Resource recommendation method and device, equipment and storage medium

Abstract

The invention provides a resource recommendation method and device, equipment and a storage medium, relates to the technical field of artificial intelligence, in particular to the technical field of deep learning, large model and natural language processing, and can be used for application scenes such as generative search, document intelligent editing, intelligent assistants, virtual assistants and intelligent e-commerce. According to the specific implementation scheme, features of a target resource and features of an author of the target resource are input into a first multi-task learning model, and a first prediction result is obtained; inputting the features of the target object into a second multi-task learning model to obtain a second prediction result; wherein the features of the target object comprise interaction features of the target object and the first recommendation stream and interaction features of the target object and an author page of the author; obtaining a third prediction result of the plurality of value factors based on the first prediction result and the second prediction result; and determining whether to add the target resource to a second recommendation stream set for the target object based on a third prediction result of the plurality of value factors.

Description

Technical Field

[0001] It involves the field of artificial intelligence technology, especially deep learning, large models and natural language processing technology. Background Technology

[0002] With resources becoming increasingly readily available, users have numerous channels for accessing them. Among these, immersive information recommendation feeds have become the mainstream method for resource recommendation. How to improve the matching degree between resources and recommended objects in the feed, and thus enhance recommendation effectiveness, is a question that the industry has consistently focused on. Summary of the Invention

[0003] This disclosure provides a resource recommendation method, apparatus, device, and storage medium.

[0004] According to one aspect of this disclosure, a resource recommendation method is provided, comprising: The features of the target resource and the features of the author of the target resource are input into the first multi-task learning model to obtain the first prediction results of multiple value factors; The features of the target object are input into the second multi-task learning model to obtain the second prediction results of multiple value factors; wherein, the features of the target object include the interaction features between the target object and the first recommendation stream and the interaction features between the target object and the author's author page; Based on the first prediction results and the second prediction results of multiple value factors, the third prediction results of multiple value factors are obtained. Based on the third prediction results of multiple value factors, determine whether to add the target resource to the second recommendation stream set for the target object.

[0005] According to another aspect of this disclosure, a resource recommendation apparatus is provided, comprising: The first input module is used to input the features of the target resource and the features of the author of the target resource into the first multi-task learning model to obtain the first prediction results of multiple value factors; The second input module is used to input the features of the target object into the second multi-task learning model to obtain the second prediction results of multiple value factors; wherein, the features of the target object include the interaction features between the target object and the first recommendation stream and the interaction features between the target object and the author's author page; The factor determination module is used to obtain a third prediction result of multiple value factors based on the first prediction result and the second prediction result of multiple value factors. The recommendation determination module is used to determine whether to add the target resource to the second recommendation stream set for the target object based on the third prediction result of multiple value factors.

[0006] According to another aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and The memory is communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the methods described in the present disclosure.

[0007] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform any of the methods according to embodiments of this disclosure.

[0008] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements any of the methods according to embodiments of this disclosure.

[0009] The embodiments disclosed herein can improve the matching degree between the second recommendation stream and the target object, thereby improving the recommendation effect and recommendation efficiency.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a schematic flowchart of a resource recommendation method provided in an embodiment of this disclosure; Figure 2 This is a schematic diagram of the processing flow of the first multi-task learning model in one embodiment of this disclosure; Figure 3 This is a schematic diagram illustrating an application example of the resource recommendation method according to an embodiment of this disclosure; Figure 4 This is a schematic block diagram of a resource recommendation device provided in an embodiment of this disclosure; Figure 5 This is a schematic block diagram of a resource recommendation device provided in another embodiment of this disclosure; Figure 6 This is a block diagram of an electronic device used to implement the resource recommendation method of the embodiments of this disclosure. Detailed Implementation

[0012] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] Figure 1 A schematic diagram of a resource recommendation method provided in an embodiment of this disclosure is shown. This method can be applied to a resource recommendation device, which can be deployed in an electronic device. The electronic device may be a single-machine or multi-machine terminal, server, or other processing device. The terminal may be a mobile device, a personal digital assistant (PDA), a handheld device, a computing device, an in-vehicle device, a wearable device, or other user equipment (UE). In some possible implementations, the method can also be implemented by a processor calling computer-readable instructions stored in memory. Figure 1 As shown, the method may include the following steps S110 to S140.

[0014] Step S110: Input the features of the target resource and the features of the author of the target resource into the first multi-task learning model to obtain the first prediction results of multiple value factors; In this embodiment of the disclosure, the target resource is a resource to be recommended to a target object, such as a video or an article. The target object is the object of the recommendation, such as a user or virtual object that needs resource recommendations. The characteristics of the target resource may include physical characteristics, such as the resource type, duration, and tags, and may also include statistical characteristics, such as the resource's play count and interaction data.

[0015] In this embodiment of the disclosure, the author of the target resource can be the entity or account that publishes the target resource. The author's characteristics can include basic information features, such as the author's field of expertise and audience size. The author's characteristics can also include content value features, such as the completion rate of the author's historical works and the revisit rate of their works in the past 30 days.

[0016] Optionally, the features of the target resource and the features of the author of the target resource can be represented by vectors. The vectors corresponding to the features of the target resource and the vectors corresponding to the features of the author of the target resource are concatenated and then input into the first multi-task learning model.

[0017] Optionally, multiple value factors can be used as labels to train a preset model to obtain a first multi-task learning model. This allows the first multi-task learning model to learn the correlation between different value factors and the features of the target resource and the features of the author. Thus, the first multi-task learning model can output the first prediction result of each of the multiple value factors.

[0018] Optionally, the first multi-task learning model can adopt a multi-gate mixture-of-experts (MMOE) model.

[0019] For example, multiple value factors may include a recommendation stream duration factor, a recommendation stream step size factor, an author page duration factor, and an author page step size factor. The recommendation stream duration factor represents the expected duration the target object will spend browsing the feed stream; the recommendation stream step size factor represents the expected number of resources the target object will browse in the feed stream; the author page duration factor represents the expected duration the target object will spend browsing the author page; and the author page step size factor represents the expected number of resources the target object will browse on the author page. Accordingly, information such as the recommendation stream duration, recommendation stream step size, author page duration, and author page step size related to the target object in the actual recommendation process can be used as labels to train a pre-defined model, resulting in a first multi-task learning model. This first multi-task learning model can then output first prediction results for the recommendation stream duration factor, recommendation stream step size factor, author page duration factor, and author page step size factor, respectively.

[0020] Step S120: Input the features of the target object into the second multi-task learning model to obtain the second prediction results of multiple value factors; wherein, the features of the target object include the interaction features between the target object and the first recommendation stream and the interaction features between the target object and the author's author page; In this embodiment, the characteristics of the target object include vertical interaction data and horizontal consumption data. Vertical interaction data refers to the interaction characteristics between the target object and the first recommendation stream, which may include the duration and number of resources browsed by the target object in the first recommendation stream and / or interaction data generated based on the first recommendation stream (likes, favorites, etc.), thereby reflecting the target object's short-term and immediate preferences for resources in the first recommendation stream. Horizontal consumption data refers to the interaction characteristics between the target object and the author's author page, which may include the target object's access data on the author page statistically analyzed by time (e.g., dwell time, historical work views, etc.), and deep association data between the target object and the author (e.g., following / unfollowing, revisit frequency, etc.), thereby reflecting the user's long-term identification and stickiness with the author. Optionally, the interaction characteristics between the target object and the author's author page may be the interaction characteristics generated when the target object swipes into the author page through the first recommendation stream, i.e., characteristics under the deep interaction scenario based on the first recommendation stream.

[0021] Optionally, the characteristics of the target object may also include basic profile characteristics of the target object, such as the target object's age, gender, region, account level, etc., and the characteristics of the target object may also include contextual characteristics of the target object, such as the target object's device, network environment, etc.

[0022] In this embodiment of the disclosure, multiple value factors are used as labels to train a preset model to obtain a second multi-task learning model. This allows the second multi-task learning model to learn the correlation between different value factors and the features of the target object. As a result, the second multi-task learning model can output the second prediction result for each of the multiple value factors.

[0023] Optionally, the second multi-task learning model can be an MMOE.

[0024] For example, multiple value factors may include a recommendation stream duration factor, a recommendation stream step size factor, an author page duration factor, and an author page step size factor. The recommendation stream duration factor represents the expected duration the target object will spend browsing the feed stream; the recommendation stream step size factor represents the expected number of resources the target object will browse in the feed stream; the author page duration factor represents the expected duration the target object will spend browsing the author page; and the author page step size factor represents the expected number of resources the target object will browse on the author page. Accordingly, information such as the recommendation stream duration, recommendation stream step size, author page duration, and author page step size related to the target object in the actual recommendation process can be used as labels to train a pre-defined model, resulting in a second multi-task learning model. This second multi-task learning model can then output second prediction results for the recommendation stream duration factor, recommendation stream step size factor, author page duration factor, and author page step size factor, respectively.

[0025] Step S130: Based on the first prediction results of multiple value factors and the second prediction results of multiple value factors, obtain the third prediction results of multiple value factors; In this embodiment of the disclosure, the first prediction results of multiple value factors and the second prediction results of multiple value factors can correspond one-to-one, that is, the first prediction result and the second prediction result of each value factor can be obtained, and the third prediction result of each value factor can be obtained based on the first prediction result and the second prediction result of each value factor.

[0026] Optionally, a third prediction result can be obtained by calculating the first and second prediction results. For example, a weighted average method can be used, where different weights are assigned to the first and second prediction results, and then a weighted average is performed to obtain the third prediction result. The weights can be adjusted according to actual business needs and data characteristics.

[0027] As an application example, based on the first prediction results of the recommendation stream duration factor, recommendation stream step length factor, author page duration factor, and author page step length factor, and the second prediction results of the recommendation stream duration factor, recommendation stream step length factor, author page duration factor, and author page step length factor, a third prediction result can be obtained for the recommendation stream duration factor, recommendation stream step length factor, author page duration factor, and author page step length factor.

[0028] Step S140: Based on the third prediction results of multiple value factors, determine whether to add the target resource to the second recommendation stream set for the target object.

[0029] In this embodiment of the disclosure, the second recommendation stream is a personalized resource recommendation stream specifically designed for the target object. Target resources whose third prediction results for multiple value factors meet preset conditions can be added to the second recommendation stream set for the target object, so that the target object can browse target resources in the second recommendation stream.

[0030] Optionally, a threshold for the third prediction result of each value factor can be preset. When the third prediction results of multiple value factors of the target resource meet the corresponding threshold conditions, the target resource is added to the second recommendation stream set for the target object.

[0031] Optionally, the weight of each value factor can be preset, and based on the third prediction result of each value factor and the weight of each value factor, it can be determined whether to add the target resource to the second recommendation stream set for the target object.

[0032] In practical application scenarios, before steps S110 to S140, a basic score can be assigned to the target object's preference for each target resource by the recommendation model. Then, steps S110 to S130 are executed to calibrate the basic score based on the third prediction results of multiple value factors of the target resource. This strengthens or weakens the target object's preference for each target resource, and more accurately recommends resources that the user likes.

[0033] Based on the resource recommendation method described above, when determining whether to add a target resource to the second recommendation stream set for the target object, the characteristics of the target object used include the interaction characteristics between the target object and the first recommendation stream, as well as the interaction characteristics between the target object and the author's page. This considers both the target object's short-term and immediate preferences, as well as its long-term identification and stickiness with the author, avoiding interest drift and recommendation homogenization caused by excessive focus on satisfying short-term interests. This improves the matching degree between the second recommendation stream and the target object, thereby enhancing the recommendation effect. Furthermore, by using an independent first multi-task learning model to process the characteristics of the target resource and the author, and a second multi-task learning model to process the characteristics of the target object, it is possible to pre-determine a static first prediction result using the first multi-task learning model. When the creation of the second recommendation stream is actually triggered, this first prediction result is read and fused with the second prediction result output by the second multi-task learning model. This improves computational efficiency, optimizes computational resources, and further enhances recommendation efficiency.

[0034] The resource recommendation method described above can be applied to video recommendation systems, where the target resource to be recommended is a video. The current mainstream video recommendation system's technical implementation path can be summarized as "single-dimensional behavior modeling + short-term interest matching": First, the core features are user long-term and short-term historical behavior sequences, including video viewing records in the feed stream, instantaneous interactive behaviors (likes, favorites, comments, and shares), basic profile features, and video physical features, but author interaction data is not considered. Second, the modeling logic uses collaborative filtering (CF) and deep learning recommendation models (such as DeepFM and DIN algorithms) to calculate recommendation priorities based on the user-video association. Some systems consider time-series models (such as Long Short-Term Memory (LSTM) networks) to capture sequence information, but user-author interaction modeling is not considered. Finally, the recommendation output uses click-through rate and interaction rate as the core optimization objectives, pushing videos with high conversion potential to users, without considering the impact of the recommendation results on user long-term interest satisfaction and author long-term value. The aforementioned video recommendation system has the following problems: 1. Lack of feature dimensions: It does not consider the user's horizontal consumption data, that is, the interaction data between the user and the author when swiping to the author's page, which makes it impossible for the recommendation system to accurately capture the user's long-term preferences and emotional identification with a specific author, making it difficult to achieve long-term matching of "user needs - author value"; 2. Short-sighted modeling goals: It does not consider the user's long-term value conversion to the author, which easily leads to the recommendation results being biased towards short-term interest satisfaction and problems such as interest drift.

[0035] Based on the resource recommendation method described above, the features of the target video and its author are input into a first multi-task learning model to obtain first prediction results for multiple value factors. The features of the target object are input into a second multi-task learning model to obtain second prediction results for multiple value factors. These features include the interaction features between the target object and the first recommendation stream, as well as the interaction features between the target object and the author's page. Based on the first and second prediction results for multiple value factors, a third prediction result for multiple value factors is obtained. Based on the third prediction result, it is determined whether to add the target video to the second recommendation stream set for the target object. Both multi-task learning models are flexible and scalable. The structure or parameters of the first and second multi-task learning models can be flexibly adjusted according to different needs and data characteristics, which is beneficial for optimizing the use of computing resources according to actual needs and achieving better model processing results. Furthermore, the video recommendation process considers the target object's preference for the author, improving the matching degree between the target object and the target video and optimizing the recommendation effect. For authors, this allows for better exploration of the value of their works and increases their exposure.

[0036] In some embodiments, the resource recommendation method may further include: Collect the target object's first consumption information for the author's resources in the third recommendation stream within a preset time window, as well as the target object's second consumption information on the author's page; Based on the first and second consumption information, labels for multiple value factors are obtained; The first preset model is trained based on labels of multiple value factors to obtain the first multi-task learning model. The second pre-defined model is trained based on labels of multiple value factors to obtain the second multi-task learning model.

[0037] In this embodiment of the disclosure, the third recommendation stream is the recommendation stream that has already been shown to the target object, i.e., the historical recommendation stream. The target object's first consumption information for the author's resources in the third recommendation stream may include the target object's viewing time, number of views, and interaction behavior (such as likes, comments, etc.) of the resources produced by the author in that recommendation stream.

[0038] In this embodiment of the disclosure, when viewing resources in the third recommended stream, the target audience can open the author's page through a swiping operation (e.g., side swipe / left / right swipe). For example, in the video list, after clicking on any video, the target audience can access the author's page by swiping left. The second consumption information of the target audience on the author's page may include the duration of the target audience's stay on the author's page, the number of works viewed, whether they follow the author, the frequency of repeat visits, etc.

[0039] Optionally, the preset time window can be adjusted according to actual needs, such as the past day, two days, a week, or a month.

[0040] In this embodiment of the disclosure, each of the multiple value factors corresponds to a label, so that the label can indicate the first preset model and the second preset model, and output the first prediction result and the second prediction result of the multiple value factors.

[0041] In practical applications, multiple value factors can include recommendation stream duration factor, recommendation stream step size factor, author page duration factor, and author page step size factor. The label for the recommendation stream duration factor can be the viewing time of the target object for author resources in the third recommendation stream within a preset time window. The label for the recommendation stream step size factor can be the number of author resources viewed by the target object in the third recommendation stream within a preset time window. The label for the author page duration factor can be the dwell time of the target object on the author page within a preset time window. The label for the author page step size factor can be the number of works viewed by the target object on the author page within a preset time window. Training the first and second preset models based on the above labels enables the first and second multi-task learning models to obtain the ability to output the first and second prediction results of multiple value factors.

[0042] According to the technical solution of this disclosure, based on the first consumption information of the target object in the third recommendation stream for the author's resources within a preset time window, and the second consumption information of the target object on the author's page, labels for multiple value factors are obtained. That is, the first consumption information and the second consumption information are used as the prediction targets of the first preset model and the second preset model, and the first preset model and the second preset model are trained to obtain the first multi-task learning model and the second multi-task learning model. This can improve the prediction accuracy of the first multi-task learning model and the second multi-task learning model, thereby improving the accuracy of resource recommendation.

[0043] In some embodiments, determining whether to add the target resource to a second recommendation stream set for the target object based on a third prediction result of multiple value factors includes: A pre-trained recommendation model is used to determine the primary relevance information between the target resource and the target object; Based on the third prediction results of multiple value factors and the first correlation information, the second correlation information is obtained; Based on the second relevance information, determine whether to add the target resource to the second recommendation stream set for the target object.

[0044] For example, when using a pre-trained recommendation model to determine the first relevance information, the recommendation model can combine the target object's historical behavior data, the semantic features of the resource, and other features for calculation. For instance, based on the target object's historical behavior such as clicks, favorites, and viewing time on resources, a profile of the resources preferred by the target object is determined. Based on the profile data of the resources preferred by the target object and the similarity between the target resource and the target object, the first relevance information between the target resource and the target object is derived. The first relevance information is used to characterize the probability or score of recommending the target resource to the target object.

[0045] In this embodiment of the disclosure, the second correlation information may be the fusion of the third prediction results of multiple value factors and the first correlation information.

[0046] Optionally, the weights of the third prediction result and the first relevance information for each of the multiple value factors can be set separately. Based on the weights of the third prediction result and the first relevance information for each value factor, the first relevance information is fused to obtain the second relevance information. The second relevance information is also used to characterize the probability or score of recommending the target resource to the target object.

[0047] Optionally, a weight can be set for each of the multiple value factors. The third prediction results of each of the multiple value factors can be fused according to the weight of each value factor to obtain the third correlation information corresponding to the multiple value factors. Then, the weights of the third correlation information and the first correlation information can be fused to obtain the second correlation information. The weights can be adjusted according to the actual business scenario and data characteristics.

[0048] Optionally, the first correlation information can be calibrated using the third prediction results based on multiple value factors to obtain the second correlation information. This model can be a learnable model or a rule-based computational model.

[0049] According to the above embodiments, the first relevance information is adjusted based on the third prediction results of multiple value factors. That is, in actual use, multiple value factors are used as weighting factors, rather than being embedded in the design of the recommendation model. This allows the recommendation model and multiple value factors to be flexibly adjusted separately, while optimizing the application of value factors. For example, it can be flexibly determined whether to use value factors and how to use them based on online strategies, thereby improving the flexibility of using horizontal and vertical consumption information of target objects and thus better solving the interest drift problem.

[0050] In some embodiments, second correlation information is obtained based on the third prediction results of multiple value factors and the first correlation information, including: Multiple first combination forms of value factors are determined based on multiple value factors; Based on the recommendation results of historical recommendation tasks, determine the first combination form applicable to the second recommendation flow from multiple first combination forms; Based on the third prediction result of the value factor corresponding to the first combination form applicable to the second recommendation flow, and the first correlation information, the second correlation information is obtained.

[0051] For example, the multiple first combination forms can include combinations of any number of value factors. For instance, the multiple value factors include value factor 1, value factor 2, and value factor 3. The first combination form can be a combination of two value factors or a combination of three value factors. For example, the multiple first combination forms include combinations of value factors 1 and 2, value factors 1 and 3, value factors 2 and 3, and value factors 1, 2, and 3. Further, the first combination form applicable to the second recommendation stream can be any one of the multiple first combination forms. According to the above embodiments, after determining the first combination form, the third prediction results of the value factors corresponding to the first combination form can be used to adjust the first relevant information to obtain the second relevant information. For example, if the first combination form used by the second recommendation stream is a combination of value factors 1 and 2, then the third prediction results of value factor 1 and value factor 2 are used to adjust the first relevant information to obtain the second relevant information.

[0052] According to the above embodiments, the first combination form used in the second recommendation stream is determined based on the recommendation performance of historical recommendation tasks. In practical applications, at predetermined time points, the recommendation performance of historical recommendation tasks can be reassessed, thereby determining the relevance information adjustment effect corresponding to different first combination forms based on the recommendation performance, and then reselecting the first combination form. This facilitates the adjustment of the application method of value factors as business characteristics change during business development, ensuring recommendation performance.

[0053] In some embodiments, based on the first prediction results of multiple value factors and the second prediction results of multiple value factors, a third prediction result of multiple value factors is obtained, including: For the first value factor among multiple value factors, a third prediction result for the time-based first value factor is obtained based on the similarity between the first prediction result and the second prediction result of the first value factor.

[0054] In some embodiments, the resource recommendation method may further include: Store the first prediction results for multiple value factors; In response to the creation instruction of the second recommendation stream, the first prediction results of multiple value factors are read.

[0055] According to the above embodiments, a static first prediction result is determined in advance using a first multi-task learning model. When the creation of the second recommendation stream is actually triggered, the first prediction result is read and fused with the second prediction result output by the second multi-task learning model, thereby improving computational efficiency, optimizing computational resources, and thus improving recommendation efficiency.

[0056] In some embodiments, the features of the target resource and the features of the author of the target resource are input into a first multi-task learning model to obtain a first prediction result of multiple value factors, including: By utilizing multiple first expert networks in the first multi-task learning model that correspond one-to-one with multiple value factors, the features of the target resource and the features of the author of the target resource are processed to obtain the first feature vector of each of the multiple value factors. The second expert network, corresponding to the second combination form of the value factor in the first multi-task learning model, is used to process the features of the target resource and the features of the author of the target resource to obtain the second feature vector of the second combination form. Based on the first feature vector of each value factor and the second feature vector of the second combination form, the first prediction result of multiple value factors is obtained.

[0057] Figure 2 A schematic diagram illustrating the processing flow of the first multi-task learning model in the above embodiments is shown. Figure 2 As shown, the first multi-task learning model 210 includes multiple first expert networks 211 corresponding one-to-one with multiple value factors 230, and a second expert network 212 corresponding to the second combination form of the value factors.

[0058] Optionally, there can be multiple second combination forms and multiple second expert networks. Figure 2 The example illustrates two second-order combination forms and two second-order expert networks. These two second-order combination forms include combinations of value factors 1 and 2, and combinations of value factors 2 and N. Figure 2 As shown, each first expert network 211 outputs a first feature vector of its corresponding value factor 230, which is applied to the calculation of the first prediction result of the value factor; in addition, each second expert network 212 outputs a second feature vector of its corresponding second combination form, which is applied to the calculation of the first prediction result of the value factor 230 corresponding to the second combination form, for example... Figure 2The first second expert network 212 is used to calculate the second feature vector of the combination of value factors 1 and 2. This second feature vector is applied to the calculation of the first prediction result of value factor 1 and the calculation of the first prediction result of value factor 2. That is, the first prediction result of the i-th value factor (i is an integer greater than or equal to 1 and less than or equal to the number of value factors) among multiple value factors is calculated based on the following two parts of information: the first feature vector output by the i-th first expert network, and the second feature vector output by the second expert network corresponding to the value factor combination form to which the i-th value factor belongs.

[0059] According to the above embodiments, the first multi-task learning model is equipped with multiple first expert networks that correspond one-to-one with multiple value factors. These networks can focus on learning the recommendation value of each value factor. At the same time, a second expert network is set up to learn the recommendation value corresponding to different combinations of value factors. This ensures that the final value factors not only have high accuracy but also incorporate value enhancement information after being combined with other value factors, thereby improving the accuracy of the value factors and thus improving the recommendation effect.

[0060] In some embodiments, the features of the target object are input into a second multi-task learning model to obtain a second prediction result of multiple value factors, including: By utilizing multiple third expert networks in the second multi-task learning model that correspond one-to-one with multiple value factors, the features of the target object are processed to obtain the third feature vector of each of the multiple value factors. The fourth expert network, corresponding to the third combination form of the value factor in the second multi-task learning model, is used to process the features of the target object and obtain the fourth feature vector of the third combination form. Based on the third eigenvector of each value factor and the fourth eigenvector of the third combination form, the second prediction results of multiple value factors are obtained.

[0061] For example, the second multi-task learning model includes multiple third expert networks corresponding one-to-one with multiple value factors, and a fourth expert network corresponding to a third combination of value factors. Optionally, the number of third combination forms and the number of fourth expert networks can also be multiple. Each third expert network outputs a third feature vector of its corresponding value factor, which is applied to the calculation of the second prediction result of that value factor; in addition, each fourth expert network outputs a fourth feature vector of its corresponding third combination form (e.g., a combination of value factors 1 and 2, or a combination of value factors 2 and 3, etc.), which is applied to the calculation of the second prediction result of the value factor corresponding to that third combination form. That is, the second prediction result of the j-th value factor (j is an integer greater than or equal to 1 and less than or equal to the number of value factors) among multiple value factors is calculated based on the following two parts of information: the third feature vector output by the j-th third expert network, and the fourth feature vector output by the fourth expert network corresponding to the value factor combination form to which the j-th value factor belongs.

[0062] According to the above embodiments, the second multi-task learning model is equipped with multiple third expert networks that correspond one-to-one with multiple value factors. These networks can focus on learning the recommendation value of each value factor. At the same time, a fourth expert network is set up to learn the recommendation value corresponding to different combinations of value factors. This ensures that the final value factors not only have high accuracy but also incorporate value enhancement information after being combined with other value factors, thereby improving the accuracy of the value factors and thus improving the recommendation effect.

[0063] Figure 3 This is a schematic diagram illustrating an application example of the resource recommendation method according to an embodiment of this disclosure. In this application example, the resource recommendation method is applied to the field of video recommendation. Based on the user's horizontal and vertical consumption data, a long-term value factor for the user's perception of the current video resource author is constructed. Specifically, the process includes the following steps: 1. Data collection: Collect the following data: 2. Feature processing: Based on the collected data, the following features are determined: (1) User characteristics, including: Basic profile features, such as age, gender, region, account level, etc.; Long-term and short-term consumption characteristics (sequence and statistical characteristics, etc.); Author interaction characteristics, such as duration of stay on author pages and frequency of repeat visits by the same author; Contextual characteristics, such as device and network environment.

[0064] (2) Author characteristics, including: Basic information features, such as author's field of expertise and fan base size; Content value characteristics, such as the completion rate of historical works and the revisit rate of works in the past 30 days.

[0065] (3) Video features, including: Physical characteristics of resources, such as resource duration and resource type; Resource statistical characteristics, such as posterior features of resources over a period of time.

[0066] 3. Tag determination: Based on the author consumed by the user at the current moment, the duration and step size of the user's continuous consumption of the same author in the feed and author page within the following 24-hour window are aggregated into the author's long-term value factor tag. The long-term value factor includes the recommendation feed duration factor, the recommendation feed step size factor, the author page duration factor, and the author page step size factor.

[0067] 4. Based on the above features and labels, train the first multi-task learning model 330 and the second multi-task learning model 340. Specifically, the video features and the video author features are vectorized (310) and then input into the first multi-task learning model 330; the user features are vectorized (320) and then input into the second multi-task model. Based on the first prediction result of the first multi-task learning model 330 and the second prediction result of the second multi-task learning model 340, determine the third prediction result for each value factor. 5. When using online strategies, the above value factors should be used as weighting factors.

[0068] When this technical solution is applied to feed stream recommendation, it considers the user's horizontal and vertical consumption model to model the user's long-term value factor to the author, avoiding the problems of interest drift and recommendation homogenization caused by excessive focus on satisfying short-term interests, and can improve the recommendation effect.

[0069] According to embodiments of this disclosure, this disclosure also provides a resource recommendation device. Figure 4 A schematic block diagram of a resource recommendation apparatus provided in an embodiment of this disclosure is shown, such as... Figure 4 As shown, the resource recommendation device includes: The first input module 410 is used to input the features of the target resource and the features of the author of the target resource into the first multi-task learning model to obtain the first prediction results of multiple value factors; The second input module 420 is used to input the features of the target object into the second multi-task learning model to obtain the second prediction results of multiple value factors; wherein, the features of the target object include the interaction features between the target object and the first recommendation stream and the interaction features between the target object and the author's author page; The factor determination module 430 is used to obtain a third prediction result of multiple value factors of the target resource based on the first prediction result of multiple value factors and the second prediction result of multiple value factors. The recommendation determination module 440 is used to determine whether to add the target resource to the second recommendation stream set for the target object based on the third prediction results of multiple value factors.

[0070] In some embodiments, such as Figure 5 As shown, the resource recommendation device may further include a model training module 510, which is used for: Collect the target object's first consumption information for the author's resources in the third recommendation stream within a preset time window, as well as the target object's second consumption information on the author's page; Based on the first and second consumption information, labels for multiple value factors are obtained; The first preset model is trained based on labels of multiple value factors to obtain the first multi-task learning model. The second pre-defined model is trained based on labels of multiple value factors to obtain the second multi-task learning model.

[0071] In some embodiments, the recommendation determination module 440 is further configured to: A pre-trained recommendation model is used to determine the primary relevance information between the target resource and the target object; Based on the third prediction results of multiple value factors and the first correlation information, the second correlation information is obtained; Based on the second relevance information, determine whether to add the target resource to the second recommendation stream set for the target object.

[0072] In some embodiments, the recommendation determination module 440 is further configured to: Multiple first combination forms of value factors are determined based on multiple value factors; Based on the recommendation results of historical recommendation tasks, determine the first combination form applicable to the second recommendation flow from multiple first combination forms; Based on the third prediction result of the value factor corresponding to the first combination form applicable to the second recommendation flow, and the first correlation information, the second correlation information is obtained.

[0073] In some embodiments, such as Figure 5 As shown, the resource recommendation device may further include a reading module 520, which is used for: Store the first prediction results for multiple value factors; In response to the creation instruction of the second recommendation stream, the first prediction results of multiple value factors are read.

[0074] In some embodiments, the first input module 420 is further configured to: By utilizing multiple first expert networks in the first multi-task learning model that correspond one-to-one with multiple value factors, the features of the target resource and the features of the author of the target resource are processed to obtain the first feature vector of each of the multiple value factors. The second expert network, corresponding to the second combination form of the value factor in the first multi-task learning model, is used to process the features of the target resource and the features of the author of the target resource to obtain the second feature vector of the second combination form. Based on the first feature vector of each value factor and the second feature vector of the second combination form, the first prediction result of multiple value factors is obtained.

[0075] In some embodiments, the second input module 410 is further configured to: By utilizing multiple third expert networks in the second multi-task learning model that correspond one-to-one with multiple value factors, the features of the target object are processed to obtain the third feature vector of each of the multiple value factors. The fourth expert network, corresponding to the third combination form of the value factor in the second multi-task learning model, is used to process the features of the target object and obtain the fourth feature vector of the third combination form. Based on the third eigenvector of each value factor and the fourth eigenvector of the third combination form, the second prediction results of multiple value factors are obtained.

[0076] The specific functions and examples of each module and submodule of the apparatus in this disclosure can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0077] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0078] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0079] Figure 6A schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0080] like Figure 6 As shown, device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 602 or loaded from storage unit 607 into random access memory (RAM) 603. RAM 603 may also store various programs and data required for the operation of device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0081] Multiple components in device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of monitors, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0082] The computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as resource recommendation methods. For example, in some embodiments, the resource recommendation method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program may be loaded and / or installed on device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the resource recommendation method described above may be performed. Alternatively, in other embodiments, the computing unit 601 may be configured to perform a resource recommendation method by any other suitable means (e.g., by means of firmware).

[0083] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0084] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0085] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory (EPROM), flash memory, optical fiber, compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0086] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a cathode ray tube (CRT) or liquid crystal display (LCD) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0087] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0088] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0089] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0090] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A resource recommendation method, comprising: The features of the target resource and the features of the author of the target resource are input into the first multi-task learning model to obtain the first prediction results of multiple value factors; The features of the target object are input into the second multi-task learning model to obtain the second prediction results of the multiple value factors; wherein, the features of the target object include the interaction features between the target object and the first recommendation stream and the interaction features between the target object and the author's author page; Based on the first prediction results and the second prediction results of the multiple value factors, a third prediction result of the multiple value factors is obtained; Based on the third prediction results of the multiple value factors, it is determined whether to add the target resource to the second recommendation stream set for the target object.

2. The method according to claim 1, further comprising: Collect the target object's first consumption information for the author's resources in the third recommendation stream within a preset time window, and the target object's second consumption information on the author's page; Based on the first consumption information and the second consumption information, the labels of the multiple value factors are obtained; The first preset model is trained based on the labels of the multiple value factors to obtain the first multi-task learning model. The second preset model is trained based on the labels of the multiple value factors to obtain the second multi-task learning model.

3. The method according to claim 1 or 2, wherein, The process of determining whether to add the target resource to the second recommendation stream set for the target object based on the third prediction result of the multiple value factors includes: A pre-trained recommendation model is used to determine the first relevance information between the target resource and the target object; Based on the third prediction results of the multiple value factors and the first correlation information, the second correlation information is obtained; Based on the second relevance information, it is determined whether to add the target resource to the second recommendation stream set for the target object.

4. The method according to claim 3, wherein, The second correlation information, obtained based on the third prediction result of the multiple value factors and the first correlation information, includes: Based on the aforementioned multiple value factors, multiple first combination forms of the value factors are determined; Based on the recommendation results of historical recommendation tasks, determine the first combination form applicable to the second recommendation flow from among the multiple first combination forms; The second correlation information is obtained based on the third prediction result of the value factor corresponding to the first combination form applicable to the second recommendation flow, and the first correlation information.

5. The method according to any one of claims 1-4, further comprising: Store the first prediction results of the multiple value factors; In response to the creation instruction of the second recommendation stream, the first prediction result of the plurality of value factors is read.

6. The method according to any one of claims 1-5, wherein, The process involves inputting the features of the target resource and the features of its author into a first multi-task learning model to obtain the first prediction results for multiple value factors, including: Using multiple first expert networks in the first multi-task learning model that correspond one-to-one with the multiple value factors, the features of the target resource and the features of the author of the target resource are processed to obtain the first feature vector of each of the multiple value factors. Using the second expert network corresponding to the second combination form of the value factor in the first multi-task learning model, the features of the target resource and the features of the author of the target resource are processed to obtain the second feature vector of the second combination form; Based on the first feature vector of each value factor and the second feature vector of the second combination form, the first prediction result of the plurality of value factors is obtained.

7. The method according to any one of claims 1-6, wherein, The step of inputting the features of the target object into the second multi-task learning model to obtain the second prediction result of the multiple value factors includes: Using multiple third expert networks in the second multi-task learning model that correspond one-to-one with the multiple value factors, the features of the target object are processed to obtain the third feature vector of each of the multiple value factors; The fourth expert network corresponding to the third combination form of the value factor in the second multi-task learning model is used to process the features of the target object and obtain the fourth feature vector of the third combination form. Based on the third feature vector of each value factor and the fourth feature vector of the third combination form, the second prediction result of the multiple value factors is obtained.

8. A resource recommendation device, comprising: The first input module is used to input the features of the target resource and the features of the author of the target resource into the first multi-task learning model to obtain the first prediction results of multiple value factors; The second input module is used to input the features of the target object into the second multi-task learning model to obtain the second prediction results of the multiple value factors; wherein, the features of the target object include the interaction features between the target object and the first recommendation stream and the interaction features between the target object and the author's author page; The factor determination module is used to obtain a third prediction result of the multiple value factors based on the first prediction result and the second prediction result of the multiple value factors. The recommendation determination module is used to determine, based on the third prediction results of the multiple value factors, whether to add the target resource to the second recommendation stream set for the target object.

9. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7.

11. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-7.

Images