Content interaction prediction method and related devices

Abstract

The application discloses a content interaction prediction method and related equipment, and related embodiments can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation and auxiliary driving; the target content can be subjected to feature extraction on an interactive operation, to obtain a target first operation feature and a target sharing operation feature of the target content on a first interactive operation; an initial prediction interaction result of the target content on the first interactive operation is determined based on the target first operation feature; the target sharing operation feature is subjected to logistic regression analysis to predict negative noise information generated by a second interactive operation on the first interactive operation; and the initial prediction interaction result is corrected based on the negative noise information to obtain a target prediction interaction result. The application can correct the prediction interaction result of the target content on the first interactive operation by capturing the negative noise information generated by the second interactive operation on the first interactive operation, thereby improving the prediction accuracy of the interaction result on the first interactive operation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a content interaction prediction method and related equipment. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, more and more application scenarios use artificial intelligence technology to recommend personalized content to users to improve the user's interaction experience.

[0003] In the process of recommending content to users, the related technology generally predicts the click probability and reading time of each candidate content through a multi-task model, and then selects the target content from the candidate content according to the predicted click probability and reading time to recommend to the user. However, the current multi-task model generally pays too much attention to the task with rich supervision signals (such as the prediction of click probability), so that the prediction effect on the task with less supervision signals (such as the prediction of reading time) is poor. SUMMARY

[0004] The embodiments of the present application provide a content interaction prediction method and related equipment, which can include a content interaction prediction device, an electronic device, a computer readable storage medium and a computer program product, and can improve the prediction accuracy of the interaction result on the first interaction operation.

[0005] The embodiments of the present application provide a content interaction prediction method, which includes:

[0006] Obtain a target content; and perform feature extraction on the target content on an interaction operation to obtain a target first operation feature and a target sharing operation feature of the target content on the first interaction operation, the target sharing operation feature being a feature shared by the first interaction operation and a second interaction operation, the second interaction operation being a preceding operation dependent on the first interaction operation;

[0007] Determine an initial predicted interaction result of the target content on the first interaction operation based on the target first operation feature;

[0008] Perform logistic regression analysis on the target sharing operation feature to predict negative noise information generated by the second interaction operation on the first interaction operation;

[0009] Based on the negative noise information, correct the initial predicted interaction result to obtain a target predicted interaction result of the target content on the first interaction operation.

[0010] Correspondingly, the embodiments of the present application provide a content interaction prediction device, which includes:

[0011] An acquisition unit is configured to acquire target content, and perform feature extraction on the target content in interactive operations to obtain target first operation features of the target content in a first interactive operation and target shared operation features, the target shared operation features being features shared by the first interactive operation and a second interactive operation, the second interactive operation being a preceding operation dependent on the first interactive operation;

[0012] A determination unit is configured to determine an initial predicted interaction result of the target content in the first interactive operation based on the target first operation features.

[0013] A prediction unit is configured to perform logistic regression analysis on the target shared operation features to predict negative noise information generated by the second interactive operation on the first interactive operation.

[0014] A correction unit is configured to correct the initial predicted interaction result based on the negative noise information to obtain a target predicted interaction result of the target content in the first interactive operation.

[0015] Optionally, in some embodiments of the present application, the acquisition unit can be specifically configured to perform feature extraction on the target content in interactive operations to obtain target first operation features of the target content in a first interactive operation, target second operation features of the target content in a second interactive operation, and target shared operation features.

[0016] The content interaction prediction apparatus can further include a result determination unit, which is configured as follows:

[0017] The result determination unit is configured to determine a predicted interaction result of the target content in the second interactive operation based on the target second operation features.

[0018] Optionally, in some embodiments of the present application, the acquisition unit can include an extraction subunit and an interaction subunit, which are configured as follows:

[0019] The extraction subunit is configured to extract first operation features of the target content in a first interactive operation, second operation features of the target content in a second interactive operation, and shared operation features.

[0020] The interaction subunit is configured to perform feature interaction processing on the first operation features, the second operation features, and the shared operation features to obtain target first operation features corresponding to the first interactive operation, target second operation features corresponding to the second interactive operation, and target shared operation features.

[0021] Optionally, in some embodiments of the present application, the interaction subunit can be specifically configured to fuse the first operation feature and the shared operation feature, and update the first operation feature based on the fused feature; fuse the second operation feature and the shared operation feature, and update the second operation feature based on the fused feature; fuse the first operation feature, the second operation feature and the shared operation feature, and update the shared operation feature based on the fused feature; return to the step of fusing the first operation feature and the shared operation feature, and updating the first operation feature based on the fused feature, until the target shared operation feature meeting the preset feature interaction condition, the target first operation feature corresponding to the first interaction operation and the target second operation feature corresponding to the second interaction operation are obtained.

[0022] Optionally, in some embodiments of the present application, the determination unit comprises a fusion subunit and a determination subunit, as follows:

[0023] The fusion subunit is configured to fuse the target shared operation feature and the target first operation feature to obtain a first fused operation feature.

[0024] The determination subunit is configured to determine an initial predicted interaction result of the target content on the first interaction operation based on the first fused operation feature.

[0025] Optionally, in some embodiments of the present application, the acquisition unit can be specifically configured to perform feature extraction on the target content on the interaction operation by using the content interaction prediction model, to obtain the target first operation feature and the target shared operation feature of the target content on the first interaction operation.

[0026] Optionally, in some embodiments of the present application, the content interaction prediction device can further comprise a training unit configured to train the content interaction prediction model; specifically, the training unit can comprise a data acquisition subunit, a feature extraction subunit, an interaction result determination subunit, an analysis subunit and an adjustment subunit, as follows:

[0027] The data acquisition subunit is configured to acquire training data, wherein the training data comprises sample content, a first expected interaction result of the sample content on the first interaction operation, and a second expected interaction result on the second interaction operation;

[0028] The feature extraction subunit is configured to perform feature extraction on the sample content on the interaction operation by using the content interaction prediction model, to obtain a target first operation feature of the sample content on the first interaction operation, a target second operation feature on the second interaction operation, and a target shared operation feature.

[0029] an interaction result determination sub-unit, configured to determine an initial first actual interaction result of the sample content on the first interaction operation and a second actual interaction result of the sample content on the second interaction operation based on the target first operation feature and the target second operation feature respectively;

[0030] an analysis sub-unit, configured to perform a logistic regression analysis on the target shared operation feature to predict actual negative noise information of the second interaction operation to the first interaction operation, and correct the initial first actual interaction result based on the actual negative noise information to obtain a target first actual interaction result;

[0031] an adjustment sub-unit, configured to adjust parameters of the content interaction prediction model according to the initial first actual interaction result, the target first actual interaction result, the first expected interaction result, the second actual interaction result and the second expected interaction result to obtain a trained content interaction prediction model.

[0032] Optionally, in some embodiments of the present application, the adjustment sub-unit can be specifically configured to calculate a first loss value between the initial first actual interaction result and the first expected interaction result, calculate a second loss value between the target first actual interaction result and the first expected interaction result, calculate a third loss value between the second actual interaction result and the second expected interaction result, and adjust the parameters of the content interaction prediction model according to the first loss value, the second loss value and the third loss value to obtain the trained content interaction prediction model.

[0033] Optionally, in some embodiments of the present application, the content interaction prediction model comprises a basic prediction module and a negative noise modeling module; and the step of adjusting the parameters of the content interaction prediction model according to the first loss value, the second loss value and the third loss value to obtain the trained content interaction prediction model can comprise:

[0034] adjusting the parameters of the content interaction prediction model according to the first loss value, the second loss value and the third loss value to obtain a pre-trained content interaction prediction model;

[0035] performing mask processing on the sample content by the basic prediction module in the pre-trained content interaction prediction model to obtain target content information meeting a preset correlation condition in terms of relevance to the second interaction operation; and performing feature extraction on the target content information in terms of interaction operation to obtain target reinforced shared operation features of the sample content;

[0036] The negative noise modeling module in the pre-trained content interaction prediction model is used to predict actual reinforcement negative noise information of the second interaction operation on the first interaction operation based on the target reinforcement shared operation feature.

[0037] The parameters of the negative noise modeling module are adjusted based on the actual reinforcement negative noise information and the first expected interaction result, to obtain a trained content interaction prediction model.

[0038] Optionally, before the step of adjusting the parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain a trained content interaction prediction model, the method can further include:

[0039] The basic prediction module in the pre-trained content interaction prediction model is used to perform feature extraction on the target content information in the interaction operation, to obtain target reinforcement first operation features of the sample content in the first interaction operation.

[0040] The initial actual reinforcement interaction result of the sample content in the first interaction operation is determined based on the target reinforcement first operation features.

[0041] The step of adjusting the parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain a trained content interaction prediction model can include:

[0042] The initial actual reinforcement interaction result is corrected based on the actual reinforcement negative noise information, to obtain target actual reinforcement interaction result of the sample content in the first interaction operation.

[0043] The parameters of the negative noise modeling module are adjusted based on the target actual reinforcement interaction result and the first expected interaction result, to obtain a trained content interaction prediction model.

[0044] An electronic device provided in an embodiment of the present application includes a processor and a memory. The memory stores a plurality of instructions. The processor loads the instructions to perform the steps in the content interaction prediction method provided in the embodiments of the present application.

[0045] An embodiment of the present application further provides a computer readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps in the content interaction prediction method provided in the embodiments of the present application are implemented.

[0046] Further, the embodiment of the present application further provides a computer program product, comprising a computer program or instructions, which, when executed by a processor, implement the steps of the content interaction prediction method provided by the embodiment of the present application.

[0047] The embodiment of the present application provides a content interaction prediction method and related equipment, which can acquire target content; and perform feature extraction on the target content in an interactive operation to obtain target first operation features of the target content in a first interactive operation and target sharing operation features, the target sharing operation features being features shared by the first interactive operation and a second interactive operation, the second interactive operation being a preceding operation dependent on the first interactive operation; determining an initial prediction interaction result of the target content in the first interactive operation based on the target first operation features; performing logistic regression analysis on the target sharing operation features to predict negative noise information generated by the second interactive operation on the first interactive operation; and correcting the initial prediction interaction result based on the negative noise information to obtain a target prediction interaction result of the target content in the first interactive operation. The present application can correct the prediction interaction result of the target content in the first interactive operation by capturing the negative noise information generated by the second interactive operation on the first interactive operation, thereby improving the prediction accuracy of the interaction result in the first interactive operation. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0049] Figure 1a is a scene schematic diagram of the content interaction prediction method provided by the embodiment of the present application;

[0050] Figure 1b is a flowchart of the content interaction prediction method provided by the embodiment of the present application;

[0051] Figure 1c is a model architecture diagram of the content interaction prediction method provided by the embodiment of the present application;

[0052] Figure 1d is another model architecture diagram of the content interaction prediction method provided by the embodiment of the present application;

[0053] Figure 1e is an explanatory diagram of the content interaction prediction method provided by the embodiment of the present application;

[0054] Figure 2is another flowchart of the content interaction prediction method provided by the embodiment of the present application;

[0055] Figure 3 is a structural schematic diagram of the content interaction prediction device provided by the embodiment of the present application;

[0056] Figure 4 is a structural schematic diagram of the electronic device provided by the embodiment of the present application. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person skilled in the art without creative work fall within the scope of protection of the present application.

[0058] The embodiment of the present application provides a content interaction prediction method and related devices, and the related devices can include a content interaction prediction device, an electronic device, a computer readable storage medium and a computer program product. The content interaction prediction device can be specifically integrated in an electronic device, and the electronic device can be a terminal or a server and the like.

[0059] It can be understood that the content interaction prediction method of the embodiment can be executed on a terminal, can be executed on a server, and can also be executed by the terminal and the server together. The above examples should not be understood as limiting the present application.

[0060] As shown in Figure 1a , taking the terminal and the server executing the content interaction prediction method together as an example. The content interaction prediction system provided by the embodiment of the present application includes a terminal 10 and a server 11 and the like; the terminal 10 and the server 11 are connected through a network, such as a wired or wireless network connection and the like, wherein the content interaction prediction device can be integrated in the server.

[0061] The server 11 can be configured to: acquire target content; and perform feature extraction on the target content in interactive operations to obtain target first operation features of the target content in a first interactive operation and target shared operation features, the target shared operation features being features shared by the first interactive operation and a second interactive operation, the second interactive operation being a preceding operation dependent on the first interactive operation; determine an initial predicted interaction result of the target content in the first interactive operation based on the target first operation features; perform logistic regression analysis on the target shared operation features to predict negative noise information generated by the second interactive operation on the first interactive operation; and correct the initial predicted interaction result based on the negative noise information to obtain a target predicted interaction result of the target content in the first interactive operation. The server 11 can be a single server, a server cluster or a cloud server composed of multiple servers. In the content interaction prediction method and device disclosed in the present application, multiple servers can be composed into a blockchain, and the servers are nodes on the blockchain.

[0062] The terminal 10 can be configured to receive the target predicted interaction result of the target content in the first interactive operation sent by the server 11 and perform content recommendation based on the target predicted interaction result. The terminal 10 can include a mobile phone, a smart voice interaction device, a smart home appliance, a vehicle-mounted terminal, an aircraft, a tablet computer, a notebook computer or a personal computer (PC, Personal Computer), etc. The terminal 10 can also be provided with a client, which can be an application program client or a browser client, etc.

[0063] The above steps of content interaction prediction performed by the server 11 can also be performed by the terminal 10.

[0064] The content interaction prediction method provided by the embodiments of the present application relates to natural language processing and machine learning in the field of artificial intelligence.

[0065] Artificial Intelligence (AI) is the use of digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is the design principle and implementation method of various intelligent machines, so that the machine has the functions of perception, reasoning and decision-making. Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software technologies. Among them, artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, automatic driving, intelligent transportation and other fields.

[0066] Nature Language processing (NLP) is an important direction in the field of computer science and artificial intelligence. It studies various theories and methods that can realize effective communication between people and computers using natural language. Natural language processing is a science that integrates linguistics, computer science and mathematics. Therefore, the research in this field will involve natural language, i.e. the language used in daily life, so it is closely related to the study of linguistics. Natural language processing technology usually includes text processing, semantic understanding, machine translation, robot question and answer, knowledge graph and other technologies.

[0067] Machine Learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It is a specialized research on how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0068] The following will be described in detail. It should be noted that the order of the following embodiments is not limited to the preferred order of the embodiments.

[0069] This embodiment will be described from the perspective of a content interaction prediction device, which can be integrated in an electronic device, which can be a server or a terminal or the like.

[0070] It can be understood that in the specific embodiments of the present application, user information such as reading time and related data are involved, and when the above embodiments of the present application are applied to specific products or technologies, user permission or consent needs to be obtained, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of the country and region.

[0071] The content interaction prediction method of the embodiments of the present application can be applied to scenarios such as content recommendation. The embodiments can be applied to various scenarios such as cloud technology, artificial intelligence, intelligent transportation, and assisted driving.

[0072] As shown in Figure 1b The specific process of the content interaction prediction method can be as follows:

[0073] 101、obtain target content; and perform feature extraction on the target content in an interactive operation to obtain target first operation features of the target content in a first interactive operation and target sharing operation features, the target sharing operation features being features shared by the first interactive operation and a second interactive operation, the second interactive operation being a preceding operation dependent on the first interactive operation.

[0074] The target content is content whose interaction result is to be predicted. The interaction result to be predicted can include an interaction result in the first interactive operation or an interaction result in the second interactive operation, which is not limited in the embodiments. It should be noted that the content form of the target content can include text, audio, image, and video, etc.

[0075] The second interactive operation is a preceding operation dependent on the first interactive operation, and the two operations have a dependent relationship in operation time. The operation time of the second interactive operation is before the operation time of the first interactive operation. For example, in a specific scenario, the target content is content recommended to a target object, the second interactive operation is a click operation on the content, and the first interactive operation is a reading time or a sharing operation on the content. The interaction result predicted in the second interactive operation can be whether the target object clicks the target content, and the interaction result predicted in the first interactive operation can be the reading time of the target object on the target content or whether the target object will share the target content. It can be understood that the target object needs to click the target content first before reading or sharing the target content, so the click is a preceding operation corresponding to the reading time or the sharing.

[0076] Since the operation times of the first interactive operation and the second interactive operation have a dependent relationship, if the same content interaction prediction model is used to predict the interaction results in the first interactive operation and the second interactive operation, the number of samples of the two operations will differ greatly in the model training process, resulting in that the first interactive operation has less supervision signal and the second interactive operation has more supervision signal.

[0077] Specifically, when the first interaction operation (such as reading duration) and the second interaction operation (such as click operation) are jointly trained, there are some task-specific problems:

[0078] First, the data sparsity problem of reading duration compared with click operation. For example, in some real recommendation systems, only a small part of exposed articles are clicked and read, which shows that the data collected for training the prediction task of reading duration is generally much lower than that of the click prediction task, which will cause a problem - the shared parameter part of the multi-task learning model (such as shared feature embedding and bottom layer shared expert layer) is mainly optimized under the supervision of the click signal. Specifically, the shared parameters account for more than 99% of the total parameters of the multi-task learning model, which shows that the optimization of the click task dominates in the multi-task model, and the optimization of the reading duration task is insufficient.

[0079] Second, the complex relationship and deep coupling between click operation and reading duration. The interaction operation of users and target content follows the behavior sequence mode of "exposure-click-reading", so there is a high dependence and serious coupling relationship between click operation and reading duration. Although the correlation between the two is very strong, there is still conflict between the two targets. For example, an article with an eye-catching title may have poor content, and the user closes the article soon after clicking it, which leads to high click but short reading duration of the article.

[0080] In the related art, the multi-task model for predicting click operation and reading duration often ignores the negative impact of click on reading duration. Due to the seesaw effect of the multi-task model - the model pays too much attention to the task with rich supervision signals (such as click prediction), which makes the effect on the task with less supervision signals (such as reading duration prediction) poor, that is, the prediction of the first interaction operation corresponding to the interaction result is not accurate, and the existing model cannot achieve good results on both click prediction and reading duration prediction.

[0081] The content interaction prediction method of the present application can provide a multi-task causal framework, which explicitly captures and removes this negative impact by introducing causal inference to improve the prediction effect of reading duration.

[0082] Among them, causal inference is a research field in statistics for analyzing the causal relationship between variables; in this embodiment, the first interaction operation depends on the second interaction operation, and the two can be regarded as having a causal relationship.

[0083] Optionally, in the embodiment, the step of "extracting features on the interaction operation of the target content to obtain target first operation features and target shared operation features of the target content in the first interaction operation" can include:

[0084] characteristics of the target content in the first interaction operation, target second operation characteristics of the target content in the second interaction operation, and target shared operation characteristics of the target content.

[0085] The target shared operation characteristics are characteristics shared by the first interaction operation and the second interaction operation; the target first operation characteristics can be regarded as unique characteristics of the first interaction operation, and the target second operation characteristics can be regarded as unique characteristics of the second interaction operation.

[0086] Optionally, in the embodiment, the step of “extracting characteristics of the target content in the interaction operations to obtain target first operation characteristics of the target content in the first interaction operation and target shared operation characteristics of the target content” can include:

[0087] extracting characteristics of the target content in the interaction operations by using a content interaction prediction model to obtain target first operation characteristics of the target content in the first interaction operation and target shared operation characteristics of the target content.

[0088] In the embodiment, the content interaction prediction model can be used to extract target first operation characteristics of the target content in the first interaction operation, target second operation characteristics of the target content in the second interaction operation, and target shared operation characteristics of the target content.

[0089] The content interaction prediction model is specifically a neural network model, which can be a residual network (ResNet) or a densely connected convolutional network (DenseNet), and it should be understood that the neural network of the embodiment is not limited to the above-mentioned several types.

[0090] Specifically, the content interaction prediction model is a multi-task model, which can include a multi-layer feature extraction network and a tower network corresponding to each task in the target multi-task; each layer of the feature extraction network includes a plurality of expert networks and a gating network corresponding to each task. For example, the content interaction prediction model can include a multi-layer feature extraction network and a tower network corresponding to both a prediction task of the first interaction operation and a prediction task of the second interaction operation.

[0091] The feature extraction network can be used to extract unique characteristics and shared characteristics of each task, and the tower network can be used to meet the specific application requirements of each task, such as a classification task or a prediction task. In each layer of the feature extraction network, a plurality of expert networks and a gating network corresponding to each task are provided, and the expert networks can include task-shared experts and task-specific experts.

[0092] In the multi-task learning process, gating networks can be used to control the weights of the expert network outputs. These output weights allow for the selection and control of the expert networks within the multi-task model. Gating networks can be used to fuse the feature data extracted by the expert networks. The expert network weights output by the gating networks differ for different tasks, resulting in variations in the fused feature data for each task at that level. Due to the weight allocation of the gating networks, different expert networks can learn different signals from different perspectives. The fused feature data for each task is then used as input to the next level of the feature extraction network for feature interaction, until finally input to the corresponding tower network for each task to obtain the processing results. In this content interaction prediction model, the gating network for each task is also multi-layered, with the number of layers matching that of the feature extraction network. In this embodiment, in scenarios involving multiple related tasks, both the specificity of the tasks and the correlation between them are distinguished, effectively improving the model's generalization ability and the accuracy of the processing results.

[0093] In this model, the expert network can be based on the same representation input, but adopt different network structures and parameters for different tasks. For the tower network of each task, the same network framework can be used, or different network frameworks can be used, making the multi-task model flexible and adaptable.

[0094] Multi-task learning (MTL) is a field within machine learning that combines multiple related tasks based on shared representation data; it is also a type of transfer learning. This approach learns relevant domain information from different tasks and then shares this information within the model. Through mutual learning and sharing among multiple tasks, the model's generalization ability and performance are improved.

[0095] In one specific embodiment, such as Figure 1c The diagram shows the model structure of a content interaction prediction model. This model consists of a bottom layer of multi-layered expert networks and a top task-specific tower network to learn high-order interactions from the embedded vectors of each input. Each expert module comprises multiple sub-networks, each referred to as an expert. Within the expert module, task-shared experts and task-specific experts are explicitly separated to avoid interference between tasks; while gating networks are used to combine the knowledge of lower-level experts.

[0096] Specifically, the content interaction prediction model can be used to predict the interaction result on the first interaction operation, denoted as task T; the content interaction prediction model can also be used to predict the interaction result on the second interaction operation, denoted as task C, and the gating network is represented by G.

[0097] wherein the gating network formula of task k in the jth layer of the expert network is shown in equation (1):

[0098] g k,j (x)=w k,j (g k,j-1 (x))S k,j (x) (1)

[0099] wherein x is the input embedded vector, w k,j is the weight function of task k, whose structure is based on a single-layer network with Softmax as the activation function, as shown in equation (2):

[0100]

[0101] wherein is the parameter matrix. Note that the gating network of the first layer is slightly different, and the formula corresponding to the gating network of the first layer is shown in equation (3):

[0102] g k,1 (x)=w k,1 (x)S k,1 (x) (3)

[0103] wherein S k,j is the selection matrix of task k in the jth layer of the network, which is composed of selection vectors, including shared experts and task k-specific experts, as shown in equation (4):

[0104]

[0105] wherein , and are the single experts specific to task k in the jth layer of the network, respectively, and there are m k task-specific experts, , and are the single experts shared by tasks in the jth layer of the network, respectively, and there are m S task-shared experts. It should be noted that the selection matrix of the shared expert module is slightly different, which is composed of all shared experts and task-specific experts.

[0106] Optionally, in the present embodiment, the step of "performing feature extraction on the target content in interactive operations to obtain target first operation features of the target content in the first interactive operation, target second operation features in the second interactive operation, and target shared operation features" can include:

[0107] extracting the first operation features of the target content in the first interactive operation, the second operation features in the second interactive operation, and the shared operation features;

[0108] The first operation feature, the second operation feature, and the shared operation feature are subjected to feature interaction processing to obtain the target first operation feature corresponding to the first interaction operation, the target second operation feature corresponding to the second interaction operation, and the target shared operation feature.

[0109] Specifically, the content information of the target content across various dimensions can be input into the content interaction prediction model. This content information can include the content title, publisher information, cover image, and the content itself. Based on the first interaction (e.g., reading time) and the second interaction (e.g., a click), the content information of the target content across these dimensions can be categorized into common features F. normal (e.g., user identification information (ID, IdentityDocument) and article identification information), exposure feature F exposure (For example, the title and cover image, which affect clicks and reading time) and post-click features F posst-click (For example, information such as article content that is only seen after clicking, which only has a direct impact on reading time), therefore the input content information f can be represented as in equation (5):

[0110] f = (F normal ,F exposure ,F post-click (5)

[0111] Among them, the content interaction prediction model extracts features of interactive operations from the content information f, such as... Figure 1c As shown, specifically, the first operational feature can be extracted by the task-specific expert T in the first layer, the second operational feature can be extracted by the task-specific expert C in the first layer, and the shared operational feature can be extracted by the task-sharing expert in the first layer. Then, the feature extraction network of the following layers performs feature interaction processing on the first operational feature, the second operational feature, and the shared operational feature, so that the task-specific expert T in the last layer outputs the target first operational feature, the task-specific expert C in the last layer outputs the target second operational feature, and the task-sharing expert in the last layer outputs the target shared operational feature.

[0112] Optionally, in this embodiment, the step "performing feature interaction processing on the first operation feature, the second operation feature, and the shared operation feature to obtain the target first operation feature corresponding to the first interaction operation, the target second operation feature corresponding to the second interaction operation, and the target shared operation feature" may include:

[0113] The first operation feature and the shared operation feature are fused together, and the first operation feature is updated based on the fused feature.

[0114] fusing the second operation feature and the shared operation feature, and updating the second operation feature based on the fused feature;

[0115] fusing the first operation feature, the second operation feature and the shared operation feature, and updating the shared operation feature based on the fused feature;

[0116] returning to the step of fusing the first operation feature and the shared operation feature, and updating the first operation feature based on the fused feature, until target shared operation feature meeting a preset feature interaction condition, target first operation feature corresponding to the first interaction operation and target second operation feature corresponding to the second interaction operation are obtained.

[0117] The preset feature interaction condition can be set according to actual conditions, and the embodiment does not limit this. For example, the preset feature interaction condition can be that the number of updates reaches a preset number. In some embodiments, the preset feature interaction condition can be determined according to the number of layers of the feature extraction network in the content interaction prediction model.

[0118] The first operation feature and the shared operation feature can be fused through a gating network; there are various fusion methods, and the embodiment does not limit this. For example, the fusion method can be a weighted operation, or can be splicing, etc. The fused feature is input into the next layer task-specific expert T, and the first operation feature is updated based on the output feature of the next layer task-specific expert T. Specifically, the output feature processed by the next layer task-specific expert T can be determined as a new first operation feature.

[0119] The second operation feature and the shared operation feature can be fused through a gating network; there are various fusion methods, and the embodiment does not limit this. For example, the fusion method can be a weighted operation, or can be splicing, etc. The fused feature is input into the next layer task-specific expert C, and the second operation feature is updated based on the output feature of the next layer task-specific expert C. Specifically, the output feature processed by the next layer task-specific expert C can be determined as a new second operation feature.

[0120] The first operation feature, the second operation feature and the shared operation feature can be fused through a gating network; there are various fusion methods, and the embodiment does not limit this. For example, the fusion method can be a weighted operation, or can be splicing, etc. The fused feature is input into the next layer task-specific expert, and the shared operation feature is updated based on the output feature of the next layer task-specific expert. Specifically, the output feature processed by the next layer task-specific expert can be determined as a new shared operation feature.

[0121] 102. determine an initial predicted interaction result of the target content on the first interaction operation based on the target first operation feature.

[0122] Optionally, in the embodiment, the step of “determining an initial predicted interaction result of the target content on the first interaction operation based on the target first operation feature” can include:

[0123] fusing the target shared operation feature and the target first operation feature to obtain a first fused operation feature;

[0124] determining an initial predicted interaction result of the target content on the first interaction operation based on the first fused operation feature.

[0125] In some embodiments, the target shared operation feature and the target first operation feature can be fused through a gating network. The fusion manner can be weighted operation or splicing processing, etc. The embodiment does not limit this. The first fused operation feature obtained by fusion is taken as an input of a tower network T in a content interaction prediction model. After processing by the tower network T, an initial predicted interaction result of the target content on the first interaction operation is obtained.

[0126] The tower network T can be a neural network structure, which can include convolutional layers and fully connected layers, etc. The embodiment does not limit this. Specifically, the tower network T can include a multilayer perceptron (MLP). The first fused operation feature is processed by the multilayer perceptron to predict probabilities of the target content belonging to each preset interaction result on the first interaction operation. According to the probabilities, an initial predicted interaction result of the target content on the first interaction operation is determined.

[0127] In some embodiments, the preset interaction result with the largest probability can be determined as the initial predicted interaction result of the target content on the first interaction operation. In other embodiments, the preset interaction result with a probability greater than a preset value can be determined as the initial predicted interaction result of the target content on the first interaction operation.

[0128] For example, when the first interaction operation is a reading duration of the target content, the reading duration interval can be divided according to actual conditions. For example, the reading duration interval can be divided into three subintervals: reading time less than 3 minutes, reading time between 3 and 10 minutes, and reading time greater than 10 minutes. The preset interaction result corresponding to the first interaction operation can include three cases corresponding to the three subintervals.

[0129] For another example, when the first interaction operation is a sharing operation of the target content, the preset interaction result corresponding to the first interaction operation can include two cases: sharing and not sharing.

[0130] Optionally, in the embodiment, the step of "performing feature extraction on the target content in interactive operations to obtain target first operation features of the target content in the first interactive operation and target shared operation features" can include:

[0131] performing feature extraction on the target content in interactive operations to obtain target first operation features of the target content in the first interactive operation, target second operation features of the target content in the second interactive operation, and target shared operation features;

[0132] The content interaction prediction method can further include:

[0133] determining a predicted interaction result of the target content in the second interactive operation based on the target second operation features.

[0134] In some embodiments, the step of "determining a predicted interaction result of the target content in the second interactive operation based on the target second operation features" can include:

[0135] fusing the target shared operation features and the target second operation features to obtain second fused operation features;

[0136] determining a predicted interaction result of the target content in the second interactive operation based on the second fused operation features.

[0137] The target shared operation features and the target second operation features can be fused through a gating network. The fusion method can be weighted operation or splicing processing, and the embodiment does not limit this. The second fused operation features obtained by fusion are taken as the input of a tower network C in the content interaction prediction model. After processing by the tower network C, a predicted interaction result of the target content in the second interactive operation is obtained.

[0138] The tower network C can be a neural network structure, which can include convolutional layers and fully connected layers, and the embodiment does not limit this. Specifically, the tower network C can include a multilayer perceptron (MLP). The second fused operation features are processed by the multilayer perceptron to predict the probability of the target content belonging to each preset interaction result in the second interactive operation. According to the probability, a predicted interaction result of the target content in the second interactive operation is determined.

[0139] In some embodiments, the preset interaction result with the largest probability can be determined as the predicted interaction result of the target content in the second interactive operation. In other embodiments, the preset interaction result with a probability greater than a preset value can be determined as the predicted interaction result of the target content in the second interactive operation.

[0140] For example, the second interaction operation is a click operation on the target content, and the preset interaction result corresponding to the second interaction operation can include two cases of clicking and not clicking.

[0141] 103. performing a logistic regression analysis on the target sharing operation feature to predict negative noise information generated by the second interaction operation on the first interaction operation.

[0142] In this embodiment, the target sharing operation feature can be subjected to logistic regression analysis by a negative noise modeling module in the content interaction prediction model. Specifically, the target sharing operation feature can be subjected to feature selection processing by a gating network, and then the processed target sharing operation feature can be input into the negative noise modeling module, which can perform logistic regression analysis thereon to predict negative noise information generated by the second interaction operation on the first interaction operation.

[0143] The logistic regression analysis can specifically include sigmoid function or tanh function operations by a hidden layer in a multilayer perceptron (MLP).

[0144] The sigmoid function, i.e., S-shaped growth curve, can be used as an activation function in a neural network or a logistic regression process to map a variable to a value range of zero to one. The tanh function, i.e., hyperbolic tangent, can be used as an activation function in a neural network in the field of deep learning.

[0145] 104. correcting the initial predicted interaction result based on the negative noise information to obtain a target predicted interaction result of the target content on the first interaction operation.

[0146] The negative noise information contains the negative influence of the second interaction operation on the first interaction operation. The correction of the initial predicted interaction result can specifically be subtracting the negative noise information from the initial predicted interaction result, so as to obtain the target predicted interaction result of the target content on the first interaction operation, which is an estimated value after subtracting the negative influence.

[0147] Specifically, if the first interaction operation is reading duration and the second interaction operation is a click operation, the negative influence can be subtracted from the original reading duration estimate (i.e., the initial predicted interaction result) by correction, so as to alleviate the negative influence of the click operation on the reading duration and strengthen the positive influence.

[0148] In a specific embodiment, the content interaction prediction method provided by the present application can be applied to a content recommendation scenario. After obtaining the target prediction interaction result of each content on the first interaction operation, the target recommendation content can be selected from the contents according to the target prediction interaction result for recommendation.

[0149] It should be noted that the content interaction prediction model can be trained by other devices and provided to the content interaction prediction device, or the content interaction prediction device can also train itself.

[0150] If the content interaction prediction device trains itself, before the step of "extracting features of the target content on the interaction operation by the content interaction prediction model to obtain the target first operation feature and the target sharing operation feature of the target content on the first interaction operation", the content interaction prediction method can further include:

[0151] obtaining training data, wherein the training data includes sample content, a first expected interaction result of the sample content on the first interaction operation, and a second expected interaction result on the second interaction operation;

[0152] extracting features of the sample content on the interaction operation by the content interaction prediction model to obtain the target first operation feature, the target second operation feature, and the target sharing operation feature of the sample content on the first interaction operation and the second interaction operation;

[0153] determining an initial first actual interaction result of the sample content on the first interaction operation and a second actual interaction result on the second interaction operation based on the target first operation feature and the target second operation feature, respectively;

[0154] performing logistic regression analysis on the target sharing operation feature to predict actual negative noise information generated by the second interaction operation on the first interaction operation; and correcting the initial first actual interaction result based on the actual negative noise information to obtain a target first actual interaction result;

[0155] adjusting parameters of the content interaction prediction model according to the initial first actual interaction result, the target first actual interaction result, the first expected interaction result, the second actual interaction result, and the second expected interaction result to obtain a trained content interaction prediction model.

[0156] The first expected interaction result can be an expected probability of the sample content belonging to each preset interaction result on the first interaction operation; and the second expected interaction result can be an expected probability of the sample content belonging to each preset interaction result on the second interaction operation.

[0157] Optionally, in the embodiment, the step of "adjusting parameters of the content interaction prediction model according to the initial first actual interaction result, the target first actual interaction result, the first expected interaction result, the second actual interaction result, and the second expected interaction result to obtain a trained content interaction prediction model" can include:

[0158] calculating a first loss value between the initial first actual interaction result and the first expected interaction result;

[0159] calculating a second loss value between the target first actual interaction result and the first expected interaction result;

[0160] calculating a third loss value between the second actual interaction result and the second expected interaction result;

[0161] adjusting parameters of the content interaction prediction model according to the first loss value, the second loss value, and the third loss value to obtain a trained content interaction prediction model.

[0162] In the training process, the parameters of the content interaction prediction model are adjusted using a back propagation algorithm, and the parameters of the content interaction prediction model are optimized based on the first loss value, the second loss value, and the third loss value, so that the first loss value, the second loss value, and the third loss value meet a preset loss condition, and a trained content interaction prediction model is obtained. Specifically, the preset loss condition can be that the sum of the first loss value, the second loss value, and the third loss value is less than a preset loss value, which can be set according to actual conditions.

[0163] The calculation method of the loss value can be various, and the embodiment does not limit it. For example, it can be a cross-entropy loss function, or it can be a mean square error loss function.

[0164] In a specific embodiment, the initial first actual interaction result is denoted as the first expected interaction result is denoted as the target first actual interaction result is denoted as the second actual interaction result is denoted as the second expected interaction result is denoted as The calculation processes of the first loss value, the second loss value, and the third loss value are shown in equations (6), (7), and (8), respectively:

[0165]

[0166]

[0167]

[0168] wherein, L T is the first loss value, L MTC is the second loss value, L C is the third loss value. S is the exposure sample dataset, S + represents the clicked sample dataset, and i represents the content of each sample.

[0169] wherein, specifically, if the second interaction operation is a click operation, then represents the true click label, usually represented by 1 for clicks and 0 for non-clicks, L C represents the loss of the click task; if the first interaction operation is reading time, the reading duration can be modeled as a multi-classification problem, and the continuous duration value is divided into multiple intervals, represents the true value vector corresponding to interval j after the reading duration is discretized into M groups, and the vector is 0 except for the value of the dimension of the group to which the true value belongs, L T represents the original multi-task reading duration loss, L MTC represents the corrected reading duration loss.

[0170] For example, the reading duration can be divided into three groups, namely, reading time less than 3 minutes, reading time 3 to 10 minutes, and reading time greater than 10 minutes.

[0171] wherein, the step of "adjusting the parameters of the content interaction prediction model according to the first loss value, the second loss value, and the third loss value to obtain the trained content interaction prediction model" can include:

[0172] fusing the first loss value, the second loss value, and the third loss value to obtain a total loss value;

[0173] adjusting the parameters of the content interaction prediction model based on the total loss value to obtain the trained content interaction prediction model.

[0174] wherein, the fusion method can be various, such as weighted fusion, etc. Specifically, the calculation process of the total loss value L can be as shown in equation (9):

[0175] L = L C + L T + L MTC (9)

[0176] In a specific embodiment, after adjusting the parameters of the content interaction prediction model based on the total loss value, the model can be further trained in the training process by weakening the input features related to the first interaction operation (such as reading duration) to better model the true negative impact.

[0177] Optionally, in the embodiment, the content interaction prediction model comprises a basic prediction module and a negative noise modeling module; the step of "adjusting the parameters of the content interaction prediction model according to the first loss value, the second loss value and the third loss value to obtain a trained content interaction prediction model" can comprise:

[0178] adjusting the parameters of the content interaction prediction model according to the first loss value, the second loss value and the third loss value to obtain a pre-trained content interaction prediction model;

[0179] performing mask processing on the sample content by the basic prediction module in the pre-trained content interaction prediction model to obtain target content information meeting a preset correlation condition in terms of relevance to the second interaction operation; performing feature extraction on the target content information in terms of interaction operation to obtain target reinforcement shared operation features of the sample content;

[0180] predicting actual reinforcement negative noise information of the second interaction operation on the first interaction operation based on the target reinforcement shared operation features by the negative noise modeling module in the pre-trained content interaction prediction model;

[0181] adjusting the parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain a trained content interaction prediction model.

[0182] The basic prediction module can be part of the content interaction prediction model other than the negative noise modeling module.

[0183] In order to make the learned negative impact of the second interaction operation more accurate, the representation of the shared expert can be intervened from the perspective of input features, so that it mainly contains the knowledge of the second interaction operation, thereby inducing the negative noise modeling module to learn higher information amount of negative impact.

[0184] The preset correlation condition can be that the content information obtained by mask processing does not include features that only have a direct impact on the first interaction operation, and the embodiment does not limit this.

[0185] In a specific embodiment, the first interaction operation is reading duration, and the second interaction operation is click operation. The content information of the sample content in each dimension can be divided into ordinary features F normal (user identification information (ID) and article identification information), exposure features F exposure (title and cover picture, which have an impact on click and reading duration) and post-click features F post-click(for example, the article content and other information that can only be seen after clicking, which only have direct influence on reading duration).

[0186] When training the parameters of the negative noise modeling module, the influence of the post-click features should be reduced, and the influence of the exposure features should be emphasized accordingly. In this way, by removing the input of the post-click features that are more positive for reading duration prediction, the negative noise modeling module can more accurately approximate the negative influence of clicks on reading duration. The specific method can be to set the post-click features F post-click MASK (masking) processing is performed, so that the target content information after processing can be as shown in equation (10):

[0187] f = (F normal ,F exposure ,[MASK]) (10)

[0188] Wherein, the mask processing can process the feature values in the post-click features into a null value -1. By weakening the input features related to reading duration in the way of mask processing, the true negative influence is approximated.

[0189] Optionally, before the step of "adjusting the parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain a trained content interaction prediction model", the embodiment can further include:

[0190] Through the basic prediction module in the pre-trained content interaction prediction model, the target content information is subjected to feature extraction in the interactive operation to obtain the target reinforced first operation feature of the sample content in the first interactive operation;

[0191] Based on the target reinforced first operation feature, the initial actual reinforcement interaction result of the sample content in the first interactive operation is determined;

[0192] The step of "adjusting the parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain a trained content interaction prediction model" can include:

[0193] Based on the actual reinforcement negative noise information, the initial actual reinforcement interaction result is corrected to obtain the target actual reinforcement interaction result of the sample content in the first interactive operation;

[0194] Based on the target actual reinforcement interaction result and the first expected interaction result, the parameters of the negative noise modeling module are adjusted to obtain a trained content interaction prediction model.

[0195] The initial actual reinforced interaction result can be corrected by subtracting the actual reinforced negative noise information from the initial actual reinforced interaction result.

[0196] The fourth loss value between the target actual reinforced interaction result and the first expected interaction result can be calculated, and the parameters of the negative noise modeling module can be adjusted based on the fourth loss value. The fourth loss value can be calculated based on a cross-entropy loss function, and the calculation process can refer to the above formula (7). Here, represents the target actual reinforced interaction result.

[0197] In a specific embodiment, referring to the model structure diagram of Figure 1d , the process of model training is shown, wherein, Figure 1d The left half part represents the process of adjusting the parameters of the content interaction prediction model based on the total loss, and the right half part represents the process of adjusting the parameters of the content interaction prediction model based on the fourth loss value after the sample content is processed by masking, and the specific description is as follows:

[0198] For the left half part, the initial first actual interaction result and the target first actual interaction result of the sample content on the first interaction operation and the second actual interaction result on the second interaction operation are calculated first, so as to calculate the first loss value, the second loss value and the third loss value, and the first loss value, the second loss value and the third loss value are fused to obtain the total loss value; and the parameters of the left side model opaque part (or the part with low transparency) are adjusted, that is, the parameters of the part of the content interaction prediction model except the negative noise modeling module are adjusted. Figure 1d For the left half part, the initial first actual interaction result and the target first actual interaction result of the sample content on the first interaction operation and the second actual interaction result on the second interaction operation are calculated first, so as to calculate the first loss value, the second loss value and the third loss value, and the first loss value, the second loss value and the third loss value are fused to obtain the total loss value; and the parameters of the left side model opaque part (or the part with low transparency) are adjusted, that is, the parameters of the part of the content interaction prediction model except the negative noise modeling module are adjusted.

[0199] Then, referring to the right half part, the target content information of the sample content is obtained by masking the post-click features, and the feature extraction on the interaction operation is performed on the target content information, so as to obtain the target actual reinforced interaction result of the sample content on the first interaction operation, and the fourth loss value is calculated based on the target actual reinforced interaction result, and the parameters of the right side model opaque part (or the part with low transparency) are adjusted, that is, the parameters of the part of the content interaction prediction model except the tower network C, the tower network T, the last layer task-specific expert C and the task-specific expert T are adjusted. Figure 1d Then, referring to the right half part, the target content information of the sample content is obtained by masking the post-click features, and the feature extraction on the interaction operation is performed on the target content information, so as to obtain the target actual reinforced interaction result of the sample content on the first interaction operation, and the fourth loss value is calculated based on the target actual reinforced interaction result, and the parameters of the right side model opaque part (or the part with low transparency) are adjusted, that is, the parameters of the part of the content interaction prediction model except the tower network C, the tower network T, the last layer task-specific expert C and the task-specific expert T are adjusted.

[0200] The content interaction prediction method provided in the application can be used in scenarios such as content recommendation, and can be used to help alleviate the interference of clicks on reading time or watching time in multi-task recommendation, and can also be used to alleviate the interference between other behaviors that exist in a dependent relationship, such as the interference of clicks on sharing. The application can capture the negative impact between the targets through a negative noise modeling module, and subtract the negative impact from the original estimated value to obtain an estimated value that alleviates the negative impact, and finally use the estimated value after subtracting the negative impact.

[0201] In a specific embodiment, as shown in Figure 1c , the embodiment can predict the interaction results of the target content on a first interaction operation (such as reading time) and on a second interaction operation (such as clicks) through the content interaction prediction model. Specifically, the negative noise modeling module is used to learn the negative impact of the reading time task due to the click task.

[0202] The negative noise modeling module can also be referred to as a negative impact modeling (NIM) module. The input of the negative noise modeling module is the output g S,L (x) of the last layer of task-shared experts, as shown in equation (11):

[0203] g S,L (x) = w S,L (g S,L-1 (x))S S,L (x) (11)

[0204] Wherein, S S,L (x) only selects the task-shared experts of the layer, and w S,L is a weight function. Finally, the prediction value of the negative impact of clicks on reading time, that is, the negative noise information , can be obtained, as shown in equation (12):

[0205]

[0206] Wherein, t NIN represents the tower network corresponding to the negative noise modeling module.

[0207] Specifically, represents the reading time prediction mainly affected by the click signal. If the original reading time output in the multi-task learning framework is represented as , the reading time estimated value after removing the click impact can be obtained, as shown in equation (13):

[0208]

[0209] Wherein, The initial prediction interaction result for reading duration The target prediction interaction result after correction.

[0210] The reading duration interval misclassified due to the negative effect of the click signal will have a higher value, and then this value is subtracted from the original estimated value , so that the correct reading duration interval will have a higher prediction probability.

[0211] The original prediction results of clicks and reading duration As shown in equation (14) and equation (15):

[0212]

[0213]

[0214] Where t C and t T are the tower networks for predicting clicks and reading duration, i.e. Figure 1c the tower network C (Tower C) and the tower network T (Tower T) in the model structure diagram of the prediction interaction result in the second interaction operation described in the above embodiments, the initial prediction interaction result in the first interaction operation described in the above embodiments.

[0215] Where the tower network can be constructed using a neural network such as a multi-layer perception (MLP); g C,L and g T,L are the respective gating networks, and L is the number of layers of the bottom expert module.

[0216] In a specific embodiment, the content interaction prediction method provided in the present application is significantly better than existing related models in predicting reading duration, and the prediction effect of the click task is also good. The test effect on the log data of a certain recommendation system is shown in Figure 1e , where MTC (Multi-Task Causal framework) is the model provided in the present application after modeling the negative effect module, MTC-fea is the model provided in the present application containing the negative effect modeling module and trained using the MASK feature, i.e., the model considering the negative effect at the feature level based on MTC; NFM, DeepFM, AutoInt, and AFN are single-task models, MMOE, AITM, and PLE are multi-task models. Compared with other models, the model MTC-fea provided in the present application achieves the best effect in all reading duration-related indicators.

[0217] The correlation indicators include mean absolute error-classification (MAE class), root mean square error-classification (RMSE class), recall, F1, mean absolute error (MAE), and root mean square error (RMSE). F1 is a comprehensive evaluation indicator, and the higher the F1 value, the better the prediction effect. AUC is a model evaluation indicator.

[0218] The estimation of reading duration is very important for a recommendation system, because a longer reading duration usually represents that the target object has greater interest in the recommended content, thereby effectively making up for the shortcoming that clicks may not reflect the real preferences of the target object, because clicks only reflect the interest of the target object in the title of the content. Accurate estimation of reading duration helps to recommend content that truly meets the interests of the target object, thereby improving user experience.

[0219] As can be seen from the above, the embodiment can obtain target content, and perform feature extraction on the target content in interactive operations to obtain target first operation features of the target content in a first interactive operation and target shared operation features. The target shared operation features are features shared by the first interactive operation and a second interactive operation, and the second interactive operation is a preceding operation dependent on the first interactive operation. Based on the target first operation features, an initial predicted interaction result of the target content in the first interactive operation is determined. The target shared operation features are subjected to logistic regression analysis to predict negative noise information generated by the second interactive operation on the first interactive operation. Based on the negative noise information, the initial predicted interaction result is corrected to obtain a target predicted interaction result of the target content in the first interactive operation. The application can correct the predicted interaction result of the target content in the first interactive operation by capturing negative noise information generated by the second interactive operation on the first interactive operation, thereby improving the prediction accuracy of the interaction result in the first interactive operation.

[0220] According to the method described in the foregoing embodiment, the following will be further described in detail with the content interaction prediction device being specifically integrated in a server as an example.

[0221] The embodiment of the application provides a content interaction prediction method, as shown in Figure 2 The specific process of the content interaction prediction method can be as follows:

[0222] 201. The server obtains target content, and performs feature extraction on the target content in interactive operations to obtain target first operation features of the target content in a first interactive operation and target shared operation features. The target shared operation features are features shared by the first interactive operation and a second interactive operation, and the second interactive operation is a preceding operation dependent on the first interactive operation.

[0223] The second interaction operation is a prerequisite operation dependent on the first interaction operation, and the two operations have a dependency relationship in operation time. The operation time of the second interaction operation is located before the operation time of the first interaction operation. For example, in a specific scenario, the target content is content recommended to a target object, the second interaction operation is a click operation on the content, and the first interaction operation is a reading time or sharing operation on the content. The predicted interaction result on the second interaction operation can be predicting whether the target object clicks the target content, and the predicted interaction result on the first interaction operation can be predicting the reading time of the target object on the target content or whether the target object will share the target content. It can be understood that the target object needs to click the target content first, and then the target object can read or share the target content. Therefore, the click is a prerequisite operation corresponding to the reading time or the sharing.

[0224] Optionally, in the embodiment, the step of "extracting features of the target content in the interaction operation to obtain target first operation features of the target content in the first interaction operation and target sharing operation features" can include:

[0225] extracting features of the target content in the interaction operation to obtain target first operation features of the target content in the first interaction operation, target second operation features of the target content in the second interaction operation, and target sharing operation features.

[0226] The target sharing operation features are features shared by the first interaction operation and the second interaction operation. The target first operation features can be regarded as unique features of the first interaction operation, and the target second operation features can be regarded as unique features of the second interaction operation.

[0227] Optionally, in the embodiment, the step of "extracting features of the target content in the interaction operation to obtain target first operation features of the target content in the first interaction operation and target sharing operation features" can include:

[0228] extracting features of the target content in the interaction operation by using the content interaction prediction model to obtain target first operation features of the target content in the first interaction operation and target sharing operation features.

[0229] In the embodiment, the content interaction prediction model can be used to extract target first operation features of the target content in the first interaction operation, target second operation features of the target content in the second interaction operation, and target sharing operation features.

[0230] Specifically, the content interaction prediction model is a multi-task model, which can include a multi-layer feature extraction network and a tower network corresponding to each task in the target multi-task; each layer of the feature extraction network includes a plurality of expert networks corresponding to each task and a gating network. For example, the content interaction prediction model can include a multi-layer feature extraction network and a tower network corresponding to both a prediction task of a first interaction operation and a prediction task of a second interaction operation.

[0231] Optionally, in the embodiment, the step of "performing feature extraction on the target content in the interaction operation to obtain target first operation features of the target content in the first interaction operation, target second operation features of the target content in the second interaction operation, and target shared operation features of the target content" can include:

[0232] extracting first operation features of the target content in the first interaction operation, second operation features of the target content in the second interaction operation, and shared operation features of the target content;

[0233] performing feature interaction processing on the first operation features, the second operation features, and the shared operation features to obtain target first operation features corresponding to the first interaction operation, target second operation features corresponding to the second interaction operation, and target shared operation features.

[0234] Optionally, in the embodiment, the step of "performing feature interaction processing on the first operation features, the second operation features, and the shared operation features to obtain target first operation features corresponding to the first interaction operation, target second operation features corresponding to the second interaction operation, and target shared operation features" can include:

[0235] fusing the first operation features and the shared operation features, and updating the first operation features based on the fused features;

[0236] fusing the second operation features and the shared operation features, and updating the second operation features based on the fused features;

[0237] fusing the first operation features, the second operation features, and the shared operation features, and updating the shared operation features based on the fused features;

[0238] Returning to the step of "fusing the first operation features and the shared operation features, and updating the first operation features based on the fused features", until target shared operation features, target first operation features corresponding to the first interaction operation, and target second operation features corresponding to the second interaction operation that satisfy a preset feature interaction condition are obtained.

[0239] The preset feature interaction condition can be set according to actual conditions, and the embodiment is not limited in this regard. For example, the preset feature interaction condition can be that the number of updates reaches a preset number. In some embodiments, the preset feature interaction condition can be determined according to the number of layers of the feature extraction network in the content interaction prediction model.

[0240] 202. The server determines an initial predicted interaction result of the target content on the first interaction operation based on the target first operation feature.

[0241] Optionally, in the embodiment, the step of "determining an initial predicted interaction result of the target content on the first interaction operation based on the target first operation feature" can include:

[0242] fusing the target shared operation feature and the target first operation feature to obtain a first fused operation feature;

[0243] determining an initial predicted interaction result of the target content on the first interaction operation based on the first fused operation feature.

[0244] The target shared operation feature and the target first operation feature can be fused through a gating network. The fusion method can be weighted operation or splicing processing, and the embodiment is not limited in this regard. The first fused operation feature obtained by fusion is input into a tower network T in the content interaction prediction model, and the initial predicted interaction result of the target content on the first interaction operation is obtained through processing of the tower network T.

[0245] The tower network T can be a neural network structure, which can include convolutional layers and fully connected layers, and the embodiment is not limited in this regard. Specifically, the tower network T can include a multilayer perceptron (MLP, Multilayer Perceptron). The first fused operation feature is processed through the multilayer perceptron to predict the probability of the target content belonging to each preset interaction result on the first interaction operation, and the initial predicted interaction result of the target content on the first interaction operation is determined according to the probability.

[0246] In some embodiments, the preset interaction result with the largest probability can be determined as the initial predicted interaction result of the target content on the first interaction operation. In other embodiments, the preset interaction result with a probability greater than a preset value can be determined as the initial predicted interaction result of the target content on the first interaction operation.

[0247] Optionally, in the embodiment, the step of "extracting features of the target content on the first interaction operation to obtain the target first operation feature and the target shared operation feature of the target content on the first interaction operation" can include:

[0248] characteristics of the target content in the first interaction operation, target second operation characteristics of the target content in the second interaction operation, and target shared operation characteristics of the target content;

[0249] The content interaction prediction method can further include:

[0250] determining, based on the target second operation characteristics, a predicted interaction result of the target content in the second interaction operation.

[0251] In some embodiments, the step of "determining, based on the target second operation characteristics, a predicted interaction result of the target content in the second interaction operation" can include:

[0252] fusing the target shared operation characteristics and the target second operation characteristics to obtain second fused operation characteristics;

[0253] determining, based on the second fused operation characteristics, a predicted interaction result of the target content in the second interaction operation.

[0254] In this embodiment, the target shared operation characteristics and the target second operation characteristics can be fused through a gating network, and the fusion method can be weighted operation or splicing processing, etc. The second fused operation characteristics obtained through the fusion are input into a tower network C in the content interaction prediction model, and the predicted interaction result of the target content in the second interaction operation is obtained through processing of the tower network C.

[0255] 203. The server performs logistic regression analysis on the target shared operation characteristics to predict negative noise information generated by the second interaction operation on the first interaction operation.

[0256] In this embodiment, the target shared operation characteristics can be subjected to logistic regression analysis through a negative noise modeling module in the content interaction prediction model. Specifically, the target shared operation characteristics can be subjected to feature selection processing through a gating network, and then the processed target shared operation characteristics are input into the negative noise modeling module, and logistic regression analysis is performed on the target shared operation characteristics through the negative noise modeling module to predict the negative noise information generated by the second interaction operation on the first interaction operation.

[0257] The logistic regression analysis can specifically include sigmoid function or tanh function operation through a hidden layer in a multilayer perceptron (MLP).

[0258] Among them, the sigmoid function, that is, the S-shaped growth curve, can be used as an activation function in a neural network or in a logistic regression process, which maps variables to a value range of zero to one. The tanh function, that is, the hyperbolic tangent, can be used as an activation function in a neural network in the field of deep learning.

[0259] 204、Server corrects the initial predicted interaction result based on the negative noise information to obtain a target predicted interaction result of the target content on the first interaction operation.

[0260] Among them, the negative noise information contains the negative influence of the second interaction operation on the first interaction operation. The correction of the initial predicted interaction result can be specifically subtracting the negative noise information from the initial predicted interaction result, that is, obtaining the target predicted interaction result of the target content on the first interaction operation, which is the estimated value after subtracting the negative influence.

[0261] It should be noted that the content interaction prediction model can be trained by other devices and provided to the content interaction prediction device, or the content interaction prediction device can also train itself.

[0262] If the content interaction prediction device trains itself, before the step of "extracting features of the target content on the interaction operation by the content interaction prediction model to obtain target first operation features and target sharing operation features of the target content on the first interaction operation", the content interaction prediction method can further include:

[0263] Obtaining training data, the training data including sample content, first expected interaction result of the sample content on the first interaction operation, and second expected interaction result on the second interaction operation;

[0264] Extracting features of the sample content on the interaction operation by the content interaction prediction model to obtain target first operation features of the sample content on the first interaction operation, target second operation features on the second interaction operation, and target sharing operation features;

[0265] Based on the target first operation features and the target second operation features respectively, determining an initial first actual interaction result of the sample content on the first interaction operation and a second actual interaction result on the second interaction operation;

[0266] performing a logistic regression analysis on the target shared operation feature to predict actual negative noise information generated by the second interaction operation on the first interaction operation; and correcting the initial first actual interaction result based on the actual negative noise information to obtain a target first actual interaction result;

[0267] adjusting parameters of the content interaction prediction model based on the initial first actual interaction result, the target first actual interaction result, the first expected interaction result, the second actual interaction result, and the second expected interaction result to obtain a trained content interaction prediction model.

[0268] The first expected interaction result can be an expected probability of the sample content belonging to each preset interaction result in the first interaction operation, and the second expected interaction result can be an expected probability of the sample content belonging to each preset interaction result in the second interaction operation.

[0269] Optionally, in the embodiment, the step of adjusting parameters of the content interaction prediction model based on the initial first actual interaction result, the target first actual interaction result, the first expected interaction result, the second actual interaction result, and the second expected interaction result to obtain a trained content interaction prediction model can include:

[0270] calculating a first loss value between the initial first actual interaction result and the first expected interaction result;

[0271] calculating a second loss value between the target first actual interaction result and the first expected interaction result;

[0272] calculating a third loss value between the second actual interaction result and the second expected interaction result;

[0273] adjusting parameters of the content interaction prediction model based on the first loss value, the second loss value, and the third loss value to obtain a trained content interaction prediction model.

[0274] The training process specifically adjusts parameters of the content interaction prediction model using a back propagation algorithm, optimizes the parameters of the content interaction prediction model based on the first loss value, the second loss value, and the third loss value, so that the first loss value, the second loss value, and the third loss value meet a preset loss condition, and obtains a trained content interaction prediction model. Specifically, the preset loss condition can be that the sum of the first loss value, the second loss value, and the third loss value is less than a preset loss value, and the preset loss value can be set according to actual conditions.

[0275] The loss value can be calculated in various ways, and the embodiments are not limited in this regard. For example, the loss value can be a cross-entropy loss function or a mean square error loss function.

[0276] In a specific embodiment, after adjusting the parameters of the content interaction prediction model based on the total loss value, the model can be further trained in the training process by weakening the input features related to the first interaction operation (such as reading duration) to make the model better model the real negative impact.

[0277] Optionally, in the embodiments, the content interaction prediction model includes a basic prediction module and a negative noise modeling module; and the step of "adjusting the parameters of the content interaction prediction model according to the first loss value, the second loss value and the third loss value to obtain a trained content interaction prediction model" can include:

[0278] adjusting the parameters of the content interaction prediction model according to the first loss value, the second loss value and the third loss value to obtain a pre-trained content interaction prediction model;

[0279] performing mask processing on the sample content through the basic prediction module in the pre-trained content interaction prediction model to obtain target content information that meets a preset correlation condition in terms of correlation with the second interaction operation; and performing feature extraction on the target content information in terms of interaction operation to obtain target reinforcement shared operation features of the sample content;

[0280] predicting actual reinforcement negative noise information of the second interaction operation on the first interaction operation based on the target reinforcement shared operation features through the negative noise modeling module in the pre-trained content interaction prediction model;

[0281] adjusting the parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain a trained content interaction prediction model.

[0282] The basic prediction module can be part of the content interaction prediction model other than the negative noise modeling module.

[0283] To make the learned negative impact of the second interaction operation more accurate, the representation of the shared expert can be intervened from the perspective of input features to mainly include knowledge of the second interaction operation, thereby inducing the negative noise modeling module to learn a higher amount of negative impact information.

[0284] The preset correlation condition can be that the content information obtained by mask processing does not include features that only have a direct impact on the first interaction operation, and the embodiments are not limited in this regard.

[0285] Optionally, before the step of "adjusting parameters of the negative noise modeling module based on the actual reinforced negative noise information and the first expected interaction result to obtain a trained content interaction prediction model", the embodiment can further include:

[0286] extracting a target reinforced first operation feature of the sample content in the first interaction operation through a basic prediction module in the pre-trained content interaction prediction model;

[0287] determining an initial actual reinforced interaction result of the sample content in the first interaction operation based on the target reinforced first operation feature;

[0288] The step of "adjusting parameters of the negative noise modeling module based on the actual reinforced negative noise information and the first expected interaction result to obtain a trained content interaction prediction model" can include:

[0289] correcting the initial actual reinforced interaction result based on the actual reinforced negative noise information to obtain a target actual reinforced interaction result of the sample content in the first interaction operation;

[0290] adjusting parameters of the negative noise modeling module based on the target actual reinforced interaction result and the first expected interaction result to obtain a trained content interaction prediction model.

[0291] The initial actual reinforced interaction result can be corrected by subtracting the actual reinforced negative noise information from the initial actual reinforced interaction result.

[0292] As can be seen from the above, the embodiment can obtain target content through a server; and extract a target first operation feature of the target content in a first interaction operation and a target shared operation feature of the target content, the target shared operation feature being a feature shared by the first interaction operation and a second interaction operation, the second interaction operation being a preceding operation dependent on the first interaction operation; determine an initial predicted interaction result of the target content in the first interaction operation based on the target first operation feature; perform a logistic regression analysis on the target shared operation feature to predict negative noise information generated by the second interaction operation on the first interaction operation; and correct the initial predicted interaction result based on the negative noise information to obtain a target predicted interaction result of the target content in the first interaction operation. The application can correct the predicted interaction result of the target content in the first interaction operation by capturing negative noise information generated by the second interaction operation on the first interaction operation, thereby improving the prediction accuracy of the interaction result in the first interaction operation.

[0293] To better implement the above method, the embodiment of the present application further provides a content interaction prediction device, as shown in the figure, which can include an acquisition unit 301, a determination unit 302, a prediction unit 303, and a correction unit 304, as follows: Figure 3

[0294] (1) Acquisition unit 301;

[0295] The acquisition unit is configured to acquire target content, and perform feature extraction on the target content in interactive operations to obtain target first operation features of the target content in a first interactive operation and target shared operation features, the target shared operation features being features shared by the first interactive operation and a second interactive operation, the second interactive operation being a preceding operation dependent on the first interactive operation.

[0296] Optionally, in some embodiments of the present application, the acquisition unit can be specifically configured to perform feature extraction on the target content in interactive operations to obtain target first operation features of the target content in a first interactive operation, target second operation features in a second interactive operation, and target shared operation features.

[0297] The content interaction prediction device can further include a result determination unit, as follows:

[0298] The result determination unit is configured to determine a predicted interaction result of the target content in the second interactive operation based on the target second operation features.

[0299] Optionally, in some embodiments of the present application, the acquisition unit can include an extraction subunit and an interaction subunit, as follows:

[0300] The extraction subunit is configured to extract first operation features of the target content in a first interactive operation, second operation features in a second interactive operation, and shared operation features.

[0301] The interaction subunit is configured to perform feature interaction processing on the first operation features, the second operation features, and the shared operation features to obtain target first operation features corresponding to the first interactive operation, target second operation features corresponding to the second interactive operation, and target shared operation features.

[0302] ​Optionally, in some embodiments of the present application, the interaction subunit can be specifically configured to fuse the first operation feature and the shared operation feature, and update the first operation feature based on the fused feature; fuse the second operation feature and the shared operation feature, and update the second operation feature based on the fused feature; fuse the first operation feature, the second operation feature and the shared operation feature, and update the shared operation feature based on the fused feature; return to the step of fusing the first operation feature and the shared operation feature, and updating the first operation feature based on the fused feature, until the target shared operation feature meeting the preset feature interaction condition, the target first operation feature corresponding to the first interaction operation and the target second operation feature corresponding to the second interaction operation are obtained.

[0303] (2) the determining unit 302;

[0304] The determining unit is configured to determine an initial predicted interaction result of the target content on the first interaction operation based on the target first operation feature.

[0305] Optionally, in some embodiments of the present application, the determining unit comprises a fusion subunit and a determining subunit, as follows:

[0306] The fusion subunit is configured to fuse the target shared operation feature and the target first operation feature to obtain a first fused operation feature.

[0307] The determining subunit is configured to determine an initial predicted interaction result of the target content on the first interaction operation based on the first fused operation feature.

[0308] (3) the predicting unit 303;

[0309] The predicting unit is configured to perform logistic regression analysis on the target shared operation feature to predict negative noise information of the second interaction operation to the first interaction operation.

[0310] (4) the correcting unit 304;

[0311] The correcting unit is configured to correct the initial predicted interaction result based on the negative noise information to obtain a target predicted interaction result of the target content on the first interaction operation.

[0312] Optionally, in some embodiments of the present application, the obtaining unit can be specifically configured to perform feature extraction on the target content on the interaction operation through a content interaction prediction model to obtain a target first operation feature and a target shared operation feature of the target content on the first interaction operation.

[0313] Optionally, in some embodiments of the present application, the content interaction prediction apparatus can further comprise a training unit configured to train the content interaction prediction model; specifically, the training unit can comprise a data acquisition subunit, a feature extraction subunit, an interaction result determination subunit, an analysis subunit and an adjustment subunit, as follows:

[0314] The data acquisition subunit is configured to acquire training data, wherein the training data comprises sample content, a first expected interaction result of the sample content on the first interaction operation, and a second expected interaction result on the second interaction operation;

[0315] The feature extraction subunit is configured to perform feature extraction on the sample content on the interaction operation by the content interaction prediction model, to obtain a target first operation feature of the sample content on the first interaction operation, a target second operation feature on the second interaction operation, and a target shared operation feature;

[0316] The interaction result determination subunit is configured to determine an initial first actual interaction result of the sample content on the first interaction operation and a second actual interaction result on the second interaction operation based on the target first operation feature and the target second operation feature, respectively;

[0317] The analysis subunit is configured to perform logistic regression analysis on the target shared operation feature to predict actual negative noise information generated by the second interaction operation on the first interaction operation; and correct the initial first actual interaction result based on the actual negative noise information to obtain a target first actual interaction result;

[0318] The adjustment subunit is configured to adjust parameters of the content interaction prediction model according to the initial first actual interaction result, the target first actual interaction result, the first expected interaction result, the second actual interaction result and the second expected interaction result, to obtain a trained content interaction prediction model.

[0319] Optionally, in some embodiments of the present application, the adjustment subunit can be specifically configured to calculate a first loss value between the initial first actual interaction result and the first expected interaction result; calculate a second loss value between the target first actual interaction result and the first expected interaction result; calculate a third loss value between the second actual interaction result and the second expected interaction result; and adjust the parameters of the content interaction prediction model according to the first loss value, the second loss value and the third loss value, to obtain a trained content interaction prediction model.

[0320] Optionally, in some embodiments of the present application, the content interaction prediction model comprises a basic prediction module and a negative noise modeling module; the step of "adjusting parameters of the content interaction prediction model according to the first loss value, the second loss value and the third loss value to obtain a pre-trained content interaction prediction model" can comprise:

[0321] adjusting parameters of the content interaction prediction model according to the first loss value, the second loss value and the third loss value to obtain a pre-trained content interaction prediction model;

[0322] performing mask processing on the sample content by the basic prediction module in the pre-trained content interaction prediction model to obtain target content information meeting a preset correlation condition in terms of relevance to the second interaction operation; performing feature extraction on the target content information in terms of interaction operation to obtain target reinforcement shared operation features of the sample content;

[0323] predicting actual reinforcement negative noise information of the second interaction operation to the first interaction operation based on the target reinforcement shared operation features by the negative noise modeling module in the pre-trained content interaction prediction model;

[0324] adjusting parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain a trained content interaction prediction model.

[0325] Optionally, in some embodiments of the present application, before the step of "adjusting parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain a trained content interaction prediction model", the method can further comprise:

[0326] performing feature extraction on the target content information in terms of interaction operation by the basic prediction module in the pre-trained content interaction prediction model to obtain target reinforcement first operation features of the sample content in the first interaction operation;

[0327] determining an initial actual reinforcement interaction result of the sample content in the first interaction operation based on the target reinforcement first operation features;

[0328] The step of "adjusting parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain a trained content interaction prediction model" can comprise:

[0329] correcting the initial actual reinforcement interaction result based on the actual reinforcement negative noise information to obtain a target actual reinforcement interaction result of the sample content in the first interaction operation;

[0330] Based on the actual enhanced interaction results of the target and the first expected interaction results, the parameters of the negative noise modeling module are adjusted to obtain the trained content interaction prediction model.

[0331] As can be seen from the above, this embodiment can acquire target content through acquisition unit 301; and extract features of the target content in terms of interactive operation to obtain the target first operation feature and target shared operation feature of the target content in the first interactive operation. The target shared operation feature is a feature shared by the first interactive operation and the second interactive operation, and the second interactive operation is a pre-operation dependent on the first interactive operation. Based on the target first operation feature, determination unit 302 determines the initial predicted interaction result of the target content in the first interactive operation. Based on the target first operation feature, prediction unit 303 performs logistic regression analysis on the target shared operation feature to predict the negative noise information generated by the second interactive operation on the first interactive operation. Based on the negative noise information, correction unit 304 corrects the initial predicted interaction result to obtain the target predicted interaction result of the target content in the first interactive operation. This application can correct the predicted interaction result of the target content in the first interactive operation by capturing the negative noise information generated by the second interactive operation on the first interactive operation, thereby improving the prediction accuracy of the interaction result in the first interactive operation.

[0332] This application also provides an electronic device, such as... Figure 4 The diagram shows a structural schematic of an electronic device involved in an embodiment of this application. This electronic device can be a terminal or a server, specifically:

[0333] The electronic device may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, and an input unit 404. Those skilled in the art will understand that... Figure 4 The electronic device structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0334] The processor 401 is the control center of the electronic device, connects each part of the entire electronic device by various interfaces and lines, and performs various functions of the electronic device and processes data by running or executing software programs and / or modules stored in the memory 402 and calling data stored in the memory 402. Optionally, the processor 401 can include one or more processing cores; preferably, the processor 401 can integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application program, etc., and the modem processor mainly processes wireless communication. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 401.

[0335] The memory 402 can be used to store software programs and modules, and the processor 401 executes various function applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), etc.; and the data storage area can store data created according to the use of the electronic device, etc. In addition, the memory 402 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory 402 can also include a memory controller to provide access for the processor 401 to the memory 402.

[0336] The electronic device also includes a power supply 403 for supplying power to each component, and preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, so as to realize functions such as management of charging, discharging, and power consumption management through the power management system. The power supply 403 can also include one or more than one direct current or alternating current power supply, a recharging system, a power failure detection circuit, a power converter or inverter, a power state indicator, etc. any component.

[0337] The electronic device can also include an input unit 404, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to user settings and function control.

[0338] Although not shown, the electronic device can also include a display unit, etc., which will not be described here. Specifically, in the present embodiment, the processor 401 in the electronic device will load the executable file corresponding to the process of one or more than one application program into the memory 402 according to the following instructions, and run the application program stored in the memory 402 by the processor 401, so as to realize various functions, as follows:

[0339] obtain target content; and perform feature extraction on the target content in interactive operations to obtain target first operation features of the target content in a first interactive operation and target shared operation features, the target shared operation features being features shared by the first interactive operation and a second interactive operation, the second interactive operation being a preceding operation dependent on the first interactive operation; determine an initial predicted interactive result of the target content in the first interactive operation based on the target first operation features; perform logistic regression analysis on the target shared operation features to predict negative noise information generated by the second interactive operation on the first interactive operation; and correct the initial predicted interactive result based on the negative noise information to obtain a target predicted interactive result of the target content in the first interactive operation.

[0340] The specific implementation of each operation can be referred to the foregoing embodiments, which will not be described here.

[0341] As can be seen from the above, the embodiment can obtain target content; and perform feature extraction on the target content in interactive operations to obtain target first operation features of the target content in a first interactive operation and target shared operation features, the target shared operation features being features shared by the first interactive operation and a second interactive operation, the second interactive operation being a preceding operation dependent on the first interactive operation; determine an initial predicted interactive result of the target content in the first interactive operation based on the target first operation features; perform logistic regression analysis on the target shared operation features to predict negative noise information generated by the second interactive operation on the first interactive operation; and correct the initial predicted interactive result based on the negative noise information to obtain a target predicted interactive result of the target content in the first interactive operation. The application can correct the predicted interactive result of the target content in the first interactive operation by capturing the negative noise information generated by the second interactive operation on the first interactive operation, thereby improving the prediction accuracy of the interactive result in the first interactive operation.

[0342] Those skilled in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions or controlled by instructions related to hardware, which can be stored in a computer readable storage medium and loaded and executed by a processor.

[0343] To this end, the embodiment of the application provides a computer readable storage medium, which stores a plurality of instructions capable of being loaded by a processor to execute steps in any content interaction prediction method provided by the embodiment of the application. For example, the instructions can execute the following steps:

[0344] obtaining target content; and performing feature extraction on the target content in an interactive operation to obtain a target first operation feature of the target content in a first interactive operation and a target shared operation feature, the target shared operation feature being a feature shared by the first interactive operation and a second interactive operation, the second interactive operation being a preceding operation dependent on the first interactive operation; determining an initial predicted interaction result of the target content in the first interactive operation based on the target first operation feature; performing logistic regression analysis on the target shared operation feature to predict negative noise information generated by the second interactive operation on the first interactive operation; and correcting the initial predicted interaction result based on the negative noise information to obtain a target predicted interaction result of the target content in the first interactive operation.

[0345] The specific implementation of each operation can refer to the foregoing embodiments, which will not be described here again.

[0346] The computer readable storage medium can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0347] Due to the instructions stored in the computer readable storage medium, the steps of any content interaction prediction method provided by the embodiments of the present application can be performed, and thus the beneficial effects of any content interaction prediction method provided by the embodiments of the present application can be achieved. Details are described in the foregoing embodiments, which will not be described here again.

[0348] According to an aspect of the present application, a computer program product or computer program is provided, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to enable the computer device to perform the method provided in the various optional implementations of the content interaction prediction aspect.

[0349] The content interaction prediction method and related device provided by the embodiments of the present application are described in detail above, and the principles and implementation manners of the present application are described by applying specific examples. The above description of the embodiments is only used to help understand the method and core idea of the present application; meanwhile, for those skilled in the art, the specific implementation manner and application range can be changed according to the idea of the present application, and the above description should not be understood as a limitation of the present application.

Claims

1. A method for predicting content interaction, the method comprising: The method comprises: obtaining target content; and extracting features of the target content on the interactive operation through a content interaction prediction model to obtain target first operation features of the target content on a first interactive operation and target shared operation features, the target shared operation features being features shared by the first interactive operation and a second interactive operation, the second interactive operation being a preceding operation dependent on the first interactive operation; determining an initial predicted interaction result of the target content on the first interactive operation based on the target first operation features; performing logistic regression analysis on the target shared operation features to predict negative noise information generated by the second interactive operation on the first interactive operation; based on the negative noise information, correcting the initial predicted interaction result to obtain a target predicted interaction result of the target content on the first interactive operation; wherein the content interaction prediction model is obtained in the following manner: obtaining training data, the training data comprising sample content, a first expected interaction result of the sample content on the first interactive operation, and a second expected interaction result on the second interactive operation; extracting features of the sample content on the interactive operation through a content interaction prediction model to obtain target first operation features of the sample content on the first interactive operation, target second operation features on the second interactive operation, and target shared operation features; determining an initial first actual interaction result of the sample content on the first interactive operation and a second actual interaction result on the second interactive operation based on the target first operation features and the target second operation features, respectively; performing logistic regression analysis on the target shared operation features to predict actual negative noise information generated by the second interactive operation on the first interactive operation; and based on the actual negative noise information, correcting the initial first actual interaction result to obtain a target first actual interaction result; adjusting parameters of the content interaction prediction model according to the initial first actual interaction result, the target first actual interaction result, the first expected interaction result, the second actual interaction result, and the second expected interaction result to obtain a trained content interaction prediction model.

2. The method of claim 1, wherein, The method further comprises: determining a predicted interaction result of the target content on the second interactive operation based on the target second operation features. The method further comprises: determining a predicted interaction result of the target content on the second interactive operation based on the target second operation features.

3. The method of claim 2, wherein, ​ extracting a first operation feature of the target content on a first interactive operation, a second operation feature of the target content on a second interactive operation, and a shared operation feature; performing feature interaction processing on the first operation feature, the second operation feature, and the shared operation feature to obtain a target first operation feature corresponding to the first interactive operation, a target second operation feature corresponding to the second interactive operation, and a target shared operation feature.

4. The method of claim 3, wherein, The feature interaction processing on the first operation feature, the second operation feature, and the shared operation feature to obtain a target first operation feature corresponding to the first interactive operation, a target second operation feature corresponding to the second interactive operation, and a target shared operation feature comprises: fusing the first operation feature and the shared operation feature, and updating the first operation feature based on the fused feature; fusing the second operation feature and the shared operation feature, and updating the second operation feature based on the fused feature; fusing the first operation feature, the second operation feature, and the shared operation feature, and updating the shared operation feature based on the fused feature; returning to the step of fusing the first operation feature and the shared operation feature, and updating the first operation feature based on the fused feature, until a target shared operation feature, a target first operation feature corresponding to the first interactive operation, and a target second operation feature corresponding to the second interactive operation that satisfy a preset feature interaction condition are obtained.

5. The method of claim 1, wherein, The determination of the initial predicted interaction result of the target content on the first interactive operation based on the target first operation feature comprises: fusing the target shared operation feature and the target first operation feature to obtain a first fused operation feature; determining the initial predicted interaction result of the target content on the first interactive operation based on the first fused operation feature.

6. The method of claim 1, wherein, The adjustment of the parameters of the content interaction prediction model according to the initial first actual interaction result, the target first actual interaction result, the first expected interaction result, the second actual interaction result, and the second expected interaction result to obtain a trained content interaction prediction model comprises: calculating a first loss value between the initial first actual interaction result and the first expected interaction result; calculating a second loss value between the target first actual interaction result and the first expected interaction result; calculating a third loss value between the second actual interaction result and the second expected interaction result; adjusting the parameters of the content interaction prediction model according to the first loss value, the second loss value, and the third loss value to obtain a trained content interaction prediction model.

7. The method of claim 6, wherein, The content interaction prediction model comprises a basic prediction module and a negative noise modeling module; and the adjustment of the parameters of the content interaction prediction model according to the first loss value, the second loss value, and the third loss value to obtain a trained content interaction prediction model comprises: According to the first loss value, the second loss value and the third loss value, the parameters of the content interaction prediction model are adjusted to obtain a pre-trained content interaction prediction model; Through a basic prediction module in the pre-trained content interaction prediction model, the sample content is subjected to mask processing to obtain target content information meeting a preset correlation condition with the second interaction operation; and the target content information is subjected to feature extraction on the interaction operation to obtain target reinforcement shared operation features of the sample content; Through a negative noise modeling module in the pre-trained content interaction prediction model, actual reinforcement negative noise information of the second interaction operation to the first interaction operation is predicted based on the target reinforcement shared operation features; Based on the actual reinforcement negative noise information and the first expected interaction result, the parameters of the negative noise modeling module are adjusted to obtain a trained content interaction prediction model.

8. The method of claim 7, wherein, Before the step of adjusting the parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain the trained content interaction prediction model, the method further includes: Through the basic prediction module in the pre-trained content interaction prediction model, the target content information is subjected to feature extraction on the interaction operation to obtain target reinforcement first operation features of the sample content on the first interaction operation; Based on the target reinforcement first operation features, an initial actual reinforcement interaction result of the sample content on the first interaction operation is determined; The step of adjusting the parameters of the negative noise modeling module based on the actual reinforcement negative noise information and the first expected interaction result to obtain the trained content interaction prediction model includes: Based on the actual reinforcement negative noise information, the initial actual reinforcement interaction result is corrected to obtain a target actual reinforcement interaction result of the sample content on the first interaction operation; Based on the target actual reinforcement interaction result and the first expected interaction result, the parameters of the negative noise modeling module are adjusted to obtain the trained content interaction prediction model.

9. A content interaction prediction apparatus characterized by comprising: The method includes: an acquisition unit configured to acquire target content; and a prediction unit configured to perform feature extraction on the target content on an interaction operation through a content interaction prediction model to obtain target first operation features of the target content on a first interaction operation and target shared operation features, the target shared operation features being features shared by the first interaction operation and a second interaction operation, the second interaction operation being a preceding operation dependent on the first interaction operation; a determination unit configured to determine an initial prediction interaction result of the target content on the first interaction operation based on the target first operation features; a prediction unit configured to perform logistic regression analysis on the target shared operation features to predict negative noise information of the second interaction operation to the first interaction operation; a correction unit configured to correct the initial prediction interaction result based on the negative noise information to obtain a target prediction interaction result of the target content on the first interaction operation; The content interaction prediction model is obtained in the following manner: obtaining training data, the training data comprising sample content, a first expected interaction result of the sample content on the first interaction operation, and a second expected interaction result on the second interaction operation; performing feature extraction on the sample content by a content interaction prediction model to obtain a target first operation feature of the sample content on the first interaction operation, a target second operation feature on the second interaction operation, and a target shared operation feature; determining an initial first actual interaction result of the sample content on the first interaction operation and a second actual interaction result on the second interaction operation based on the target first operation feature and the target second operation feature, respectively; performing logistic regression analysis on the target shared operation feature to predict actual negative noise information generated by the second interaction operation on the first interaction operation, and correcting the initial first actual interaction result based on the actual negative noise information to obtain a target first actual interaction result; adjusting parameters of the content interaction prediction model according to the initial first actual interaction result, the target first actual interaction result, the first expected interaction result, the second actual interaction result, and the second expected interaction result to obtain a trained content interaction prediction model.

10. An electronic device, comprising: The computer readable storage medium stores a plurality of instructions adapted to be loaded by the processor to execute the steps of the content interaction prediction method of any one of claims 1 to 8.

11. A computer readable storage medium, characterized in that, The computer readable storage medium stores a plurality of instructions adapted to be loaded by the processor to execute the steps of the content interaction prediction method of any one of claims 1 to 8.

12. A computer program product comprising computer programs or instructions, characterized in that, The computer program or instructions, when executed by the processor, implement the steps of the content interaction prediction method of any one of claims 1 to 8.

Images