Method and device for determining information pushing opportunity and electronic equipment
By analyzing the historical behavioral characteristics and predicted activity of target users, personalized information push strategies and timings are determined, which solves the problem of treating information push strategies equally in the existing technology and improves user response efficiency and experience.
Patent Information
- Application Number
- CN202411374829.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-29
- Publication Date
- 2025-01-24
AI Technical Summary
The prior art implements the same information push strategy for all users, resulting in low response efficiency for users to the information push content and poor user experience.
By obtaining the historical behavior characteristics of target users, enter to the loss classification model and activity prediction model, determine the target information push strategy and timing, combine information push content characteristics and environmental characteristics, optimize the timing of information push.
The response efficiency and user experience of the user's information push content have been improved, and the optimal distribution of resources has been achieved.
Smart Images
Figure CN119357465A_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the field of artificial intelligence technology, and specifically relates to a method, device and electronic device for determining the timing of information push. Background Art
[0002] The current recommendation system generally includes many links such as user login, user browsing, and user purchase. In addition to end-to-end observation of overall revenue and transaction status, each link is optimized separately, which plays an important role in improving user experience and cultivating user mindset. When performing multi-task optimization, the current recommendation system treats all users equally, which will lead to the optimization seesaw phenomenon. For example: allocating a limited budget to improve the click-through rate in the browsing stage will increase the click-through rate while causing a decrease in the success rate in the purchase stage; or, investing a limited budget in the purchase stage will make it difficult to maintain the overall click-through rate in the browsing stage. Different user groups have different preferences for different information push timings. Implementing the same information push strategy for all users cannot achieve the global optimal allocation of resources, resulting in low user response efficiency to information push content and poor user experience. Summary of the invention
[0003] The purpose of the embodiments of the present application is to provide a method, device and electronic device for determining the timing of information push, which can solve the problem that the prior art implements the same information push strategy for all users, resulting in low efficiency in users' response to information push content.
[0004] In a first aspect, an embodiment of the present application provides a method for determining an information push timing, including:
[0005] Obtain the historical behavior characteristics of the target user within a preset time period;
[0006] Inputting the historical behavior characteristics into a churn classification model to obtain a churn classification result;
[0007] Inputting the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain an activity prediction result;
[0008] Determining a target information push strategy from the N information push strategies according to the historical behavior characteristics, the activity prediction result, and N information push strategies, where N is an integer greater than or equal to 2;
[0009] According to the target information push strategy, a target information push timing corresponding to the target user is determined.
[0010] Optionally, determining a target information push strategy from the N information push strategies according to the historical behavior characteristics, the activity prediction result, and the N information push strategies includes:
[0011] Obtain information push content characteristics and current environment characteristics;
[0012] According to the information push content features, the current environment features, the historical behavior features, the activity prediction results and the i-th information push strategy, calculate the user preference value corresponding to the i-th information push strategy to obtain N user preference values corresponding to the N information push strategies, where i takes the value of each integer from 2 to N in turn;
[0013] The information push strategy corresponding to the maximum value among the N user preference degree values is determined as the target information push strategy.
[0014] Optionally, the calculating the user preference value corresponding to the ith information push strategy according to the information push content feature, the current environment feature, the historical behavior feature, the activity prediction result and the ith information push strategy includes:
[0015] Obtaining characteristic values of each information push opportunity in the i-th information push strategy;
[0016] The characteristic values of each information push opportunity in the i-th information push strategy, the historical behavior characteristics, the activity prediction results, the information push content characteristics, and the current environment characteristics are input into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy.
[0017] Optionally, the step of inputting the characteristic values of each information push opportunity in the i-th information push strategy, the historical behavior characteristics, the activity prediction result, the information push content characteristics, and the current environment characteristics into a preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy includes:
[0018] Superimposing the historical behavior feature, the activity prediction result, and the information push content feature to obtain a first superposition result;
[0019] Superimposing the characteristic values of each information push opportunity in the i-th information push strategy, the first superposition result, and the current environment characteristics to obtain a second superposition result corresponding to the i-th information push strategy;
[0020] The second superposition result corresponding to the i-th information push strategy is input into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy.
[0021] Optionally, inputting the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain an activity prediction result includes:
[0022] Inputting the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain a target probability value of the target user generating a target behavior;
[0023] The activity prediction result is determined according to the target probability value.
[0024] Optionally, determining the activity prediction result according to the target probability value includes:
[0025] When the target probability value is less than a preset probability value, determining the activity prediction result to be the target probability value;
[0026] When the target probability value is greater than or equal to the preset probability value, the average duration of the target user's target behavior within the preset time period, the target mean and the target standard deviation in the target activity prediction model are obtained, and the activity prediction result is calculated based on the average duration, the target mean, the target standard deviation, and the target probability value.
[0027] Optionally, the target activity prediction model is obtained in the following manner:
[0028] Get the probability density function of the lognormal distribution;
[0029] Performing negative logarithm calculation on the probability density function to obtain a loss function of a log-normal distribution;
[0030] Obtain a target loss function according to the binary cross entropy loss function of the Bernoulli distribution and the loss function of the lognormal distribution;
[0031] The activity prediction model is trained based on the target loss function to obtain a target activity prediction model.
[0032] In a second aspect, an embodiment of the present application provides a device for determining an information push timing, including:
[0033] The first acquisition module is used to acquire the historical behavior characteristics of the target user within a preset time period;
[0034] A first processing module, used for inputting the historical behavior features into a churn classification model to obtain a churn classification result;
[0035] A second processing module, configured to input the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain an activity prediction result;
[0036] A first determination module, configured to determine a target information push strategy from the N information push strategies according to the historical behavior characteristics, the activity prediction result, and the N information push strategies, where N is an integer greater than or equal to 2;
[0037] The second determining module is used to determine the target information push timing corresponding to the target user according to the target information push strategy.
[0038] In a third aspect, an embodiment of the present application further provides an electronic device, including a memory, a transceiver, and a processor:
[0039] A memory for storing a computer program; a transceiver for sending and receiving data under the control of a processor; and a processor for reading the computer program in the memory and executing the method as described in the first aspect above.
[0040] In a fourth aspect, an embodiment of the present application further provides a processor-readable storage medium, wherein the processor-readable storage medium stores a computer program, and the computer program is used to enable the processor to execute the method described in the first aspect above.
[0041] In an embodiment of the present application, historical behavior characteristics of a target user within a preset time period are obtained, and the historical behavior characteristics are input into a churn classification model to obtain a churn classification result, and the historical behavior characteristics are input into a target activity prediction model corresponding to the churn classification result to obtain an activity prediction result. According to the historical behavior characteristics of the target user, different target activity prediction models corresponding to different churn classification results are proposed for the zero inflation problem, that is, non-churn users and churn users use different target activity prediction models to obtain activity prediction results, which can solve the problem of poor fitting of the head and tail extreme values and improve the area under ROC (Area Under Curve, AUC) of the target activity prediction model, wherein ROC refers to the receiver operating characteristic curve (receiver operating characteristic curve, ROC). Furthermore, based on the historical behavior characteristics, the activity prediction results and N information push strategies, a target information push strategy is determined from the N information push strategies. The target information push strategy is used to determine the target information push timing corresponding to the target user. Each information push strategy is combined with the historical behavior characteristics and the activity prediction results to identify the target user's preference for each information push strategy, and then the target information push strategy suitable for the target user is determined. The target information push timing of the target user can be determined through the target information push strategy, thereby improving the user's response efficiency to the information push content and the user experience. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 is a bell-shaped curve diagram of user value changing over time provided by an embodiment of the present application;
[0043] Figure 2 is a flow chart of a method for determining information push timing provided by an embodiment of the present application;
[0044] Figure 3 This is a life cycle data distribution diagram of a target user provided by an embodiment of the present application;
[0045] Figure 4 It is a flow chart of the process of determining the target information push strategy provided by the embodiment of the present application;
[0046] Figure 5 is a curve relationship diagram of different means and standard deviations provided in the embodiments of the present application;
[0047] Figure 6 It is a structural schematic diagram of a device for determining information push timing provided in an embodiment of the present application;
[0048] Figure 7 It is a structural block diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0049] The following will be combined with the drawings in the embodiments of the present application to clearly describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application belong to the scope of protection of this application.
[0050] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way are interchangeable where appropriate, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of one type, and the number of objects is not limited. For example, the first object can be one or more. In addition, "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.
[0051] Figure 1 It is a bell-shaped curve of user value changing over time. The horizontal axis is time and the vertical axis is activity. User value is evaluated by activity, and the user life cycle is divided into five stages: introduction, growth, maturity, dormancy, and churn through changes in slope.
[0052] There are significant differences in preferences among users at different stages. For example: (1) For users in the introduction stage, the order volume and browsing duration are relatively low. (2) For users in the growth stage, the order volume and browsing duration increase significantly, with the increase in browsing duration significantly greater than the increase in order volume. From a behavioral perspective, users at this stage tend to browse and click around, but their conversion rate to purchase still needs to be improved. (3) For users in the mature stage, the probability of order completion is the highest, and their browsing duration decreases slightly, indicating that users at this stage have a strong and clear desire to buy. (4) For users in the dormant stage and the churn stage, the average browsing duration of the former is much greater than that of the latter.
[0053] An important part of the information push strategy is to choose the right time to push information and implement the information push strategy. Taking online shopping as an example, the information push timing includes: (1) login; (2) browsing; (3) purchase; (4) logout, etc. The user response effect is different when information is pushed at different information push timings. Generally speaking, it is suitable to implement the information push strategy on the login interface for users in the introduction stage, because users at this stage need to use marketing methods to improve user stickiness. For mature users, it is more suitable to choose the "purchase" stage as the best contact time, which can be targeted and improve the shopping experience of mature users.
[0054] Therefore, the present application provides a method, device and electronic device for determining the timing of information push, which combines each information push strategy with historical behavior characteristics and activity prediction results, identifies the target user's preference for each information push strategy, and then determines the target information push timing of the target information push strategy suitable for the target user, thereby improving the user's response efficiency to the information push content and user experience.
[0055] The following describes the method for determining the timing of information push provided by the embodiment of the present application through specific embodiments and application scenarios in conjunction with the accompanying drawings.
[0056] Specifically, Figure 2 As shown, the embodiment of the present application provides a method for determining the timing of information push, which may specifically include the following steps:
[0057] Step 201, obtaining historical behavior characteristics of a target user within a preset time period.
[0058] Specifically, if the target user is a new user, since the new user has no historical behavior characteristics, the historical behavior characteristics of the target user are set to zero. If the target user is an old user, the historical behavior characteristics of the target user within a preset time period (such as 30 days) are collected. For example, the historical behavior characteristics include the number of logins, the number of clicks, the amount of consumption, etc., which are recorded as a vector T1 = {t1, t2, t3...}, where T1 represents the historical behavior characteristics of the target user, t1 represents the number of logins, t2 represents the number of clicks, and t3 represents the amount of consumption.
[0059] Step 202: input the historical behavior features into a churn classification model to obtain a churn classification result.
[0060] Specifically, in the target user's life cycle data distribution, a variety of distributions such as Poisson and exponential are mixed, such as Figure 3 As shown in the figure, the data distribution has the following three main characteristics: (1) zero-inflated Poisson distribution; there are many users who have lost for more than 30 days; (2) there are relatively few mature users, but this small number of people contributes to the vast majority of logins, activity, and browsing; (3) the overall activity of the majority of the remaining people conforms to the exponential distribution. The user life cycle prediction problem is generally modeled using regression, and the loss function uses the mean squared error loss function (MSE). However, the characteristics of MSE cannot accurately handle the zero-inflated problem and are sensitive to extreme values. As a result, the commonly used modeling methods cannot solve the prediction problem of mature users and long-term lost users. The prediction results of the model will be biased by the extreme distributions at both ends, resulting in inaccurate distribution of the predicted data.
[0061] Therefore, the user life cycle distribution is considered to be two parts, one is an extremely inactive group consisting of long-term churn users, and the other is a normal active group close to the exponential distribution. The historical behavior characteristics of the target user are input into the churn classification model to obtain the churn prediction result P1. If the churn prediction result P1 is greater than the preset probability value Thres1, it means that the target user is a churn user; conversely, if the churn prediction result P1 is less than or equal to the preset probability value Thres1, it means that the target user is a non-churn user. The churn classification result can be used to determine whether the target user is a churn user or a non-churn user.
[0062] It should be noted that the churn classification model can adopt the DeepFM model structure, using the FM structure to extract low-order features and the Deep Neural Networks (DNN) structure to extract high-order features. DeepFM is a recommendation system model that combines Factorization Machines (FM) and Deep Neural Networks.
[0063] Step 203: input the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain an activity prediction result.
[0064] Specifically, different churn classification results correspond to different target activity prediction models. Non-churn users correspond to target activity prediction model B, and churn users correspond to target activity prediction model C. If the churn classification result is a non-churn user, the historical behavior characteristics of the target user are input into the target activity prediction model B to obtain the activity prediction result P corresponding to the non-churn user. active , activity prediction result P active It is used to indicate the activity level of the target user. If the churn classification result is a churned user, the historical behavior characteristics of the target user are input into the target activity prediction model C to obtain the activity prediction result P corresponding to the churned user. active That is, the churn classification model acts as a gating network to route the users to be detected to the appropriate target activity prediction model. In this way, the target activity prediction model can focus only on specific types of users, solve the problem of poor fitting of the head and tail extreme values, and improve the AUC of the target activity prediction model.
[0065] Step 204: Determine a target information push strategy from the N information push strategies according to the historical behavior characteristics, the activity prediction result, and the N information push strategies, where N is an integer greater than or equal to 2.
[0066] Specifically, based on the historical behavior characteristics, activity prediction results and N information push strategies, a target information push strategy suitable for the target user is determined from the N information push strategies, and information is pushed to the target user through the target information push strategy. That is, each information push strategy is combined with the historical behavior characteristics and activity prediction results to identify the target user's preference for each information push strategy, and then a target information push strategy suitable for the target user is determined. Information is pushed to the target user through the target information push strategy, which can improve the user's response efficiency to information push content and user experience.
[0067] Step 205: Determine the target information push timing corresponding to the target user according to the target information push strategy.
[0068] Specifically, the target information push strategy includes a target information push timing, and the target information push strategy is used to determine the target information push timing for pushing information to the target user.
[0069] In an embodiment of the present application, the historical behavior characteristics of the target user within a preset time period are obtained, the historical behavior characteristics are input into the churn classification model to obtain the churn classification result, the historical behavior characteristics are input into the target activity prediction model corresponding to the churn classification result to obtain the activity prediction result, and according to the historical behavior characteristics of the target user, different target activity prediction models corresponding to different churn classification results are proposed for the zero inflation problem, that is, non-churn users and churn users use different target activity prediction models to obtain activity prediction results, which can solve the problem of poor fitting of the head and tail extreme values and improve the AUC of the target activity prediction model. In addition, according to the historical behavior characteristics, the activity prediction result and N information push strategies, a target information push strategy is determined from the N information push strategies, each information push strategy is combined with the historical behavior characteristics and the activity prediction result, the preference of the target user for each information push strategy is identified, and then the target information push strategy suitable for the target user is determined, and the target information push timing of the target user can be determined through the target information push strategy to improve the user's response efficiency and user experience to the information push content.
[0070] As an optional specific embodiment, the step 204 determines a target information push strategy from the N information push strategies according to the historical behavior characteristics, the activity prediction result, and the N information push strategies, and specifically includes steps 2041 to 2043:
[0071] Step 2041, obtaining information push content characteristics and current environment characteristics.
[0072] Specifically, the current environment features {M1, M2, M3...} are obtained, where M1, M2, ... represent a variety of different environment features, for example, M1 represents the weather, M2 represents the temperature, M3 represents whether it is a promotion day, etc. Information push content features T3 = {S1, S2, ...}, where S1, S2, ... represent a variety of different marketing activities, for example, S1 represents the amount of marketing coupons, and S2 represents whether to send text messages, etc.
[0073] Step 2042, based on the information push content characteristics, the current environment characteristics, the historical behavior characteristics, the activity prediction results and the i-th information push strategy, calculate the user preference value corresponding to the i-th information push strategy to obtain N user preference values corresponding to the N information push strategies, where i takes the value of each integer from 2 to N in turn.
[0074] Specifically, for the i-th information push strategy among the N information push strategies: based on the information push content features, current environment features, historical behavior features, activity prediction results, and the i-th information push strategy, the user preference value corresponding to the i-th information push strategy is calculated. The user preference value corresponding to each of the N information push strategies is calculated by the above-mentioned calculation method of the user preference value corresponding to the i-th information push strategy, thereby obtaining N user preference values corresponding to the N information push strategies.
[0075] The user preference value corresponding to each information push strategy represents the success rate of the target user responding to the information push content when the information push operation is performed under the information push strategy.
[0076] Step 2043: Determine the information push strategy corresponding to the maximum value among the N user preference values as the target information push strategy.
[0077] Specifically, N user preference values are sorted in descending order, and the maximum value among the N user preference values is selected as the best information push strategy for the target user, that is, the target information push strategy. Information is pushed to the target user through the target information push strategy, which can improve the user's response efficiency to the information push content and user experience.
[0078] As an optional specific embodiment, the step 2042 calculates the user preference value corresponding to the i-th information push strategy according to the information push content characteristics, the current environment characteristics, the historical behavior characteristics, the activity prediction result and the i-th information push strategy, specifically including:
[0079] Obtaining characteristic values of each information push opportunity in the i-th information push strategy;
[0080] The characteristic values of each information push opportunity in the i-th information push strategy, the historical behavior characteristics, the activity prediction results, the information push content characteristics, and the current environment characteristics are input into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy.
[0081] Specifically, the i-th information push strategy includes multiple information push opportunities, each information push opportunity indicates that information is pushed at the opportunity, and the characteristic value of each information push opportunity in the i-th information push strategy is obtained. If information is pushed at an information push opportunity, the characteristic value of the information push opportunity is 1, and conversely, if information is not pushed at an information push opportunity, the characteristic value of the information push opportunity is 0. The characteristic values, historical behavior characteristics, activity prediction results, information push content characteristics, and current environment characteristics of each information push opportunity in the i-th information push strategy are input into the preference degree prediction model, and the user preference degree value corresponding to the i-th information push strategy is calculated through the preference program prediction model. The user preference degree value corresponding to each information push strategy in the N information push strategies is calculated by the calculation method of the user preference degree value corresponding to the i-th information push strategy, thereby obtaining the N user preference degree values corresponding to the N information push strategies.
[0082] Example: Information push timing is divided into four categories: (A) login; (B) browsing; (C) purchase; (D) logout. The login timing is represented by C1, the browsing timing is represented by C2, the purchase timing is represented by C3, and the logout timing is represented by C4. If the information push action is taken at the login timing (such as: publishing marketing activity content), then C1 = 1, otherwise C1 = 0. If the information push action is taken at the browsing timing, then C2 = 1, otherwise C2 = 0. If the information push action is taken at the purchase timing, then C3 = 1, otherwise C3 = 0. If the information push action is taken at the logout timing, then C4 = 1, otherwise C4 = 0.
[0083] Furthermore, the step of inputting the characteristic values of each information push opportunity in the i-th information push strategy, the historical behavior characteristics, the activity prediction result, the information push content characteristics, and the current environment characteristics into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy specifically includes:
[0084] Superimposing the historical behavior feature, the activity prediction result, and the information push content feature to obtain a first superposition result;
[0085] Superimposing the characteristic values of each information push opportunity in the i-th information push strategy, the first superposition result, and the current environment characteristics to obtain a second superposition result corresponding to the i-th information push strategy;
[0086] The second superposition result corresponding to the i-th information push strategy is input into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy.
[0087] Specifically, based on the characteristic values of each information push opportunity in the i-th information push strategy and the current environmental characteristics, construct T4={C1, C2, C3…, M1, M2, M3…}, where {C1, C2, C3…} represents the characteristic values of each information push opportunity, and {M1, M2, M3…} represents the current environmental characteristics.
[0088] The historical behavior characteristics are superimposed with the activity prediction results to obtain T2, and T2 is superimposed with the information push content characteristics T3 to obtain the first superposition result. The first superposition result is superimposed with T4 of the i-th information push strategy to obtain the second superposition result corresponding to the i-th information push strategy. The second superposition result is input into the preference prediction model, and the user preference value corresponding to the i-th information push strategy is calculated by the preference prediction model.
[0089] It should be noted that the characteristic value of the first information push opportunity among multiple information push opportunities is set to 1, and the characteristic values of other information push opportunities are set to 0, that is, T4 = {1, 0, 0…, M1, M2, M3…}, thereby obtaining an information push strategy, which indicates that information push is performed at the first information push opportunity. The characteristic value of the second information push opportunity among multiple information push opportunities is set to 1, and the characteristic values of other information push opportunities are set to 0, that is, T4 = {0, 1, 0…, M1, M2, M3…}, thereby obtaining another information push strategy, which indicates that information push is performed at the second information push opportunity. The characteristic value of the third information push opportunity among multiple information push opportunities is set to 1, and the characteristic values of other information push opportunities are set to 0, that is, T4 = {0, 0, 1…, M1, M2, M3…}, thereby obtaining another information push strategy, which indicates that information push is performed at the third information push opportunity. By analogy, N different information push strategies are obtained, and the i-th information push strategy is any one of them.
[0090] Among them, the preference degree prediction model can adopt a multi-gate mixture-of-experts (MMoE) structure.
[0091] like Figure 4 As shown, the process of determining the target information push strategy for the target user is described below through a specific embodiment:
[0092] The value of N is 4, and the four information push strategies are the first information push strategy (i.e., information push at login time), the second information push strategy (i.e., information push at browsing time), the third information push strategy (i.e., information push at purchase time), and the fourth information push strategy (i.e., information push at logout time). The historical behavior characteristics of the target user are obtained, and the historical behavior characteristics are input into the churn classification model to obtain the churn classification result. It is determined whether the churn classification result is a churn user. If it is a churn user, the historical behavior characteristics are input into the target activity prediction model C to obtain the activity prediction result; if it is a non-churn user, the historical behavior characteristics are input into the target activity prediction model B to obtain the activity prediction result. The activity prediction result is superimposed with the historical behavior characteristics to obtain T2, and T2 is superimposed with the information push content feature T3 to obtain the first superposition result. The first superposition result is superimposed with T4={1,0,0,0,M1,M2,M3...} corresponding to the first information push strategy to obtain a second superposition result; the first superposition result is superimposed with T4={0,1,0,0,M1,M2,M3...} corresponding to the second information push strategy to obtain another second superposition result; the first superposition result is superimposed with T4={0,0,1,0,M1,M2,M3...} corresponding to the third information push strategy to obtain another second superposition result; the first superposition result is superimposed with T4={0,0,0,1,M1,M2,M3...} corresponding to the fourth information push strategy to obtain another second superposition result, thereby obtaining four second superposition results corresponding to the four information push strategies. The four second superposition results are input into the preference degree prediction model to obtain four user preference degree values R={R1, R2, R3, R4}. Among them, R1 represents the user preference value corresponding to the first information push strategy, R2 represents the user preference value corresponding to the second information push strategy, R3 represents the user preference value corresponding to the third information push strategy, and R4 represents the user preference value corresponding to the fourth information push strategy. If the maximum value among R1, R2, R3, and R4 is R2, the target information push strategy is the second information push strategy, that is, information push is performed at the browsing time.
[0093] The following is a specific description of the method for obtaining the target activity prediction model in step 203:
[0094] Get the probability density function of the lognormal distribution;
[0095] Performing negative logarithm calculation on the probability density function to obtain a loss function of a log-normal distribution;
[0096] Obtain a target loss function according to the binary cross entropy loss function of the Bernoulli distribution and the loss function of the lognormal distribution;
[0097] The activity prediction model is trained based on the target loss function to obtain a target activity prediction model.
[0098] Furthermore, the objective loss function is specifically obtained by the following formula:
[0099] L ZILN (x; μ, σ) = -Sign {x＝0} log(1-p)-Sign {x＞0} (logp-L lognormal (x; μ, σ)
[0100]
[0101] Among them, x represents the duration of the sample user's target behavior;
[0102] μ and σ represent the mean and standard deviation of the log-normal distribution corresponding to the duration of the target behavior of the sample users;
[0103] L ZILN (x; μ, σ) represents the target loss function;
[0104] -Sign {x＝0} log(1-p)-Sign {x＞0} (logp) represents the binary cross entropy loss function of Bernoulli distribution;
[0105] L lognormal (x; μ, σ) represents the loss function of the log-normal distribution;
[0106] p represents the probability value that the sample user will generate the target behavior.
[0107] Specifically, first obtain the probability density function of the lognormal distribution:
[0108]
[0109] Among them, x represents the duration of the sample user's target behavior;
[0110] μ and σ represent the mean and standard deviation of the log-normal distribution corresponding to the duration of the sample users' target behavior (e.g., browsing behavior); Figure 5 As shown, S1 is the curve when μ is 0 and σ is 0.25, S2 is the curve when μ is 0 and σ is 0.5, and S3 is the curve when μ is 0 and σ is 1;
[0111] f(x; μ, σ) represents the probability density function corresponding to x.
[0112] By performing negative logarithm calculation on the above probability density function, we can obtain the loss function of log-normal distribution:
[0113]
[0114] It should be noted that in the zero-inflated + log-normal distribution ZILN distribution, in addition to considering the non-zero log-normal distribution (i.e., the duration distribution of the target behavior of the sample user), it is also necessary to consider an additional Bernoulli distribution (i.e., whether the sample user will generate the target behavior); therefore, the target loss function of the activity prediction model includes the binary cross entropy loss function of the Bernoulli distribution and the loss function of the log-normal distribution:
[0115] L ZILN (x; μ, σ) = -Sign {x＝0} log(1-p)-Sign {x>0} (logp-L lognormal (x; μ, σ)
[0116] Compared with the conventional MSE loss function, the objective function of the activity prediction model does not become abnormally large when the estimated value is abnormally large. This avoids the problem of MSE loss becoming invalid when the zero-inflation condition occurs.
[0117] The activity prediction model is trained through the target loss function to optimize the mean μ and standard deviation σ in the activity prediction model. When the activity prediction model converges, the optimized target mean μ′ and target standard deviation σ′ are obtained, and the trained target activity prediction model is obtained.
[0118] It should be noted that before training the target activity prediction model, the historical behavior characteristics of the sample users are obtained and input into the churn classification model, and the label of the sample users is "whether active"; the input of activity prediction model B and activity prediction model C is the historical behavior characteristics of the sample users, and the label of the sample users is the corresponding activity prediction results; the input of the preference prediction model is the historical behavior characteristics of the sample users, the corresponding information push content characteristics, the corresponding environmental characteristics, the corresponding activity prediction results, and the characteristic values of each information push timing in the corresponding information push strategy. The label of the user sample is 1 (positive sample) or 0 (negative sample), and the basis for judging positive and negative samples is whether the actual activity results are improved after the information push strategy is implemented.
[0119] The following is a specific example to illustrate how to determine positive samples and negative samples:
[0120] Assume that a sample user is currently in the growth stage, and his activity (i.e., the actual activity result) is 3 views per day. Through marketing activities, a 5-yuan marketing activity voucher is issued in the second information push strategy, i.e., the "browsing" link, and the sample user's activity is increased to 5 views per day, and he is converted into a mature user. Since the user successfully transformed from the "growth stage" to the "mature stage", the activity increased from 3 to 5; therefore, the user conversion result is 1, indicating a positive sample. As shown in Table 1 below, comparing the changes in activity before and after information push, when the sample user's activity decreases under a certain information push strategy, it is a negative sample, and when it increases, it is a positive sample.
[0121] Table 1
[0122]
[0123] For example: a sample user logged in for a total of 5 days in the past 30 days, and the browsing time each day was 8 minutes, 9 minutes, 10 minutes, 11 minutes, and 12 minutes respectively, and the historical average browsing time was 10 minutes; after this login, after information was pushed through the first information push strategy, the browsing time was 5 minutes, then the browsing time in the past 30 days was updated to: 9 minutes, 10 minutes, 11 minutes, 12 minutes, 5 minutes, and the historical average browsing time was updated to 9.4 minutes; by comparison, it was found that the activity level decreased, so the response efficiency of the sample user for information push under this first information push strategy is a negative sample.
[0124] As an optional specific embodiment, the step 203 inputs the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain an activity prediction result, specifically including:
[0125] Inputting the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain a target probability value of the target user generating a target behavior;
[0126] The activity prediction result is determined according to the target probability value.
[0127] Specifically, the historical behavior features are input into the target activity prediction model corresponding to the churn classification results to obtain the target probability value p that the target user will perform the target behavior (such as browsing behavior). The target probability value p can be used to determine whether the target user will perform the target behavior and the possibility of performing the target behavior, thereby determining the activity prediction result of the target user.
[0128] Furthermore, the step of determining the activity prediction result according to the target probability value specifically includes:
[0129] When the target probability value is less than a preset probability value, determining the activity prediction result to be the target probability value;
[0130] When the target probability value is greater than or equal to the preset probability value, the average duration of the target user's target behavior within the preset time period, the target mean and the target standard deviation in the target activity prediction model are obtained, and the activity prediction result is calculated based on the average duration, the target mean, the target standard deviation, and the target probability value.
[0131] Specifically, the historical behavior features are input into the target activity prediction model corresponding to the churn classification result to obtain the target probability value p that the target user will generate the target behavior (such as browsing behavior). p is compared with the preset probability value (such as 0.5). If p is less than the preset probability value, it means that the target user will not generate the target behavior. The activity prediction result P active =p. If p is greater than or equal to the preset probability value, it means that the target user will generate the target behavior, and the estimated duration of the target behavior is obtained by the target mean μ′ and target standard deviation σ′ in the target activity prediction model. Therefore, the activity prediction result is:
[0132]
[0133] Among them, x′ represents the average duration of the target user's target behavior within the preset time period;
[0134] μ′ and σ′ represent the target mean and target standard deviation in the target activity prediction model;
[0135] p′ represents the target probability value.
[0136] In summary, the above-mentioned embodiment of the present application proposes different target activity prediction models corresponding to different churn classification results for the zero-inflation problem based on the historical behavior characteristics of the target user, that is, non-churn users and churn users use different target activity prediction models to obtain activity prediction results, which can solve the problem of poor fitting of the head and tail extreme values and improve the AUC of the target activity prediction model. In addition, each information push strategy is combined with the historical behavior characteristics and activity prediction results to identify the target user's preference for each information push strategy, and then determine the information push strategy with the highest preference as the target information push strategy for the target user. The target information push timing corresponding to the target user can be determined through the target information push strategy, and information is pushed to the target user at the target information push timing, thereby improving the user's response efficiency to the information push content and user experience.
[0137] The method for determining information push timing provided in the embodiment of the present application can be executed by a device for determining information push timing. In the embodiment of the present application, the device for determining information push timing is taken as an example to illustrate the method for determining information push timing provided in the embodiment of the present application.
[0138] like Figure 6 As shown, the embodiment of the present application further provides a device 600 for determining an information push timing, which specifically includes:
[0139] The first acquisition module 601 is used to acquire the historical behavior characteristics of the target user within a preset time period;
[0140] The first processing module 602 is used to input the historical behavior characteristics into the churn classification model to obtain a churn classification result;
[0141] The second processing module 603 is used to input the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain an activity prediction result;
[0142] A first determination module 604 is configured to determine a target information push strategy from the N information push strategies according to the historical behavior characteristics, the activity prediction result, and the N information push strategies, where N is an integer greater than or equal to 2;
[0143] The first determining module 605 is used to determine the target information push timing corresponding to the target user according to the target information push strategy.
[0144] Optionally, the first determining module 604 is specifically configured to:
[0145] Obtain information push content characteristics and current environment characteristics;
[0146] According to the information push content features, the current environment features, the historical behavior features, the activity prediction results and the i-th information push strategy, calculate the user preference value corresponding to the i-th information push strategy to obtain N user preference values corresponding to the N information push strategies, where i takes the value of each integer from 2 to N in turn;
[0147] The information push strategy corresponding to the maximum value among the N user preference degree values is determined as the target information push strategy.
[0148] Optionally, when the determination module 604 calculates the user preference value corresponding to the i-th information push strategy according to the information push content characteristics, the current environment characteristics, the historical behavior characteristics, the activity prediction result and the i-th information push strategy, it is specifically used to:
[0149] Obtaining characteristic values of each information push opportunity in the i-th information push strategy;
[0150] The characteristic values of each information push opportunity in the i-th information push strategy, the historical behavior characteristics, the activity prediction results, the information push content characteristics, and the current environment characteristics are input into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy.
[0151] Optionally, when the determining module 604 inputs the characteristic values of each information push opportunity in the i-th information push strategy, the historical behavior characteristics, the activity prediction result, the information push content characteristics, and the current environment characteristics into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy, it is specifically used to:
[0152] Superimposing the historical behavior feature, the activity prediction result, and the information push content feature to obtain a first superposition result;
[0153] Superimposing the characteristic values of each information push opportunity in the i-th information push strategy, the first superposition result, and the current environment characteristics to obtain a second superposition result corresponding to the i-th information push strategy;
[0154] The second superposition result corresponding to the i-th information push strategy is input into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy.
[0155] Optionally, the second processing module 603 is specifically configured to:
[0156] Inputting the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain a target probability value of the target user generating a target behavior;
[0157] The activity prediction result is determined according to the target probability value.
[0158] Optionally, when determining the activity prediction result according to the target probability value, the second processing module 603 is specifically configured to:
[0159] When the target probability value is less than a preset probability value, determining the activity prediction result to be the target probability value;
[0160] When the target probability value is greater than or equal to the preset probability value, the average duration of the target user's target behavior within the preset time period, the target mean and the target standard deviation in the target activity prediction model are obtained, and the activity prediction result is calculated based on the average duration, the target mean, the target standard deviation, and the target probability value.
[0161] Optionally, the target activity prediction model is obtained in the following manner:
[0162] Get the probability density function of the lognormal distribution;
[0163] Performing negative logarithm calculation on the probability density function to obtain a loss function of a log-normal distribution;
[0164] Obtain a target loss function according to the binary cross entropy loss function of the Bernoulli distribution and the loss function of the lognormal distribution;
[0165] The activity prediction model is trained based on the target loss function to obtain a target activity prediction model.
[0166] In summary, the above-mentioned embodiment of the present application proposes different target activity prediction models corresponding to different churn classification results for the zero-inflation problem based on the historical behavior characteristics of the target user, that is, non-churn users and churn users use different target activity prediction models to obtain activity prediction results, which can solve the problem of poor fitting of the head and tail extreme values and improve the AUC of the target activity prediction model. In addition, each information push strategy is combined with the historical behavior characteristics and activity prediction results to identify the target user's preference for each information push strategy, and then determine the information push strategy with the highest preference as the target information push strategy for the target user. The target information push timing corresponding to the target user can be determined through the target information push strategy, and information is pushed to the target user at the target information push timing, thereby improving the user's response efficiency to the information push content and user experience.
[0167] It should be noted here that the device for determining the timing of information push provided in the embodiment of the present application can implement all the method steps implemented in the embodiment of the method for determining the timing of information push, and can achieve the same technical effect. The parts and beneficial effects of this embodiment that are the same as those of the method embodiment will not be described in detail here.
[0168] It should be noted that the division of units in the embodiments of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional unit in each embodiment of the present application may be integrated into a processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0169] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store program code.
[0170] like Figure 7 As shown, an embodiment of the present application further provides an electronic device, including a memory 720, a transceiver 710, and a processor 700:
[0171] A memory 720, for storing computer programs;
[0172] A transceiver 710, configured to transmit and receive data under the control of the processor;
[0173] The processor 700 is configured to read the computer program in the memory and execute the steps of the method for determining the information push timing as described in any of the above embodiments.
[0174] Among them, Figure 7In the embodiment, the bus architecture may include any number of interconnected buses and bridges, specifically linking together various circuits of one or more processors represented by processor 700 and memory represented by memory 720. The bus architecture may also link together various other circuits such as peripherals, voltage regulators, and power management circuits, which are well known in the art and are therefore not further described herein. The bus interface provides an interface. The transceiver 710 may be a plurality of components, namely, a transmitter and a receiver, providing a unit for communicating with various other devices on a transmission medium, which may include a wireless channel, a wired channel, an optical cable, and other transmission media. The processor 700 is responsible for managing the bus architecture and general processing, and the memory 720 may store data used by the processor 700 when performing operations.
[0175] The processor 700 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or a complex programmable logic device (CPLD). The processor may also adopt a multi-core architecture.
[0176] The processor calls the computer program stored in the memory to execute the method for determining the information push timing provided in the embodiment of the present application according to the obtained executable instructions. The processor and the memory can also be arranged physically separately.
[0177] It should be noted here that the above-mentioned electronic device provided in the embodiment of the present application can implement all the method steps implemented in the above-mentioned method embodiment for determining the timing of information push, and can achieve the same technical effect. The parts and beneficial effects of this embodiment that are the same as the method embodiment will not be described in detail here.
[0178] An embodiment of the present application further provides a processor-readable storage medium, wherein the processor-readable storage medium stores a computer program, and the computer program is used to enable the processor to execute the above-mentioned method for determining the timing of information push.
[0179] The processor-readable storage medium can be any available medium or data storage device that can be accessed by the processor, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MO), etc.), optical storage (such as CD, DVD, BD, HVD, etc.), and semiconductor storage (such as ROM, EPROM, EEPROM, non-volatile memory (NANDFLASH), solid-state drive (SSD)), etc.
[0180] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) that contain computer-usable program code.
[0181] The present application 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 the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer executable instructions. These computer executable instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing 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.
[0182] These processor executable instructions may also be stored in a processor readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the processor 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.
[0183] These processor-executable instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational 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.
[0184] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.
Claims
1. A method for determining information push timing, characterized in that: include: Obtain the historical behavior characteristics of the target user within a preset time period; Inputting the historical behavior characteristics into a churn classification model to obtain a churn classification result; Inputting the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain an activity prediction result; Determining a target information push strategy from the N information push strategies according to the historical behavior characteristics, the activity prediction result, and N information push strategies, where N is an integer greater than or equal to 2; According to the target information push strategy, a target information push timing corresponding to the target user is determined.
2. The method according to claim 1, characterized in that The determining a target information push strategy from the N information push strategies according to the historical behavior characteristics, the activity prediction result, and the N information push strategies includes: Obtain information push content characteristics and current environment characteristics; According to the information push content features, the current environment features, the historical behavior features, the activity prediction results and the i-th information push strategy, calculate the user preference value corresponding to the i-th information push strategy to obtain N user preference values corresponding to the N information push strategies, where i takes the value of each integer from 2 to N in turn; The information push strategy corresponding to the maximum value among the N user preference degree values is determined as the target information push strategy.
3. The method according to claim 2, characterized in that The calculating the user preference value corresponding to the i-th information push strategy according to the information push content feature, the current environment feature, the historical behavior feature, the activity prediction result and the i-th information push strategy includes: Obtaining characteristic values of each information push opportunity in the i-th information push strategy; The characteristic values of each information push opportunity in the i-th information push strategy, the historical behavior characteristics, the activity prediction results, the information push content characteristics, and the current environment characteristics are input into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy.
4. The method according to claim 3, characterized in that The step of inputting the characteristic values of each information push opportunity in the i-th information push strategy, the historical behavior characteristics, the activity prediction result, the information push content characteristics, and the current environment characteristics into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy includes: Superimposing the historical behavior feature, the activity prediction result, and the information push content feature to obtain a first superposition result; Superimposing the characteristic values of each information push opportunity in the i-th information push strategy, the first superposition result, and the current environment characteristics to obtain a second superposition result corresponding to the i-th information push strategy; The second superposition result corresponding to the i-th information push strategy is input into the preference degree prediction model to obtain the user preference degree value corresponding to the i-th information push strategy.
5. The method according to claim 1, characterized in that The step of inputting the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain an activity prediction result includes: Inputting the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain a target probability value of the target user generating a target behavior; The activity prediction result is determined according to the target probability value.
6. The method according to claim 5, characterized in that The determining the activity prediction result according to the target probability value includes: When the target probability value is less than a preset probability value, determining the activity prediction result to be the target probability value; When the target probability value is greater than or equal to the preset probability value, the average duration of the target user's target behavior within the preset time period, the target mean and the target standard deviation in the target activity prediction model are obtained, and the activity prediction result is calculated based on the average duration, the target mean, the target standard deviation, and the target probability value.
7. The method according to claim 1, characterized in that The target activity prediction model is specifically obtained in the following way: Get the probability density function of the lognormal distribution; Performing negative logarithm calculation on the probability density function to obtain a loss function of a log-normal distribution; Obtain a target loss function according to the binary cross entropy loss function of the Bernoulli distribution and the loss function of the lognormal distribution; The activity prediction model is trained based on the target loss function to obtain a target activity prediction model.
8. A device for determining information push timing, characterized in that: include: The first acquisition module is used to acquire the historical behavior characteristics of the target user within a preset time period; A first processing module, used for inputting the historical behavior features into a churn classification model to obtain a churn classification result; A second processing module, configured to input the historical behavior features into a target activity prediction model corresponding to the churn classification result to obtain an activity prediction result; A first determination module, configured to determine a target information push strategy from the N information push strategies according to the historical behavior characteristics, the activity prediction result, and the N information push strategies, where N is an integer greater than or equal to 2; The second determining module is used to determine the target information push timing corresponding to the target user according to the target information push strategy.
9. An electronic device, characterized in that: Including memory, transceiver, processor: Memory for storing computer programs; a transceiver, for transmitting and receiving data under the control of the processor; A processor, configured to read the computer program in the memory and execute the method for determining the information push timing as described in any one of claims 1 to 7.
10. A processor-readable storage medium, characterized in that: The processor-readable storage medium stores a computer program, and the computer program is used to enable the processor to execute the method for determining the information push timing according to any one of claims 1 to 7.