Resource recommendation method, recommendation model training method and device

Abstract

The invention provides a resource recommendation method, a recommendation model training method and a recommendation model training device, and relates to the field of data processing, in particular to the field of artificial intelligence. According to the specific implementation scheme, user features corresponding to a target user to be subjected to resource recommendation are obtained; a resource feature corresponding to each candidate resource included in the candidate resource pool is obtained, and the resource features are used for representing the matching degree between the preference information of the target user and the candidate resources; determining index scores corresponding to the candidate resources on the plurality of recommendation indexes; for any candidate resource, inputting the user feature, the resource feature corresponding to the candidate resource and the index score corresponding to the candidate resource into a target recommendation model to obtain a recommendation score corresponding to the candidate resource; and sorting the candidate resources based on the recommendation score to obtain a target recommendation list, and performing resource recommendation to the target user based on the target recommendation list.

Description

Technical Field

[0001] The present disclosure relates to the field of data processing, in particular to the field of artificial intelligence, and specifically to a resource recommendation method, a training method for a recommendation model, and a device thereof. Background Art

[0002] With the rapid development of the short video industry, users' demand for video content is increasing. How to recommend videos to users more flexibly and accurately has become a key issue in improving user experience. Summary of the Invention

[0003] The present disclosure provides a resource recommendation method, a training method for a recommendation model, an apparatus, a device, and a storage medium.

[0004] According to one aspect of the present disclosure, a resource recommendation method is provided, comprising: obtaining user features corresponding to a target user for whom resource recommendation is to be made; obtaining resource features corresponding to each candidate resource included in a candidate resource pool, wherein the resource features are used to represent the degree of match between the target user's preference information and the candidate resource; determining the indicator scores corresponding to the candidate resource on a plurality of recommendation indicators; for any candidate resource, inputting the user features, the resource features corresponding to the candidate resource, and the indicator score corresponding to the candidate resource into a target recommendation model to obtain a recommendation score corresponding to the candidate resource; sorting the candidate resources based on the recommendation scores to obtain a target recommendation list, and recommending resources to the target user based on the target recommendation list.

[0005] In this application, each candidate resource is assigned a resource feature that represents the degree of match between the target user's preference information and the candidate resource, which can more accurately identify the target user's preference for different types of resources. By combining the scoring of candidate resources on multiple recommendation indicators and combining user characteristics, the accuracy of resource recommendations is improved, the user experience is enhanced, and user satisfaction is increased.

[0006] According to another aspect of the present disclosure, a method for training a recommendation model is provided, comprising: collecting log data corresponding to a plurality of user groups, wherein the user groups correspond one-to-one to candidate recommendation models, and the log data records at least the length of time that users included in the user groups stay on each recommended resource corresponding to the users; calculating an evaluation index value corresponding to each candidate recommendation model based on the log data; determining an optimization strategy corresponding to each candidate recommendation model based on the evaluation index value, and optimizing the candidate recommendation model based on the optimization strategy to obtain an optimized candidate recommendation model.

[0007] According to another aspect of the present disclosure, a resource recommendation device is provided, including: a first acquisition module for acquiring user characteristics corresponding to a target user to whom resource recommendation is to be made; a second acquisition module for acquiring resource characteristics corresponding to each candidate resource included in a candidate resource pool, wherein the resource characteristics are used to represent the degree of match between the target user's preference information and the candidate resource; a first determination module for determining the indicator scores corresponding to the candidate resource on multiple recommendation indicators; a second determination module for inputting the user characteristics, the resource characteristics corresponding to the candidate resource, and the indicator score corresponding to the candidate resource into a target recommendation model for any candidate resource to obtain a recommendation score corresponding to the candidate resource; a resource recommendation module for sorting the candidate resources based on the recommendation score to obtain a target recommendation list, and recommending resources to the target user based on the target recommendation list.

[0008] According to another aspect of the present disclosure, a training device for a recommendation model is provided, comprising: a collection module for collecting log data corresponding to a plurality of user groups, wherein a user group corresponds one-to-one to a candidate recommendation model, and the log data records at least the length of time that users included in the user group stay on each recommended resource corresponding to the user; a calculation module for calculating an evaluation index value corresponding to each candidate recommendation model based on the log data; and an optimization module for determining an optimization strategy corresponding to each candidate recommendation model based on the evaluation index value, and optimizing the candidate recommendation model based on the optimization strategy to obtain an optimized candidate recommendation model.

[0009] According to another aspect of the present disclosure, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above-mentioned resource recommendation method or recommendation model training method.

[0010] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the above-mentioned resource recommendation method or recommendation model training method.

[0011] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, which implements the above-mentioned resource recommendation method or recommendation model training method when executed by a processor.

[0012] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The accompanying drawings are used to better understand the present invention and do not constitute a limitation of the present invention.

[0014] Figure 1 It is a schematic diagram of an exemplary implementation of a resource recommendation method according to an exemplary embodiment of the present disclosure.

[0015] Figure 2 It is a schematic diagram of an exemplary implementation of a resource recommendation method according to an exemplary embodiment of the present disclosure.

[0016] Figure 3 It is a schematic diagram of an exemplary implementation of a resource recommendation method according to an exemplary embodiment of the present disclosure.

[0017] Figure 4 This is an overall introduction diagram of a target recommendation model according to an exemplary embodiment of the present disclosure.

[0018] Figure 5 3 is an exemplary schematic diagram of a training method for a recommendation model according to an exemplary embodiment of the present disclosure.

[0019] Figure 6 It is an overall framework diagram of a training method for a recommendation model according to an exemplary embodiment of the present disclosure.

[0020] Figure 7 is an exemplary schematic diagram of a resource recommendation device according to an exemplary embodiment of the present disclosure.

[0021] Figure 8 2 is an exemplary schematic diagram of a training device for a recommendation model according to an exemplary embodiment of the present disclosure.

[0022] Figure 9 is a schematic diagram of an electronic device according to an exemplary embodiment of the present disclosure. DETAILED DESCRIPTION

[0023] The following description of exemplary embodiments of the present disclosure is made in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding. These details should be considered as merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0024] Artificial Intelligence (AI) is the study of how computers can simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include computer vision, speech recognition, natural language processing, as well as deep learning / learning, big data processing, and knowledge graphs.

[0025] In the technical solutions disclosed herein, the acquisition, storage, and application of user personal information involved comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0026] Figure 1 This is a schematic diagram of an exemplary implementation of a resource recommendation method shown in this application. Figure 1 As shown, the resource recommendation method includes the following steps:

[0027] S101: Obtain user characteristics corresponding to the target user for whom resource recommendation is to be performed.

[0028] After receiving the resource recommendation request, the target user to whom the resource recommendation is to be made may be determined based on the user identifier carried in the resource recommendation request.

[0029] Among them, the resources mentioned in this application may include various types of resources such as video resources, audio resources, document resources, etc., that is, it can be understood that the resource recommendation method introduced in this application can be applied to various scenarios related to resource recommendation such as video resource recommendation, audio resource recommendation, document resource recommendation, etc.

[0030] The user features may include user attribute features corresponding to the target user, user history browsing resource sequences, etc. For example, in a video recommendation scenario, the user features may include user attribute features corresponding to the target user, user history browsing video sequences, etc.

[0031] S102: Obtain resource features corresponding to each candidate resource included in the candidate resource pool, wherein the resource features are used to indicate a matching degree between the target user's preference information and the candidate resource.

[0032] Among them, resource characteristics are used to indicate the matching degree between the target user's preference information and the candidate resources. For example, the resource characteristics of a candidate resource can be used to indicate whether the candidate resource meets the target user's preferred resource type, and whether the candidate resource meets the target user's preferred duration and other characteristics.

[0033] S103: Determine the index scores corresponding to the candidate resources on the multiple recommendation indicators.

[0034] Among them, multiple recommendation indicators can be pre-set based on actual conditions. For example, the recommendation indicators may include positive feedback indicators, negative feedback indicators, completion indicators, and other recommendation indicators.

[0035] For each candidate resource, each recommendation indicator corresponds to an indicator score. Assuming that 6 recommendation indicators are set, each candidate resource corresponds to 6 indicator scores.

[0036] S104: For any candidate resource, the user characteristics, the resource characteristics corresponding to the candidate resource, and the index score corresponding to the candidate resource are input into the target recommendation model to obtain the recommendation score corresponding to the candidate resource.

[0037] Assume that the candidate resource pool includes 1000 candidate resources, namely candidate resource 1, candidate resource 2, candidate resource 3, and so on to candidate resource 1000. After determining the user characteristics corresponding to the target user, the resource characteristics corresponding to each of the 1000 candidate resources, and the indicator scores corresponding to each of the 1000 candidate resources, for each candidate resource, the user characteristics corresponding to the target user, the resource characteristics corresponding to the candidate resource, and the indicator score corresponding to the candidate resource are input into the pre-trained target recommendation model to obtain the recommendation score corresponding to the candidate resource.

[0038] Exemplarily, user features corresponding to the target user, resource features corresponding to candidate resource 1, and index scores corresponding to candidate resource 1 are input into a pre-trained target recommendation model to obtain a recommendation score corresponding to candidate resource 1.

[0039] Exemplarily, the user features corresponding to the target user, the resource features corresponding to the candidate resource 2, and the index score corresponding to the candidate resource 2 are input into a pre-trained target recommendation model to obtain the recommendation score corresponding to the candidate resource 2.

[0040] This process is deduced in this way until the recommendation scores corresponding to the 1,000 candidate resources in the candidate resource pool are obtained, wherein the candidate resources and the recommendation scores are in a one-to-one correspondence.

[0041] S105 , sorting candidate resources based on the recommendation scores to obtain a target recommendation list, and recommending resources to the target user based on the target recommendation list.

[0042] After determining the recommendation score corresponding to each candidate resource in the candidate resource pool, the candidate resources are sorted in descending order of the recommendation score to generate a target recommendation list, and resources are recommended to the target user based on the target recommendation list.

[0043] The embodiment of the present application proposes a resource recommendation method, including: obtaining user characteristics corresponding to the target user to whom resource recommendation is to be made; obtaining resource characteristics corresponding to each candidate resource included in the candidate resource pool, wherein the resource characteristics are used to represent the degree of match between the target user's preference information and the candidate resource; determining the indicator scores corresponding to the candidate resource on multiple recommendation indicators; for any candidate resource, inputting the user characteristics, the resource characteristics corresponding to the candidate resource, and the indicator score corresponding to the candidate resource into the target recommendation model to obtain the recommendation score corresponding to the candidate resource; sorting the candidate resources based on the recommendation score to obtain a target recommendation list, and recommending resources to the target user based on the target recommendation list. In the present application, each candidate resource is assigned a resource characteristic for representing the degree of match between the target user's preference information and the candidate resource, which can more accurately identify the target user's preference for different types of resources. By combining the scores of the candidate resources on multiple recommendation indicators and combining user characteristics, the accuracy of resource recommendations is improved, the user experience is enhanced, and user satisfaction is increased.

[0044] Figure 2 This is a schematic diagram of an exemplary implementation of a resource recommendation method shown in this application. Figure 2 As shown, the resource recommendation method includes the following steps:

[0045] S201: Obtain user characteristics corresponding to the target user for whom resource recommendation is to be performed.

[0046] S202: Obtain resource features corresponding to each candidate resource included in the candidate resource pool, wherein the resource features are used to indicate a matching degree between the target user's preference information and the candidate resource.

[0047] Regarding the specific implementation of steps S201 to S202, reference may be made to the detailed introduction of the relevant parts in the above embodiment, which will not be elaborated here.

[0048] S203: Determine the index scores corresponding to the candidate resources on the multiple recommendation indicators.

[0049] In the present application, the recommendation index at least includes a sequence fusion index. For any candidate resource, the method for determining the index score corresponding to the sequence fusion index corresponding to the candidate resource includes: obtaining the total number of candidate resources included in the candidate resource pool; estimating the target user's estimated completion rate for each candidate resource (taking the video recommendation scenario as an example, the estimated completion rate refers to the ratio of the target user's stay time on a candidate resource to the total stay time of the candidate resource), and estimating the target user's estimated stay time for each candidate resource; determining the first reference value and the second reference value corresponding to each candidate resource based on the total number, the estimated completion rate and the estimated stay time; for any candidate resource, performing a weighted summation on the first reference value and the second reference value corresponding to the candidate resource to obtain the sequence fusion score corresponding to the candidate resource in the sequence fusion index.

[0050] Among them, the above-mentioned determination of the first reference value and the second reference value corresponding to each candidate resource based on the total number, the estimated completion rate and the estimated stay time includes: sorting the candidate resources in descending order according to the estimated completion rate to obtain the first sequence identifier corresponding to each candidate resource (the first sequence identifier may be the sequence number of the candidate resource after sorting the candidate resources in descending order according to the estimated completion rate); sorting the candidate resources in descending order according to the estimated stay time to obtain the second sequence identifier corresponding to each candidate resource (the second sequence identifier may be the sequence number of the candidate resource after sorting the candidate resources in descending order according to the estimated stay time); for any candidate resource, determining the first reference value corresponding to the candidate resource based on the total number and the first sequence identifier corresponding to the candidate resource; for any candidate resource, determining the second reference value corresponding to the candidate resource based on the total number and the second sequence identifier corresponding to the candidate resource.

[0051] The above describes the specific ideas for obtaining the index scores corresponding to the order fusion indicators corresponding to the candidate resources. The following describes the specific calculation formula for determining the index scores corresponding to the order fusion indicators corresponding to the candidate resources.

[0052] In some embodiments, for any candidate resource, a calculation formula for determining a first reference value corresponding to the candidate resource based on the total number and the first sequence identifier corresponding to the candidate resource is:

[0053] A score =(log(N)-log(A rank )+0.1) / (log(N)+0.1)

[0054] In the above formula, A score represents the first reference value corresponding to the current candidate resource, N represents the total number of candidate resources included in the candidate resource pool, A rankRepresents the first sequence identifier corresponding to the current candidate resource (for example, if the candidate resources are sorted in descending order of estimated completion rate, and the sequence number of a candidate resource is 100, when calculating the first reference value corresponding to the candidate resource, A rank The value entered is 100).

[0055] Furthermore, for any candidate resource, a calculation formula for determining the second reference value corresponding to the candidate resource based on the total number and the second sequence identifier corresponding to the candidate resource is:

[0056] B score =(log(N)-log(B rank )+0.1) / (log(N)+0.1)

[0057] In the above formula, B score represents the second reference value corresponding to the current candidate resource, N represents the total number of candidate resources included in the candidate resource pool, and B rank Represents the second sequence identifier corresponding to the current candidate resource (for example, if the candidate resources are sorted in descending order of estimated stay duration and the sequence number of a candidate resource is 200, when calculating the second reference value corresponding to the candidate resource, B rank The value entered is 200).

[0058] Furthermore, for any candidate resource, after obtaining the first reference value and the second reference value corresponding to the candidate resource as described above, the first reference value and the second reference value corresponding to the candidate resource are weightedly summed to obtain the order fusion score corresponding to the order fusion index of the candidate resource:

[0059] F=a*A score +b*B score

[0060] In the above formula, F represents the order fusion score of the current candidate resource corresponding to the order fusion index; A score Represents the first reference value corresponding to the current candidate resource; B score Represents the second reference value corresponding to the current candidate resource; a and b are both set hyperparameters.

[0061] The calculation method for the sequence fusion score corresponding to the sequence fusion indicator introduced above enables the subsequent model to introduce the sequence relationship corresponding to the candidate resources when calculating the recommendation score. This is equivalent to smoothing the subsequent model when fusing multiple recommendation indicators, thereby alleviating the differences in single usage value fusion at multiple target scales and reducing the noise impact caused by extreme values. At the same time, it can balance the ranking goals of different targets and improve the diversity of the model.

[0062] The above describes the sequence fusion index in the recommendation index. Furthermore, the recommendation index may also include a completion rate index, a stay index, a completion index, a negative feedback index, and a positive feedback index.

[0063] Among them, the indicator score corresponding to the completion rate indicator can be the target user's estimated completion rate of the candidate resources. Taking the video recommendation scenario as an example, the estimated completion rate refers to the estimated ratio of the target user's stay time on the candidate resources to the total stay time of the candidate resources.

[0064] Among them, the indicator score corresponding to the stay indicator can be the estimated stay time of the target user on the candidate resource. Taking the video recommendation scenario as an example, the estimated stay time refers to the estimated stay time of the target user on the candidate resource.

[0065] Among them, the indicator score corresponding to the completion indicator can be the target user's estimated completion probability of the candidate resource. Taking the video recommendation scenario as an example, the estimated completion probability refers to the probability estimate that the target user's stay time on the candidate resource is greater than or equal to the total length of the candidate resource.

[0066] The score corresponding to the negative feedback indicator can be the estimated probability of a target user providing negative feedback on a candidate resource. For example, in the video recommendation scenario, the estimated negative feedback probability refers to the estimated probability that the target user's stay on a candidate resource is less than a preset stay time. The preset stay time can be set based on actual circumstances, for example, 3 seconds.

[0067] The score corresponding to the positive feedback indicator can be the estimated probability of a target user giving positive feedback on a candidate resource. For example, in the video recommendation scenario, the estimated positive feedback probability refers to the estimated probability of a target user giving positive feedback on a candidate resource, such as liking, following, commenting, sharing, or adding it to a favorite.

[0068] Among them, for the recommendation indicators such as completion rate indicator, stay indicator, completion indicator, negative feedback indicator and positive feedback indicator, the indicator scores of the candidate resources corresponding to multiple recommendation indicators are determined, including: for any recommendation indicator, obtaining the indicator user-related characteristics and indicator resource-related characteristics related to the recommendation indicator, and estimating the indicator score of the candidate resource corresponding to the recommendation indicator based on the indicator user-related characteristics and indicator resource-related characteristics.

[0069] The above details the calculation process of each recommendation indicator and its score. Multi-dimensional indicators can improve the accuracy of model recommendations and increase user satisfaction.

[0070] S204: Determine the target user group to which the target user belongs.

[0071] In some embodiments, multiple user groups may be set, each of which corresponds to a candidate recommendation model. Each candidate recommendation model may be set with different model parameters.

[0072] In some embodiments, after the target user is determined as described above, it is determined whether the target user is a new user. If the target user is a new user, the target user group to which the target user belongs is determined from multiple user groups according to the user quantity balance principle; if the target user is an old user, the user group to which the target user was previously divided can be used as the target user group to which the target user belongs.

[0073] S205 , determining a target recommendation model corresponding to the target user group from multiple candidate recommendation models, where the target recommendation model includes a multi-layer perceptron and an online fusion layer.

[0074] After determining the target user group to which the target user belongs, the candidate recommendation model corresponding to the target user group is used as the target recommendation model. By assigning candidate recommendation models to multiple user groups, feedback data from different user groups can be collected during subsequent training of the candidate recommendation models. This feedback data can be used to evaluate the performance differences of the multiple candidate recommendation models, thereby optimizing the candidate recommendation models.

[0075] S206: For any candidate resource, the user features and the resource features corresponding to the candidate resource are input into a multi-layer perceptron to obtain the impact factor corresponding to each recommendation indicator, wherein the impact factor corresponds to the recommendation indicator one-to-one.

[0076] For example, taking candidate resource 1 as an example, the user features corresponding to the target user, the resource features corresponding to candidate resource 1, and the index score corresponding to candidate resource 1 are input into the multi-layer perceptron included in the target recommendation model, and the multi-layer perceptron outputs the impact factor corresponding to each recommendation index.

[0077] In some implementations, the multilayer perceptron can be designed as two layers.

[0078] S207: Input the impact factors and the index scores corresponding to the candidate resources into the online fusion layer for weighted fusion to obtain the recommendation scores corresponding to the candidate resources.

[0079] For example, continuing to take candidate resource 1 as an example, if candidate resource 1 corresponds to 6 recommendation indicators, it is not difficult to understand that the above multi-layer perceptron outputs a total of 6 impact factors corresponding to the 6 recommendation indicators corresponding to candidate resource 1, and then the total of 6 impact factors corresponding to candidate resource 1 and the indicator scores corresponding to the 6 recommendation indicators are input into the online fusion layer for weighted fusion to obtain the recommendation score corresponding to candidate resource 1.

[0080] S208 , sorting the candidate resources based on the recommendation scores to obtain a target recommendation list, and recommending resources to the target user based on the target recommendation list.

[0081] After determining the recommendation score corresponding to each candidate resource in the candidate resource pool, the candidate resources are sorted in descending order of the recommendation score to generate a target recommendation list, and resources are recommended to the target user based on the target recommendation list.

[0082] The embodiments of the present application describe in detail the internal structure of the target recommendation model and the specific process of the target recommendation model outputting the recommendation scores corresponding to the candidate resources. By combining user characteristics, resource characteristics, and the scores of candidate resources on multiple recommendation indicators, the target recommendation model can more accurately recommend resources to users, thereby improving user experience and increasing user satisfaction.

[0083] Figure 3 This is a schematic diagram of an exemplary implementation of a resource recommendation method shown in this application. Figure 3 As shown, the resource recommendation method includes the following steps:

[0084] S301: Receive a resource recommendation request, and determine a target user to whom resource recommendation is to be made based on the resource recommendation request.

[0085] S302: Determine user request features based on request-related information corresponding to the resource recommendation request.

[0086] Among them, the request-related information may include the network type currently used by the device when initiating the resource recommendation request (such as 5G, 4G or WiFi), the number of refreshes requested, and other information.

[0087] After obtaining the request-related information, feature extraction may be performed on the request-related information to obtain user request features.

[0088] S303: Determine user portrait features based on user-related information corresponding to the target user.

[0089] The user-related information includes user attribute information and user historical interaction information.

[0090] The user attribute information may include the target user's corresponding attribute information such as age, gender, and education level.

[0091] The user history interaction information includes the target user's corresponding historical browsing resources, historical browsing timestamps, historical browsing resource categories and other information.

[0092] Among them, determining the user portrait features based on the user-related information corresponding to the target user includes: determining the user attribute features corresponding to the target user based on the user attribute information; determining the user interaction statistical features and user interaction sequence features corresponding to the target user based on the user historical interaction information; and constructing the user portrait features corresponding to the target user based on the user attribute features, user interaction statistical features and user interaction sequence features.

[0093] The determining of the user attribute features corresponding to the target user based on the user attribute information includes: extracting features from the user attribute information to obtain the user attribute features.

[0094] Determining the user interaction statistical features corresponding to the target user based on historical user interaction information includes: performing statistics on the user's historical interaction information and extracting features to obtain the user interaction statistical features. The user interaction statistical features are used to reflect statistically significant features such as the target user's historical resource browsing volume and the target user's cumulative weekly visits.

[0095] Determining the user interaction sequence features corresponding to the target user based on the user's historical interaction information includes extracting features from the user's historical interaction information to obtain the user interaction sequence features. The user interaction sequence features are used to reflect features with sequence significance, such as the target user's historical resource browsing sequence, historical satisfactory resource browsing sequence, and historical satisfactory resource category sequence.

[0096] S304: Determine user scenario characteristics based on scenario-related information corresponding to the resource recommendation request.

[0097] The scenario-related information may include scenario information such as whether it is a working day or leisure time when the resource recommendation request is initiated.

[0098] After acquiring the scene-related information, feature extraction is performed on the scene-related information to obtain user scene features.

[0099] S305: Construct user features corresponding to the target user based on the user request features, user portrait features, and user scenario features.

[0100] S306: Obtain preference information corresponding to the target user, where the preference information at least includes the user's preferred resource type and the user's preferred resource duration.

[0101] Among them, user preferred resource types refer to the resource types that target users like to browse, such as movies, TV shows, news, etc.

[0102] Among them, the user-preferred resource length refers to the length of resources that the target user likes to browse. For example, the target user prefers short videos of less than 20 seconds or short drama videos of more than 30 minutes.

[0103] S307: Obtain resource information of the candidate resource.

[0104] For example, the resource information may include the resource type, resource duration, resource tag, interaction volume, and other information of the candidate resource.

[0105] S308: Determine resource characteristics corresponding to the candidate resource based on the preference information and the resource information.

[0106] Among them, resource characteristics are used to indicate the matching degree between the target user's preference information and the candidate resources. For example, the resource characteristics of a candidate resource can be used to indicate whether the candidate resource meets the target user's preferred resource type, and whether the candidate resource meets the target user's preferred duration and other characteristics.

[0107] S309: Determine the index scores corresponding to the candidate resources on the multiple recommendation indicators.

[0108] Among them, multiple recommendation indicators can include completion rate indicator, stay indicator, completion indicator, negative feedback indicator, positive feedback indicator and sequence fusion indicator. The specific method of obtaining the indicator scores of these 6 indicators can refer to the introduction of the relevant parts in the above embodiment, and will not be repeated here.

[0109] S310: For any candidate resource, user characteristics, resource characteristics corresponding to the candidate resource, and indicator scores corresponding to the candidate resource are input into a target recommendation model to obtain a recommendation score corresponding to the candidate resource.

[0110] Figure 4 This is an overall introduction diagram of a target recommendation model shown in this application, such as Figure 4 As shown, the target recommendation model includes a multi-layer perceptron and an online fusion layer. For any candidate resource, user features (including user request features, user attribute features, user interaction statistics, user interaction sequence features, and user scenario features) and the resource features corresponding to the candidate resource are input into the multi-layer perceptron to obtain the impact factor corresponding to each recommendation indicator. The impact factor and the indicator score corresponding to the candidate resource are input into the online fusion layer for weighted fusion to obtain the recommendation score corresponding to the candidate resource.

[0111] Combine Figure 4 To understand, the completion rate index, stay index, completion index, negative feedback index, positive feedback index and sequence fusion index are used to calculate the index scores of these 6 recommendation indicators (in Figure 4The above six recommendation indicators are respectively represented as the estimated completion rate, estimated stay time, estimated completion probability, estimated negative feedback probability, and sequence fusion score) and are recorded as A, B, C, D, E, and F, respectively. The impact factors corresponding to these six recommendation indicators are recorded as f1, f2, f3, f4, f5, and f6. Among them, the impact factors correspond to the recommendation indicators one by one. In some embodiments, the corresponding calculation formula for weighted fusion in the online fusion layer is:

[0112] final_score=A f1 *N f2 *C f3 *D f4 *E f5 *F f6

[0113] In the above formula, final_score represents the final recommendation score corresponding to a candidate resource; A, B, C, D, E, and F represent the scores corresponding to the six recommendation indicators: completion rate indicator, retention indicator, broadcast completion indicator, negative feedback indicator, positive feedback indicator, and sequence fusion indicator, respectively; f1, f2, f3, f4, f5, and f6 represent the impact factors corresponding to the six recommendation indicators: completion rate indicator, retention indicator, broadcast completion indicator, negative feedback indicator, positive feedback indicator, and sequence fusion indicator, respectively.

[0114] S311 , sorting candidate resources based on the recommendation scores to obtain a target recommendation list, and recommending resources to the target user based on the target recommendation list.

[0115] The embodiments of the present application provide a detailed introduction to the specific content and acquisition process of user features. For example, user features include user request features, user attribute features, user interaction statistical features, user interaction sequence features, and user scenario features. The embodiments of the present application also provide a detailed introduction to the specific content and acquisition process of resource features. By acquiring multi-dimensional features, a more accurate recommendation score can be output when the subsequent target recommendation model calculates the recommendation score, thereby more accurately recommending resources to users, improving user experience and increasing user satisfaction.

[0116] Figure 5 This is an exemplary diagram of a training method for a recommendation model shown in this application, such as Figure 5 As shown in FIG, the training method of the recommendation model includes the following steps:

[0117] S501 , collecting log data corresponding to a plurality of user groups, wherein the user groups correspond to the candidate recommendation models one-to-one, and the log data records at least the length of time that the users included in the user group stay at each recommended resource corresponding to the user.

[0118] Among them, the candidate recommendation model is used to obtain the recommendation score corresponding to the candidate resource based on the input user characteristics, resource characteristics corresponding to the candidate resource, and the indicator score corresponding to the candidate resource, so that resources can be recommended to users based on the recommendation score.

[0119] The following combination Figure 6 Make an introduction, Figure 6 This is the overall framework diagram of a training method for a recommendation model shown in this application, such as Figure 6 As shown, user groups correspond one-to-one to candidate recommendation models. When making online recommendations, after determining the user group corresponding to the target user to whom resources are to be recommended, the candidate recommendation model corresponding to the target user's user group is used as the target recommendation model for the target user. The target user's user characteristics, resource characteristics of the candidate resources, and the indicator scores of the candidate resources are then input into the target recommendation model to obtain the recommendation score corresponding to the candidate resource. Based on the recommendation score, resources are recommended to the target user. After a period of time, log data corresponding to multiple user groups is collected.

[0120] S502: Calculate the evaluation index value corresponding to each candidate recommendation model based on the log data.

[0121] In some embodiments, taking a video recommendation scenario as an example, the evaluation index value may be the total playback time of resources for all users in the user group corresponding to the candidate recommendation model.

[0122] S503 , determining an optimization strategy corresponding to each candidate recommendation model based on the evaluation index value, and optimizing the candidate recommendation model based on the optimization strategy to obtain an optimized candidate recommendation model.

[0123] Among them, the candidate recommendation model can be optimized by offline optimization.

[0124] The embodiment of the present application introduces a training method for a recommendation model. By calculating the evaluation index value based on the log data of different user groups, the recommendation effects of different candidate recommendation models can be accurately evaluated to optimize the model. Moreover, since the optimization process in this solution is based on log data feedback, the candidate recommendation model is continuously optimized with the user's interaction during the actual application process, the model effect is continuously improved, and a high recommendation quality and user satisfaction are maintained.

[0125] Furthermore, in some embodiments, an optimization strategy corresponding to each candidate recommendation model is determined based on the evaluation index value, and the candidate recommendation model is optimized based on the optimization strategy to obtain an optimized candidate recommendation model, including: determining the optimal user group and the worst user group from multiple user groups based on the evaluation index value; obtaining the difference between the evaluation index value corresponding to the optimal user group and the evaluation index value corresponding to the worst user group; in response to the difference being greater than or equal to a preset difference threshold, determining the optimal recommendation model corresponding to the optimal user group from the candidate recommendation models, and obtaining the optimal model parameters corresponding to the optimal recommendation model; for the remaining candidate recommendation models other than the optimal recommendation model, after randomly perturbing the optimal model parameters, updating the remaining candidate recommendation models based on the model parameters after the random perturbation to obtain an updated candidate recommendation model; wherein the optimal model parameters corresponding to the optimal recommendation model remain unchanged. After obtaining the candidate recommendation models of the current round as described above, a new round of log collection and model optimization process is performed based on the current candidate recommendation model.

[0126] For example, assume that there are 10 user groups in total. Since user groups correspond to candidate recommendation models one-to-one, that is, there are 10 candidate recommendation models in total. For each user group, the total playback time of resources of all users included in the user group within the target time period is obtained as the evaluation index value of the candidate recommendation model corresponding to the user group. Then, the user group corresponding to the largest evaluation index value is taken as the optimal user group, and the user group corresponding to the smallest evaluation index value is taken as the worst user group. Assuming that the optimal user group is user group 1 and the worst user group is user group 10, the difference between the evaluation index value corresponding to user group 1 and the evaluation index value corresponding to user group 10 is obtained. If the difference is greater than or equal to the preset difference threshold, the candidate recommendation model 1 corresponding to user group 1 is taken as the optimal recommendation model, and the model parameters of candidate recommendation model 1 are taken as the optimal model parameters for the current round. For the remaining candidate recommendation models except the optimal recommendation model, that is, for candidate recommendation models 2 to candidate recommendation models 10, after randomly perturbing the optimal model parameters respectively, the remaining candidate recommendation models are updated based on the model parameters after random perturbation to obtain updated candidate recommendation models.

[0127] Furthermore, if in a certain round of optimization, the difference between the evaluation index value corresponding to the optimal user group and the evaluation index value corresponding to the worst user group is less than the preset difference threshold, it means that the model training has reached a convergence state. At this time, the optimal recommendation model corresponding to the optimal user group in the current round can be determined from the candidate recommendation models, and the optimal model parameters corresponding to the optimal recommendation model can be obtained; the model parameters corresponding to all candidate recommendation models are updated to the optimal model parameters.

[0128] The multi-round optimization process based on log data and the model convergence judgment process introduced above enable the candidate recommendation model to be continuously optimized with user interaction during actual application. The unsupervised training method is used to continuously improve the model's effectiveness, maintaining high recommendation quality and user satisfaction.

[0129] Figure 7 This is an exemplary schematic diagram of a resource recommendation device shown in this application. Figure 7 As shown, the resource recommendation device 700 includes a first acquisition module 701, a second acquisition module 702, a first determination module 703, a second determination module 704 and a resource recommendation module 705, wherein:

[0130] The first acquisition module 701 is used to acquire user characteristics corresponding to the target user for whom resource recommendation is to be performed.

[0131] The second acquisition module 702 is configured to acquire resource features corresponding to each candidate resource included in the candidate resource pool, wherein the resource features are used to indicate a matching degree between the target user's preference information and the candidate resource.

[0132] The first determination module 703 is configured to determine the index scores corresponding to the candidate resources on the multiple recommendation indexes.

[0133] The second determination module 704 is configured to input the user characteristics, the resource characteristics corresponding to the candidate resource, and the index score corresponding to the candidate resource into the target recommendation model for any candidate resource to obtain a recommendation score corresponding to the candidate resource.

[0134] The resource recommendation module 705 is used to sort the candidate resources based on the recommendation scores, obtain a target recommendation list, and recommend resources to the target user based on the target recommendation list.

[0135] In this device, each candidate resource is assigned a resource feature that represents the degree of match between the target user's preference information and the candidate resource, which can more accurately identify the target user's preference for different types of resources. By combining the scoring of candidate resources on multiple recommendation indicators and combining user characteristics, the accuracy of resource recommendations is improved, the user experience is enhanced, and user satisfaction is increased.

[0136] Furthermore, the second determining module 704 is further configured to: determine the target user group to which the target user belongs; and determine a target recommendation model corresponding to the target user group from a plurality of candidate recommendation models.

[0137] Furthermore, the target recommendation model includes a multi-layer perceptron and an online fusion layer, and the second determination module 704 is also used to: input user features and resource features corresponding to candidate resources into the multi-layer perceptron to obtain the impact factors corresponding to each recommendation indicator, wherein the impact factors correspond to the recommendation indicators one-to-one; input the impact factors and the indicator scores corresponding to the candidate resources into the online fusion layer for weighted fusion to obtain the recommendation scores corresponding to the candidate resources.

[0138] Furthermore, the recommendation index includes at least a sequence fusion index. The first determination module 703 is also used to: obtain the total number of candidate resources included in the candidate resource pool; estimate the target user's estimated completion rate for each candidate resource, and estimate the target user's estimated stay time for each candidate resource; determine the first reference value and the second reference value corresponding to each candidate resource based on the total number, the estimated completion rate and the estimated stay time; for any candidate resource, perform a weighted sum of the first reference value and the second reference value corresponding to the candidate resource to obtain the sequence fusion score corresponding to the candidate resource in the sequence fusion index.

[0139] Furthermore, the first determination module 703 is also used to: sort the candidate resources in descending order according to the estimated completion rate to obtain a first sequence identifier corresponding to each candidate resource; sort the candidate resources in descending order according to the estimated stay time to obtain a second sequence identifier corresponding to each candidate resource; for any candidate resource, determine a first reference value corresponding to the candidate resource based on the total number and the first sequence identifier corresponding to the candidate resource; for any candidate resource, determine a second reference value corresponding to the candidate resource based on the total number and the second sequence identifier corresponding to the candidate resource.

[0140] Furthermore, the first acquisition module 701 is also used to: receive a resource recommendation request, and determine the target user to whom resource recommendation is to be made based on the resource recommendation request; determine user request characteristics based on request-related information corresponding to the resource recommendation request; determine user portrait characteristics based on user-related information corresponding to the target user; determine user scenario characteristics based on scenario-related information corresponding to the resource recommendation request; and construct user characteristics corresponding to the target user based on user request characteristics, user portrait characteristics, and user scenario characteristics.

[0141] Furthermore, the user-related information includes user attribute information and user historical interaction information. The first acquisition module 701 is also used to: determine the user attribute characteristics corresponding to the target user based on the user attribute information; determine the user interaction statistical characteristics and user interaction sequence characteristics corresponding to the target user based on the user historical interaction information; and construct the user portrait characteristics corresponding to the target user based on the user attribute characteristics, user interaction statistical characteristics and user interaction sequence characteristics.

[0142] Furthermore, the second acquisition module 702 is also used to: obtain preference information corresponding to the target user, the preference information at least including the user's preferred resource type and the user's preferred resource duration; obtain resource information of the candidate resource; and determine resource characteristics corresponding to the candidate resource based on the preference information and resource information.

[0143] Furthermore, the recommendation indicators also include completion rate indicators, retention indicators, completion indicators, negative feedback indicators and positive feedback indicators.

[0144] Furthermore, the first determination module 703 is also used to: for any recommendation indicator, obtain indicator user-related features and indicator resource-related features related to the recommendation indicator, and estimate the indicator score of the candidate resource corresponding to the recommendation indicator based on the indicator user-related features and indicator resource-related features.

[0145] Figure 8 This is an exemplary schematic diagram of a training device for a recommendation model shown in this application. Figure 8 As shown, the training device 800 of the recommendation model includes a collection module 801, a calculation module 802 and an optimization module 803, wherein:

[0146] The collection module 801 is used to collect log data corresponding to multiple user groups, wherein the user groups correspond one-to-one to the candidate recommendation models, and the log data records at least the length of time the users included in the user group stay on each recommended resource corresponding to the user.

[0147] The calculation module 802 is used to calculate the evaluation index value corresponding to each candidate recommendation model based on the log data.

[0148] The optimization module 803 is configured to determine an optimization strategy corresponding to each candidate recommendation model based on the evaluation index value, and optimize the candidate recommendation model based on the optimization strategy to obtain an optimized candidate recommendation model.

[0149] The candidate recommendation model is used to obtain the recommendation score corresponding to the candidate resource based on the input user features, the resource features corresponding to the candidate resource, and the indicator score corresponding to the candidate resource.

[0150] By calculating the evaluation index value based on the log data of different user groups, this device can accurately evaluate the recommendation effects of different candidate recommendation models to optimize the models. Moreover, since the optimization process in this scheme is based on log data feedback, the candidate recommendation model is continuously optimized with the user's interaction during the actual application process, the model effect is continuously improved, and a high recommendation quality and user satisfaction are maintained.

[0151] Furthermore, the optimization module 803 is also used to: determine the optimal user group and the worst user group from multiple user groups based on the evaluation index value; obtain the difference between the evaluation index value corresponding to the optimal user group and the evaluation index value corresponding to the worst user group; in response to the difference being greater than or equal to a preset difference threshold, determine the optimal recommendation model corresponding to the optimal user group from the candidate recommendation models, and obtain the optimal model parameters corresponding to the optimal recommendation model; for the remaining candidate recommendation models other than the optimal recommendation model, after randomly perturbing the optimal model parameters, update the remaining candidate recommendation models based on the model parameters after the random perturbation to obtain updated candidate recommendation models; wherein, the optimal model parameters corresponding to the optimal recommendation model remain unchanged.

[0152] Furthermore, the optimization module 803 is also used to: in response to the difference being less than a preset difference threshold, determine the optimal recommendation model corresponding to the optimal user group from the candidate recommendation models, and obtain the optimal model parameters corresponding to the optimal recommendation model; and update the model parameters corresponding to all candidate recommendation models to the optimal model parameters.

[0153] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0154] Figure 9 A schematic block diagram of an example electronic device 900 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 can 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 provided as examples only and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0155] like Figure 9 As shown, the device 900 includes a computing unit 901, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 902 or a computer program loaded from a storage unit 908 into a random access memory (RAM) 903. Various programs and data required for the operation of the device 900 can also be stored in the RAM 903. The computing unit 901, the ROM 902, and the RAM 903 are connected to each other via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.

[0156] Various components in the device 900 are connected to the I / O interface 905, including an input unit 906, such as a keyboard, a mouse, etc.; an output unit 907, such as various types of displays, speakers, etc.; a storage unit 908, such as a magnetic disk, an optical disk, etc.; and a communication unit 909, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 909 allows the device 900 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0157] The computing unit 901 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 901 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 901 performs the various methods and processes described above, such as the resource recommendation method or the training method of the recommendation model. For example, in some embodiments, the resource recommendation method or the training method of the recommendation model can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as the storage unit 908. In some embodiments, part or all of the computer program can be loaded and / or installed on the device 900 via the ROM 902 and / or the communication unit 909. When the computer program is loaded into the RAM 903 and executed by the computing unit 901, one or more steps of the resource recommendation method or the training method of the recommendation model described above can be performed. Alternatively, in other embodiments, the computing unit 901 may be configured to execute the resource recommendation method or the recommendation model training method in any other appropriate manner (for example, by means of firmware).

[0158] Various embodiments of the systems and techniques described 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), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0159] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0160] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0161] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the 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 acoustic input, voice input, or tactile input).

[0162] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0163] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0164] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not a limitation herein.

[0165] The above specific embodiments do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure shall be included within the scope of protection of this disclosure.

Claims

1. A resource recommendation method, comprising: Obtain user characteristics corresponding to the target user for whom resource recommendation is to be made; Obtaining resource features corresponding to each candidate resource included in the candidate resource pool, wherein the resource features are used to indicate a matching degree between the target user's preference information and the candidate resource; Determining the indicator scores corresponding to the candidate resource on multiple recommendation indicators; For any candidate resource, input the user characteristics, the resource characteristics corresponding to the candidate resource, and the indicator score corresponding to the candidate resource into a target recommendation model to obtain a recommendation score corresponding to the candidate resource; The candidate resources are sorted based on the recommendation scores to obtain a target recommendation list, and resources are recommended to the target user based on the target recommendation list.

2. The method according to claim 1, wherein Before inputting the user characteristics, the resource characteristics corresponding to the candidate resources, and the indicator scores corresponding to the candidate resources into the target recommendation model, the method further includes: Determining a target user group to which the target user belongs; A target recommendation model corresponding to the target user group is determined from a plurality of candidate recommendation models.

3. The method according to claim 2, wherein: The target recommendation model includes a multi-layer perceptron and an online fusion layer. The user features, resource features corresponding to the candidate resources, and indicator scores corresponding to the candidate resources are input into the target recommendation model to obtain the recommendation scores corresponding to the candidate resources, including: Inputting the user features and the resource features corresponding to the candidate resources into the multi-layer perceptron to obtain the impact factor corresponding to each of the recommendation indicators, wherein the impact factor corresponds to the recommendation indicator in a one-to-one manner; The impact factors and the index scores corresponding to the candidate resources are input into the online fusion layer for weighted fusion to obtain the recommendation scores corresponding to the candidate resources.

4. The method according to claim 3, wherein: The recommendation index includes at least a sequence fusion index, and determining the index score of the candidate resource corresponding to the sequence fusion index includes: Obtaining the total number of candidate resources included in the candidate resource pool; estimating the target user's estimated completion rate for each of the candidate resources, and estimating the target user's estimated stay time for each of the candidate resources; Determine a first reference value and a second reference value respectively corresponding to each candidate resource based on the total number, the estimated completion rate, and the estimated stay duration; For any candidate resource, a weighted sum is performed on the first reference value and the second reference value corresponding to the candidate resource to obtain a sequence fusion score corresponding to the sequence fusion index of the candidate resource.

5. The method according to claim 4, wherein The determining, based on the total number, the estimated completion rate, and the estimated stay duration, respectively corresponding to each candidate resource, of a first reference value and a second reference value includes: Sort the candidate resources in descending order of the estimated completion rate to obtain a first sequence identifier corresponding to each candidate resource; Sort the candidate resources in descending order of the estimated stay durations to obtain a second sequence identifier corresponding to each candidate resource; For any of the candidate resources, determining a first reference value corresponding to the candidate resource based on the total number and a first sequence identifier corresponding to the candidate resource; For any of the candidate resources, a second reference value corresponding to the candidate resource is determined based on the total number and a second sequence identifier corresponding to the candidate resource.

6. The method according to any one of claims 1 to 5, wherein The step of obtaining user characteristics corresponding to the target user for whom resource recommendation is to be performed includes: receiving a resource recommendation request, and determining a target user for whom resource recommendation is to be made based on the resource recommendation request; determining a user request feature based on request related information corresponding to the resource recommendation request; Determine user profile features based on user-related information corresponding to the target user; Determining user scenario characteristics based on scenario-related information corresponding to the resource recommendation request; The user features corresponding to the target user are constructed based on the user request features, the user portrait features and the user scenario features.

7. The method according to claim 6, wherein: The user-related information includes user attribute information and user historical interaction information. The determining of user portrait features based on the user-related information corresponding to the target user includes: Determining a user attribute feature corresponding to the target user based on the user attribute information; Determining user interaction statistical features and user interaction sequence features corresponding to the target user based on the user historical interaction information; A user portrait feature corresponding to the target user is constructed based on the user attribute feature, the user interaction statistical feature, and the user interaction sequence feature.

8. The method according to claim 1, wherein The obtaining of resource characteristics corresponding to each candidate resource included in the candidate resource pool includes: Obtaining preference information corresponding to the target user, the preference information including at least the user's preferred resource type and the user's preferred resource duration; Obtaining resource information of the candidate resource; The resource characteristics corresponding to the candidate resource are determined based on the preference information and the resource information.

9. The method according to claim 4, wherein: The recommendation indicators also include a completion rate indicator, a stay indicator, a completion indicator, a negative feedback indicator, and a positive feedback indicator.

10. The method according to claim 9, wherein: Determining the indicator scores corresponding to the candidate resources on the multiple recommendation indicators includes: For any of the recommendation indicators, obtain indicator user-related features and indicator resource-related features related to the recommendation indicator, and estimate the indicator score of the candidate resource corresponding to the recommendation indicator based on the indicator user-related features and the indicator resource-related features.

11. A method for training a recommendation model, comprising: Collecting log data corresponding to a plurality of user groups, wherein the user groups correspond to candidate recommendation models one-to-one, and the log data records at least the length of time that users included in the user groups spend on each recommended resource corresponding to the users; Calculate the evaluation index value corresponding to each candidate recommendation model based on the log data; Determining an optimization strategy corresponding to each candidate recommendation model based on the evaluation index value, and optimizing the candidate recommendation model based on the optimization strategy to obtain an optimized candidate recommendation model; The candidate recommendation model is used to obtain a recommendation score corresponding to the candidate resource based on input user features, resource features corresponding to the candidate resource, and an indicator score corresponding to the candidate resource.

12. The method according to claim 11, wherein The step of determining an optimization strategy corresponding to each candidate recommendation model based on the evaluation index value, and optimizing the candidate recommendation model based on the optimization strategy to obtain an optimized candidate recommendation model includes: Determining an optimal user group and a worst user group from the plurality of user groups based on the evaluation index values; Obtaining a difference between the evaluation index value corresponding to the optimal user group and the evaluation index value corresponding to the worst user group; In response to the difference being greater than or equal to a preset difference threshold, determining an optimal recommendation model corresponding to the optimal user group from the candidate recommendation models, and obtaining optimal model parameters corresponding to the optimal recommendation model; For the remaining candidate recommendation models except the optimal recommendation model, after randomly perturbing the optimal model parameters, updating the remaining candidate recommendation models based on the model parameters after the random perturbation to obtain updated candidate recommendation models; The optimal model parameters corresponding to the optimal recommendation model remain unchanged.

13. The method according to claim 12, wherein: The method further comprises: In response to the difference being less than the preset difference threshold, determining an optimal recommendation model corresponding to the optimal user group from the candidate recommendation models, and obtaining optimal model parameters corresponding to the optimal recommendation model; The model parameters corresponding to all the candidate recommendation models are updated to the optimal model parameters.

14. A resource recommendation device, comprising: The first acquisition module is used to obtain user characteristics corresponding to the target user for whom resource recommendation is to be made; A second acquisition module is configured to acquire resource features corresponding to each candidate resource included in the candidate resource pool, wherein the resource features are used to represent a matching degree between the target user's preference information and the candidate resource; A first determination module is used to determine the indicator scores corresponding to the candidate resources on multiple recommendation indicators; A second determination module is configured to input the user characteristics, the resource characteristics corresponding to the candidate resource, and the index score corresponding to the candidate resource into a target recommendation model for any candidate resource to obtain a recommendation score corresponding to the candidate resource; The resource recommendation module is used to sort the candidate resources based on the recommendation scores, obtain a target recommendation list, and recommend resources to the target user based on the target recommendation list.

15. A training device for a recommendation model, comprising: A collection module, configured to collect log data corresponding to a plurality of user groups, wherein the user groups correspond one-to-one to the candidate recommendation models, and the log data records at least the length of time that users included in the user groups spend on each recommended resource corresponding to the users; A calculation module, configured to calculate, based on the log data, an evaluation index value corresponding to each of the candidate recommendation models; an optimization module, configured to determine an optimization strategy corresponding to each candidate recommendation model based on the evaluation index value, and optimize the candidate recommendation model based on the optimization strategy to obtain an optimized candidate recommendation model; The candidate recommendation model is used to obtain a recommendation score corresponding to the candidate resource based on input user features, resource features corresponding to the candidate resource, and an indicator score corresponding to the candidate resource.

16. 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, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-10 or 11-13.

17. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to cause the computer to execute the method according to any one of claims 1-10 or 11-13.

18. A computer program product comprising a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1 to 10 or 11 to 13.

Images