Information recommendation method and related equipment

Abstract

The embodiment of the invention provides an information recommendation method and device, electronic equipment, a computer readable storage medium and a computer program product, and the method comprises the steps: obtaining user features of a first user, and obtaining information features of first recommendation information in a first recommendation information set; according to the user characteristics of the first user and the information characteristics of the first recommendation information, determining first probability information that the first user accesses the first recommendation information; according to the weight parameter information of the selection model and a pre-established corresponding relationship between weight parameters and recommendation information, determining second probability information of the selection model for selecting the first recommendation information; and determining target recommendation information in the first recommendation information set according to the first probability information and the second probability information, and pushing the target recommendation information to the first user. Therefore, the accuracy of information recommendation can be improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to an information recommendation method, device, electronic device, computer-readable storage medium, and computer program product. Background Art

[0002] In the information recommendation scenario, in order to ensure stable recommendation benefits, it is often necessary to test which of the multiple candidate recommendation information has better user feedback in a small range before officially launching a large amount of recommendation information. Manually observing the small-scale test process for a period of time and selecting the recommended information to be finally launched is simple to implement, but it is too dependent on manual and subjective judgment, time-consuming and labor-intensive. In addition, when the user feedback of multiple recommendation information is not much different, the accuracy of manually selecting the most suitable recommendation information to be launched may be low. Summary of the invention

[0003] Embodiments of the present application provide an information recommendation method, device, electronic device, computer-readable storage medium, and computer program product, which can improve the accuracy of information recommendation.

[0004] In a first aspect, an embodiment of the present application provides an information recommendation method, comprising: Acquire user characteristics of the first user, and acquire information characteristics of the first recommended information in the first recommended information set; determining first probability information of the first user accessing the first recommended information according to a user feature of the first user and an information feature of the first recommended information; Determining second probability information that the selection model selects the first recommendation information according to weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommendation information; According to the first probability information and the second probability information, target recommended information is determined in the first recommended information set, and the target recommended information is pushed to the first user.

[0005] In a second aspect, an embodiment of the present application provides an information recommendation device, including: an acquisition unit, configured to acquire user characteristics of the first user, and acquire information characteristics of the first recommended information in the first recommended information set; a determining unit, configured to determine first probability information of the first user accessing the first recommended information according to a user feature of the first user and an information feature of the first recommended information; The determining unit is further configured to determine second probability information that the selection model selects the first recommendation information according to weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommendation information; A pushing unit is used to determine target recommended information in the first recommendation information set according to the first probability information and the second probability information, and push the target recommended information to the first user.

[0006] In a third aspect, an embodiment of the present application provides an electronic device, comprising: a processor; and a memory configured to store computer executable instructions, wherein when the computer executable instructions are executed, the processor executes the information recommendation method as described in the first aspect.

[0007] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium for storing computer-executable instructions, which, when executed by a processor, implement the information recommendation method as described in the first aspect.

[0008] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the information recommendation method as described in the first aspect.

[0009] It can be seen that in the embodiment of the present application, first, the user characteristics of the first user are obtained, and the information characteristics of the first recommended information in the first recommended information set are obtained; then, based on the user characteristics of the first user and the information characteristics of the first recommended information, the first probability information that the first user accesses the first recommended information is determined; then, based on the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommended information, the second probability information that the selection model selects the first recommended information is determined; finally, based on the first probability information and the second probability information, the target recommended information is determined in the first recommended information set, and the target recommended information is pushed to the first user. In this way, in the process of determining the target recommended information to be pushed to the first user among the various first recommended information included in the first recommended information set, on the one hand, the user characteristics and information characteristics can be used as a reference to predict the possibility of the first user accessing each first recommended information, that is, to determine the first probability information; on the other hand, the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommended information can be used to determine the possibility of the selection model selecting each first recommended information, that is, to determine the second probability information; and then, the first probability information and the second probability information are used together to determine the target recommended information in the first recommended information set and push it to the first user. By combining the probability information from two different sources for information recommendation, it is beneficial to improve the prediction accuracy, so that the target recommended information pushed to the first user accurately matches the user's needs, interests or preferences, that is, the accuracy of information recommendation is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative labor. Figure 1 A schematic diagram of an implementation environment of an information recommendation method provided in an embodiment of the present application; Figure 2 A processing flow chart of an information recommendation method provided in an embodiment of the present application; Figure 3A The first type of first recommendation information of the target recommendation object provided in the embodiment of the present application; Figure 3B The second first recommendation information of the target recommendation object provided in the embodiment of the present application; Figure 3C The third type of first recommendation information of the target recommendation object provided in the embodiment of the present application; Figure 4 A schematic diagram of a model accuracy stage in an information recommendation method provided in an embodiment of the present application; Figure 5 A schematic diagram of an information recommendation stage in an information recommendation method provided in an embodiment of the present application; Figure 6 A schematic diagram of a model correction principle provided in an embodiment of the present application; Figure 7 A schematic diagram of an information recommendation device provided in an embodiment of the present application; Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0011] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0012] An information recommendation method embodiment provided in this specification: In the information recommendation scenario, the sender of the recommendation information and the receiver of the recommendation information may not agree on the aesthetics. Considering the huge differences in aesthetics between people, it is possible that the sender is very confident in the visual appeal of the recommendation information designed by itself, but the receiver does not buy it and thinks that the recommendation information looks too rustic. Therefore, for the same recommendation object, the sender often needs to pre-configure multiple recommendation information with similar content but different forms of expression for recommending the recommendation object, compare each recommendation information through A / B testing, and select the recommendation information with the best user feedback to go online.

[0013] A / B testing, in simple terms, is to develop two plans for the same goal, for example, two pages, let some users use plan A, and other users use plan B, record the user's usage, and see which plan is more in line with the design.

[0014] During the A / B testing process, operators often need to spend a lot of time and energy to observe which recommended information has better user feedback. However, this process is time-consuming and labor-intensive, and overly dependent on the subjective judgment of operators. When the feedback of multiple recommended information is similar, the recommended information judged by operators to be the most effective may not be accurate.

[0015] Therefore, in order to solve the above problems, an embodiment of the present application provides an information recommendation method.

[0016] The information recommendation method provided in one or more embodiments of the present specification may be applicable to an implementation environment of the information recommendation method, which implementation environment at least includes a server 101 for determining target recommended information.

[0017] The server 101 may be a single server, or a server cluster consisting of a plurality of servers, or one or more cloud servers in a cloud computing platform, for information recommendation.

[0018] In this implementation environment, during the process of information recommendation, the server 101 first obtains the user characteristics of the first user and the information characteristics of the first recommended information in the first recommended information set; then, based on the user characteristics of the first user and the information characteristics of the first recommended information, determines the first probability information that the first user accesses the first recommended information; then, based on the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommended information, determines the second probability information that the selection model selects the first recommended information; finally, based on the first probability information and the second probability information, determines the target recommended information in the first recommended information set, and pushes the target recommended information to the first user. In this way, in the process of determining the target recommended information to be pushed to the first user among the various first recommended information included in the first recommended information set, on the one hand, the user characteristics and information characteristics can be used as a reference to predict the possibility of the first user accessing each first recommended information, that is, to determine the first probability information; on the other hand, the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommended information can be used to determine the possibility of the selection model selecting each first recommended information, that is, to determine the second probability information; and then, the first probability information and the second probability information are used together to determine the target recommended information in the first recommended information set and push it to the first user. By combining the probability information from two different sources for information recommendation, it is beneficial to improve the prediction accuracy, so that the target recommended information pushed to the first user accurately matches the user's needs, interests or preferences, that is, the accuracy of information recommendation is improved.

[0019] Figure 2 A processing flow chart of an information recommendation method provided in an embodiment of the present application. Figure 2 The information recommendation method provided in this embodiment specifically includes steps S202 to S208: Step S202: obtaining user characteristics of the first user, and obtaining information characteristics of the first recommended information in the first recommended information set.

[0020] The number of the first users may be one or more.

[0021] The system or electronic device that executes the information recommendation method belongs to the recommendation issuer, and correspondingly, each first user can be the recommendation receiver.

[0022] In this specification, ordinal numbers such as "first" and "second" are used to distinguish different technical features and have no actual meaning, so they will not be elaborated below.

[0023] User characteristics refer to the description of the user's basic attributes, information needs, information behavior, psychological state, environmental influences, etc. These characteristics can help understand the characteristics and needs of users in order to better design and optimize products or services. User characteristics can be obtained through user research and analysis. They summarize the overall characteristics of a user group or a certain type of people and are extensive.

[0024] For example, user characteristics may include one or more of the following categories: (a1) Natural attributes: including name, gender, age, region, etc. This information is usually provided directly by the user.

[0025] (a2) Profile features: These are features inferred or calculated based on the user’s historical behavior, such as interests, behaviors, etc.

[0026] (a3) Relationship characteristics: involving crowd attributes, attention relationships, intimacy, etc.

[0027] Acquiring the user characteristics of the first user may include acquiring a first user set, and acquiring the user characteristics of each first user in the first user set. The first user set may be a user set custom-configured by the recommendation issuer, and the first user set includes one or more first users, each of which is a recommendation recipient.

[0028] For example, if the recommendation issuer is a financial service platform, n registered users of the financial service platform of a specified category are added to a preset empty set in advance to obtain a first user set. n is an integer greater than or equal to 1.

[0029] In specific implementation, a corresponding user identifier may be pre-configured for each first user in the first user set, and the correspondence between the user identifier of the first user and the user feature of the first user, that is, the correspondence between the user identifier and the user feature, may be stored.

[0030] On this basis, the user characteristics of the first user may be obtained by querying the user characteristics of the first user according to the correspondence between the user identifier and the user characteristics and the user identifier of the first user.

[0031] Considering that some user characteristics may be private information, obtaining the user characteristics of the first user may be obtaining the user characteristics of the first user under the premise of authorization by the first user.

[0032] Acquiring the user features of the first user may also be acquiring a user feature set of the first user, and determining one or more user features required for information recommendation in the user feature set according to the target recommendation object.

[0033] The first recommendation information set may include one or more first recommendation information. Each first recommendation information in the first recommendation information set may correspond to the same recommendation object, ie, the target recommendation object, or each first recommendation information in the first recommendation information set may correspond to different recommendation objects.

[0034] For example, the first recommendation information set includes: Advertisement 1, Advertisement 2, and Advertisement 3, among which Advertisement 1 corresponds to the recommendation object "Product A", Advertisement 2 corresponds to the recommendation object "Product A", and Advertisement 3 corresponds to the recommendation object "Product A", that is, Advertisement 1, Advertisement 2, and Advertisement 3 are all advertising information recommending Product A to users. The difference between the three is the advertising style. The advertising style of Advertisement 1 is simple, the advertising style of Advertisement 2 is gorgeous, and the advertising style of Advertisement 3 is suspenseful.

[0035] For another example, the first recommendation information set includes: advertisement 4, advertisement 5, and advertisement 6, wherein advertisement 4 corresponds to the recommendation object "activity A", advertisement 5 corresponds to the recommendation object "activity B", and advertisement 6 corresponds to the recommendation object "activity C".

[0036] For another example, the first recommendation information set includes: Advertisement 7, Advertisement 8, Advertisement 9, where Advertisement 7 corresponds to the recommended object "Service A", Advertisement 8 corresponds to the recommended object "Service B", and Advertisement 9 corresponds to the recommended object "Service B", that is, Advertisement 7 is used to recommend a service to the user, and Advertisement 8 and Advertisement 9 recommend another service to the user. The advertising content of Advertisement 8 and Advertisement 9 are two different versions, one focusing on emphasizing the advantages of Service B, and the other focusing on emphasizing that the limited-time offer of Service B is about to end.

[0037] Information recommendation usually refers to the content, products, services or other resources recommended to users by the system or platform based on the user's interests, behaviors, historical data or other relevant factors. These recommendations are intended to help users discover new things they may be interested in, improve user experience, increase user engagement, or promote sales and conversions.

[0038] Specifically in the embodiment of the present application, the first recommendation information refers to information used by the system or platform to recommend content, products, services or other resources to the first user based on the first user's interests, behaviors, historical data or other relevant factors. For example, the first recommendation information can be an advertisement.

[0039] The following is an exemplary explanation using the case where each first recommendation information in the first recommendation information set corresponds to the same recommendation object. In the case where each first recommendation information in the first recommendation information set corresponds to different recommendation objects, reference may be made to the corresponding explanation part where each first recommendation information in the first recommendation information set corresponds to the same recommendation object.

[0040] The recommended object may be a product / service / activity, etc., recommended to the user through the information recommendation processing method. The embodiment of the present application does not limit the specific content of the recommended object. The target recommended object may be a product / service / activity, etc., recommended to the first user through the information recommendation processing method.

[0041] For example, the resource lending business wants to recommend the newly launched financial product A to the first user. After receiving the recommendation, the first user may purchase it or ignore it, so "financial product A" is the target recommendation object. Alternatively, the transaction business wants to recommend a limited-time discount event for the first type of goods to the first user. After receiving the recommendation, the first user may participate in the event or not, so "limited-time discount event for the first type of goods" is the target recommendation object.

[0042] For the same recommended object, multiple recommendation information with similar contents but different presentation forms for recommending the recommended object may be pre-configured, and the multiple recommendation information together constitute a recommendation information set for the recommended object.

[0043] Specifically in the embodiment of the present application, first recommendation information with similar content but different presentation forms for recommending a target recommendation object may be pre-configured, and the multiple first recommendation information together constitute a first recommendation information set for the target recommendation object.

[0044] The information feature may describe the display manner of the first recommended information from one or more dimensions.

[0045] In specific implementation, a corresponding information identifier may be pre-configured for each first recommended information in the first recommended information set, and the correspondence between the information identifier of the first recommended information and the information feature of the first recommended information, that is, the correspondence between the information identifier and the user feature, may be stored.

[0046] Acquiring the information feature of the first recommended information in the first recommended information set may be, for each first recommended information in the recommended information set, querying and obtaining the information feature of the first recommended information according to the correspondence between the information identifier and the user feature, and the information identifier of the first recommended information.

[0047] In a specific implementation, the information feature of the first recommendation information may include at least one of multimedia data, format information and tag information.

[0048] The information feature may be multimedia data, format information, tag information, or may include at least two of the above three types of data.

[0049] Multimedia is a combination of multiple media, generally including text, sound, image and other media forms. In computer systems, multimedia refers to a human-computer interactive information exchange and communication medium that combines two or more media. The media used include text, pictures, photos, sound, animation and video, as well as the interactive functions provided by the program.

[0050] Multimedia data may include one or more of the following: text data, audio data, video data, and the like.

[0051] For example, the recommended information set of the credit limit increase activity includes the first recommended information 1 and the first recommended information 2. The text of the first recommended information 1 is: Congratulations, XXXXXX. The text of the first recommended information 2 is: Good news, XXXXXXX.

[0052] The format information may reflect the display mode of each element in the first recommendation information.

[0053] The first recommendation information may include one or more elements, which may be text, a pattern, a control, and the like.

[0054] The format information may include typesetting information. For example, in a display page of the first recommended information, element 1 is located in a first preset area, and element 2 is located in a second preset area.

[0055] The format information may include text style information. For example, in the first recommendation information, the text 1 "XXXXXX" uses a first type of font, and the font color is black, and the text 2 "XX" uses a second type of font, with an underline, and the font color is bright red.

[0056] The format information may include material theme information. For example, for the same service to be recommended, the first recommendation information 1 uses a picture material with a theme of "gift box", and the first recommendation information 2 uses a picture material with a theme of "Spring Festival".

[0057] The format information may include element type information. For example, for product 1, first recommendation information 1 uses a first type of element, and first recommendation information 2 uses a first type of element and a second type of element.

[0058] The various format information listed above are merely exemplary, and any information that can reflect the display method of the elements in the first recommendation information can be used as the format information in the embodiment of the present application.

[0059] The label information may be obtained by pre-labeling the first recommendation information manually or by using a machine learning model. The machine learning model may be a classification model, a text generation model, or other models with labeling capabilities.

[0060] The tag information may reflect the characteristics of the first recommendation information in one or more dimensions.

[0061] For example, the tag information of the first recommended information 1 includes: simple style, and the tag information of the first recommended information 2 includes: retro style, festive style.

[0062] In different first recommendation information of the same target recommendation object: (b1) Different texts / audio / video may be used. For example, if the first recommendation information is an advertisement, the different advertising language does not affect the fact that they are advertising the same product / service / activity.

[0063] (b2) Different formats may be used: black font, bold font, flashing effects, typesetting, etc.

[0064] (b3) They may be labeled with different tags: Taking the first recommended information as an advertisement as an example, each tag can represent an advertising style. For example, Advertisement 1 is labeled with a simple style, Advertisement 2 is labeled with an exaggerated style, Advertisement 3 is labeled with a retro style, and so on.

[0065] In order to more intuitively understand the difference between each first recommended information in the first recommended information set, you can refer to Figure 3A-3C .

[0066] Figure 3A The first type of first recommendation information of the target recommendation object provided in the embodiment of the present application; Figure 3B The second first recommendation information of the target recommendation object provided in the embodiment of the present application; Figure 3C This is the third type of first recommendation information of the target recommendation object provided in the embodiment of the present application.

[0067] like Figure 3A As shown, the text of the first recommendation information 302 includes: "XX Festival Limited Time Surprise", "Congratulations on Being Selected", "XXXXXX", "Happy to Accept". Among them, "Congratulations on Being Selected" uses the first type of font.

[0068] like Figure 3B As shown, the text of the first recommendation information 304 includes: "Congratulations on being selected", "XXXXXX", "Use immediately". Among them, "Congratulations on being selected" uses the second type of font. The first recommendation information 304 uses a picture material with the theme of "gift box".

[0069] like Figure 3C As shown, the text of the first recommendation information 306 includes: "Congratulations on being selected", "XXXXXX", "Use immediately". Among them, "Congratulations on being selected" uses the third type of font. The first recommendation information 306 uses a picture material with the theme of "Spring Festival".

[0070] Step S204: determining first probability information of the first user accessing the first recommended information according to the user characteristics of the first user and the information characteristics of the first recommended information.

[0071] The first user accessing the first recommendation information refers to the first user obtaining or viewing the specific content of the first recommendation information generated by the system or platform through a preset access method (for example, clicking, browsing, searching, etc.). For example, the first user is user 1, the first recommendation information is advertisement 1, and user 1 clicks advertisement 1 to enter the advertisement page of advertisement 1.

[0072] The first probability information of the first user accessing the first recommended information may reflect the possibility of the first user accessing the first recommended information.

[0073] Determining first probability information of the first user accessing the first recommended information based on user characteristics of the first user and information characteristics of the first recommended information can be accomplished by fusing the user characteristics with the information characteristics to obtain comprehensive characteristics, and predicting the behavior of the first user based on the comprehensive characteristics to obtain a probability value representing the likelihood of the first user accessing the first recommended information, i.e., the first probability information.

[0074] User features can be represented in vector form, and information features can be represented in vector form.

[0075] The user feature and the information feature are fused to obtain the comprehensive feature. The first vector representing the user feature and the second vector representing the information feature are concatenated to obtain a first comprehensive vector, and the first comprehensive vector is used as the comprehensive feature.

[0076] Alternatively, the user feature and the information feature are fused to obtain a comprehensive feature, or a first vector representing the user feature and a second vector representing the information feature are weighted summed to obtain a second comprehensive vector, and the second comprehensive vector is used as the comprehensive feature.

[0077] In the process of predicting the behavior of the first user, a classification model, a regression model, a deep learning model, etc. can be used.

[0078] The classification model is used as an example for illustrative explanation: the comprehensive features are input into the classification model for classification processing to obtain first probability information, and the classification result is determined based on the numerical comparison result of the first probability information and the preset probability threshold. The classification result is one of the following two categories: Category 1: the first user accesses the first recommended information; Category 2: the first user does not access the first recommended information.

[0079] The input data of the classification model may include comprehensive features, and the output data of the classification model may include first probability information. Assuming that p1 represents the first probability information, the value of p1 is greater than or equal to 0 and less than or equal to 1. The closer the value of p1 is to 1, the higher the possibility that the first user accesses the first recommended information. The closer the value of p1 is to 0, the lower the possibility that the first user accesses the first recommended information.

[0080] The classification model can be a binary classification neural network model, a forward neural network model, a tree model, a linear model, or any other model with classification function.

[0081] The following takes the classification model as a target binary classification neural network model as an example to illustrate the training method of the classification model: Construct user features and information features based on historical data, and splice user features and information features to obtain training samples. The training samples carry a first label. The first label may include an exposure sub-label that characterizes whether the recommended information corresponding to the information feature has been exposed before the user corresponding to the user feature. The first label may also include a click sub-label that characterizes whether the recommended information corresponding to the information feature has been clicked by the user corresponding to the user feature. The training samples include positive samples and negative samples. The positive samples are click samples, and the negative samples are only exposed but not clicked samples. The initial binary classification neural network model to be trained with the training sample input parameters is trained to obtain a target binary classification neural network model. The target binary classification neural network model can be used to predict the user's click rate on the recommended information. If the target binary classification neural network model is used to perform step S204 in the embodiment of the present application, the target binary classification neural network model can be used to predict the click rate of the first user on the first recommended information. The click rate is the first probability information of the first user visiting the first recommended information.

[0082] In addition, based on a similar concept, the user features and the information features may also be input into a classification model for classification processing to obtain the first probability information.

[0083] In the process of determining the first probability information of the first user accessing the first recommended information based on the user characteristics of the first user and the information characteristics of the first recommended information, any model or other technical means that can be used to predict the behavior of the first user can be used in the embodiments of the present application. For details, please refer to the corresponding description part of the classification model.

[0084] In addition, based on the user characteristics of the first user and the information characteristics of the first recommended information, the first probability information of the first user accessing the first recommended information is determined, or the user characteristics of the first user and the information characteristics of the first recommended information can be combined to obtain a target information combination; based on the correspondence between the target information combination and the pre-configured information combination and probability information, the target probability information corresponding to the target information combination is determined, and the target probability information is determined as the first probability information of the first user accessing the first recommended information.

[0085] Step S206: Determine second probability information that the selection model selects the first recommendation information according to the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommendation information.

[0086] The selection model may be a machine learning model that employs a reinforcement learning algorithm.

[0087] RL (Reinforcement Learning) is a machine learning method that aims to maximize cumulative rewards by learning the optimal strategy through the interaction between the agent and the environment. The core idea of ​​reinforcement learning is to learn through trial and error. The agent takes actions in the environment and adjusts its strategy based on the feedback (reward or punishment) from the environment.

[0088] Exemplarily, the reinforcement learning algorithm may be a MAB (Multi-armed bandit) algorithm.

[0089] The MAB problem is a classic optimization problem that is used to balance exploration and exploitation. In this problem, there are multiple "arms", each of which provides a different, usually unknown, rate of return. The goal is to find the best arm, the one that provides the greatest reward, through a series of attempts. "Arm" can be thought of as a metaphor for the actions or strategies that a decision maker can choose.

[0090] The selection model in the embodiment of the present application may have multiple arms, and the model parameters of the selection model include weight parameter information. There is a corresponding relationship between the weight parameter information and the multiple arms of the selection model.

[0091] For example, the selection model has n arms, and the weight parameter information includes n weight parameters: k1, k2, ..., kn. The n arms correspond to the n weight parameters one by one. n is an integer greater than 1. Among the n arms, for any arm, the larger the value of the weight parameter corresponding to the arm, the higher the possibility that the arm is selected by the selection model; the smaller the value of the weight parameter corresponding to the arm, the lower the possibility that the arm is selected by the selection model.

[0092] Before step S206 is executed, a correspondence between the weight parameter and the recommendation information may be established according to the weight parameter information of the selection model and the first recommendation information set.

[0093] For example, the weight parameter information includes three weight parameters: k1, k2, k3, and the first recommendation information set includes three first recommendation information: first recommendation information 1, first recommendation information 2, and first recommendation information 3. A correspondence between the weight parameters and the recommendation information is established, and the correspondence includes: k1 corresponds to the first recommended information 1; k2 corresponds to the first recommended information 2; k3 corresponds to the first recommended information 3.

[0094] For another example, the information recommendation set includes 3 advertisements, and the first user set includes 1,000 first users. The weight parameter information of the selection model includes 3 weight parameters: k1, k2, and k3. In the pre-established correspondence, advertisement 1 corresponds to k1, advertisement 2 corresponds to k2, and advertisement 3 corresponds to k3. This correspondence remains unchanged before the information recommendation is completed for these 1,000 first users.

[0095] After the corresponding relationship between the weight parameter and the recommended information is established, the selection model can be used to perform selection processing in the first recommended information set to obtain a selection result.

[0096] It should be noted that the subsequent steps do not need the selection result output by the selection model, but the second probability information generated in the process of obtaining the selection result by performing a selection process on the selection model.

[0097] Specifically, for user 1, taking the information recommendation set including advertisement 1, advertisement 2 and advertisement 3 as an example, the selection model can perform a selection prediction among the three advertisements, and output three second probability information at the same time: the second probability information that the selection model selects advertisement 1, the second probability information that the selection model selects advertisement 2, and the second probability information that the selection model selects advertisement 3. The second probability information of advertisement 1, the second probability information of advertisement 2 and the second probability information of advertisement 3 can be used to determine the selection result of the selection model, which is one or more of advertisement 1, advertisement 2 and advertisement 3. In the subsequent steps, only the second probability information of each advertisement is needed, and the selection result is not needed.

[0098] Taking the example that the first recommendation information set includes n first recommendation information, during the process of the selection model performing a selection process, the selection model can generate n second probability information, and the n first recommendation information corresponds one-to-one to the n second probability information.

[0099] It should be noted that the selection model can have two states. In the process of determining the second probability information of the selection model selecting the first recommendation information based on the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommendation information, the two states can adopt different specific implementation methods.

[0100] In a specific implementation, the second probability information that the selection model selects the first recommendation information is determined based on the weight parameter information of the selection model and the pre-established correspondence between the weight parameters and the recommendation information. The target state of the selection model can be determined in a preset state set of the selection model, and the second probability information that the selection model selects the first recommendation information is determined under the target state based on the weight parameter information of the selection model and the pre-established correspondence between the weight parameters and the recommendation information.

[0101] Exemplarily, the preset state set of the selection model may include an exploration state and an utilization state. The target state may be one of the exploration state and the utilization state.

[0102] When the selection model is in the utilization state, the selection model selects the first recommendation information corresponding to the arm with the highest weight parameter.

[0103] When the selection model is in the exploration state, the selection model excludes a preset number of arms with the highest weight parameters and then selects the first recommendation information corresponding to one of the remaining arms.

[0104] The target state is determined in the preset state set of the selection model by generating a random number within a preset numerical range, comparing the random number with a preset numerical threshold to obtain a comparison result, and determining the target state in the preset state set according to the comparison result.

[0105] For example, the preset numerical range is [0,1], and a random number is generated. The random number can be a real number. If the random number is less than the preset numerical threshold 0.8, the target state is determined to be the utilization state. If the random number is greater than or equal to the preset numerical threshold 0.8, the target state is determined to be the exploration state.

[0106] In the case where the selection model adopts the LinUCB (linear upper confidence bound) algorithm, the target state is determined in the preset state set of the selection model. The number of first users corresponding to the exploration state can be determined according to the confidence parameter of the selection model and the number of first users, and the first users corresponding to the exploration state are selected from the first users according to the number of first users, and the unselected first users are used as the first users corresponding to the utilization state. In this way, the target state corresponding to any first user can be learned.

[0107] LinUCB algorithm is used to make the best decision when the decision maker needs to choose multiple options at the same time. In each iteration, LinUCB algorithm will dynamically adjust its evaluation of each option based on the previous cumulative benefits, and finally select the option with the highest expected benefit.

[0108] Specifically, the LinUCB algorithm uses a linear regression model as a decision model, which assumes that the benefit of each option is related to a set of features such as feature vectors. In each iteration, the algorithm updates this set of features based on previous observations and uses them to build a linear regression model. For each option, the algorithm calculates the confidence interval of its expected benefit, and then selects the option with the largest upper bound of the confidence interval as the choice for this iteration.

[0109] The core idea of ​​the LinUCB algorithm is to dynamically adjust the evaluation of options so that the expected return value of each option gradually approaches the true value. This process involves two important elements: (a1) Feature representation: The LinUCB algorithm associates the benefit of each option with a set of features such as feature vectors, which can reflect the advantages and disadvantages, preferences and other factors of the option.

[0110] (a2) Uncertainty estimation: The LinUCB algorithm uses the upper bound of the confidence interval as the upper bound of the expected return of each option. This approach can reduce the decision risk between options by exploring options with large uncertainty.

[0111] The LinUCB algorithm is an online training algorithm that continuously updates model parameters based on new observations, that is, after each decision, the model parameters are updated based on new user feedback or reward information. By continuously updating model parameters, LinUCB can dynamically adapt to changes in user connection habits over time.

[0112] Specifically in the embodiment of the present application, when the selection model adopts the LinUCB algorithm, user behavior data can be collected after the information recommendation is completed to the first user to adjust the weight parameter information of the selection model.

[0113] In a specific implementation, the information recommendation method also includes: obtaining user characteristics of multiple second users, and obtaining information characteristics of second recommended information in a second recommended information set, the second users include users to whom recommended information has been pushed, and the second recommended information includes recommended information that has been pushed to the users; determining third probability information of the second user accessing the second recommended information based on the user characteristics of the second user and the information characteristics of the second recommended information; determining average probability information of the second user accessing the second recommended information based on the third probability information; and configuring weight parameter information of the selection model based on the average probability information.

[0114] Before putting the selection model into use, you can do some model-related data preparation, such as configuring the model parameters of the selection model. The model parameters of the selection model include weight parameter information.

[0115] The second user may be a user who participates in the information recommendation process in the historical data. The number of the second users may be one or more. The system or electronic device that executes the information recommendation method belongs to the recommendation issuer, and correspondingly, each second user may be the recommendation receiver.

[0116] The second recommendation information set may include one or more second recommendation information. Each second recommendation information in the second recommendation information set may correspond to the same recommendation object, ie, the target recommendation object, or each second recommendation information in the second recommendation information set may correspond to different recommendation objects.

[0117] The second recommendation information set can be obtained from historical data.

[0118] Taking into account that some products may be launched for a limited time, some services may be provided for a limited time, and some activities may be open for a limited time, the target recommendation object may be recommended multiple times in different time periods. For example, the target recommendation object is a credit limit increase activity, which is open for a limited time of one week each time. A credit limit increase activity was opened three months ago, and it is open again today.

[0119] The second recommendation information set of the target recommendation object may be a recommendation information set used when the target recommendation object is recommended in the historical data.

[0120] Since the technical concepts are similar, the acquisition of user features of multiple second users and the acquisition of information features of the second recommendation information in the second recommendation information set may refer to the corresponding description part of step S202.

[0121] According to the user characteristics of the second user and the information characteristics of the second recommended information, third probability information of the second user accessing the second recommended information is determined, and reference may be made to the corresponding description of the aforementioned step S204.

[0122] The third probability information of the second user accessing the second recommended information may refer to the corresponding description of the first probability information of the first user accessing the first recommended information.

[0123] There are multiple second users, and each second user has corresponding third probability information. Further, the average probability information of the second user accessing the second recommended information can be determined according to the third probability information.

[0124] In a specific implementation, when the second recommendation information set includes m second recommendation information, in the process of processing the i-th second recommendation information, firstly, according to the user characteristics of the second user u and the information characteristics of the i-th second recommendation information, the third predicted probability of the second user u accessing the i-th second recommendation information is determined. , and then determine the average probability information of the second user accessing the second recommended information based on the third probability information . Wherein, m is an integer greater than 0, i is an integer greater than 0 and less than or equal to m, , I represents the second recommended information set. u is an integer greater than 0, , U represents a second user set, and the second user set U includes n second users, where n is an integer greater than 0.

[0125] Configuring the weight parameter information of the selection model according to the average probability information may be to use the average probability information corresponding to the second recommendation information as a weight parameter of the selection model, and there is a corresponding relationship between the weight parameter and the second recommendation information.

[0126] Configuring the weight parameter information can be understood as configuring corresponding initial values ​​for the weight parameters corresponding to each arm of the selection model before the selection model is put into application.

[0127] In a specific implementation, the information recommendation method also includes: obtaining historical access behavior data for a historical recommendation information set, and constructing a test sample based on the historical access behavior data; inputting the test sample into a selection model for selection processing to obtain a selection result of the test sample; determining whether the loss of the selection model converges based on the selection result; if so, determining that the selection model passes the test; if not, determining that the selection model fails the test, and adjusting the preset parameters of the selection model.

[0128] Before executing step S206, an offline verification may be performed to determine whether the values ​​of the preset parameters of the selected model are reasonable.

[0129] In specific implementation, the historical recommended information set with the highest similarity to the first information recommendation set can be selected. For example, if the current activity to be launched is a "credit limit increase" activity, the exposure and click events of similar "credit limit increase" activities in history are selected, and a queue is created with the second recommended information as the dimension. For example, if there are n second recommended information of similar "credit limit increase" activities in history, n queues are created. For a certain queue, the exposure and click events corresponding to the second recommended information are constructed in chronological order to simulate the migration and convergence of the probability of each arm of the selection model offline. n is a natural number greater than 0.

[0130] The preset parameter of the selection model can be an artificially configured hyperparameter. When the value of the hyperparameter is reasonable, the loss convergence of the selection model can be determined based on the selection results of the test samples, and then it can be determined that the selection model has passed the test, and steps S206-S208 in the information recommendation method provided in the embodiment of the present application can be executed. When the value of the hyperparameter is unreasonable, it can be determined based on the interactive test results that the loss of the selection model does not converge, and then it can be determined that the selection model has failed the test. When it is determined that the selection model has failed the test, a prompt message can be generated to prompt the user to modify the preset parameter, or the preset parameter can be automatically adjusted according to the preset adjustment direction.

[0131] In addition, during the offline testing phase, the classification model can also be tested using offline binary classification evaluation indicators such as AUC (Area Under Curve).

[0132] In a specific implementation, second probability information of the selection model selecting the first recommended information is determined based on weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommended information, including: when the selection model is in a first state, selecting multiple first weight parameters whose values ​​are higher than a first threshold in the weight parameter information; deleting multiple first recommended information from the first information recommendation set based on the multiple first weight parameters and the correspondence to obtain a third recommended information set, wherein the number of deleted first recommended information is the same as the number of the multiple first weight parameters; determining the first recommended information selected by the selection model in the third recommended information set; configuring the second probability information of the first recommended information selected by the selection model according to a first preset value, and configuring the second probability information of the first recommended information not selected in the third recommendation set according to a second preset value.

[0133] The first state may be the aforementioned exploration state.

[0134] When the selection model is in the first state, a plurality of first weight parameters having values ​​higher than a first threshold value may be selected from the weight parameter information.

[0135] Alternatively, when the selection model is in the first state, one or more first weight parameters with the largest values ​​may be selected from the weight parameter information.

[0136] According to the correspondence between the multiple first weight parameters and the weight parameters and the recommended information, the multiple first recommended information are deleted from the first information recommendation set to obtain a third recommended information set, and the number of the deleted first recommended information is the same as the number of the multiple first weight parameters.

[0137] For example, the first information recommendation set includes first recommendation information 1, first recommendation information 2, and first recommendation information 3. In the correspondence between weight parameters and recommendation information, first recommendation information 1 corresponds to weight parameter k1, first recommendation information 2 corresponds to weight parameter k2, and first recommendation information 3 corresponds to weight parameter k3. k1, k2, and k3 are all real numbers greater than or equal to 0 and less than or equal to 1, and k1 < k2 < k3. Selecting multiple first weight parameters with values higher than the first threshold in the weight parameter information includes: k2 and k3. According to the multiple first weight parameters and the correspondence between weight parameters and recommendation information, delete first recommendation information 2 and first recommendation information 3 in the first information recommendation set to obtain a third recommendation information set, and this third recommendation information set includes: first recommendation information 1.

[0138] In the third recommendation information set, determine the first recommendation information selected by the selected model, which can be to determine the first recommendation information selected by the selected model in the third recommendation information set through the Thompson sampling method.

[0139] Thompson sampling is a decision-making algorithm based on the Bayesian method. By maintaining and updating the posterior distribution of the reward probability distribution of each arm, it dynamically balances exploration and exploitation, so as to maximize the cumulative reward in the MAB problem.

[0140] In the third recommendation information set, determine the first recommendation information selected by the selected model, which can also be to determine the first recommendation information selected by the selected model in the third recommendation information set through the ε-greedy strategy.

[0141] The ε-greedy strategy is a simple algorithm that balances exploration and exploitation by introducing the parameter ε. It maximizes the cumulative reward in the MAB problem by selecting the arm with the highest currently estimated reward in most cases (i.e., in the exploitation state) and randomly selecting an arm in a few cases (i.e., in the exploration state).

[0142] In the third recommendation information set, determine the first recommendation information selected by the selected model, which can also be to determine the first recommendation information selected by the selected model in the third recommendation information set through the UCB (upper confidence bound) strategy.

[0143] The UCB strategy considers the return and uncertainty of each arm and selects the arm with the highest upper confidence bound.

[0144] Configure the second probability information of the first recommendation information selected by the selected model according to the first preset value, and configure the second probability information of the first recommendation information not selected in the third recommendation set according to the second preset value.

[0145] Exemplarily, the first preset value may be 1, and the second preset value may be 0.

[0146] For example, the first recommendation information set includes: first recommendation information 1, first recommendation information 2, first recommendation 3, and first recommendation information 4. The first recommendation information selected by the selection model is the first recommendation information 1, and the first recommendation information that is not selected includes: first recommendation information 2, first recommendation 3, and first recommendation information 4. In the weight parameter information of the selection model, the initial value of the weight parameter k1 corresponding to the first recommendation information 1 is configured to 1; the initial value of the weight parameter k2 corresponding to the first recommendation information 2 is configured to 0; the initial value of the weight parameter k3 corresponding to the first recommendation information 3 is configured to 0; and the initial value of the weight parameter k4 corresponding to the first recommendation information 4 is configured to 0.

[0147] When the selection model is in the first state, a part of the first recommendation information with higher values ​​of the corresponding weight parameters is excluded, and the first recommendation information selected by the selection model is determined from the remaining first recommendation information, and then the second probability information of each first recommendation information in the first recommendation information set is determined, which is conducive to exploring non-optimal choices and can better adapt to changes in user behavior.

[0148] For example, the weight parameter of Ad 1 has the highest value, which means that compared with Ad 2 and Ad 3, User 1 is likely to prefer the visual style of Ad 1 and is more likely to click on Ad 1. However, in fact, user behavior is not static. If you always recommend the ad with the highest weight parameter, it will be difficult to detect changes in user preferences. In the exploration state, you can tentatively push Ad 2 or Ad 3 to the user to see how the user reacts.

[0149] In a specific implementation, second probability information of the selection model selecting the first recommendation information is determined based on weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommendation information, including: when the selection model is in the second state, selecting a second weight parameter in the weight parameter information; determining the first recommendation information corresponding to the second weight parameter based on the second weight parameter and the correspondence between the weight parameter and the recommendation information; and configuring the second probability information of the first recommendation information corresponding to the second weight parameter based on the second weight parameter.

[0150] The weight parameter information includes multiple weight parameters, and the processing method of each weight parameter is exactly the same. The second weight parameter can be any weight parameter in the weight parameter information.

[0151] For example, the weight parameter information includes: k1, k2, k3. In the process of performing a selection process when the selection model is in the second state: A second weight parameter k1 is selected from the weight parameter information; the first recommendation information X1 corresponding to k1 is determined according to the correspondence between k1, the weight parameter and the recommendation information; and the second probability information of X1 is configured according to k1.

[0152] A second weight parameter k2 is selected from the weight parameter information; based on the correspondence between k2, the weight parameter and the recommendation information, the first recommendation information X2 corresponding to k2 is determined; and based on k2, the second probability information of X2 is configured.

[0153] A second weight parameter k3 is selected from the weight parameter information; the first recommendation information X3 corresponding to k3 is determined according to the correspondence between k3, the weight parameter and the recommendation information; and the second probability information of X3 is configured according to k3.

[0154] The second state may be the aforementioned utilization state.

[0155] According to the second weight parameter, the second probability information of the first recommended information corresponding to the second weight parameter is configured. The second weight parameter can be determined as the second probability information of the first recommended information corresponding to the second weight parameter, or the second weight parameter can be normalized and the second weight parameter after normalization is determined as the second probability information of the first recommended information corresponding to the second weight parameter.

[0156] Step S208: determining target recommended information in the first recommended information set according to the first probability information and the second probability information, and pushing the target recommended information to the first user.

[0157] For each first user, after executing step S202 to step S206, n first probability information and n second probability information may be obtained, where n is the number of first recommended information in the first recommended information set, and n is an integer greater than 0.

[0158] Determining the target recommended information in the first recommended information set based on the first probability information and the second probability information may be determining a comprehensive evaluation score of each first recommended information based on the first probability information and the second probability information, and determining one or more first recommended information with the highest comprehensive evaluation score as the target recommended information.

[0159] For example, the first probability information of the first user accessing the first recommended information 1 is p1, the first probability information of the first user accessing the first recommended information 2 is p2, and the first probability information of the first user accessing the first recommended information 3 is p3. The selection model selects the second probability information of the first recommended information 1 as p1', the selection model selects the second probability information of the first recommended information 2 as p2', and the selection model selects the second probability information of the first recommended information 3 as p3. Based on p1 and p1', the comprehensive evaluation score P1 of the first recommended information 1 can be determined; based on p2 and p2', the comprehensive evaluation score P2 of the first recommended information 2 can be determined; based on p3 and p3', the comprehensive evaluation score P3 of the first recommended information 3 can be determined. Assuming that P1<P2<P3, the first recommended information 3 corresponding to P3 can be determined as the target recommended information.

[0160] Determining the comprehensive evaluation score of each first recommendation information according to the first probability information and the second probability information may be to perform weighted summation of the first probability information and the second probability information for any first recommendation information in the first recommendation information set to obtain the comprehensive evaluation score of the first recommendation information.

[0161] For example, for the i-th first recommended information in the first recommended information set, according to the first probability information With the second probability information Perform weighted summation: , where 0≤ ≤1, It is a custom configuration parameter, i is an integer greater than 0.

[0162] The comprehensive evaluation score of each first recommendation information is determined according to the first probability information and the second probability information. Alternatively, for any first recommendation information in the first recommendation information set, the first probability information and the second probability information are averaged to obtain the comprehensive evaluation score of the first recommendation information.

[0163] The calculation method of the comprehensive evaluation score can be customized, and any calculation method that refers to the first probability information and the second probability information during the calculation process can be used.

[0164] According to the first probability information and the second probability information, the target recommended information is determined in the first recommended information set, or the first recommended information set is screened according to the first probability information, the second probability information and the pre-configured recommendation conditions of each first recommended information to obtain the target recommended information.

[0165] The recommendation conditions of each first recommendation information are pre-configured. For example, the recommendation condition of the first recommendation information 1 is: the first probability information is greater than or equal to the specified threshold value p1, and the second probability information is greater than or equal to the specified threshold value p2. The recommendation condition of the first recommendation information 2 is: the first probability information or the second probability information is greater than the specified threshold value p3. The recommendation condition of the first recommendation information 3 is: the average value of the first probability information and the second probability information is greater than the specified threshold value p4, and so on.

[0166] After determining the target recommendation information, the target recommendation information can be pushed to the first user, and then the first user's interaction data for the target recommendation information can be collected. The interaction data can be regarded as feedback information of the first user on the information recommendation process, and can be used to adjust the weight parameters of the selection model.

[0167] For example, when the first user clicks on the target recommendation information, the weight parameter corresponding to the target recommendation information in the weight parameter information of the selection model may be triggered to automatically increase to a preset value.

[0168] Taking the recommendation of 3 different versions of advertisements to a first user set consisting of multiple first users as an example, the working principle of the selection model is illustrated: at the beginning, the selection model does not know which advertisement is popular, and it is basically a blind guess. In the weight parameter information of the selection model, the initial value of the weight parameter corresponding to each advertisement can be the same or a custom configuration. After the selection model is put into the information recommendation method provided in the embodiment of the present application, each time after the target recommendation information is recommended to a first user in the first user set, the behavior data of the first user can be collected. If the first user turns on the advertisement, the value of the corresponding weight parameter is increased a little. If the first user turns off the advertisement, the value of the corresponding weight parameter is reduced a little or remains unchanged. In this way, each time an advertisement is recommended to the first user, it is equivalent to a trial. After accumulating multiple trials, the weight parameters corresponding to each advertisement are very different from the initial value. The weight parameters of some advertisements have higher values, and the weight parameters of some advertisements have lower values. Furthermore, the selection of the selection model will be more and more in line with the popular trend, that is, the more popular an advertisement is, the more people will recommend it to.

[0169] In the case where there are multiple first users, the interaction data of any one of the multiple first users for the target recommendation information may trigger the parameter update of the selection model. Therefore, the second probability information determined by the selection model can reflect the group preference of the first users, and the aforementioned first probability information can reflect the individual behavior prediction results of the first users. Therefore, the first probability information and the second probability information can provide references for determining the target recommendation information in the first recommendation information set in two different dimensions. Compared with only referring to the individual behavior prediction results, the recommendation accuracy is relatively higher, and compared with only using the selection model, the recommendation accuracy is relatively higher.

[0170] Taking the A / B test scenario of information recommendation as an example, in this scenario, operators often need to spend a lot of time and energy to observe which recommended information has better user feedback, but this process is both time-consuming and labor-intensive, and overly dependent on the subjective judgment of operators. When the feedback of multiple recommended information is similar, the recommended information determined by the operator to be the most effective may not be accurate. If the information recommendation method provided by the embodiment of the present application is used in the A / B test process, the model parameters of the selection model can be continuously adjusted when executing the information recommendation process. Compared with manually observing user feedback, it saves manpower and material resources, and is conducive to better distinguishing the first recommended information with closer user feedback, thereby improving the accuracy of information recommendation.

[0171] In a specific implementation, the first probability information is generated in the process of inputting the user characteristics of the first user and the information characteristics of the first recommended information into the first classification model for classification processing; the information recommendation method also includes: pushing the first information recommendation set to multiple third users on average; collecting the access behavior data of the third user for the first recommended information; when the data volume of the access behavior data is higher than the second threshold, constructing the input data of the second classification model according to the user characteristics of the third user, the information characteristics of the first recommended information and the access behavior data, and inputting the input data into the second classification model for classification processing to obtain the second classification result; inputting the user characteristics of the first user and the information characteristics of the first recommended information into the second classification model for classification processing to obtain the first classification result; the second classification model is obtained by inputting the training samples constructed based on historical data into the first classification model for iterative training; generating the training loss of the first classification model according to the first classification result and the second classification result; the training loss is used to adjust the model parameters of the first classification model.

[0172] The number of the plurality of third users may be predetermined.

[0173] The first information recommendation set is evenly distributed and pushed to the multiple third users.

[0174] For example, the first information recommendation set includes 5 first recommendation information, and the number of third users is 100. Each first recommendation information is distributed to 20 third users, and each third user receives only one first recommendation information.

[0175] The third user may be a user having similar user characteristics to the first user, and may refer to the corresponding description part of the first user.

[0176] Average distribution means that the number of first users assigned the same first recommendation information is determined by the number of third users and the number of first recommendation information in the first information recommendation set, and the number of first users assigned the same first recommendation information is a fixed value. However, the specific first user to which each first recommendation information is assigned is randomly determined.

[0177] The access behavior data of the third user to the first recommended information may be behavior data indicating whether the third user accesses the first recommended information after the first recommended information is pushed to the third user, for example, the third user clicks on the first recommended information.

[0178] In this implementation, the parameter update of the first classification model is not performed continuously, but is performed once every certain period of time. The condition for triggering the parameter update may be that the access behavior data of the third user for the first recommendation information accumulates to a preset data volume threshold, i.e., the second threshold.

[0179] The second classification model may be a classification model whose parameters are to be trained. The first classification model may be obtained by inputting training samples constructed from historical data into the second classification model for iterative training.

[0180] Unbiased networks and biased networks are relative concepts: A biased network is a network that allows or introduces a certain degree of systematic bias during the design and implementation process. A biased network can be understood as a network with systematic bias. In contrast, an unbiased network is a network that minimizes or eliminates systematic bias during the design and implementation process. The goal of an unbiased network is to ensure that the output results or decisions are as close to the true value or optimal solution as possible without being affected by any systematic bias. An unbiased network can be understood as a network without systematic bias.

[0181] In the aforementioned step S204, the first classification model can be used in the process of determining the first probability information. The first classification model can be obtained by inputting the training samples constructed by historical data into the second classification model for iterative training. Considering that the recommendation information in the historical data and the first recommendation information when executing step S204 may be similar in style but different in specific content, the first classification model can be regarded as a biased network, that is, a network with systematic deviations. In the training process of the second classification model, the first recommendation information set is used, and in the execution process of the information recommendation method provided in the embodiment of the present application, "the first recommendation information set is pushed to multiple third users on average" can be executed synchronously, so that in the subsequent process of constructing the input data of the second classification model and inputting the input data into the second classification model for classification processing, the systematic deviation is low, and it can be approximately regarded as a network without systematic deviation, that is, an unbiased network.

[0182] The network parameters of the unbiased network may be closer to the real expression. However, considering that the number of samples of the unbiased network is much smaller than that of the biased network compared with the model training in the model preparation stage of the first classification model, in the process of determining the target recommendation information, the biased network with higher reliability is actually used. The unbiased network is used to constrain the biased network, which improves the classification accuracy of the biased network.

[0183] Using the unbiased network to constrain the biased network can be understood as: the purpose of using the comprehensive training loss generated by the biased and unbiased networks is to adjust the training loss of the biased network, that is, to use the output of the unbiased network to constrain the training loss of the biased network.

[0184] The input data of the second classification model is constructed based on the user characteristics of the third user, the information characteristics of the first recommendation information, and the access behavior data. The user characteristics of the third user and the information characteristics of the first recommendation information are used to construct the training sample, and the access behavior data is used to construct the label information of the training sample. The input data is input into the second classification model for classification processing to obtain a second classification result.

[0185] The user characteristics of the third user and the information characteristics of the first recommended information are input into the second classification model for classification processing to obtain the second classification result. The corresponding description part of the aforementioned step S204 can be referred to. The difference is that the data required by the aforementioned step S204 is the second probability information generated in the process of generating the first classification result, while the data required in this implementation method is the second classification result.

[0186] The training loss of the first classification model is generated according to the first classification result and the second classification result. The first loss value is generated according to the first classification result, the second loss value is generated according to the second classification result, and the first loss value and the second loss value are weightedly summed to obtain the training loss of the first classification model.

[0187] Alternatively, the training loss of the first classification model can also be obtained in the following manner: obtaining user representation information reflecting user characteristics generated in the process of generating the first classification result, which can be referred to as first representation information for the sake of distinction; obtaining user representation information reflecting user characteristics generated in the process of generating the second classification result, which can be referred to as second representation information for the sake of distinction; generating the training loss of the first classification model based on the first representation information and the second representation information.

[0188] User representation information refers to the features or representation information extracted from the user's historical behavior, features, and context through various methods and models for recommendation systems. These representation information can capture the user's interests, preferences, behavior patterns, etc.

[0189] The unbiased network can be trained using a data set of an unbiased random experiment to obtain the first representation information without systematic bias, while the biased user network is trained using the full data set to obtain the second representation information with systematic bias. The information tower can be used to train to obtain the feature representation information of the first recommendation information. When the sample size reaches a certain scale, the network parameters obtained through unbiased sample learning are closer to the true expression, and it is used as a benchmark to constrain the biased samples. Exemplarily, Mse-Loss is used to constrain the user expressions of the two networks so that they are as close as possible to achieve the effect of correcting the bias. In some implementations, the unbiased first representation information can be frozen when training Mse-Loss. The final loss function is defined as follows, where is the coefficient for adjusting the two parts of the Loss: (1) in, is the second representation information, is the first representation information, It is the feature representation information of the first recommended information. Represents the sample characteristics and whether the user clicks in an unbiased random experiment. Tuning parameters for custom configurations. The function is an activation function. Used to indicate parameter freeze.

[0190] In such Figure 2In the illustrated embodiment, first, user characteristics of the first user are obtained, and information characteristics of the first recommended information in the first recommended information set are obtained; then, based on the user characteristics of the first user and the information characteristics of the first recommended information, first probability information that the first user accesses the first recommended information is determined; then, based on weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommended information, second probability information that the selection model selects the first recommended information is determined; finally, based on the first probability information and the second probability information, target recommended information is determined in the first recommended information set, and the target recommended information is pushed to the first user. In this way, in the process of determining the target recommended information to be pushed to the first user among the various first recommended information included in the first recommended information set, on the one hand, the user characteristics and information characteristics can be used as a reference to predict the possibility of the first user accessing each first recommended information, that is, to determine the first probability information; on the other hand, the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommended information can be used to determine the possibility of the selection model selecting each first recommended information, that is, to determine the second probability information; and then, the first probability information and the second probability information are used together to determine the target recommended information in the first recommended information set and push it to the first user. By combining the probability information from two different sources for information recommendation, it is beneficial to improve the prediction accuracy, so that the target recommended information pushed to the first user accurately matches the user's needs, interests or preferences, that is, the accuracy of information recommendation is improved.

[0191] The following takes the application of an information recommendation method provided by this embodiment in an A / B test scenario as an example. Figure 4-Figure 6 The information recommendation method provided by this embodiment is further described. Figure 4 A schematic diagram of a model accuracy stage in an information recommendation method provided in an embodiment of the present application; Figure 5 A schematic diagram of an information recommendation stage in an information recommendation method provided in an embodiment of the present application; Figure 6 A schematic diagram of a model correction principle provided in an embodiment of the present application.

[0192] like Figure 4 As shown, the historical similar activity data 402 can be used to train the point prediction model 404. The point prediction model 404 can be a classification model, and the specific reference can be made to the classification model mentioned in the corresponding description of the aforementioned step S204 and the aforementioned first classification model. User features 4022 and information features 4024 can be extracted from the historical similar activity data 402, and positive and negative samples 4026 can also be constructed based on the historical similar activity data 402.

[0193] The statistics module 406 includes an activation function 4064 and a selection model 4066 .

[0194] During the model preparation phase, the statistics module 406 may use the first interaction score 4062 generated by the point prediction model 404 to configure the weight parameters of the selection model 4066 .

[0195] Specifically, in the statistical module 406, the first interaction score 4062 generated by the point prediction model 404 can be normalized by the activation function 4064, and then the normalized first interaction score 4062 is selected by the selection model 4066 to obtain a first sub-result. The first sub-result can be used to configure the weight parameter of the selection model 4066. The first sub-result can refer to the aforementioned second probability information.

[0196] The personalization module 408 includes an activation function 4084 .

[0197] During the model preparation phase, the personalization module 408 may use the first interaction score 4082 generated by the point prediction model 404 for normalization.

[0198] It should be emphasized that in Figure 4 In the figure, the statistical module 406 includes a first interaction score 4062, and the personalization module 408 includes a first interaction score 4082. The two are essentially the same data and are both generated by the point prediction model 404. It can be understood that after the point prediction model 404 generates a piece of data, the data is copied into two copies, one for the statistical module 406 and the other for the personalization module 408. Therefore, in order to facilitate distinction, the data used for the statistical module 406 is named the first interaction score 4062, and the data used for the personalization module 408 is named the first interaction score 4082.

[0199] Specifically, in the personalization module 408, the second interaction score 4082 generated by the point prediction model 404 may be normalized by the activation function 4084 to obtain a second sub-result. The second sub-result may refer to the aforementioned first probability information.

[0200] The first sub-result of the statistics module 406 and the second sub-result of the personalization module 408 may be used together to determine the target recommendation information 410 .

[0201] In addition, the statistical module 406 may also be subjected to offline simulation processing through the historical recommendation information set 412 .

[0202] like Figure 5 As shown, unbiased random experiment accumulation data 504 can be obtained by using the current activity unbiased random experiment 502. The unbiased random experiment accumulation data 504 can be used to generate the first recommended information click rate 5048, and can also be used to generate user features 5042, information features 5044, and positive and negative samples 5046.

[0203] The first recommended information click rate 5048 can be used in the statistical module 506. Figure 4 Similarly, please refer to the corresponding description section.

[0204] User features 5042, information features 5044, and positive and negative samples 5046 can be used in the personalization module 508. Figure 4 Similarly, please refer to the corresponding description section.

[0205] In addition, it needs to be emphasized that Figure 4 and Figure 5 The difference is: in the model preparation stage, both the statistical module and the personalized module need to use the same point prediction model. In the offline evaluation and online adjustment stages, only the personalized module needs to use the point prediction model, and the statistical module does not need to use the point prediction model. This is because, in the process of selecting the model, only the initial value of the configuration weight parameter is related to the point prediction model, and the subsequent multiple rounds of selection no longer use the point prediction model.

[0206] like Figure 6 As shown, the biased user network 602 generates first representation information 614 through the multi-layer perceptron 608. The unbiased user network 604 generates second representation information 616 through the multi-layer perceptron 610.

[0207] The aforementioned first classification model may include a biased user network 602 and a recommendation information network 606 .

[0208] The aforementioned second classification model may include an unbiased user network 604 and a recommendation information network 606 .

[0209] The first representation information 614 and the second representation information 616 may constitute a constraint loss.

[0210] The recommendation information network 606 generates feature representation information 618 of the first recommendation information through the multi-layer perceptron 612 .

[0211] Through the activation function 620, a biased estimate 624 can be generated based on the first representation information 614 and the feature representation information 618 of the first recommended information; through the activation function 622, an unbiased estimate 626 can be generated based on the second representation information 616 and the feature representation information 618 of the first recommended information.

[0212] A comprehensive training loss can be generated using the real clicks 628 , the biased estimate 624 , and the unbiased estimate 626 , which can be used to adjust the model parameters of the biased user network 602 .

[0213] Since the technical concept is the same, the description in this embodiment is relatively simple, and the relevant parts may refer to the corresponding description of the method embodiment provided above.

[0214] An example of an information recommendation device provided in this specification is as follows: In the above-mentioned embodiment, an information recommendation method is provided. Based on the same technical concept, the embodiment of the present application also provides an information recommendation device, which is described below in conjunction with the accompanying drawings.

[0215] Figure 7 This is a schematic diagram of an information recommendation device provided in an embodiment of the present application. This embodiment provides an information recommendation device 700, including: An acquiring unit 702 is configured to acquire a user feature of a first user and an information feature of a first recommended information in a first recommended information set; A determining unit 704 is configured to determine first probability information of the first user accessing the first recommended information according to a user feature of the first user and an information feature of the first recommended information; The determining unit 704 is further configured to determine second probability information of the selection model selecting the first recommendation information according to weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommendation information; The pushing unit 706 is configured to determine target recommended information in the first recommendation information set according to the first probability information and the second probability information, and push the target recommended information to the first user.

[0216] Optionally, the information recommendation device 700 further includes a configuration unit; The acquisition unit 702 is further configured to acquire user characteristics of a plurality of second users, and to acquire information characteristics of second recommendation information in a second recommendation information set, wherein the second users include users to which recommendation information has been pushed, and the second recommendation information includes recommendation information that has been pushed to users; The determining unit 704 is further configured to determine third probability information of the second user accessing the second recommended information according to a user feature of the second user and an information feature of the second recommended information; The determining unit 704 is further configured to determine average probability information of the second user accessing the second recommended information according to the third probability information; The configuration unit is used to configure the weight parameter information of the selection model according to the average probability information.

[0217] Optionally, when determining the second probability information that the selection model selects the first recommendation information according to the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommendation information, the determination unit 704 performs the following steps: When the selection model is in the first state, selecting a plurality of first weight parameters whose values ​​are higher than a first threshold value from the weight parameter information; According to the multiple first weight parameters and the corresponding relationship, multiple first recommended information are deleted from the first information recommendation set to obtain a third recommended information set, wherein the number of the deleted first recommended information is the same as the number of the multiple first weight parameters; Determining, in the third recommendation information set, first recommendation information selected by the selection model; The second probability information of the first recommendation information selected by the selection model is configured according to the first preset value, and the second probability information of the first recommendation information not selected in the third recommendation set is configured according to the second preset value.

[0218] Optionally, the first probability information is generated in the process of inputting the user characteristics of the first user and the information characteristics of the first recommendation information into a first classification model for classification processing to obtain a first classification result; the information recommendation device further includes: a collection unit, a classification unit and a training unit; The pushing unit is further used to push the first information recommendation set to multiple third users on average; The collecting unit is used to collect the access behavior data of the third user with respect to the first recommendation information; The classification unit is used to construct input data of a second classification model according to the user characteristics of the third user, the information characteristics of the first recommendation information and the access behavior data when the data volume of the access behavior data is higher than a second threshold, input the input data into the second classification model for classification processing, and obtain a second classification result; The classification unit is further used to input the user characteristics of the first user and the information characteristics of the first recommendation information into the second classification model for classification processing to obtain the first classification result; the second classification model is obtained by inputting the training samples constructed based on historical data into the first classification model for iterative training; A training unit is used to generate a training loss of the first classification model according to the first classification result and the second classification result; the training loss is used to adjust the model parameters of the first classification model.

[0219] Optionally, the information recommendation device 700 further includes: a selection unit and a judgment unit; The acquisition unit 702 is further used to acquire historical access behavior data for the historical recommendation information set, and construct a test sample according to the historical access behavior data; The selection unit is used to input the test sample into the selection model for selection processing to obtain a selection result of the test sample; The judging unit is used to determine whether the loss of the selected model converges according to the selection result; If yes, it is determined that the selection model passes the test; If not, it is determined that the selection model fails the test, and the preset parameters of the selection model are adjusted.

[0220] Optionally, the target state is the second state; the second interaction prediction result includes a second interaction score; when the determination unit 706 determines the second interaction prediction result of the first recommendation information being selected by the selection model under the target state according to the weight parameter assigned by the selection model to the first recommendation information, the determination unit 706 performs the following steps: A second interaction score of the first recommendation information is configured according to the weight parameter of the first recommendation information.

[0221] In an embodiment of the present application, an acquisition unit is used to acquire user characteristics of a first user, and to acquire information characteristics of the first recommended information in a first recommended information set; a determination unit is used to determine first probability information that the first user accesses the first recommended information based on the user characteristics of the first user and the information characteristics of the first recommended information; the determination unit is further used to determine second probability information that the selection model selects the first recommended information based on weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommended information; and a push unit is used to determine target recommended information in the first recommended information set based on the first probability information and the second probability information, and to push the target recommended information to the first user. In this way, in the process of determining the target recommended information to be pushed to the first user among the various first recommended information included in the first recommended information set, on the one hand, the user characteristics and information characteristics can be used as a reference to predict the possibility of the first user accessing each first recommended information, that is, to determine the first probability information; on the other hand, the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommended information can be used to determine the possibility of the selection model selecting each first recommended information, that is, to determine the second probability information; and then, the first probability information and the second probability information are used together to determine the target recommended information in the first recommended information set and push it to the first user. By combining the probability information from two different sources for information recommendation, it is beneficial to improve the prediction accuracy, so that the target recommended information pushed to the first user accurately matches the user's needs, interests or preferences, that is, the accuracy of information recommendation is improved.

[0222] Corresponding to the information recommendation method described above, based on the same technical concept, an embodiment of the present application further provides an electronic device, which is used to execute the information recommendation method provided above. Figure 8 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

[0223] like Figure 8 As shown, the electronic device may have relatively large differences due to different configurations or performances, and may include one or more processors 801 and memory 802, and the memory 802 may store one or more storage applications or data. Among them, the memory 802 may be a temporary storage or a permanent storage. The application stored in the memory 802 may include one or more modules (not shown in the figure), and each module may include a series of computer executable instructions in the electronic device. Furthermore, the processor 801 may be configured to communicate with the memory 802 to execute a series of computer executable instructions in the memory 802 on the electronic device. The electronic device may also include one or more power supplies 803, one or more wired or wireless network interfaces 804, one or more input / output interfaces 805, one or more keyboards 806, etc.

[0224] In a specific embodiment, the electronic device includes a memory and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer executable instructions for the electronic device, and the one or more programs configured to be executed by one or more processors include the following computer executable instructions: Acquire user characteristics of the first user, and acquire information characteristics of the first recommended information in the first recommended information set; determining first probability information of the first user accessing the first recommended information according to a user feature of the first user and an information feature of the first recommended information; Determining second probability information that the selection model selects the first recommendation information according to weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommendation information; According to the first probability information and the second probability information, target recommended information is determined in the first recommended information set, and the target recommended information is pushed to the first user.

[0225] The computer-readable storage medium provided in this embodiment is used to store computer-executable instructions. When the computer-executable instructions are executed by a processor, the following process is implemented: Acquire user characteristics of the first user, and acquire information characteristics of the first recommended information in the first recommended information set; determining first probability information of the first user accessing the first recommended information according to a user feature of the first user and an information feature of the first recommended information; Determining second probability information that the selection model selects the first recommendation information according to weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommendation information; According to the first probability information and the second probability information, target recommended information is determined in the first recommended information set, and the target recommended information is pushed to the first user.

[0226] It should be noted that the embodiment of the computer-readable storage medium in this specification and the embodiment of the information recommendation method in this specification are based on the same inventive concept, so the specific implementation of this embodiment can refer to the implementation of the aforementioned corresponding method, and the repeated parts will not be repeated.

[0227] Another embodiment of the present disclosure further provides a computer program product, the computer program product comprising a computer program, and when the computer program is executed by a processor, the following process is implemented: Acquire user characteristics of the first user, and acquire information characteristics of the first recommended information in the first recommended information set; determining first probability information of the first user accessing the first recommended information according to a user feature of the first user and an information feature of the first recommended information; Determining second probability information that the selection model selects the first recommendation information according to weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommendation information; According to the first probability information and the second probability information, target recommended information is determined in the first recommended information set, and the target recommended information is pushed to the first user.

[0228] The computer program product in the embodiments of the present disclosure can implement the various processes of the above-mentioned information recommendation method embodiments and achieve the same effects and functions, which will not be repeated here.

[0229] In an embodiment of the present application, first, user characteristics of the first user are obtained, and information characteristics of the first recommended information in the first recommended information set are obtained; then, based on the user characteristics of the first user and the information characteristics of the first recommended information, first probability information that the first user accesses the first recommended information is determined; then, based on weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommended information, second probability information that the selection model selects the first recommended information is determined; finally, based on the first probability information and the second probability information, target recommended information is determined in the first recommended information set, and the target recommended information is pushed to the first user. In this way, in the process of determining the target recommended information to be pushed to the first user among the various first recommended information included in the first recommended information set, on the one hand, the user characteristics and information characteristics can be used as a reference to predict the possibility of the first user accessing each first recommended information, that is, to determine the first probability information; on the other hand, the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommended information can be used to determine the possibility of the selection model selecting each first recommended information, that is, to determine the second probability information; and then, the first probability information and the second probability information are used together to determine the target recommended information in the first recommended information set and push it to the first user. By combining the probability information from two different sources for information recommendation, it is beneficial to improve the prediction accuracy, so that the target recommended information pushed to the first user accurately matches the user's needs, interests or preferences, that is, the accuracy of information recommendation is improved.

[0230] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0231] It should be understood by those skilled in the art that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, the embodiments of the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, this specification may take the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0232] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable electronic device to produce a machine, so that the instructions executed by the processor of the computer or other programmable electronic device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0233] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable electronic device to work in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0234] These computer program instructions may also be loaded onto a computer or other programmable electronic device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0235] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0236] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0237] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined in this article, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0238] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0239] Embodiments of the present application may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. One or more embodiments of the present specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communications network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0240] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0241] The above description is only an embodiment of this document and is not intended to limit this document. For those skilled in the art, this document may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this document should be included in the scope of the claims of this document.

Claims

1. An information recommendation method, characterized in that: include: Acquire user characteristics of the first user, and acquire information characteristics of the first recommended information in the first recommended information set; determining first probability information of the first user accessing the first recommended information according to a user feature of the first user and an information feature of the first recommended information; Determining second probability information that the selection model selects the first recommendation information according to weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommendation information; According to the first probability information and the second probability information, target recommended information is determined in the first recommended information set, and the target recommended information is pushed to the first user.

2. The method according to claim 1, characterized in that: The method further comprises: Acquire user characteristics of a plurality of second users, and acquire information characteristics of second recommended information in a second recommended information set, wherein the second users include users to which recommended information has been pushed, and the second recommended information includes recommended information that has been pushed to users; determining third probability information that the second user accesses the second recommended information according to the user characteristics of the second user and the information characteristics of the second recommended information; Determining average probability information of the second user accessing the second recommended information according to the third probability information; The weight parameter information of the selection model is configured according to the average probability information.

3. The method according to claim 1, characterized in that The determining, based on the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommendation information, second probability information that the selection model selects the first recommendation information includes: When the selection model is in the first state, selecting a plurality of first weight parameters whose values ​​are higher than a first threshold value from the weight parameter information; According to the multiple first weight parameters and the corresponding relationship, multiple first recommended information are deleted from the first information recommendation set to obtain a third recommended information set, wherein the number of the deleted first recommended information is the same as the number of the multiple first weight parameters; Determining, in the third recommendation information set, first recommendation information selected by the selection model; The second probability information of the first recommendation information selected by the selection model is configured according to the first preset value, and the second probability information of the first recommendation information not selected in the third recommendation set is configured according to the second preset value.

4. The method according to claim 1, characterized in that The first probability information is generated in the process of inputting the user characteristics of the first user and the information characteristics of the first recommendation information into a first classification model for classification processing to obtain a first classification result; The method further comprises: Pushing the first information recommendation set to multiple third users on average; Collecting access behavior data of the third user with respect to the first recommendation information; When the amount of the access behavior data is higher than a second threshold, constructing input data of a second classification model according to the user characteristics of the third user, the information characteristics of the first recommendation information, and the access behavior data, inputting the input data into the second classification model for classification processing, and obtaining a second classification result; Inputting the user features of the first user and the information features of the first recommendation information into the second classification model for classification processing to obtain the first classification result; the second classification model is obtained by inputting the training samples constructed based on historical data into the first classification model for iterative training; A training loss of the first classification model is generated according to the first classification result and the second classification result; the training loss is used to adjust model parameters of the first classification model.

5. The method according to claim 1, characterized in that The method further comprises: Acquire historical access behavior data for a historical recommendation information set, and construct a test sample based on the historical access behavior data; Inputting the test sample into the selection model for selection processing to obtain a selection result of the test sample; Determining whether the loss of the selected model converges according to the selection result; If yes, it is determined that the selection model passes the test; If not, it is determined that the selection model fails the test, and the preset parameters of the selection model are adjusted.

6. The method according to claim 1, characterized in that The determining, based on the weight parameter information of the selection model and the pre-established correspondence between the weight parameter and the recommendation information, second probability information that the selection model selects the first recommendation information includes: When the selection model is in the second state, selecting a second weight parameter from the weight parameter information; determining, according to the second weight parameter and the correspondence between the weight parameter and the recommendation information, the first recommendation information corresponding to the second weight parameter; According to the second weight parameter, second probability information of the first recommendation information corresponding to the second weight parameter is configured.

7. An information recommendation device, characterized in that: include: an acquisition unit, configured to acquire user characteristics of the first user, and acquire information characteristics of the first recommended information in the first recommended information set; a determining unit, configured to determine first probability information of the first user accessing the first recommended information according to a user feature of the first user and an information feature of the first recommended information; The determining unit is further configured to determine second probability information that the selection model selects the first recommendation information according to weight parameter information of the selection model and a pre-established correspondence between the weight parameter and the recommendation information; A pushing unit is used to determine target recommended information in the first recommendation information set according to the first probability information and the second probability information, and push the target recommended information to the first user.

8. An electronic device, characterized in that: include: processor; And, a memory configured to store computer executable instructions, which, when executed, cause the processor to perform the information recommendation method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store computer-executable instructions, and the computer-executable instructions, when executed by a processor, implement the information recommendation method according to any one of claims 1 to 6.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the information recommendation method according to any one of claims 1 to 6.

Images