multimedia resource in multiple dimensions are determined, and a first feature set is obtained; According to the first feature set, a first future predicted publishing result of the first multimedia resource at the target time is determined; When a value of a target loss function determined according to the first future predicted publishing result and the first future actual publishing result does not satisfy a preset condition, parameters in the publishing result prediction model are adjusted to obtain a first publishing result prediction model to be trained.

11. The method of claim 10, wherein, According to the first media resource information and the first target abnormal publishing result sequence, features of the first multimedia resource in multiple dimensions are determined, and a first feature set is obtained, including at least two of the following steps: According to text information of the first multimedia resource, a text feature of the first multimedia resource is determined, wherein the first media resource information includes text information of the first multimedia resource, and the first feature set includes the text feature of the first multimedia resource; determining visual features of the first multimedia resource according to picture information of the first multimedia resource, wherein the first media resource information comprises the picture information of the first multimedia resource, and the first feature set comprises the visual features of the first multimedia resource; determining sequence features of the first multimedia resource according to the first target abnormal publishing result sequence, wherein the first feature set comprises the sequence features of the first multimedia resource; determining attribute features of the first multimedia resource according to attribute information of the first multimedia resource, wherein the first media resource information comprises the attribute information of the first multimedia resource, and the first feature set comprises the attribute features of the first multimedia resource.

12. The method of claim 10, wherein, The method further comprises: obtaining second media resource information of a published second multimedia resource, a second historical actual publishing result sequence of the second multimedia resource, and a second future actual publishing result of the second multimedia resource at the target time, wherein the second historical actual publishing result sequence comprises a set of actual publishing results of the second multimedia resource within the first time period; obtaining a second target conversion probability randomly generated by the sequence conversion module, wherein the second target conversion probability is used to indicate whether the second historical actual publishing result sequence is converted into a second target abnormal publishing result sequence, and at least part of publishing results in the second target abnormal publishing result sequence are abnormal publishing results; when the second target conversion probability indicates that the second historical actual publishing result sequence is converted into the second target abnormal publishing result sequence, converting the second historical actual publishing result sequence into the second target abnormal publishing result sequence, and training the first publishing result prediction model using the second media resource information, the second target abnormal publishing result sequence, and the second future actual publishing result, wherein the first publishing result prediction model is used to generate a second future predicted publishing result of the second multimedia resource at the target time; when the second target conversion probability indicates that the second historical actual publishing result sequence is not converted into the second target abnormal publishing result sequence, training the first publishing result prediction model using the second media resource information, the second historical actual publishing result sequence, and the second future actual publishing result.

13. The method according to any one of claims 1 to 5, characterized in that, The training of the to-be-trained publishing result prediction model using the first media resource information, the first historical actual publishing result sequence, and the first future actual publishing result comprises: The sharing rate prediction model is trained using the first media resource information, the first historical actual publishing result sequence, and the first future actual publishing result, wherein the publishing result prediction model is the sharing rate prediction model, the sharing rate prediction model is used to generate a first future predicted sharing rate of the first multimedia resource at the target time according to the first media resource information and the first historical actual publishing result sequence, the first future predicted sharing rate represents a predicted number of times of sharing of the first multimedia resource divided by a number of times of exposure, and the first future actual publishing result is a first future actual sharing rate of the first multimedia resource at the target time, the first future actual sharing rate representing an actual number of times of sharing of the first multimedia resource divided by a number of times of exposure.

14. A computer readable storage medium, characterized in that, The computer readable storage medium comprises a stored program, wherein the program performs the method described in any one of claims 1 to 13 when executed.

15. An electronic device comprising a memory and a processor, characterized in that The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 13 by using the computer program.

Images