Online marketing method and device based on deep network model, and terminal equipment

Abstract

The invention is suitable for the technical field of intelligent marketing, and provides an online marketing method and device based on a deep network model, and terminal equipment, and the method comprises the steps: obtaining user features of users in a user group; inputting the user features and the specific intervention features into a pre-trained target deep network model, obtaining prediction response results of the user under different specific intervention features through the deep network model, and determining a promotion value of the user according to the prediction response results of the user under different intervention features, the promotion value is used for evaluating response changes of the user after the user accepts the different intervention features; and determining a target user in the user group based on the promotion value, and performing online marketing on the target user. By adopting the method and the device, the corresponding crowd of the user can be accurately identified, and the effectiveness and the accuracy of marketing recommendation are improved.

Description

Technical Field

[0001] The present application relates to the field of intelligent marketing technology, and in particular to an online marketing method, apparatus and terminal device based on a deep network model. Background Art

[0002] Marketing recommendations have been applied to e-commerce, precision medicine, etc. Marketing recommendations aim to accurately identify the corresponding population of each user (including intervention-sensitive population and intervention-insensitive population) based on the user's past behavioral characteristics, so as to decide whether to implement intervention measures (such as red envelopes, coupons, etc.) to obtain the best conversion benefits (such as maximizing clicks or purchases).

[0003] In view of this, how to accurately identify the user's corresponding group and improve the effectiveness and accuracy of marketing recommendations is a problem that needs to be considered currently. Summary of the invention

[0004] The embodiments of the present application provide an online marketing method, apparatus, and terminal device based on a deep network model, which can accurately identify the user's corresponding group and improve the effectiveness and accuracy of marketing recommendations.

[0005] In a first aspect, the embodiments of the present application provide an online marketing method based on a deep network model, comprising:

[0006] Obtain user characteristics of users in the user group;

[0007] Inputting the user characteristics and specific intervention characteristics into a pre-trained target deep network model, obtaining predicted response results of the user under different specific intervention characteristics through the deep network model, and determining the user's lift value according to the predicted response results of the user under different intervention characteristics, wherein the lift value is used to evaluate the response change of the user after receiving the different intervention characteristics;

[0008] Target users in the user group are determined based on the lift value, and online marketing is performed to the target users.

[0009] In a possible implementation manner of the first aspect, the step of obtaining predicted response results of the user under different specific intervention features includes:

[0010] Dividing the user group according to the user characteristics and the specific intervention characteristics to obtain a plurality of user subgroups corresponding to the user group;

[0011] Determining a target grouping of users in each of the user subgroups;

[0012] The user is predicted based on the specific intervention feature and the target group to obtain predicted response results of the user under different specific intervention features.

[0013] In a possible implementation manner of the first aspect, before the step of inputting the user features and the specific intervention features into the pre-trained target deep network model, the method further includes:

[0014] Obtaining an original sample set, wherein the original sample set includes sample features, intervention feature labels, and response feature labels of sample users;

[0015] Constructing a deep network model, wherein the deep network model includes a group division module, an automatic selector, and a main network module;

[0016] The original sample set is input into the group division module, and the group division module divides the sample users in the original sample set according to the sample features, the intervention feature labels and the response feature labels to obtain several sample subsets:

[0017] Inputting the sample subsets into corresponding automatic selectors respectively, and determining sample groups of sample users in the sample subsets through the automatic selectors;

[0018] Predicting the sample user based on the intervention feature label and the sample grouping through the main network module, obtaining the response results of the sample user under different intervention feature labels, and determining the sample lift value of the sample user according to the response results of the sample user under different intervention feature labels;

[0019] Inputting the sample user intervention feature label, the response feature label and the sample improvement value into a preset evaluation mechanism for evaluation to obtain the improvement effect of the deep network model;

[0020] The model parameters of the deep network model are adjusted based on the preset loss function of the main network module, and iterations are repeated. When the improvement effect does not improve continuously within a preset number of times, a trained target deep network model is obtained.

[0021] In a possible implementation manner of the first aspect, the method further includes:

[0022] Obtaining constraint conditions corresponding to the sample grouping;

[0023] The main network module predicts the sample users in the sample group based on the intervention feature labels under the constraints to obtain the response results of the sample users under different intervention feature labels; and determines the improvement value of the user under the constraints according to the predicted response results of the user under different intervention features.

[0024] In a possible implementation manner of the first aspect, before the step of inputting the user features and the specific intervention features into the pre-trained target deep network model, the method further includes:

[0025] Obtaining an original sample set, wherein the original sample set includes sample features, intervention feature labels, and response feature labels of sample users;

[0026] Constructing a deep network model, wherein the deep network model includes a group division module, an automatic selector, and a main network module;

[0027] The original sample set is input into the group division module, and the group division module divides the sample users in the original sample set according to the sample features, the intervention feature labels and the response feature labels to obtain several sample subsets:

[0028] Inputting the sample subsets into corresponding automatic selectors respectively, and determining sample groups of sample users in the sample subsets through the automatic selectors;

[0029] The sample user is predicted by the main network module based on the intervention feature label and the sample grouping, the response results of the sample user under different intervention feature labels are obtained, and the sample lift value of the sample user is determined according to the response results of the sample user under different intervention feature labels; according to the specific relationship between the sample lift values ​​in different sample groups, samples are selected from different sample groups to construct sample pairs, and implicit lift information is captured based on the sample pairs; each sample pair includes sample users from different sample groups;

[0030] Inputting the sample user intervention feature label, the response feature label and the sample improvement value into a preset evaluation mechanism for evaluation to obtain the improvement effect of the deep network model;

[0031] The model parameters of the deep network model are adjusted based on the preset loss function of the main network module, and iterations are repeated. When the improvement effect does not improve continuously within a preset number of times, a trained target deep network model is obtained.

[0032] In a possible implementation manner of the first aspect, the method further includes:

[0033] Obtaining constraint conditions corresponding to the sample grouping;

[0034] During the training process of the deep network model, the main network module predicts the sample users in the sample group based on the intervention feature labels under the constraints to obtain the response results of the sample users under different intervention feature labels; based on the predicted response results of the users under different intervention features, the improvement value of the users under the constraints is determined.

[0035] In a second aspect, an embodiment of the present application provides an online marketing device based on a deep network model, comprising:

[0036] A feature acquisition unit, used to acquire user features of users in a user group;

[0037] a lift value determination unit, configured to input the user characteristics and specific intervention characteristics into a pre-trained target deep network model, obtain predicted response results of the user under different specific intervention characteristics through the deep network model, and determine the user's lift value according to the predicted response results of the user under different intervention characteristics, wherein the lift value is used to evaluate the response change of the user after receiving the different intervention characteristics;

[0038] An online marketing unit is used to determine target users in the user group based on the boost value and perform online marketing to the target users.

[0039] In a third aspect, an embodiment of the present application provides a terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the online marketing method based on the deep network model as described in the first aspect above is implemented.

[0040] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the online marketing method based on the deep network model as described in the first aspect above is implemented.

[0041] In a fifth aspect, an embodiment of the present application provides a computer program product, which, when executed on a terminal device, enables the terminal device to execute the online marketing method based on the deep network model as described in the first aspect above.

[0042] In an embodiment of the present application, user characteristics of users in a user group are obtained, the user characteristics and specific intervention characteristics are input into a pre-trained target deep network model, and the predicted response results of the user under different specific intervention characteristics are obtained through the deep network model. Based on the predicted response results of the user under different intervention characteristics, the lift value of the user is accurately determined, and then the lift value is used to evaluate the response changes of the user after receiving the different intervention characteristics. This is conducive to segmenting the user group based on the lift value, accurately identifying the user's corresponding population, and then effectively determining the target user from the user group, and then conducting precise online marketing to the target user, which can improve the effectiveness and accuracy of marketing recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0044] Figure 1 It is a flowchart of an implementation of an online marketing method based on a deep network model provided in an embodiment of the present application;

[0045] Figure 2 It is a specific implementation flow chart of obtaining predicted response results in the online marketing method based on the deep network model provided in the embodiment of the present application;

[0046] Figure 3 It is a specific implementation flow chart of constructing and training a target deep network model in an online marketing method based on a deep network model provided in an embodiment of the present application;

[0047] Figure 3.1 It is a schematic diagram of the architecture of a deep network model in the online marketing method based on the deep network model provided in the embodiment of the present application;

[0048] Figure 4 It is another specific implementation flow chart of constructing and training a target deep network model in an online marketing method based on a deep network model provided in an embodiment of the present application;

[0049] Figure 5 It is a structural block diagram of an online marketing device based on a deep network model provided in an embodiment of the present application;

[0050] Figure 6 It is a schematic diagram of a terminal device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0051] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0052] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0053] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0054] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0055] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0056] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0057] As an example but not a limitation, the online marketing method based on the deep network model provided in the embodiment of the present application can be applied to terminal devices and servers that need to perform various types of data for marketing recommendations, and the terminal devices can specifically include vehicle-mounted terminals, mobile phones, tablet computers, laptop computers, ultra-mobile personal computers (UMPCs), desktop computers, etc. The embodiment of the present application does not impose any restrictions on the specific types of terminal devices.

[0058] Figure 1 The implementation process of the online marketing method based on the deep network model provided in the embodiment of the present application is shown, and the method flow includes steps S101 to S103. The specific implementation principle of each step is as follows:

[0059] Step S101: Obtain user characteristics of users in a user group.

[0060] In this embodiment, user characteristics may include one or more of the user's basic information, behavior characteristics, preference characteristics, social characteristics, and technical characteristics. Among them, basic information includes age, gender, geographic location, income level, etc.; behavior characteristics include historical browsing records, historical purchase records, etc.; preference characteristics include user preferences, etc.; social characteristics include information such as the user's friends, followed users, or followed users on social media; technical characteristics include the type of device used by the user, operating system, browser, network connection type, speed, etc.

[0061] As one of the input information of the target deep network model, user characteristics play a vital role in the model's prediction results and subsequent target user determination. User characteristics usually cover multiple dimensions, aiming to fully reflect the user's attributes, behaviors, preferences, etc., so as to conduct more accurate personalized marketing.

[0062] Step S102: Input the user characteristics and specific intervention characteristics into a pre-trained target deep network model, obtain the predicted response results of the user under different specific intervention characteristics through the deep network model, and determine the user's improvement value based on the predicted response results of the user under different intervention characteristics.

[0063] The improvement value is used to evaluate the change in the user's response after receiving the different intervention features.

[0064] There are counterfactual results for marketing recommendations. The results that can be observed when an individual is intervened or not intervened are called factual results, and the results that cannot be observed are called counterfactual results. Since it is impossible to observe both factual results and counterfactual results at the same time, it is impossible to determine the changes in the user's response after receiving the different intervention features. In this embodiment, a pre-trained deep network model is used to obtain the lift value of each user in the user group, and based on the lift value, the changes in the user's response after receiving the different intervention features can be effectively evaluated.

[0065] The above-mentioned specific intervention feature is a binary feature, such as 1 and 0. The specific intervention feature is 1, which means intervention, and the specific intervention feature is 0, which means no intervention. If the specific intervention feature is 1, the lift value is the difference between the predicted response result (factual result) of the model under the intervention condition and the predicted response result (counterfactual result) of the model under the non-intervention condition; if the specific intervention feature is 0, the lift value is the difference between the predicted response result (counterfactual result) of the model under the intervention condition and the predicted response result (factual result) of the model under the non-intervention condition.

[0066] As a possible implementation of this application, Figure 2 A specific implementation process of obtaining the predicted response results of the user under different specific intervention characteristics in the online marketing method based on the deep network model provided in the embodiment of the present application is shown, and is detailed as follows:

[0067] A1: The user group is divided according to the user characteristics and the specific intervention characteristics to obtain a plurality of user subgroups corresponding to the user group. The user subgroups include a non-negative boost value subgroup and a non-positive boost value subgroup.

[0068] A2: Determine the target grouping of users in each of the user subgroups. The target groups include: intervention-sensitive groups (positive promotion groups): conversion benefits can be brought in as long as intervention measures are implemented; natural conversion groups (no promotion groups): conversion benefits can be brought in regardless of whether intervention measures are implemented or not; indifferent groups (no promotion groups): conversion benefits cannot be brought in regardless of whether intervention measures are implemented or not; counter-effect groups (negative promotion groups): conversion benefits can be brought in without intervention measures.

[0069] A3: Predict the user based on the specific intervention feature and the target group, and obtain the predicted response results of the user under different specific intervention features. When the specific intervention feature is 1, obtain the predicted response results (factual results) of the model under the intervention condition and the predicted response results (counterfactual results) of the model under the non-intervention condition; when the specific intervention feature is 0, obtain the predicted response results (counterfactual results) of the model under the intervention condition and the predicted response results (factual results) of the model under the non-intervention condition.

[0070] As an embodiment of the present application, Figure 3 A specific implementation process of constructing and training a target deep network model in an online marketing method based on a deep network model provided in an embodiment of the present application is shown, and is described in detail as follows:

[0071] Step S201: Acquire an original sample set, wherein the original sample set includes sample features, intervention feature labels, and response feature labels of sample users.

[0072] Step S202: construct a deep network model, wherein the deep network model includes a group division module, an automatic selector and a main network module.

[0073] Exemplarily, in this embodiment, the structure of the deep network model is as follows Figure 3.1 As shown, there may be more than one automatic selector.

[0074] Step S203: input the original sample set into the group division module, and the group division module divides the sample users in the original sample set according to the sample characteristics, the intervention characteristic labels and the response characteristic labels to obtain a plurality of sample subsets.

[0075] In this embodiment, the sample features, intervention feature labels (binary intervention 0, 1) and response feature labels (binary response 0, 1) of the sample users in the original sample set are input into the deep network model, and the original sample set can be divided according to different combinations of possible values ​​of the intervention feature labels and possible values ​​of the response feature labels through the group division module. The division example is as follows:

[0076] g1: The intervention feature label is 1, the response feature label is 1, the sample users are intervened and have conversion benefits. The sample users may belong to the intervention-sensitive population or the natural conversion population (actively promoted or non-promoted population). Therefore, the promotion value of the sample users in the g1 group is non-negative;

[0077] g2: The intervention feature label is 0, the response feature label is 1, and the sample users are not intervened and have conversion benefits. The sample users may belong to the natural conversion population or the reaction population (negative promotion or no promotion population). Therefore, the promotion value of the sample users in the g2 group is not positive;

[0078] g3: The intervention feature label is 1, the response feature label is 0, and the sample users have no conversion benefits after the intervention. The sample users belong to the indifferent group or the reaction group (negative improvement or no improvement group). Therefore, the improvement value of the sample users in the g3 group is not positive;

[0079] g4: The intervention feature label is 0, and the response feature label is also 0. The sample users have not received any intervention and have no conversion benefits. The sample users may belong to the intervention-sensitive population or the indifferent population (actively improved or not improved population). Therefore, the improvement value of the sample users in the g4 group is non-negative.

[0080] From the above analysis, we can see that the improvement values ​​of the sample users in groups g1 and g4 are non-negative, so they are combined into sample subset g 14 , the improvement values ​​of sample users in groups g2 and g3 are non-positive, so they are merged into sample subset g 23 Therefore, the original sample user set is divided into two sample subsets g 14 and g 23 .

[0081] Step S204: inputting the sample subsets into corresponding automatic selectors respectively, and determining sample groups of sample users in the sample subsets through the automatic selectors.

[0082] In this example, two selector MLPs consisting of a MLP multilayer perceptron and an output layer are constructed. 14 and MLP 23 , the sample subset g 14 Input Selector MLP 14 , sample subset g 23 Input Selector MLP 23 To automatically learn to group sample users.

[0083] In a possible implementation, the sample subset g is calculated according to the following formula (1): 14 The sample users in are grouped and the sample subset g is calculated according to the following formula (2): 23 The sample users in are grouped into

[0084]

[0085] Among them, σ represents the sigmoid function, and t1 and t2 represent hyperparameters. Indicates that in the sample subset g 14 The probability that the i-th sample user belongs to the actively promoted group, Indicates that in the sample subset g 23 The probability that the i-th sample user belongs to the negative promotion group.

[0086] In this embodiment, according to and At the same time, the probability P of the two sample subsets belonging to the non-improvement population can be obtained W , as follows:

[0087]

[0088] in, and Zero vectors representing different batch lengths.

[0089] Step S205: The main network module predicts the sample user based on the intervention feature label and the sample grouping, obtains the response results of the sample user under different intervention feature labels, and determines the sample lift value of the sample user according to the response results of the sample user under different intervention feature labels.

[0090] In this embodiment, the main network module is used to predict the response results of sample users under different intervention feature labels.

[0091] In one possible implementation, the main network module can adopt two types of neural network structures: in the first type of neural network structure, only a response prediction model is established, and the intervention feature enters the model as a feature for model training. By changing the intervention feature label variable, the predicted fact result is obtained. and counterfactual results, and the factual results The difference between the counterfactual result and the improvement value is obtained, such as the S learner. In the second type of neural network structure, the response prediction models of intervention and non-intervention are established respectively, and the predicted factual results are obtained through different sub-networks. and counterfactual results, and use their difference to obtain the boost value τ, such as T learner. Most boost networks use the second neural network structure.

[0092] The calculation formula (4) is as follows:

[0093]

[0094] in, represents the fact result predicted by the i-th sample user, τ i represents the improvement value predicted by the i-th sample user, g i represents the i-th sample subset, f(·) represents the specific response prediction model, and θ represents the model learnable parameters.

[0095] Step S206: input the sample user intervention feature label, the response feature label and the sample improvement value into a preset evaluation mechanism for evaluation to obtain the improvement effect of the deep network model.

[0096] The above preset evaluation mechanism is used to evaluate the network performance of the deep network model based on the evaluation index and determine the improvement effect of the deep network model. The larger the value of the evaluation index, the better the ranking effect of the method improvement value and the more accurate the prediction of the target population.

[0097] Step S207: adjusting the model parameters of the deep network model based on the preset loss function of the main network module, and repeatedly iterating to evaluate the effect of the deep network model according to the evaluation index. When the effect does not improve within a preset number of consecutive times, a trained target deep network model is obtained.

[0098] The preset number of times can be set according to the needs. For example, the preset number of times is set to 5 times. When the improvement effect is not improved within 5 consecutive times, overfitting is prevented, the model training is completed, and the trained target deep improvement network model is obtained.

[0099] In this embodiment, the cross entropy in the main network module is used as the preset loss function for the prediction response result regression task in the deep network model. The specific expression of the preset loss function L is as follows:

[0100]

[0101] where y i express The corresponding true response feature label.

[0102] In a possible implementation, constraints corresponding to the sample grouping are obtained, and the main network module predicts sample users in the sample grouping based on the intervention feature labels under the constraints to obtain response results of the sample users under different intervention feature labels; based on the predicted response results of the user under different intervention features, the improvement value of the user under the constraints is determined.

[0103] The above constraints are used in the model training or prediction process, and certain constraints are imposed on the lift values ​​of sample users in different sample groups. In this embodiment, by imposing different constraints on the lift values ​​of different groups of people, the deep network model of the above output lift value is optimized.

[0104] In some implementations, the constraint is: L G : Increase the improvement value of samples belonging to the positive improvement population; L R : Reduce the lift value of samples belonging to the negative population; L W : Limit the improvement value of samples belonging to the non-improvement population to zero. Constraint L G , L R , L W The calculation formulas (6)-(8) are as follows:

[0105]

[0106] Among them, bs (batch size) represents the amount of data processed by the model at one time during training, τ is the improvement value, and P Gi It represents the probability that the i-th sample user belongs to the actively promoted group. Ri represents the probability that the i-th sample user belongs to the negative promotion group, P Wi It represents the probability that the i-th sample user belongs to the group without improvement.

[0107] In this embodiment, the above constraints are used during the training process to ensure that the deep network model is more fair, accurate and explainable when predicting user response results, thereby improving the accuracy of the lift value and thereby improving the effectiveness and accuracy of marketing recommendations.

[0108] As an embodiment of the present application, Figure 4 Another specific implementation process of constructing and training a target deep network model in an online marketing method based on a deep network model provided in an embodiment of the present application is shown, and is described in detail as follows:

[0109] Step S301: Acquire an original sample set, which includes sample features, intervention feature labels, and response feature labels of sample users.

[0110] Step S302: construct a deep network model, wherein the deep network model includes a group division module, an automatic selector and a main network module.

[0111] The main network module in the deep network model constructed in this embodiment also uses the implicit boosting information in the constructed sample pairs to guide model training.

[0112] Step S303: input the original sample set into the group division module, and the group division module divides the sample users in the original sample set according to the sample characteristics, the intervention characteristic labels and the response characteristic labels to obtain a number of sample subsets.

[0113] Step S304: input the sample subsets into corresponding automatic selectors respectively, and determine the sample grouping of sample users in the sample subsets through the automatic selectors.

[0114] In this embodiment, a sample deterministic grouping is added, and the sample users are determined to be in a certain sample group according to a grouping threshold.

[0115] Deterministic grouping is performed according to the following calculation formulas (9)-(11):

[0116]

[0117] The above (9) and (10) represent the discretization of probability values, P GWhen it is greater than the grouping threshold The probability value is 1, otherwise The probability value is 0. R When it is greater than the grouping threshold The probability value is 1, otherwise The probability value is 0.

[0118] Step S305: The main network module predicts the sample user based on the intervention feature label and the sample grouping, obtains the response results of the sample user under different intervention feature labels, and determines the sample lift value of the sample user according to the response results of the sample user under different intervention feature labels.

[0119] In this embodiment, the specific steps of step S301, step S303, and step S305 refer to step S201, step S202, and step S205 in the above embodiment, which will not be repeated here.

[0120] Step S306: selecting samples from different sample groups to construct sample pairs according to the specific relationship between the sample lift values ​​in different sample groups through the main network module, and capturing implicit lift information based on the sample pairs.

[0121] Each sample pair contains sample users from different sample groups.

[0122] The sample groups include G positive promotion group, R negative promotion group and W no promotion group. The promotion value of the sample in G is greater than the promotion value of the sample in W, and the promotion value of the sample in W is greater than the promotion value of the sample in R. In this embodiment, the sample pair includes sample users from different sample groups. Exemplarily, the method of constructing the sample pair is:<x1,x2> :<G,W> ,<G,R> ,<W,R> ,<G∪W,R> ,<G,W∪R> , <<G∪W> ,<W∪R> >.

[0123] Consider using the implicit lift information between sample pairs to guide model training to obtain more accurate lift values.

[0124] Step S307: input the sample user intervention feature label, the response feature label and the sample improvement value into a preset evaluation mechanism for evaluation to obtain the improvement effect of the deep network model.

[0125] The above preset evaluation mechanism is used to evaluate the network performance of the deep network model based on the evaluation index and determine the improvement effect of the deep network model. The larger the value of the evaluation index, the better the ranking effect of the method improvement value and the more accurate the prediction of the target population.

[0126] Step S308: adjusting the model parameters of the deep network model based on the preset loss function of the main network module, and repeatedly iterating to evaluate the effect of the deep network model according to the evaluation index. When the effect does not improve within a preset number of consecutive times, a trained target deep network model is obtained.

[0127] The preset number of times can be set according to the needs. For example, the preset number of times is set to 5 times. When the improvement effect is not improved for 5 consecutive times, overfitting is prevented, the model training is completed, and the trained target deep improvement network model is obtained.

[0128] In this embodiment, during model training, according to the preset loss function of the main network module, the main network module is iteratively trained. When the effect of the deep network model is evaluated according to the evaluation index and there is no improvement within a preset number of consecutive times, the trained target deep network model is obtained.

[0129] In a possible implementation, the constraint conditions are determined based on the lift value differences corresponding to sample users in different sample groups; during the training process of the deep network model, the main network module predicts the sample users in the sample groups based on the intervention feature labels under the constraint conditions to obtain response results of the sample users under different intervention feature labels; and the lift value of the user under the constraint conditions is determined based on the predicted response results of the user under different intervention features.

[0130] In this embodiment, the model is optimized by using the difference in lift values ​​between different groups of people as a constraint.

[0131] For sample pairs<G,W> The constraints are as follows:

[0132]

[0133]

[0134] For sample pairs<G,R> The constraints are as follows:

[0135]

[0136]

[0137] For sample pairs<W,R> The constraints are as follows:

[0138]

[0139]

[0140] For sample pairs<G∪W,R> The constraints are as follows:

[0141]

[0142]

[0143] For sample pairs<G,W∪R> The constraints are as follows:

[0144]

[0145] For sample pairs<<G∪W> ,<W∪R> The constraints of > are calculated as follows:

[0146]

[0147]

[0148] In this embodiment, four evaluation mechanisms can be used to evaluate the model during the model training process.

[0149] The first type: the lift score in the top h percentile (LIFT@h), which is calculated as follows (25):

[0150]

[0151] According to the sorting of the lift values, the sample users in the top h percentiles are selected, and the sum of intervention response values ​​S t , the number of samples in the intervention group N t , the total response value of the control group S c , the number of samples in the control group is N c , h is set to 30 in the experiment.

[0152] The second type: weighted average improvement (WAU), calculated as follows (26):

[0153]

[0154] The total sample users are divided into k groups according to the improvement value, and the sum of the intervention group response values ​​in the kth group The sum of the control group response values ​​within the kth group The number of intervention group samples in group k The number of control group samples for all groups and In the experiments, k is set to 10.

[0155] The third type: the normalized area under the QINI curve (QINI), calculated as follows (27):

[0156]

[0157] The sum of the response values ​​of the intervention group among the first m sample users The sum of the control group response results and values ​​among the first m sample users The number of sample users in the intervention group among the first m sample users The sum of the control group response results and values ​​among the first m sample users

[0158] The fourth method is to use the normalized area under the lift curve (AUUC), which is calculated as follows (28):

[0159]

[0160] The sum of the response values ​​of the intervention group among the first m sample users The sum of the control group response results and values ​​among the first m sample users The number of sample users in the intervention group among the first m sample users The sum of the control group response results and values ​​among the first m sample users

[0161] In this embodiment, a larger value of the evaluation index indicates a better method improvement ranking effect, and a more accurate prediction of the target population.

[0162] Exemplarily, the QINI evaluation mechanism is used for evaluation, and the maximum value of the QINI evaluation is the optimal value. The current network performance value of the deep network model is updated according to the maximum evaluation value, and the early stopping mechanism is used to prevent the model from overfitting. If the current optimal performance value is not less than the current network performance value, the early stopping counter is increased by 1. If the current optimal performance value is less than the current network performance value, the current network performance value is updated. If the iteration stop condition is met and the early stopping counter is greater than 5, the optimized target deep network model is obtained, otherwise the current optimal performance value is continued to be compared to be not less than the current network performance value.

[0163] Step S103: determining target users in the user group based on the boost value, and performing online marketing to the target users.

[0164] By comparing the predicted response results of users under different intervention features, we can calculate the lift value of each user. The lift value helps us effectively segment the user group and identify those users who respond most actively to specific intervention features, that is, target users. When conducting online marketing to target users, we can improve the accuracy and effectiveness of online marketing recommendations.

[0165] As can be seen from the above, in the embodiment of the present application, by obtaining the user characteristics of users in the user group, the user characteristics and specific intervention characteristics are input into the pre-trained target deep network model, and the predicted response results of the user under different specific intervention characteristics are obtained through the deep network model, and based on the predicted response results of the user under different intervention characteristics, the lift value of the user is accurately determined, and then the lift value is used to evaluate the response changes of the user after accepting the different intervention characteristics, which is conducive to segmenting the user group based on the lift value, accurately identifying the corresponding population to which the user belongs, and then effectively determining the target user from the user group, and then conducting precise online marketing to the target user, which can improve the effectiveness and accuracy of marketing recommendations.

[0166] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0167] Corresponding to the online marketing method based on the deep network model described in the above embodiment, Figure 5 A structural block diagram of an online marketing device based on a deep network model provided in an embodiment of the present application is shown. For ease of explanation, only the parts related to the embodiment of the present application are shown.

[0168] Reference Figure 5 The online marketing device based on the deep network model includes: a feature acquisition unit 51, a boost value determination unit 52, and an online marketing unit 53, wherein:

[0169] A feature acquisition unit 51 is used to acquire user features of users in a user group;

[0170] A lift value determination unit 52 is used to input the user characteristics and specific intervention characteristics into a pre-trained target deep network model, obtain the predicted response results of the user under different specific intervention characteristics through the deep network model, and determine the user's lift value according to the predicted response results of the user under different intervention characteristics, wherein the lift value is used to evaluate the response change of the user after receiving the different intervention characteristics;

[0171] The online marketing unit 53 is used to determine the target users in the user group based on the boost value and perform online marketing to the target users.

[0172] As a possible implementation manner of the present application, the step of obtaining the predicted response results of the user under different specific intervention features includes:

[0173] Dividing the user group according to the user characteristics and the specific intervention characteristics to obtain a plurality of user subgroups corresponding to the user group;

[0174] determining a target grouping of users in each of the user subgroups;

[0175] The user is predicted based on the specific intervention feature and the target group to obtain predicted response results of the user under different specific intervention features.

[0176] As a possible implementation manner of the present application, the above-mentioned device further includes a first model building training unit, which is used to:

[0177] Obtaining an original sample set, wherein the original sample set includes sample features, intervention feature labels, and response feature labels of sample users;

[0178] Constructing a deep network model, wherein the deep network model includes a group division module, an automatic selector, and a main network module;

[0179] The original sample set is input into the group division module, and the group division module divides the sample users in the original sample set according to the sample features, the intervention feature labels and the response feature labels to obtain several sample subsets:

[0180] Inputting the sample subsets into corresponding automatic selectors respectively, and determining sample groups of sample users in the sample subsets through the automatic selectors;

[0181] Predicting the sample user based on the intervention feature label and the sample grouping through the main network module, obtaining the response results of the sample user under different intervention feature labels, and determining the sample lift value of the sample user according to the response results of the sample user under different intervention feature labels;

[0182] Inputting the sample user intervention feature label, the response feature label and the sample improvement value into a preset evaluation mechanism for evaluation to obtain the improvement effect of the deep network model;

[0183] The model parameters of the deep network model are adjusted based on the preset loss function of the main network module, and iterations are repeated. When the improvement effect does not improve continuously within a preset number of times, a trained target deep network model is obtained.

[0184] As a possible implementation of the present application, the first model building training unit is further used for:

[0185] Obtaining constraint conditions corresponding to the sample grouping;

[0186] The main network module predicts the sample users in the sample group based on the intervention feature labels under the constraints to obtain the response results of the sample users under different intervention feature labels; and determines the improvement value of the user under the constraints according to the predicted response results of the user under different intervention features.

[0187] As a possible implementation manner of the present application, the above-mentioned device further includes a second model building training unit, which is used to:

[0188] Obtaining an original sample set, wherein the original sample set includes sample features, intervention feature labels, and response feature labels of sample users;

[0189] Constructing a deep network model, wherein the deep network model includes a group division module, an automatic selector, and a main network module;

[0190] The original sample set is input into the group division module, and the group division module divides the sample users in the original sample set according to the sample features, the intervention feature labels and the response feature labels to obtain several sample subsets:

[0191] Inputting the sample subsets into corresponding automatic selectors respectively, and determining sample groups of sample users in the sample subsets through the automatic selectors;

[0192] Predicting the sample user based on the intervention feature label and the sample grouping through the main network module, obtaining the response results of the sample user under different intervention feature labels, and determining the sample lift value of the sample user according to the response results of the sample user under different intervention feature labels;

[0193] The main network module selects samples from different sample groups to construct sample pairs according to the specific relationship between the sample lift values ​​in different sample groups, and captures implicit lift information based on the sample pairs; each sample pair includes sample users from different sample groups;

[0194] Inputting the sample user intervention feature label, the response feature label and the sample improvement value into a preset evaluation mechanism for evaluation to obtain the improvement effect of the deep network model;

[0195] The model parameters of the deep network model are adjusted based on the preset loss function of the main network module, and iterations are repeated. When the improvement effect does not improve continuously within a preset number of times, a trained target deep network model is obtained.

[0196] As a possible implementation of the present application, the second model building training unit is further used for:

[0197] Obtaining constraint conditions corresponding to the sample grouping;

[0198] During the training process of the deep network model, the main network module predicts the sample users in the sample group based on the intervention feature labels under the constraints to obtain the response results of the sample users under different intervention feature labels; based on the predicted response results of the users under different intervention features, the improvement value of the users under the constraints is determined.

[0199] As can be seen from the above, in the embodiment of the present application, by obtaining the user characteristics of users in the user group, the user characteristics and specific intervention characteristics are input into the pre-trained target deep network model, and the predicted response results of the user under different specific intervention characteristics are obtained through the deep network model, and based on the predicted response results of the user under different intervention characteristics, the lift value of the user is accurately determined, and then the lift value is used to evaluate the response changes of the user after accepting the different intervention characteristics, which is conducive to segmenting the user group based on the lift value, accurately identifying the corresponding population to which the user belongs, and then effectively determining the target user from the user group, and then conducting precise online marketing to the target user, which can improve the effectiveness and accuracy of marketing recommendations.

[0200] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0201] The present application also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, Figures 1 to 4 Represents the steps of any online marketing method based on a deep network model.

[0202] The embodiment of the present application also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, Figures 1 to 4 Represents the steps of any online marketing method based on a deep network model.

[0203] The embodiment of the present application also provides a computer program product, when the computer program product is run on a terminal device, the terminal device executes the following Figures 1 to 4 Represents the steps of any online marketing method based on a deep network model.

[0204] Figure 6Schematic diagram of a terminal device provided by an embodiment of the present application. Figure 6 As shown, the terminal device 6 of this embodiment includes: a processor 60, a memory 61, and a computer program 62 stored in the memory 61 and executable on the processor 60. When the processor 60 executes the computer program 62, the steps in the above-mentioned embodiments of the online marketing method based on the deep network model are implemented, for example Figure 1 Alternatively, when the processor 60 executes the computer program 62, the functions of each module / unit in the above-mentioned device embodiments are realized, for example Figure 5 The functions of units 51 to 53 are shown.

[0205] Exemplarily, the computer program 62 may be divided into one or more modules / units, which are stored in the memory 61 and executed by the processor 60 to complete the present application. The one or more modules / units may be a series of computer-readable instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program 62 in the terminal device 6.

[0206] The terminal device 6 may include, but is not limited to, a processor 60 and a memory 61. Those skilled in the art will appreciate that Figure 6 It is only an example of the terminal device 6 and does not constitute a limitation of the terminal device 6. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, the terminal device 6 may also include input and output devices, network access devices, buses, etc.

[0207] The processor 60 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0208] The memory 61 may be an internal storage unit of the terminal device 6, such as a hard disk or memory of the terminal device 6. The memory 61 may also be an external storage device of the terminal device 6, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device 6. Further, the memory 61 may also include both an internal storage unit and an external storage device of the terminal device 6. The memory 61 is used to store the computer program and other programs and data required by the terminal device. The memory 61 may also be used to temporarily store data that has been output or is to be output.

[0209] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0210] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0211] 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 computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device that can carry the computer program code to the device / terminal device, a recording medium, a computer memory, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), an electric carrier signal, a telecommunication signal, and a software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0212] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0213] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. An online marketing method based on a deep network model, characterized in that: include: Obtain user characteristics of users in the user group; Inputting the user characteristics and specific intervention characteristics into a pre-trained target deep network model, obtaining predicted response results of the user under different specific intervention characteristics through the deep network model, and determining the user's lift value according to the predicted response results of the user under different intervention characteristics, wherein the lift value is used to evaluate the response change of the user after receiving the different intervention characteristics; Target users in the user group are determined based on the lift value, and online marketing is performed to the target users.

2. The method according to claim 1, characterized in that The step of obtaining the predicted response results of the user under different specific intervention characteristics includes: Dividing the user group according to the user characteristics and the specific intervention characteristics to obtain a plurality of user subgroups corresponding to the user group; determining a target grouping of users in each of the user subgroups; The user is predicted based on the specific intervention feature and the target group to obtain predicted response results of the user under different specific intervention features.

3. The method according to claim 1, characterized in that Before the step of inputting the user features and the specific intervention features into the pre-trained target deep network model, the method further includes: Obtaining an original sample set, wherein the original sample set includes sample features, intervention feature labels, and response feature labels of sample users; Constructing a deep network model, wherein the deep network model includes a group division module, an automatic selector, and a main network module; The original sample set is input into the group division module, and the group division module divides the sample users in the original sample set according to the sample features, the intervention feature labels and the response feature labels to obtain several sample subsets: Inputting the sample subsets into corresponding automatic selectors respectively, and determining sample groups of sample users in the sample subsets through the automatic selectors; Predicting the sample user based on the intervention feature label and the sample grouping through the main network module, obtaining the response results of the sample user under different intervention feature labels, and determining the sample lift value of the sample user according to the response results of the sample user under different intervention feature labels; Inputting the sample user intervention feature label, the response feature label and the sample improvement value into a preset evaluation mechanism for evaluation to obtain the improvement effect of the deep network model; The model parameters of the deep network model are adjusted based on the preset loss function of the main network module, and iterations are repeated. When the improvement effect does not improve continuously within a preset number of times, a trained target deep network model is obtained.

4. The method according to claim 3, characterized in that The method further comprises: Obtaining constraint conditions corresponding to the sample grouping; During the training process of the deep network model, the main network module predicts the sample users in the sample group based on the intervention feature labels under the constraints to obtain the response results of the sample users under different intervention feature labels; based on the predicted response results of the users under different intervention features, the improvement value of the users under the constraints is determined.

5. The method according to claim 1, characterized in that Before the step of inputting the user features and the specific intervention features into the pre-trained target deep network model, the method further includes: Obtaining an original sample set, wherein the original sample set includes sample features, intervention feature labels, and response feature labels of sample users; Constructing a deep network model, wherein the deep network model includes a group division module, an automatic selector, and a main network module; The original sample set is input into the group division module, and the group division module divides the sample users in the original sample set according to the sample features, the intervention feature labels and the response feature labels to obtain several sample subsets: Inputting the sample subsets into corresponding automatic selectors respectively, and determining sample groups of sample users in the sample subsets through the automatic selectors; Predicting the sample user based on the intervention feature label and the sample grouping through the main network module, obtaining the response results of the sample user under different intervention feature labels, and determining the sample lift value of the sample user according to the response results of the sample user under different intervention feature labels; The main network module selects samples from different sample groups to construct sample pairs according to the specific relationship between the sample lift values ​​in different sample groups, and captures implicit lift information based on the sample pairs; each sample pair includes sample users from different sample groups; Inputting the sample user intervention feature label, the response feature label and the sample improvement value into a preset evaluation mechanism for evaluation to obtain the improvement effect of the deep network model; The model parameters of the deep network model are adjusted based on the preset loss function of the main network module, and iterations are repeated. When the improvement effect does not improve continuously within a preset number of times, a trained target deep network model is obtained.

6. The method according to claim 5, characterized in that The method further comprises: Determine the constraint conditions based on the lift value differences corresponding to sample users in different sample groups; During the training process of the deep network model, the main network module predicts the sample users in the sample group based on the intervention feature labels under the constraints to obtain the response results of the sample users under different intervention feature labels; based on the predicted response results of the users under different intervention features, the improvement value of the users under the constraints is determined.

7. An online marketing device based on a deep network model, characterized in that: include: A feature acquisition unit, used to acquire user features of users in a user group; a lift value determination unit, configured to input the user characteristics and specific intervention characteristics into a pre-trained target deep network model, obtain predicted response results of the user under different specific intervention characteristics through the deep network model, and determine the user's lift value according to the predicted response results of the user under different intervention characteristics, wherein the lift value is used to evaluate the response change of the user after receiving the different intervention characteristics; An online marketing unit is used to determine target users in the user group based on the boost value and perform online marketing to the target users.

8. The device according to claim 7, characterized in that The step of obtaining the predicted response results of the user under different specific intervention characteristics includes: Dividing the user group according to the user characteristics and the specific intervention characteristics to obtain a plurality of user subgroups corresponding to the user group; determining a target grouping of users in each of the user subgroups; The user is predicted based on the specific intervention feature and the target group to obtain predicted response results of the user under different specific intervention features.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the online marketing method based on the deep network model as described in any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the online marketing method based on the deep network model as described in any one of claims 1 to 6 is implemented.

Images