Content recommendation method and device, electronic equipment and computer readable storage medium
By using interaction duration in the content recommendation model for sample classification and confidence loss calculation, the model is updated to improve recommendation efficiency, and the problem of complex and low efficiency of interaction duration features in the prior art is solved, and more accurate content is achieved recommend.
Patent Information
- Application Number
- CN202310595601.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-24
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2043-05-24
AI Technical Summary
When the existing content recommendation method uses interaction duration as a feature, the model training is complex and the information gain is small, which increases resource requirements, and relies on manual experience to lead to insufficient recommendation efficiency.
By obtaining the current training set of the content recommendation model, the recommendation probability of the interactive content samples is predicted, and the samples are classified according to the interaction duration, the interaction confidence loss is calculated, the recommendation results are adjusted, and the content recommendation model is updated to improve the recommendation efficiency.
显式地学习交互持续时间的偏序关系,精准刻画对象与内容之间的交互行为,提升内容推荐模型的精度和效率。
Smart Images

Figure CN119025743A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of content recommendation, and in particular to a content recommendation method, device, electronic device and computer-readable storage medium. Background Art
[0002] In recent years, with the rapid development of Internet technology, more and more content of various types have appeared on the Internet. In order to increase the effective exposure or conversion of these contents, these contents can be recommended to different objects for display. When an object interacts with different content, the duration of continuous interaction is often different. For example, taking click as an example, after the object clicks on the recommended content, the length of stay for different content can often be different. In order to improve the recommendation efficiency of content recommendation, it is necessary to consider the factor of interaction duration. Existing content recommendation methods often use the interaction duration as a training feature of the content recommendation model to recommend content, or use the predicted interaction duration to correct the predicted recommendation probability, and recommend content based on the corrected recommendation probability.
[0003] In the process of research and practice of the prior art, the inventors of the present application found that when the interaction duration is introduced as a feature, since the training of the model often includes many features, the features are acted on the final estimated value through multiple layers of neural networks, and the effect is not direct enough, resulting in a small information gain. The use of interaction duration for weighted correction requires the deployment of an interaction duration prediction model, which increases the resources for model deployment. In addition, the use of interaction duration for weighted correction of recommendation probability often relies on manual experience, and the effect of the correction will be compromised, thereby resulting in insufficient efficiency in content recommendation. Summary of the invention
[0004] Embodiments of the present invention provide a content recommendation method, device, electronic device, and computer-readable storage medium, which can improve the recommendation efficiency of content recommendation.
[0005] A content recommendation method, comprising:
[0006] Acquire a current training set of the content recommendation model, wherein the current training set includes an interaction content sample of each object in the object set and an interaction duration corresponding to the interaction content sample;
[0007] Using the content recommendation model, predicting the recommendation probability of the interactive content sample for the object, and obtaining a predicted recommendation result;
[0008] Classifying the interaction content samples according to the interaction duration, and comparing the predicted recommendation results corresponding to the interaction content samples of different categories to obtain the interaction confidence loss;
[0009] Determining a target recommendation loss of the content recommendation model based on the interaction confidence loss and the predicted recommendation result;
[0010] The content recommendation model is updated according to the target recommendation loss, and the updated content recommendation model is used to recommend the content to be recommended to the target object in the object set.
[0011] Accordingly, an embodiment of the present invention provides a content recommendation device, including:
[0012] An acquisition unit, configured to acquire a current training set of a content recommendation model, wherein the current training set includes an interaction content sample of each object in the object set and an interaction duration corresponding to the interaction content sample;
[0013] A prediction unit, configured to use the content recommendation model to predict the recommendation probability of the interactive content sample for the object, and obtain a predicted recommendation result;
[0014] A comparison unit, configured to classify the interaction content samples according to the interaction duration, and compare the predicted recommendation results corresponding to the interaction content samples of different categories to obtain an interaction confidence loss;
[0015] A determination unit, configured to determine a target recommendation loss of the content recommendation model based on the interaction confidence loss and the predicted recommendation result;
[0016] A recommendation unit is used to update the content recommendation model according to the target recommendation loss, and use the updated content recommendation model to recommend the content to be recommended to the target object in the object set.
[0017] In some embodiments, the comparison unit may be specifically configured to obtain an interaction duration threshold corresponding to the current training set; and classify the interaction content sample based on the interaction duration threshold and the interaction duration.
[0018] In some embodiments, the comparison unit can be specifically used to obtain the historical interaction duration of each object interacting with at least one historical interaction content; determine the time mean of the historical interaction duration, and use the time mean as the initial interaction duration threshold corresponding to each object; and use the initial interaction duration threshold as the interaction duration threshold corresponding to the current training set.
[0019] In some embodiments, the comparison unit can be specifically used to compare the interaction duration with the interaction duration threshold; take the interaction content samples whose interaction duration is greater than or equal to the interaction duration threshold as first category samples to obtain a first category sample set; take the interaction content samples whose interaction duration is less than the interaction duration threshold as second category samples to obtain a second category sample set.
[0020] In some embodiments, the comparison unit can be specifically used to construct a recommendation result pair based on the first category sample set and the second category sample set; calculate the difference between the recommendation results in the recommendation result pair to obtain the recommendation difference; and correct the recommendation difference to obtain the interaction confidence loss.
[0021] In some embodiments, the comparison unit can be specifically used to select one interactive content sample from the first category sample set and one interactive content sample from the second category sample set to obtain a target sample pair; filter out the predicted recommendation result corresponding to the interactive content sample in the target sample pair from the predicted recommendation result to obtain the current recommendation result; and construct the recommendation result pair based on the current recommendation result and the target sample pair.
[0022] In some embodiments, the comparison unit can be specifically used to obtain historical recommendation results corresponding to candidate objects, where the candidate objects include objects corresponding to the interactive content samples in the target sample pair; when the target sample pair contains interactive content samples of the same object, the current recommendation result and the historical recommendation result are combined to obtain the recommendation result pair; when the target sample pair contains interactive content samples of different objects, the current recommendation result is adjusted based on the historical recommendation results to obtain the recommendation result pair.
[0023] In some embodiments, the comparison unit can be specifically used to screen out candidate objects corresponding to the interactive content samples in the target sample pair in the object set; obtain at least one historical recommendation probability of the candidate object, and the historical recommendation probability includes using the content recommendation model to predict the recommendation probability of historical interactive content for the candidate object; determine the probability mean of the historical recommendation probabilities, and use the probability mean as the historical recommendation result corresponding to the candidate object.
[0024] In some embodiments, the comparison unit may be specifically configured to calculate the difference between the historical recommendation result and the current recommendation result to obtain an adjusted current recommendation result; and combine the adjusted current recommendation results to obtain the recommendation result pair.
[0025] In some embodiments, the comparison unit may be specifically configured to combine the current recommendation results to obtain the recommendation result pair when the target sample pair includes interactive content samples of the same object.
[0026] In some embodiments, the comparison unit can be specifically used to sort the recommendation results in the recommendation result pair based on the interaction duration and the interaction duration threshold; and calculate the difference between the recommendation results in the recommendation result pair according to the sorting result to obtain the recommendation difference.
[0027] In some embodiments, the comparison unit can be specifically used to obtain a training penalty value corresponding to the current training set; merge the training penalty value with the recommendation difference to obtain a target recommendation difference; compare the target recommendation difference with a preset recommendation difference, and determine the interaction confidence loss based on the comparison result.
[0028] In some embodiments, the determination unit can be specifically used to obtain the annotated recommendation result corresponding to the interactive content sample; compare the annotated recommendation result with the corresponding predicted recommendation result to obtain an initial recommendation loss; and merge the initial recommendation loss with the interaction confidence loss to obtain the target recommendation loss.
[0029] In some embodiments, the recommendation unit can be specifically used to obtain the content to be recommended, and use the updated content recommendation model to predict the current recommendation probability of the content to be recommended for each object in the object set; based on the current recommendation probability, the target object corresponding to the content to be recommended is screened out in the object set, and the content to be recommended is recommended to the target object.
[0030] In addition, an embodiment of the present invention further provides an electronic device, including a processor and a memory, wherein the memory stores an application program, and the processor is used to run the application program in the memory to implement the content recommendation method provided by the embodiment of the present invention.
[0031] In addition, an embodiment of the present invention further provides a computer-readable storage medium, which stores a plurality of instructions, and the instructions are suitable for a processor to load to execute the steps in any content recommendation method provided by the embodiment of the present invention.
[0032] In addition, an embodiment of the present application also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implements the steps in the content recommendation method provided in the embodiment of the present application.
[0033] After obtaining the current training set of the content recommendation model, the embodiment of the present invention includes the interaction content samples of each object in the object set and the interaction duration corresponding to the interaction content samples. The content recommendation model is used to predict the recommendation probability of the interaction content samples for the object to obtain the predicted recommendation result. Then, the interaction content samples are classified according to the interaction duration, and the predicted recommendation results corresponding to the interaction content samples of different categories are compared to obtain the interaction confidence loss. Then, based on the interaction confidence loss and the predicted recommendation result, the target recommendation loss of the content recommendation model is determined. Then, the content recommendation model is updated according to the target recommendation loss, and the updated content recommendation model is used to recommend the content to be recommended to the target object in the object set. Since the scheme can measure the confidence of the interaction behavior by the interaction duration, and the interaction content samples are classified by the interaction duration, and then the predicted recommendation results corresponding to the interaction content samples of different categories are compared to obtain the interaction confidence loss, the content recommendation model can explicitly learn the partial order relationship based on the interaction duration through the interaction confidence loss, and can accurately describe the true intention of the interaction behavior between the object and the content, improve the accuracy of the content recommendation model, and then improve the recommendation efficiency of the content recommendation. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.
[0035] Figure 1 is a schematic diagram of a scenario of a content recommendation method provided by an embodiment of the present invention;
[0036] Figure 2 is a flowchart of a content recommendation method provided by an embodiment of the present invention;
[0037] Figure 3 is a schematic diagram of a comparison of training methods provided by an embodiment of the present invention;
[0038] Figure 4 is a schematic diagram of displaying content to be recommended provided by an embodiment of the present invention;
[0039] Figure 5 is another flowchart of the content recommendation method provided by an embodiment of the present invention;
[0040] Figure 6 is a structural diagram of a content recommendation device provided by an embodiment of the present invention;
[0041] Figure 7It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.
[0043] The embodiment of the present invention provides a content recommendation method, device and computer-readable storage medium. The content recommendation device can be integrated in an electronic device, which can be a server or a terminal.
[0044] Among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, network acceleration services (Content Delivery Network, CDN), and basic cloud computing services such as big data and artificial intelligence platforms. The terminal can be a smart phone, tablet computer, laptop computer, desktop computer, smart speaker, smart watch, etc., but is not limited to this. The terminal and the server can be directly or indirectly connected via wired or wireless communication, and this application does not limit this.
[0045] For example, see Figure 1 Taking the content recommendation device integrated in an electronic device as an example, the electronic device obtains a current training set of a content recommendation model, which may include an interaction content sample of each object in an object set and an interaction duration corresponding to the interaction content sample. Then, the content recommendation model is used to predict the recommendation probability of the interaction content sample for the object to obtain a predicted recommendation result. The interaction content sample is classified according to the interaction duration, and the predicted recommendation results corresponding to interaction content samples of different categories are compared to obtain an interaction confidence loss. Then, based on the interaction confidence loss and the predicted recommendation result, a target recommendation loss of the content recommendation model is determined, the content recommendation model is updated according to the target recommendation loss, and the updated content recommendation model is used to recommend the content to be recommended to the target object in the object set, thereby improving the recommendation efficiency of content recommendation.
[0046] Among them, the content recommendation method provided in the embodiment of the present application relates to the machine learning direction in artificial intelligence. The embodiment of the present application can adjust the partial order relationship of the predicted recommendation results of the interactive content samples by the interaction duration, so that the model can explicitly learn the partial order relationship, thereby improving the training accuracy of the content recommendation model.
[0047] Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses 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 in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.
[0048] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.
[0049] Among them, machine learning (ML) is a multi-disciplinary interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.
[0050] It can be understood that in the specific implementation of the present application, related data such as the interaction content samples, interaction duration, historical interaction content or historical interaction duration of the object, when the following embodiments of the present application are applied to specific products or technologies, permission or consent is required, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.
[0051] It should be noted that the description order of the following embodiments is not intended to limit the preferred order of the embodiments.
[0052] This embodiment will be described from the perspective of a content recommendation device, which can be specifically integrated into an electronic device, which can be a server or a terminal. The terminal can include a tablet computer, a laptop computer, a personal computer (PC), a wearable device, a virtual reality device, or other smart devices that can perform content recommendation.
[0053] A content recommendation method, comprising:
[0054] A current training set of a content recommendation model is obtained, where the current training set includes interaction content samples of each object in an object set and interaction durations corresponding to the interaction content samples. The content recommendation model is used to predict the recommendation probability of the interaction content samples for the objects to obtain predicted recommendation results. The interaction content samples are classified according to the interaction durations, and the predicted recommendation results corresponding to interaction content samples of different categories are compared to obtain interaction confidence loss. A target recommendation loss of the content recommendation model is determined based on the interaction confidence loss and the predicted recommendation results. The content recommendation model is updated according to the target recommendation loss, and the updated content recommendation model is used to recommend the content to be recommended to the target object in the object set.
[0055] like Figure 2 As shown, the specific process of the content recommendation method is as follows:
[0056] 101. Obtain the current training set of the content recommendation model.
[0057] The content recommendation model may be a model for recommending content to objects in the object set. The content recommendation model may include an untrained initial content recommendation model or a model trained using a historical training set corresponding to the object set.
[0058] Among them, the current training set may include the interaction content samples of each object in the object set and the interaction duration corresponding to the interaction content samples. The so-called object set may include at least one user who uses the content recommendation model to recommend content. The corresponding interaction content sample may be a content sample that interacts with the user. There may be many types of interactions, for example, clicks and conversions. Clicks may include opening, forwarding, liking, commenting, collecting, screenshots, copying or content selection, etc. Conversions may include registering an account, downloading an APP, following a public account or submitting various materials, etc. It should be noted that the interaction content samples may include positive samples and negative samples, so-called positive samples may be content samples that have interacted with the object, and negative samples may be content samples that have not interacted with the object. There may be many types of content for the interaction content samples, for example, text, audio and video or other types of content.
[0059] The interaction duration corresponding to the interactive content sample can be the duration of the interaction between the object and the positive sample. Taking click as an example, the interaction duration can include the time the object stays on the content sample after the click, which can also be called the landing page residence time. For example, object a exits 0.1s after clicking on ad a. At this time, the residence time for ad a can be 0.1s.
[0060] There are many ways to obtain the current training set of the content recommendation model, which can be as follows:
[0061] For example, the current training set of the content recommendation model uploaded by the terminal or client can be directly received, or the interaction record of each object in the object set can be obtained, and based on the interaction record, the interaction content sample of each object is screened out in the network or content database, and the interaction duration corresponding to the interaction content sample is counted in the interaction record, so as to obtain the current training set of the content recommendation model, or the interaction content corresponding to at least one object is screened out in the network or database to obtain the interaction content sample, and the interaction duration corresponding to the interaction content sample is determined, so as to obtain the current training set of the content recommendation model, or, when the memory of the current training set is large or the number of samples is large, an update request for the content recommendation model can also be received, and the update request can include the storage address of the current training set, and the current training set is obtained based on the storage address, and so on.
[0062] 102. Use the content recommendation model to predict the recommendation probability of the interactive content sample for the object and obtain the predicted recommendation result.
[0063] The predicted recommendation result may represent the probability or degree of recommending the interactive content sample to the corresponding object.
[0064] There are many ways to use the content recommendation model to predict the recommendation probability of the interactive content sample for the object, which can be as follows:
[0065] For example, a content recommendation model can be used to extract features from interactive content samples to obtain content features of multiple dimensions, and the content features can be fused to obtain recommendation features of the interactive content samples for corresponding objects in the object set. Based on the recommendation features, the recommendation probability of the interactive content samples for the object can be determined, thereby obtaining the predicted recommendation results corresponding to the interactive content samples.
[0066] Among them, the recommendation probability may include recommending the interactive content sample to the corresponding object, and the probability that the object will click or convert the interactive content and other interactive operations. For example, when the interaction is a click, the recommendation probability can be the probability that the object clicks on the interactive content, which can also be called the click-through rate for the interactive content sample.
[0067] 103. According to the interaction duration, the interaction content samples are classified, and the predicted recommendation results corresponding to the interaction content samples of different categories are compared to obtain the interaction confidence loss.
[0068] Among them, the interaction confidence loss can characterize the loss information of the confidence of the object in interacting with the interactive content sample. The so-called interaction confidence can be used to evaluate the authenticity of the interaction. A good observation indicator for measuring the confidence of the interaction can include the duration of the interaction. Taking the interaction as a click as an example, the interaction duration can be the dwell time of the click or the dwell time of the landing page. The confidence of the interaction can be measured by the dwell time of the click. For example, for an object u, it exits 0.1s after clicking on ad a and exits 5s after clicking on ad b. Then for ad a, it may be a wrong click, that is, the confidence of this click is not high or the credibility of the click behavior in conveying the interaction intention is not high.
[0069] The process of classifying the interaction content samples according to the interaction duration and comparing the prediction recommendation results corresponding to the interaction content samples of different categories to obtain the interaction confidence loss can be specifically as follows:
[0070] S1. Classify the interaction content samples according to the interaction duration.
[0071] For example, an interaction duration threshold corresponding to the current training set may be obtained, and the interaction content samples may be classified based on the interaction duration threshold and the interaction duration.
[0072] The interaction duration threshold may be a threshold for dividing the interaction duration into long and short ones. For example, if it is greater than or equal to the interaction duration threshold, it indicates that the interaction duration of the interaction content sample is long, or if it is less than the interaction duration threshold, it indicates that the interaction duration of the interaction content sample is short. There may be multiple ways to obtain the interaction duration threshold corresponding to the current training set, for example, the historical interaction duration of each object interacting with at least one historical interaction content may be obtained, the time mean of the historical interaction duration may be determined, and the time mean may be used as the initial interaction duration threshold corresponding to each object, and the initial interaction duration threshold may be used as the interaction duration threshold corresponding to the current training set, or a preset interaction duration threshold may be obtained, and the preset interaction duration threshold may be used as the interaction duration threshold corresponding to the current training set, or the content type of the interaction content sample in the current training set may be obtained, and the time threshold corresponding to the content type may be screened out from the preset time threshold set, thereby obtaining the interaction duration threshold corresponding to the current training set, or the object identifier of each object in the object set may be used, and the time threshold corresponding to the object identifier may be screened out from the preset time threshold set, thereby obtaining the interaction duration threshold, and so on.
[0073] The historical interaction content may include at least one content in which each object interacts within a historical time period.
[0074] After obtaining the interaction duration threshold corresponding to the current training set, the interaction content samples can be classified based on the interaction duration threshold and the interaction duration. There are many ways to classify, for example, comparing the interaction duration with the interaction duration, taking the interaction content samples whose interaction duration is greater than or equal to the interaction duration threshold as the first category samples to obtain the first category sample set, and taking the interaction content samples whose interaction duration is less than the interaction duration threshold as the second category samples to obtain the second category sample set.
[0075] S2. Compare the predicted recommendation results corresponding to different categories of interactive content samples to obtain the interactive confidence loss.
[0076] For example, based on the first category sample set and the second category sample set, a recommendation result pair can be constructed, the difference between the recommendation results in the recommendation result pair is calculated to obtain a recommendation difference, and the recommendation difference is corrected to obtain an interaction confidence loss, which can be specifically as follows:
[0077] C1. Construct a pair of recommendation results based on the first category sample set and the second category sample set.
[0078] The recommendation result pair may include recommendation results of different categories of interactive content samples, where the recommendation results may include predicted recommendation results, historical predicted results, revised predicted recommendation results, etc. The recommendation result pair may also be called a pair, which is used to characterize the partial order relationship of the interaction duration.
[0079] There are multiple ways to construct recommendation result pairs based on the first category sample set and the second category sample set, which can be as follows:
[0080] For example, one interactive content sample can be selected from each of the first category sample set and the second category sample set to obtain a target sample pair, and the predicted recommendation result corresponding to the interactive content sample in the target sample pair can be filtered out in the predicted recommendation results to obtain the current recommendation result, and a recommendation result pair can be constructed based on the current recommendation result and the target sample pair.
[0081] The target sample pairs may include interactive content samples of different categories.
[0082] Among them, the current recommendation result can be the predicted recommendation result corresponding to each interactive content sample in the target sample pair. Based on the current recommendation result and the target sample pair, there can be multiple ways to construct the recommendation result pair. For example, the historical recommendation results corresponding to the candidate object can be obtained. When the target sample pair contains interactive content samples of the same object, the current recommendation result and the historical recommendation result are combined to obtain the recommendation result pair. When the target sample pair contains interactive content samples of different objects, the current recommendation result is adjusted based on the historical recommendation result to obtain the recommendation result pair. The specific method can be as follows:
[0083] (1) Obtain the historical recommendation results corresponding to the candidate object.
[0084] The candidate objects include objects corresponding to the interactive content samples in the target sample pair.
[0085] There are many ways to obtain the historical recommendation results corresponding to the candidate object, which can be as follows:
[0086] For example, the candidate objects corresponding to the interactive content samples in the target sample pair can be screened out in the object set, at least one historical recommendation probability of the candidate objects can be obtained, the probability mean of the historical recommendation probabilities can be determined, and the probability mean can be used as the historical recommendation result corresponding to the candidate objects.
[0087] The historical recommendation probability may include using a content recommendation model to predict the recommendation probability of historical interaction content for the candidate object. There are multiple ways to obtain at least one historical recommendation probability of the candidate object. For example, at least one historical interaction content corresponding to the candidate object may be obtained, and the content recommendation model may be used to predict the recommendation probability corresponding to the historical interaction content to obtain the historical recommendation probability.
[0088] It should be noted that the historical interaction content here can be the same as or different from the historical interaction content when determining the interaction duration threshold. However, the interaction time of these two historical interaction contents can be the same interaction time period. In this way, it can be ensured that the interaction duration threshold and the historical recommendation probability determined based on the historical interaction content are in the same dimension, thereby ensuring the update accuracy of the content recommendation model. It should also be noted that each object in the object set corresponds to a historical recommendation result, and the recommendation time period of the historical recommendation results of different objects can be the same.
[0089] (2) When the target sample pair contains interactive content samples of the same object, the current recommendation result and the historical recommendation result are combined to obtain a recommendation result pair.
[0090] For example, when the target sample pair includes interactive content samples of the same object, the current recommendation result of each interactive content sample in the target sample pair is combined with the historical recommendation result to obtain a recommendation result pair.
[0091] Among them, since the target sample pair includes two interactive content samples, the recommendation result pair can include the recommendation result pair consisting of each interactive content sample and the historical recommendation result. For example, taking the case where the target sample pair includes interactive content sample A and interactive content sample B corresponding to object u, the current recommendation result of interactive content sample A is recommendation result A, the current recommendation result of interactive content sample B is recommendation result B, and the historical recommendation result corresponding to object u is recommendation result C, at this time, the recommendation result pair can include the recommendation result pair consisting of recommendation result A and recommendation result C, and can also include the recommendation result pair consisting of recommendation result B and recommendation result C.
[0092] (3) When the target sample pair contains interactive content samples of different objects, the current recommendation result is adjusted based on the historical recommendation results to obtain a recommendation result pair.
[0093] For example, the difference between the historical recommendation result and the current recommendation result may be calculated to obtain the adjusted current recommendation result, and the adjusted current recommendation result may be combined to obtain a recommendation result pair.
[0094] There are many ways to calculate the difference between the historical recommendation result and the current recommendation result. For example, the current recommendation result can be subtracted from the historical recommendation result to obtain the difference between the historical recommendation result and the current recommendation result, and the difference is used as the current recommendation result. Specifically, it can be shown in formula (1):
[0095]
[0096] in, is the current recommendation result after adjustment, p(u,a) is the current recommendation result, is the historical recommendation result, u is the object identifier, a is the interactive content, and the interactive content of the current recommendation result can be the same as or different from the interactive content in the historical recommendation content.
[0097] Among them, it should be noted that since the target sample pair contains interactive content samples of different objects, and the differences in some behavioral habits of different objects will lead to large differences between the corresponding predicted recommendation results. Since the predicted recommendation results corresponding to the interactive content samples of different objects need to be compared in the subsequent interactive confidence loss, if the predicted recommendation results corresponding to different objects themselves are greatly different, they will not be able to be compared, or the comparison difference is large. Therefore, when constructing the recommendation result pairs corresponding to different objects, the behavioral differences of different objects can be smoothed by normalization, so that the predicted recommendation results corresponding to different objects are in the same calculation dimension. The normalization here can be understood as normalizing the predicted recommendation results of the interactive content samples according to the average estimated value (historical recommendation results), so that the current recommendation results corresponding to different objects in the target sample pair are comparable, thereby obtaining a more accurate interactive confidence loss.
[0098] Optionally, in some embodiments, based on the current recommendation result and the target sample pair, the method of constructing a recommendation result pair may also include: when the target sample pair contains interactive content samples of the same object, combining the current recommendation results to obtain a recommendation result pair. For example, taking the target sample pair containing two different interactive content samples A and B of the same object u, and the current recommendation result corresponding to interactive content sample A is recommendation result A, and the current recommendation result corresponding to interactive content sample B is recommendation result B, recommendation result A and recommendation result B can be combined to obtain a recommendation result pair {recommendation result A, recommendation result B}, and so on.
[0099] C2. Calculate the difference between the recommended results in the recommended result pair to obtain the recommended difference.
[0100] The recommendation difference may be the difference between the recommendation results. For example, taking click as an example, the recommendation result may be the click rate estimate, and the corresponding recommendation difference may be the difference between the click rate estimates.
[0101] There are many ways to calculate the difference between the recommendation results in the recommendation result pair, which can be as follows:
[0102] For example, the recommendation results in the recommendation result pair may be sorted based on the interaction duration and the interaction duration threshold, and the difference between the recommendation results in the recommendation result pair may be calculated according to the sorting result to obtain the recommendation difference.
[0103] Among them, there are multiple ways to sort the recommended results in the recommended result pair based on the interaction duration and the interaction duration threshold. For example, the recommended results in the recommended result pair can be sorted according to the size of the interaction duration corresponding to the recommended result in the recommended result pair. For example, when the recommended result pair includes the current recommended result, the current recommended result can be sorted according to the size of the interaction duration corresponding to the current recommended result. When the recommended result pair includes the adjusted current recommended result, the adjusted current recommended result can be sorted according to the interaction duration corresponding to the adjusted current recommended result before adjustment. When the recommended result pair includes the current recommended result and the historical recommended result, the interaction duration threshold can be used as the interaction duration corresponding to the historical recommended result, and then the current recommended result and the historical recommended result can be sorted according to the size of the interaction duration corresponding to the current recommended result and the historical recommended result, respectively.
[0104] After the recommendation results in the recommendation result pair are sorted, the difference between the recommendation results in the recommendation result pair can be calculated according to the sorting result to obtain the recommendation difference. There are many ways to calculate the difference between the recommendation results in the recommendation result pair. For example, based on the sorting result, a target recommendation result with a short interaction duration can be screened out in the recommendation result pair, and the target recommendation result is used as a minuend, and the other recommendation result in the recommendation result pair is used as a subtrahend. Based on the minuend and the subtrahend, the recommendation difference between the recommendation results is calculated.
[0105] Among them, taking the example that the recommendation result pair includes recommendation results A and B corresponding to the same object, and the interaction duration corresponding to recommendation result A is less than the interaction duration corresponding to recommendation result B, the recommendation difference can be the difference between recommendation result A and recommendation result B, or, taking the example that the recommendation result pair includes recommendation result A and historical recommendation result B corresponding to the same object, and the interaction duration corresponding to recommendation result A is less than the interaction duration threshold corresponding to historical recommendation result B, the recommendation difference can be the difference between recommendation result A and historical recommendation result B, or, taking the example that the recommendation result pair includes adjusted recommendation result A and adjusted recommendation result B corresponding to different objects, and the interaction duration corresponding to the adjusted recommendation result A before adjustment is less than the interaction duration corresponding to the adjusted recommendation result B before adjustment, the recommendation difference can be the difference between the adjusted recommendation result A and the adjusted recommendation result B, and so on.
[0106] C3. Correct the recommendation difference to obtain the interaction confidence loss.
[0107] For example, the training penalty value corresponding to the current training set can be obtained, the training penalty value can be merged with the recommendation difference to obtain the target recommendation difference, the target recommendation difference can be compared with the preset recommendation difference, and the interaction confidence loss can be determined based on the comparison result.
[0108] The training penalty value can be understood as a penalty term in the model training process, or as a supplementary adjustment of the loss function, which is mainly used to alleviate overfitting. The training penalty value can be a pre-set penalty term or a randomly generated penalty term. The training penalty value can be 0 or any other value. The training penalty values corresponding to different training sets can be the same or different. There are many ways to merge the training penalty value with the recommendation difference value. For example, the training penalty value can be directly added to the recommendation difference value to obtain the target recommendation difference value.
[0109] Among them, the preset recommendation difference can be a pre-set recommendation difference, which is used to compare with the calculated target recommendation difference, so as to select the recommendation difference for determining the interaction confidence loss. After comparing the target recommendation difference with the preset recommendation difference, there can be multiple ways to determine the interaction confidence loss based on the comparison result. For example, when the target recommendation difference exceeds the preset recommendation difference, the target recommendation difference is merged with the preset confidence parameter to obtain the interaction confidence loss. When the target recommendation difference does not exceed the preset recommendation difference, the preset recommendation difference is merged with the preset confidence parameter to obtain the interaction confidence loss.
[0110] Among them, taking the recommendation result pair including the recommendation results A and B corresponding to two different interaction content samples of the same object u, and the interaction duration corresponding to the recommendation result A is less than the interaction duration corresponding to the recommendation result B, and the preset recommendation difference is 0 as an example, the interaction confidence loss corresponding to the recommendation result pair can be shown as formula (2), which can be specifically as follows:
[0111] L1=λ|max(f clicj (v - ,w c )-f click (v + ,w c )+τ),0| (2)
[0112] Among them, L1 is the interaction confidence loss, f click (v _ ,w c ) is the recommendation result A, f click (v + ,w c ) is the recommendation result B, v - is the interactive content sample whose interaction duration is less than the interaction duration threshold, v + is the interactive content sample whose interaction duration is greater than or equal to the interaction duration threshold, w c is the training weight, and τ is the training penalty value.
[0113] Among them, taking the recommendation result pair including the current recommendation result A and the historical recommendation result B of the same object u, and the preset recommendation difference value is 0 as an example, when the interaction duration corresponding to the current recommendation result A is less than the interaction duration threshold, the interaction confidence loss corresponding to the recommendation result pair can be shown as formula (3), which can be specifically as follows:
[0114]
[0115] Among them, L1 is the interaction confidence loss, f click (u,w c ) is the current recommendation result A, is the historical recommendation result B, u is the corresponding object, τ is the training penalty value, and w c is the training weight.
[0116] Optionally, when the interaction duration corresponding to the current recommendation result A is greater than the interaction duration threshold, the interaction confidence loss corresponding to the recommendation result A may be as shown in formula (4), which may be as follows:
[0117]
[0118] Among them, L1 is the interaction confidence loss, f click(u,w c ) is the current recommendation result A, is the historical recommendation result B, u is the corresponding object, τ is the training penalty value, and w c is the training weight.
[0119] Among them, taking the recommendation result pair including the adjusted recommendation results A and B corresponding to the interactive content samples of different objects, the interaction duration corresponding to the adjusted recommendation result A before adjustment is less than the interaction duration corresponding to the adjusted recommendation result B before adjustment, and the preset recommendation difference is 0 as an example, the interaction confidence loss corresponding to the recommendation result pair can be shown as formula (5), which can be specifically as follows:
[0120]
[0121] Among them, L1 is the interaction confidence loss, is the adjusted recommendation result A, is the adjusted recommendation result B,u - is the object corresponding to the recommendation result A, u + is the object corresponding to the recommendation result B, a1 is the interactive content sample corresponding to the recommendation result A, and a2 is the interactive content sample corresponding to the recommendation result B.
[0122] 104. Based on the interaction confidence loss and the predicted recommendation results, determine the target recommendation loss of the content recommendation model.
[0123] The target recommendation loss may include the loss generated when the content recommendation model performs content recommendation on the interactive content samples in the current training set.
[0124] There are many ways to determine the target recommendation loss of the content recommendation model based on the interaction confidence loss and the predicted recommendation result, which can be as follows:
[0125] For example, the annotated recommendation results corresponding to the interactive content samples can be obtained, and the annotated recommendation results can be compared with the corresponding predicted recommendation results to obtain the initial recommendation loss, and the initial recommendation loss and the interactive confidence loss can be fused to obtain the target recommendation loss.
[0126] The annotated recommendation result may be a recommendation result annotated on the interactive content sample, indicating whether the interactive content sample is actually recommended to the corresponding object, or may also indicate whether the object actually clicks on the interactive content sample or converts based on the interactive content sample after the interactive content sample is pushed to the corresponding object. The annotated recommendation result may be a binary label, for example, 1 if the object clicks on the interactive content sample, and 0 if the object does not click on the interactive content sample. The annotated recommendation result may also be a multivariate label, which characterizes the degree to which the object actually clicks on or converts the interactive content sample through a multivariate recommendation probability value, for example, it may be any value between [0,1], the larger the value, the higher the degree to which the object actually clicks on or converts the user, and so on.
[0127] The initial recommendation loss may be a recommendation loss obtained by taking click or conversion as a supervisory signal, which is a supervised learning loss. There are many ways to compare the labeled recommendation results with the corresponding predicted recommendation results to obtain the initial recommendation loss. For example, a cross entropy loss function may be used to compare the labeled results with the corresponding predicted recommendation results to obtain the initial recommendation loss. Alternatively, other types of loss functions may be used to compare the labeled results with the corresponding predicted recommendation results to obtain the initial recommendation loss.
[0128] After comparing the labeled recommendation results with the corresponding predicted recommendation results to obtain the initial recommendation loss, the initial recommendation loss can be fused with the interaction confidence loss to obtain the target recommendation loss. There are many ways to fuse the initial recommendation loss with the interaction confidence loss. For example, the initial recommendation loss of each interactive content sample in the target sample pair corresponding to the interaction confidence loss can be screened out from the initial recommendation loss to obtain the candidate recommendation loss corresponding to the interaction confidence, and the candidate recommendation loss is added to the interaction confidence loss to obtain the current recommendation loss corresponding to the interactive content sample in the target sample pair or the cross-target sample pair. The current recommendation loss is accumulated to obtain the target recommendation loss corresponding to the current training set. Alternatively, the initial recommendation loss of each interactive content sample in the target sample pair corresponding to the interaction confidence loss can be screened out from the initial recommendation loss to obtain the candidate recommendation loss corresponding to the interaction confidence, and the training weights are obtained. Based on the training weights, the candidate recommendation loss and the interaction confidence loss are weighted respectively, and the weighted candidate recommendation loss and the weighted interaction confidence loss are added to obtain the current recommendation loss corresponding to the target sample pair. The current recommendation loss is accumulated to obtain the target recommendation loss corresponding to the current training set, and so on.
[0129] The current recommendation loss may include the initial recommendation loss corresponding to the interactive content sample in the target sample pair and the interactive confidence loss of the corresponding recommendation result pair. Taking the loss function corresponding to the initial recommendation loss as the cross entropy loss as an example, the specific formula (6) may be as follows:
[0130] L=L CE (y click ,f click (v,w c ))+L1(6)
[0131] Among them, L is the current recommendation loss, L CE (y click ,f click (v,w c )) is the initial recommendation loss, y click is the annotation recommendation result, f click (v,w c ) is the predicted recommendation result, v is the content feature (embedding) of the interactive content sample, and w c is the training weight, and L1 is the interaction confidence loss corresponding to the target sample pair.
[0132] It should be noted that, for the initial recommendation loss, taking the example that the target sample pair corresponding to the interaction confidence loss includes the interaction content sample A and the interaction content sample B, the initial recommendation loss corresponding to the interaction confidence loss can include the initial recommendation loss corresponding to these two interaction content samples. In addition, it should be noted that the target sample pair can correspond to one recommendation result pair or two recommendation result pairs, so the interaction confidence loss corresponding to the target sample pair can also be one or more. For example, the target sample pair includes interactive content sample A and interactive content sample B, the recommendation result A corresponding to interactive content sample A, and the recommendation result B corresponding to interactive content sample B. The corresponding recommendation result pair may include {recommendation result A, recommendation result B}, or may include {recommendation result A, historical recommendation result} and {recommendation result B, historical recommendation result}. When the recommendation result pair is {recommendation result A, recommendation result B}, the corresponding initial recommendation loss may include the initial recommendation losses corresponding to the interactive content sample A and the interactive content sample B respectively. When the recommendation result pair is {recommendation result A, historical recommendation result} and {recommendation result B, historical recommendation result}, the initial recommendation loss corresponding to {recommendation result A, historical recommendation result} may be the initial recommendation loss corresponding to the interactive content sample A, and the initial recommendation loss corresponding to {recommendation result B, historical recommendation result} may be the initial recommendation loss corresponding to the interactive content sample B. When determining the current recommendation loss, the initial recommendation loss of the same interactive content sample and the corresponding interactive confidence loss can be added to obtain the current recommendation loss corresponding to the interactive content sample. Alternatively, the initial recommendation loss of the interactive content sample in the same target sample pair and the corresponding interactive confidence loss can be added to obtain the current recommendation loss corresponding to the target sample pair.
[0133] After determining the current recommendation loss, the current recommendation loss can be accumulated to obtain the target recommendation loss of the current training set, and the target recommendation loss is used as the target recommendation loss of the content recommendation model.
[0134] Among them, in the training of content recommendation models, a supervised learning loss function (i.e., the initial recommendation loss corresponding to the cross entropy loss function) is generally used. On this basis, in this scheme, a new contrastive learning / self-supervised learning (SL) loss function (i.e., the corresponding interactive confidence loss) can be added. Taking the content recommendation model as DNN as an example, the comparison between the existing training method and the training method of this scheme can be specifically as follows: Figure 3As shown, the partial order relationship based on the interaction duration is explicitly learned, so that the predicted recommendation results of the interaction content samples corresponding to different interaction durations can be adjusted, so that the predicted recommendation results corresponding to the interaction content samples with short interaction durations are smaller than the predicted recommendation results corresponding to the interaction content samples with long interaction durations. This can more accurately characterize the true intention of the object's interaction behavior, improve the training accuracy of the content recommendation model, and further improve the recommendation efficiency of content recommendation.
[0135] 105. Update the content recommendation model according to the target recommendation loss, and use the updated content recommendation model to recommend the content to be recommended to the target object in the object set.
[0136] The content to be recommended may be content that needs to be recommended, and there may be various types of content to be recommended, for example, it may include text, audio and video, hyperlinks, or other content that can be pushed or exposed, and so on.
[0137] There are many ways to update the content recommendation model according to the target recommendation loss, which can be as follows:
[0138] For example, a gradient descent algorithm may be used to update the network parameters of the content recommendation model based on the target recommendation loss to obtain an updated content recommendation model, and the updated content recommendation model is used as the content recommendation model, and the step of obtaining the current training set of the content recommendation model is returned to execute until the content recommendation model converges to obtain an updated target recommendation model. Alternatively, other network parameter updating algorithms may be used to update the network parameters of the content recommendation model based on the target recommendation loss to obtain an updated content recommendation model, and the updated content recommendation model is used as the content recommendation model, and the step of obtaining the current training set of the content recommendation model is returned to execute until the content recommendation model converges to obtain an updated target recommendation model.
[0139] After the content recommendation model is updated, the updated content recommendation model can be used to recommend the content to be recommended to the target object in the object set. There are many ways to recommend the recommended content to the target object in the object set. For example, the content to be recommended can be obtained, and the updated content recommendation model can be used to predict the current recommendation probability of the content to be recommended for each object in the object set. Based on the current recommendation probability, the target object corresponding to the content to be recommended is screened out in the object set, and the content to be recommended is recommended to the target object.
[0140] There are many ways to obtain the content to be recommended. For example, at least one content uploaded by the terminal or client can be directly obtained to obtain the content to be recommended, or at least one content can be screened out from the network or content library to obtain the content to be recommended, or at least one published content can be obtained on the content interaction platform to obtain the content to be recommended, or, when the number of contents to be recommended is large or the memory is large, a content recommendation request can be received, which carries the storage address of the content to be recommended, and the content to be recommended is obtained based on the storage address, and so on.
[0141] After obtaining the content to be recommended, the updated content recommendation model is used to predict the current recommendation probability of the content to be recommended for each object in the object set. The method of predicting the current recommendation probability can be similar to the method of predicting the recommendation probability of the interactive content sample for the object, as described above, and will not be repeated here.
[0142] After predicting the recommendation probability, the target object corresponding to the content to be recommended can be screened out from the object set based on the current recommendation probability. There are many ways to screen out the target object. For example, at least one object whose current recommendation probability exceeds a preset general threshold can be screened out from the object set to obtain the target object, or the objects in the object set can be classified based on the current recommendation probability, and at least one target object corresponding to a preset sorting interval (for example, Top N, where N is an integer greater than 0) can be screened out from the object set based on the classification result to obtain the target object, and so on.
[0143] After the target object is screened, the content to be recommended can be recommended to the target object. There are many ways to recommend to the target object, for example, the content to be recommended can be directly pushed to the terminal corresponding to the target object, or the content to be recommended can be displayed on the terminal corresponding to the target object. Figure 4 As shown, etc.
[0144] There are many ways to display the content to be recommended on the terminal or client of the target object. For example, it can be displayed in the form of a patch content in the page currently browsed by the target object, or it can be directly displayed in a preset area in the terminal or client of the target object, or the content to be displayed can be displayed in the form of a hyperlink on the terminal or client of the target object, and the terminal or client responds to the trigger response to the hyperlink and displays the content to be recommended. Alternatively, the content to be recommended can also be displayed on the terminal or client of the target object. When the preset time threshold is exceeded and the terminal or client does not receive the trigger operation of the target object, the content to be recommended is withdrawn, set to transparent, or the display of the content to be recommended is stopped.
[0145] After receiving the content to be recommended, the target object can interact with the content to be recommended through a terminal or a client. There are many ways of interaction, such as clicking or converting, etc.
[0146] From the above, it can be seen that after obtaining the current training set of the content recommendation model, the embodiment of the present application includes the interaction content sample of each object in the object set and the interaction duration corresponding to the interaction content sample, and adopts the content recommendation model to predict the recommendation probability of the interaction content sample for the object to obtain the predicted recommendation result, and then, according to the interaction duration, the interaction content sample is classified, and the predicted recommendation results corresponding to the interaction content samples of different categories are compared to obtain the interaction confidence loss, and then, based on the interaction confidence loss and the predicted recommendation result, the target recommendation loss of the content recommendation model is determined, and then, the content recommendation model is recommended according to the target recommendation loss. The recommendation model is updated, and the updated content recommendation model is used to recommend the content to be recommended to the target object in the object set; since the scheme can measure the confidence of the interaction behavior by the interaction duration, and classify the interaction content samples by the interaction duration, and then compare the predicted recommendation results corresponding to the interaction content samples of different categories, so as to obtain the interaction confidence loss. Through the interaction confidence loss, the content recommendation model can explicitly learn the partial order relationship based on the interaction duration, which can accurately characterize the true intention of the interaction behavior between the object and the content, improve the accuracy of the content recommendation model, and then improve the recommendation efficiency of the content recommendation.
[0147] According to the method described in the above embodiment, the following is further described in detail with examples.
[0148] In this embodiment, the content recommendation device is specifically integrated in an electronic device, the electronic device is a server, the object set is a user set, the object is a user, the interactive content sample is an advertising sample, the interaction is a click, the interaction duration is the click dwell time, the content to be recommended is an advertisement, the content recommendation model is an advertising recommendation model, the recommendation probability is the click-through rate, and the predicted recommendation result is the click-through rate estimate.
[0149] like Figure 5 As shown, a content recommendation method, the specific process is as follows:
[0150] 201. The server obtains a current training set of the advertisement recommendation model.
[0151] For example, the server can directly receive the current training set of the advertising recommendation model uploaded by the terminal or client, or obtain the click record of each user in the user set, and based on the click record, screen out the advertising samples corresponding to each user in the network or content database, and count the click dwell time corresponding to the advertising sample in the click record, so as to obtain the current training set of the advertising recommendation model, or screen out the advertisement corresponding to at least one user in the network or database, obtain the advertising sample, and determine the click dwell time corresponding to the advertising sample, so as to obtain the current training set of the advertising recommendation model, or, when the memory of the current training set is large or the number of samples is large, it can also receive an update request for the advertising recommendation model, and the update request can include the storage address of the current training set, and based on the storage address, obtain the current training set, and so on.
[0152] 202. Use the advertising recommendation model to predict the click rate of the advertising sample for the user and obtain the estimated click rate value.
[0153] For example, the server can use an advertising recommendation model to extract features from advertising samples, obtain content features in multiple dimensions, fuse the content features, obtain click features of the advertising samples for the corresponding users in the user set, and determine the click rate of the advertising samples for the user based on the click features, thereby obtaining an estimated click rate corresponding to the advertising samples.
[0154] 203. The server classifies the advertisement samples according to the click dwell time.
[0155] For example, the server can obtain the historical click dwell time of each user and at least one historical advertising sample, determine the time mean of the historical click dwell time, and use the time mean as the initial click dwell time threshold corresponding to each user, and use the initial click dwell time threshold as the click dwell time threshold corresponding to the current training set, or obtain a preset click dwell time threshold, and use the preset click dwell time threshold as the click dwell time threshold corresponding to the current training set, or obtain the advertising type of the advertising sample in the current training set, filter out the time threshold corresponding to the content type in the preset time threshold set, and thus obtain the click dwell time threshold corresponding to the current training set, or, based on the user ID of each user in the user set, filter out the time threshold corresponding to the user ID in the preset time threshold set, and obtain the click dwell time threshold, and so on.
[0156] The server compares the click dwell time with the click dwell time, takes the advertising samples whose click dwell time is greater than or equal to the click dwell time threshold as the first category samples to obtain the first category sample set, and takes the advertising samples whose click dwell time is less than the click dwell time threshold as the second category samples to obtain the second category sample set.
[0157] 204. The server constructs an estimation value pair based on the first category sample set and the second category sample set.
[0158] For example, the server can select one advertising sample from the first category sample set and one advertising sample from the second category sample set to obtain a target sample pair, and filter out the estimated click rate estimate corresponding to the advertising sample in the target sample pair from the estimated click rate estimate to obtain the current estimate.
[0159] The server can filter out candidate users corresponding to the advertisement samples in the target sample pair from the user set, obtain at least one historical advertisement sample corresponding to the candidate user, use the advertisement recommendation model to predict the click-through rate corresponding to the historical advertisement sample, and obtain the historical click-through rate. Determine the probability mean of the historical click-through rate, and use the probability mean as the historical estimated value corresponding to the candidate user.
[0160] When the target sample pair includes advertising samples of the same user, the server combines the current estimated value of each advertising sample in the target sample pair with the historical estimated value to obtain an estimated value pair; or, when the target sample pair includes advertising samples of different users, the server calculates the difference between the historical estimated value and the current estimated value to obtain an adjusted current estimated value, and combines the adjusted current estimated values to obtain an estimated value pair; or, when the target sample pair includes advertising samples of the same user, the server combines the current estimated values to obtain an estimated value pair.
[0161] 205. The server calculates the difference between the estimated values in the estimated value pair to obtain an estimated difference.
[0162] For example, when the estimate pair includes the current estimate, the server can sort the current estimate according to the click dwell time corresponding to the current estimate; when the estimate pair includes the adjusted current estimate, the server can sort the adjusted current estimate according to the click dwell time corresponding to the adjusted current estimate before adjustment; when the estimate pair includes the current estimate and historical estimate, the click dwell time threshold can be used as the click dwell time corresponding to the historical estimate, and then the current estimate and the historical estimate can be sorted according to the click dwell time corresponding to each of them.
[0163] Based on the sorting result, the server can filter out the target estimated value with a smaller click dwell time from the estimated value pair, use the target estimated value as the minuend, use the other estimated value in the estimated value pair as the subtrahend, and calculate the estimated difference between the estimated values based on the minuend and the subtrahend.
[0164] 206. The server corrects the estimated difference to obtain the interaction confidence loss.
[0165] For example, the server can obtain the training penalty value corresponding to the current training set, add the training penalty value to the estimated difference, and thus obtain the target estimated difference. The target estimated difference is compared with the preset estimated difference. When the target estimated difference exceeds the preset estimated difference, the target estimated difference is merged with the preset confidence parameter to obtain the interactive confidence loss. When the target estimated difference does not exceed the preset estimated difference, the preset estimated difference is merged with the preset confidence parameter to obtain the interactive confidence loss. Specifically, it can be shown in formulas (2), (3), (4) and (5). See the above description for details, and will not be repeated here.
[0166] 207. The server determines a target click loss of the advertising recommendation model based on the interaction confidence loss and the click rate estimate.
[0167] For example, the server may obtain the annotation estimate corresponding to the advertisement sample, and use the cross entropy loss function to compare the annotation result with the corresponding estimate click rate estimate, thereby obtaining the initial recommendation loss. Alternatively, other types of loss functions may be used to compare the annotation result with the corresponding click rate estimate, thereby obtaining the initial recommendation loss.
[0168] The server selects the initial recommendation loss of each advertisement sample in the target sample pair corresponding to the interaction confidence loss from the initial recommendation loss, obtains the candidate recommendation loss corresponding to the interaction confidence, adds the candidate recommendation loss to the interaction confidence loss, thereby obtaining the current recommendation loss corresponding to the advertisement sample in the target sample pair or the cross-target sample pair, and accumulates the current recommendation losses to obtain the target recommendation loss corresponding to the current training set. Alternatively, the server may select the initial recommendation loss of each advertisement sample in the target sample pair corresponding to the interaction confidence loss from the initial recommendation loss, obtain the candidate recommendation loss corresponding to the interaction confidence, obtain the training weight, and weight the candidate recommendation loss and the interaction confidence loss based on the training weight, respectively, and add the weighted candidate recommendation loss and the weighted interaction confidence loss to obtain the current recommendation loss corresponding to the target sample pair, and accumulate the current recommendation losses to obtain the target recommendation loss corresponding to the current training set. Specifically, it can be shown in formula (6). Please refer to the above description for details, which will not be repeated here.
[0169] 208. The server updates the advertisement recommendation model according to the target recommendation loss.
[0170] For example, the server can use a gradient descent algorithm to update the network parameters of the advertising recommendation model based on the target recommendation loss to obtain an updated advertising recommendation model, and use the updated advertising recommendation model as the advertising recommendation model, and return to execute the step of obtaining the current training set of the advertising recommendation model until the advertising recommendation model converges to obtain the updated target recommendation model. Alternatively, other network parameter updating algorithms can be used to update the network parameters of the advertising recommendation model based on the target recommendation loss to obtain an updated advertising recommendation model, and use the updated advertising recommendation model as the advertising recommendation model, and return to execute the step of obtaining the current training set of the advertising recommendation model until the advertising recommendation model converges to obtain the updated target recommendation model.
[0171] 209. The server uses the updated advertisement recommendation model to recommend the advertisement to the target users in the user set.
[0172] For example, the server can directly obtain at least one content uploaded by the terminal or client to obtain an advertisement, or can filter out at least one content in the network or content library to obtain an advertisement, or can obtain at least one published content on the content interaction platform to obtain an advertisement, or, when the number of advertisements is large or the memory is large, it can also receive a content recommendation request that carries the storage address of the advertisement, and obtain the advertisement based on the storage address, and so on.
[0173] The server uses the updated advertisement recommendation model to predict the current click rate of the advertisement for each user in the user set. At least one user whose current click rate exceeds a preset general threshold is selected from the user set to obtain the target user. Alternatively, the users in the user set may be classified based on the current click rate, and at least one target user corresponding to a preset sorting interval (e.g., Top N, where N is an integer greater than 0) is selected from the user set based on the classification result to obtain the target user, and so on.
[0174] The server may directly push the advertisement to the terminal corresponding to the target user, or may display the advertisement on the terminal corresponding to the target user, and so on.
[0175] After receiving the advertisement, the target user can click on or convert the advertisement through the terminal or client.
[0176] From the above, it can be seen that after the server of this embodiment obtains the current training set of the advertising recommendation model, the current training set includes the advertising samples of each user in the user set and the click-stay time corresponding to the advertising samples, and uses the advertising recommendation model to predict the click-through rate of the advertising samples for the object to obtain the click-through rate estimate. Then, according to the click-stay time, the advertising samples are classified, and the click-through rate estimates corresponding to the advertising samples of different categories are compared to obtain the interaction confidence loss. Then, based on the interaction confidence loss and the predicted recommendation result, the target recommendation loss of the advertising recommendation model is determined, and then, the advertising recommendation model is adjusted according to the target recommendation loss. The updated advertising recommendation model is used to recommend the content to the target users in the object set. Since the scheme can measure the confidence of the click behavior by the click dwell time, and classify the advertising samples by the click dwell time, and then compare the click estimates corresponding to the advertising samples of different categories, the interactive confidence loss is obtained. Through the interactive confidence loss, the advertising recommendation model can explicitly learn the partial order relationship based on the click dwell time, which can accurately characterize the true intention of the click behavior between the user and the advertisement, improve the accuracy of the advertising recommendation model, and thus improve the recommendation efficiency of the advertising recommendation.
[0177] In order to better implement the above method, an embodiment of the present invention further provides a content recommendation device, which can be integrated into an electronic device, such as a server or a terminal, and the terminal can include a tablet computer, a laptop computer and / or a personal computer.
[0178] For example, Figure 6 As shown, the content recommendation device may include an acquisition unit 301, a prediction unit 302, a comparison unit 303, a determination unit 304 and a recommendation unit 305, as follows:
[0179] (1) Acquisition unit 301;
[0180] The acquisition unit 301 is used to acquire a current training set of a content recommendation model, where the current training set includes an interaction content sample of each object in an object set and an interaction duration corresponding to the interaction content sample.
[0181] For example, the acquisition unit 301 can be specifically used to receive the current training set of the content recommendation model uploaded by the terminal or the client, or it can obtain the interaction record of each object in the object set, and based on the interaction record, screen out the interaction content sample of each object in the network or content database, and count the interaction duration corresponding to the interaction content sample in the interaction record, so as to obtain the current training set of the content recommendation model, or screen out the interaction content corresponding to at least one object in the network or database, obtain the interaction content sample, and determine the interaction duration corresponding to the interaction content sample, so as to obtain the current training set of the content recommendation model, or, when the memory of the current training set is large or the number of samples is large, it can also receive an update request for the content recommendation model, and the update request can include the storage address of the current training set, and based on the storage address, the current training set is obtained, and so on.
[0182] (2) prediction unit 302;
[0183] The prediction unit 302 is used to use the content recommendation model to predict the recommendation probability of the interactive content sample for the object and obtain the predicted recommendation result.
[0184] For example, the prediction unit 302 can be specifically used to use a content recommendation model to extract features from the interactive content sample, obtain content features of multiple dimensions, fuse the content features, obtain recommendation features of the interactive content sample for the corresponding object in the object set, and determine the recommendation probability of the interactive content sample for the object based on the recommendation features, thereby obtaining a predicted recommendation result corresponding to the interactive content sample.
[0185] (3) a comparison unit 303;
[0186] The comparison unit 303 is used to classify the interaction content samples according to the interaction duration, and compare the predicted recommendation results corresponding to the interaction content samples of different categories to obtain the interaction confidence loss.
[0187] For example, the comparison unit 303 can be specifically used to obtain an interaction duration threshold corresponding to the current training set, classify the interaction content samples based on the interaction duration threshold and the interaction duration, construct a recommendation result pair based on the classified first category sample set and second category sample set, calculate the difference between the recommendation results in the recommendation result pair, obtain the recommendation difference, and correct the recommendation difference to obtain the interaction confidence loss.
[0188] (4) determining unit 304;
[0189] The determination unit 304 is used to determine the target recommendation loss of the content recommendation model based on the interaction confidence loss and the predicted recommendation result.
[0190] For example, the determination unit 304 can be specifically used to obtain the annotated recommendation result corresponding to the interactive content sample, compare the annotated recommendation result with the corresponding predicted recommendation result to obtain an initial recommendation loss, and fuse the initial recommendation loss with the interactive confidence loss to obtain a target recommendation loss.
[0191] (5) Recommendation unit 305;
[0192] The recommendation unit 305 is used to update the content recommendation model according to the target recommendation loss, and use the updated content recommendation model to recommend the content to be recommended to the target object in the object set.
[0193] For example, the recommendation unit 305 can be specifically used to update the content recommendation model according to the target recommendation loss, obtain the content to be recommended, and use the updated content recommendation model to predict the current recommendation probability of the content to be recommended for each object in the object set, and based on the current recommendation probability, filter out the target object corresponding to the content to be recommended in the object set, and recommend the content to be recommended to the target object.
[0194] In specific implementation, the above units can be implemented as independent entities, or can be arbitrarily combined to be implemented as the same or several entities. The specific implementation of the above units can refer to the previous method embodiments, which will not be repeated here.
[0195] As can be seen from the above, in this embodiment, after the acquisition unit 301 acquires the current training set of the content recommendation model, the current training set includes the interaction content samples of each object in the object set and the interaction duration corresponding to the interaction content samples. The prediction unit 302 adopts the content recommendation model to predict the recommendation probability of the interaction content samples for the object to obtain the predicted recommendation result. Then, the comparison unit 303 classifies the interaction content samples according to the interaction duration, and compares the predicted recommendation results corresponding to the interaction content samples of different categories to obtain the interaction confidence loss. Then, the determination unit 304 determines the target recommendation loss of the content recommendation model based on the interaction confidence loss and the predicted recommendation result. Then, the recommendation unit 304 determines the target recommendation loss of the content recommendation model based on the interaction confidence loss and the predicted recommendation result. Yuan 305 updates the content recommendation model according to the target recommendation loss, and uses the updated content recommendation model to recommend the content to be recommended to the target object in the object set; since the scheme can measure the confidence of the interaction behavior by the interaction duration, and classify the interaction content samples by the interaction duration, and then compare the predicted recommendation results corresponding to the interaction content samples of different categories, so as to obtain the interaction confidence loss, the interaction confidence loss allows the content recommendation model to explicitly learn the partial order relationship based on the interaction duration, which can accurately characterize the true intention of the interaction behavior between the object and the content, improve the accuracy of the content recommendation model, and thus improve the recommendation efficiency of the content recommendation.
[0196] The embodiment of the present invention further provides an electronic device, such as Figure 7 As shown, it shows a schematic diagram of the structure of an electronic device involved in an embodiment of the present invention, specifically:
[0197] 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 appreciate that Figure 7 The electronic device structure shown in the figure does not constitute a limitation on the electronic device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.
[0198] The processor 401 is the control center of the electronic device, and uses various interfaces and lines to connect various parts of the entire electronic device. By running or executing software programs and / or modules stored in the memory 402, and calling data stored in the memory 402, the processor 401 performs various functions of the electronic device and processes data. Optionally, the processor 401 may include one or more processing cores; preferably, the processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 401.
[0199] The memory 402 can be used to store software programs and modules. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the electronic device, etc. In addition, the memory 402 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. Accordingly, the memory 402 may also include a memory controller to provide the processor 401 with access to the memory 402.
[0200] The electronic device also includes a power supply 403 for supplying power to each component. Preferably, the power supply 403 can be logically connected to the processor 401 through a power management system, so as to manage charging, discharging, power consumption and other functions through the power management system. The power supply 403 can also include one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, power status indicators and other arbitrary components.
[0201] The electronic device may further include an input unit 404, which may be used to receive input digital or character information and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.
[0202] Although not shown, the electronic device may further include a display unit, etc., which will not be described in detail herein. Specifically in this embodiment, the processor 401 in the electronic device will load the executable files corresponding to the processes of one or more application programs into the memory 402 according to the following instructions, and the processor 401 will run the application programs stored in the memory 402, thereby realizing various functions, as follows:
[0203] A current training set of a content recommendation model is obtained, where the current training set includes interaction content samples of each object in an object set and interaction durations corresponding to the interaction content samples. The content recommendation model is used to predict the recommendation probability of the interaction content samples for the objects to obtain predicted recommendation results. The interaction content samples are classified according to the interaction durations, and the predicted recommendation results corresponding to interaction content samples of different categories are compared to obtain interaction confidence loss. A target recommendation loss of the content recommendation model is determined based on the interaction confidence loss and the predicted recommendation results. The content recommendation model is updated according to the target recommendation loss, and the updated content recommendation model is used to recommend the content to be recommended to the target object in the object set.
[0204] For example, the electronic device may obtain a current training set of a content recommendation model, which includes an interactive content sample of each object in an object set and an interactive duration corresponding to the interactive content sample. The content recommendation model is used to extract features of the interactive content sample to obtain content features of multiple dimensions, and the content features are fused to obtain recommendation features of the interactive content sample for the corresponding object in the object set. Based on the recommendation features, the recommendation probability of the interactive content sample for the object is determined, thereby obtaining a predicted recommendation result corresponding to the interactive content sample. An interaction duration threshold corresponding to the current training set is obtained, and the interactive content sample is classified based on the interaction duration threshold and the interaction duration. A recommendation result pair is constructed based on the classified first category sample set and second category sample set, and the difference between the recommended results in the recommendation result pair is calculated to obtain a recommendation difference, and the recommendation difference is corrected to obtain an interaction confidence loss. An annotated recommendation result corresponding to the interactive content sample is obtained, and the annotated recommendation result is compared with the corresponding predicted recommendation result to obtain an initial recommendation loss, and the initial recommendation loss and the interaction confidence loss are fused to obtain a target recommendation loss. The content recommendation model is updated according to the target recommendation loss to obtain the content to be recommended. The updated content recommendation model is used to predict the current recommendation probability of the content to be recommended for each object in the object set. Based on the current recommendation probability, the target object corresponding to the content to be recommended is screened out in the object set, and the content to be recommended is recommended to the target object.
[0205] The specific implementation of the above operations can be found in the previous embodiments and will not be described in detail here.
[0206] From the above, it can be seen that after obtaining the current training set of the content recommendation model, the embodiment of the present application includes the interaction content sample of each object in the object set and the interaction duration corresponding to the interaction content sample, and adopts the content recommendation model to predict the recommendation probability of the interaction content sample for the object to obtain the predicted recommendation result, and then, according to the interaction duration, the interaction content sample is classified, and the predicted recommendation results corresponding to the interaction content samples of different categories are compared to obtain the interaction confidence loss, and then, based on the interaction confidence loss and the predicted recommendation result, the target recommendation loss of the content recommendation model is determined, and then, the content recommendation model is recommended according to the target recommendation loss. The recommendation model is updated, and the updated content recommendation model is used to recommend the content to be recommended to the target object in the object set; since the scheme can measure the confidence of the interaction behavior by the interaction duration, and classify the interaction content samples by the interaction duration, and then compare the predicted recommendation results corresponding to the interaction content samples of different categories, so as to obtain the interaction confidence loss. Through the interaction confidence loss, the content recommendation model can explicitly learn the partial order relationship based on the interaction duration, which can accurately characterize the true intention of the interaction behavior between the object and the content, improve the accuracy of the content recommendation model, and then improve the recommendation efficiency of the content recommendation.
[0207] A person of ordinary skill in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be completed by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.
[0208] To this end, an embodiment of the present invention provides a computer-readable storage medium, in which a plurality of instructions are stored, and the instructions can be loaded by a processor to execute the steps in any content recommendation method provided in the embodiment of the present invention. For example, the instructions can execute the following steps:
[0209] A current training set of a content recommendation model is obtained, where the current training set includes interaction content samples of each object in an object set and interaction durations corresponding to the interaction content samples. The content recommendation model is used to predict the recommendation probability of the interaction content samples for the objects to obtain predicted recommendation results. The interaction content samples are classified according to the interaction durations, and the predicted recommendation results corresponding to interaction content samples of different categories are compared to obtain interaction confidence loss. A target recommendation loss of the content recommendation model is determined based on the interaction confidence loss and the predicted recommendation results. The content recommendation model is updated according to the target recommendation loss, and the updated content recommendation model is used to recommend the content to be recommended to the target object in the object set.
[0210] For example, the electronic device may obtain a current training set of a content recommendation model, which includes an interactive content sample of each object in an object set and an interactive duration corresponding to the interactive content sample. The content recommendation model is used to extract features of the interactive content sample to obtain content features of multiple dimensions, and the content features are fused to obtain recommendation features of the interactive content sample for the corresponding object in the object set. Based on the recommendation features, the recommendation probability of the interactive content sample for the object is determined, thereby obtaining a predicted recommendation result corresponding to the interactive content sample. An interaction duration threshold corresponding to the current training set is obtained, and the interactive content sample is classified based on the interaction duration threshold and the interaction duration. A recommendation result pair is constructed based on the classified first category sample set and second category sample set, and the difference between the recommended results in the recommendation result pair is calculated to obtain a recommendation difference, and the recommendation difference is corrected to obtain an interaction confidence loss. An annotated recommendation result corresponding to the interactive content sample is obtained, and the annotated recommendation result is compared with the corresponding predicted recommendation result to obtain an initial recommendation loss, and the initial recommendation loss and the interaction confidence loss are fused to obtain a target recommendation loss. The content recommendation model is updated according to the target recommendation loss to obtain the content to be recommended. The updated content recommendation model is used to predict the current recommendation probability of the content to be recommended for each object in the object set. Based on the current recommendation probability, the target object corresponding to the content to be recommended is screened out in the object set, and the content to be recommended is recommended to the target object.
[0211] The specific implementation of the above operations can be found in the previous embodiments, which will not be described in detail here.
[0212] The computer-readable storage medium may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.
[0213] Since the instructions stored in the computer-readable storage medium can execute the steps in any content recommendation method provided in the embodiments of the present invention, the beneficial effects that can be achieved by any content recommendation method provided in the embodiments of the present invention can be achieved. Please refer to the previous embodiments for details and will not be repeated here.
[0214] According to one aspect of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, and the computer instructions are 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, so that the computer device executes the methods provided in various optional implementations of the above-mentioned content recommendation or content push aspects.
[0215] The above is a detailed introduction to a content recommendation method, device, electronic device and computer-readable storage medium provided by an embodiment of the present invention. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.
Claims
1. A content recommendation method, It is characterized in that include: Acquire a current training set of the content recommendation model, wherein the current training set includes an interaction content sample of each object in the object set and an interaction duration corresponding to the interaction content sample; Using the content recommendation model, predicting the recommendation probability of the interactive content sample for the object, and obtaining a predicted recommendation result; Classifying the interaction content samples according to the interaction duration, and comparing the predicted recommendation results corresponding to the interaction content samples of different categories to obtain the interaction confidence loss; Determining a target recommendation loss of the content recommendation model based on the interaction confidence loss and the predicted recommendation result; The content recommendation model is updated according to the target recommendation loss, and the updated content recommendation model is used to recommend the content to be recommended to the target object in the object set.
2. The content recommendation method according to claim 1, It is characterized in that The classifying the interaction content sample according to the interaction duration includes: Obtaining an interaction duration threshold corresponding to the current training set; The interaction content sample is classified based on the interaction duration threshold and the interaction duration.
3. The content recommendation method according to claim 2, It is characterized in that The obtaining the interaction duration threshold corresponding to the current training set includes: Obtaining a historical interaction duration of each object interacting with at least one historical interaction content; Determine a time average of the historical interaction durations, and use the time average as an initial interaction duration threshold corresponding to each object; The initial interaction duration threshold is used as the interaction duration threshold corresponding to the current training set.
4. The content recommendation method according to claim 2, It is characterized in that The classifying the interaction content sample based on the interaction duration threshold and the interaction duration includes: comparing the interaction duration to the interaction duration threshold; Taking the interaction content samples whose interaction duration is greater than or equal to the interaction duration threshold as first category samples, to obtain a first category sample set; The interaction content samples whose interaction duration is less than the interaction duration threshold are taken as second category samples to obtain a second category sample set.
5. The content recommendation method according to claim 4, It is characterized in that The comparing the predicted recommendation results corresponding to the interactive content samples of different categories to obtain the interactive confidence loss includes: Based on the first category sample set and the second category sample set, construct a recommendation result pair; Calculating the difference between the recommended results in the recommended result pair to obtain a recommended difference; The recommendation difference is corrected to obtain the interaction confidence loss.
6. The content recommendation method according to claim 5, It is characterized in that The step of constructing a recommendation result pair based on the first category sample set and the second category sample set includes: Selecting one interactive content sample from each of the first category sample set and the second category sample set to obtain a target sample pair; Filtering out the predicted recommendation result corresponding to the interactive content sample in the target sample pair from the predicted recommendation results to obtain the current recommendation result; The recommendation result pair is constructed based on the current recommendation result and the target sample pair.
7. The content recommendation method according to claim 6, It is characterized in that The constructing the recommendation result pair based on the current recommendation result and the target sample pair includes: Obtaining historical recommendation results corresponding to candidate objects, wherein the candidate objects include objects corresponding to the interactive content samples in the target sample pair; When the target sample pair includes interactive content samples of the same object, combining the current recommendation result and the historical recommendation result to obtain the recommendation result pair; When the target sample pair includes interactive content samples of different objects, the current recommendation result is adjusted based on the historical recommendation result to obtain the recommendation result pair.
8. The content recommendation method according to claim 7, It is characterized in that The obtaining of the historical recommendation results corresponding to the candidate object includes: Screening out candidate objects corresponding to the interactive content samples in the target sample pair from the object set; Obtaining at least one historical recommendation probability of the candidate object, the historical recommendation probability comprising predicting a recommendation probability of historical interaction content for the candidate object using the content recommendation model; A probability mean of the historical recommendation probabilities is determined, and the probability mean is used as the historical recommendation result corresponding to the candidate object.
9. The content recommendation method according to claim 7, It is characterized in that The adjusting the current recommendation result based on the historical recommendation result to obtain the recommendation result pair includes: Calculating the difference between the historical recommendation result and the current recommendation result to obtain an adjusted current recommendation result; The adjusted current recommendation results are combined to obtain the recommendation result pair.
10. The content recommendation method according to claim 6, It is characterized in that The constructing the recommendation result pair based on the current recommendation result and the target sample pair includes: When the target sample pair includes interactive content samples of the same object, the current recommendation results are combined to obtain the recommendation result pair.
11. The content recommendation method according to claim 5, It is characterized in that The calculating the difference between the recommendation results in the recommendation result pair to obtain the recommendation difference includes: sorting the recommendation results in the recommendation result pairs based on the interaction duration and the interaction duration threshold; According to the sorting result, the difference between the recommendation results in the recommendation result pair is calculated to obtain the recommendation difference.
12. The content recommendation method according to claim 5, It is characterized in that The correcting the recommendation difference to obtain the interaction confidence loss includes: Obtaining a training penalty value corresponding to the current training set; The training penalty value is combined with the recommendation difference to obtain a target recommendation difference; The target recommendation difference is compared with a preset recommendation difference, and based on the comparison result, the interaction confidence loss is determined.
13. The content recommendation method according to any one of claims 1 to 12, It is characterized in that The determining, based on the interaction confidence loss and the predicted recommendation result, a target recommendation loss of the content recommendation model includes: Obtaining a labeling recommendation result corresponding to the interactive content sample; Comparing the annotated recommendation result with the corresponding predicted recommendation result to obtain an initial recommendation loss; The initial recommendation loss is fused with the interaction confidence loss to obtain the target recommendation loss.
14. The content recommendation method according to claim 13, It is characterized in that The adopting the updated content recommendation model to recommend the content to be recommended to the target object in the object set includes: Acquire the content to be recommended, and use the updated content recommendation model to predict the current recommendation probability of the content to be recommended for each object in the object set; Based on the current recommendation probability, a target object corresponding to the content to be recommended is screened out from the object set, and the content to be recommended is recommended to the target object.
15. A content recommendation device, It is characterized in that include: An acquisition unit, configured to acquire a current training set of a content recommendation model, wherein the current training set includes an interaction content sample of each object in the object set and an interaction duration corresponding to the interaction content sample; A prediction unit, configured to use the content recommendation model to predict the recommendation probability of the interactive content sample for the object, and obtain a predicted recommendation result; A comparison unit, configured to classify the interaction content samples according to the interaction duration, and compare the predicted recommendation results corresponding to the interaction content samples of different categories to obtain an interaction confidence loss; A determination unit, configured to determine a target recommendation loss of the content recommendation model based on the interaction confidence loss and the predicted recommendation result; A recommendation unit is used to update the content recommendation model according to the target recommendation loss, and use the updated content recommendation model to recommend the content to be recommended to the target object in the object set.
16. An electronic device, It is characterized in that It comprises a processor and a memory, wherein the memory stores an application program, and the processor is used to run the application program in the memory to execute the steps in the content recommendation method according to any one of claims 1 to 14.
17. A computer program product comprising a computer program / instructions, It is characterized in that When the computer program / instructions are executed by a processor, the steps of the content recommendation method according to any one of claims 1 to 14 are implemented.
18. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps in the content recommendation method according to any one of claims 1 to 14.
Citation Information
Patent Citations
Video recommendation method and device, equipment and storage medium
CN111107435A
Search recommendation model generation method and device based on behavior log and storage medium
CN111738805A
Book sorting pushing model training method, electronic equipment and storage medium
CN112417290A
Training method of game battle recommendation model and related device
CN115487508A
Model training and information recommendation method and device, storage medium and equipment
CN116091163A
