Click feedback method and device for push message

By collecting and analyzing user click feedback information, the prediction system is trained to predict the probability of click time and content type, and personalized message push is realized, solving the problem of insufficient message push in the existing technology and improving the user experience.

CN112422413BActive Publication Date: 2024-11-19GUANGZHOU QISHENG INFORMATION TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202011245574.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-11-10
Publication Date
2024-11-19
Estimated Expiration
2040-11-10

AI Technical Summary

Technical Problem

The lack of feedback mechanisms in the prior art, and the failure to effectively consider the user's usage habits, usage experience, and the time and occasion of message opening, resulting in the unpredictable message pushing is not accurate enough, reducing the user experience.

Method used

By collecting user click feedback information on push messages, generating training data and training prediction systems, obtaining user feature data to predict the probability of click time and content type, personalized message push for user habits and feelings is realized.

Benefits of technology

Pushing the messages you really need to the user at the right time improves the accurate positioning of message push, enhances the user's satisfaction during the period when he wants to receive the message, and reduces the user's sense of being harassed.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112422413B_ABST
    Figure CN112422413B_ABST
Patent Text Reader

Abstract

The present invention discloses a click feedback method and device for push messages. The click feedback method comprises: collecting click feedback information of users on push messages, wherein the click feedback information comprises: click time information, click content type information, and feature information associated with click time and / or click content type; using the click feedback information to generate training data, and training a prediction system according to the generated training data; extracting multiple users from a database, obtaining user feature data corresponding to the multiple users, and inputting the user feature data into the prediction system as test data to obtain prediction results, wherein the prediction results comprise: the probability of click time enumeration, the probability of click content type enumeration. Taking into account the user's usage habits, usage experience, and the time and occasion of opening the message, the purpose of accurately locating the user for the push message is achieved, and the user receives the message within the time period in which the user hopes to receive the message.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of Internet, and in particular to a click feedback method and device for pushing messages. Background Art

[0002] Information push is a new technology that reduces information overload by regularly transmitting information needed by users on the Internet through certain technical standards or protocols. In the operation process, enterprises have two main tasks, one is to generate new users, and the other is to retain existing users. To complete these two tasks, in addition to developing new products, it is also necessary to actively push new product information to users at a regular and quantitative time.

[0003] However, from the user's perspective, passively accepting messages is more likely to cause users to feel negative emotions than actively seeking information, thereby reducing the user's experience of using the product.

[0004] Therefore, with respect to the active message push mechanism of related technologies, there is currently a lack of a feedback mechanism to consider the user's usage habits, usage experience, and the time and occasion of opening the message. Summary of the invention

[0005] The main purpose of the present invention is to disclose a click feedback method and device for push messages, so as to at least solve the message active push mechanism used in the related technology. At present, there is still a lack of a feedback mechanism to take into account the user's usage habits, usage experience, and the time and occasion of opening the message.

[0006] According to one aspect of the present invention, a click feedback method for a push message is provided.

[0007] The click feedback method for push messages according to the present invention includes: collecting user click feedback information on push messages, wherein the click feedback information includes: click time information, click content type information, and feature information associated with the click time and / or click content type; using the click feedback information to generate training data, and training a prediction system based on the generated training data; extracting multiple users from a database, obtaining user feature data corresponding to the multiple users, and inputting the user feature data as test data into the prediction system to obtain prediction results, wherein the prediction results include: the probability of click time enumeration and the probability of click content type enumeration.

[0008] According to another aspect of the present invention, a click feedback device for push messages is provided.

[0009] The click feedback device for push messages according to the present invention comprises: an information collection module, used to collect user click feedback information on push messages, wherein the click feedback information comprises: click time information, click content type information, and feature information associated with click time and / or click content type; a system training module, used to generate training data using the click feedback information, and train a prediction system based on the generated training data; a result acquisition module, used to retrieve multiple users from a database, obtain user feature data corresponding to the multiple users, and input the user feature data as test data into the prediction system to obtain prediction results, wherein the prediction results comprise: the probability of click time enumeration and the probability of click content type enumeration.

[0010] With respect to the active message push mechanism adopted in the related art, the present invention provides a feedback mechanism, which takes into account the user's usage habits, usage experience, and the time and occasion of opening the message, and pushes the messages that the user really needs at the appropriate time, thereby achieving the purpose of accurately locating the user for push message and enabling the user to receive the message within the time period when he or she hopes to receive the message. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 is a flow chart of a click feedback method for push messages according to an embodiment of the present invention;

[0012] Figure 2 is a flow chart of a user characterization method according to a preferred embodiment of the present invention;

[0013] Figure 3 is a one-hot feature interval diagram of click time enumeration of pushed content according to a preferred embodiment of the present invention;

[0014] Figure 4 is a one-hot feature interval diagram enumerating click content types of pushed content according to a preferred embodiment of the present invention;

[0015] Figure 5 is a one-hot interval diagram of user characteristics of pushed content according to a preferred embodiment of the present invention;

[0016] Figure 6 is a flow chart of a method for generating training data training prediction system according to a preferred embodiment of the present invention;

[0017] Figure 7 is a diagram showing the training working principle of a prediction system according to a preferred embodiment of the present invention;

[0018] Figure 8 is a flow chart of using a prediction system to generate prediction results according to a preferred embodiment of the present invention;

[0019] Fig. 9 is a diagram of the prediction working principle of the prediction system according to a preferred embodiment of the present invention;

[0020] Fig.10 is a general flow chart of a click prediction method for a Push system according to a preferred embodiment of the present invention;

[0021] Fig.11 is a structural block diagram of a click feedback device for push messages according to an embodiment of the present invention;

[0022] Fig.12 It is a structural block diagram of a click feedback device for push messages according to a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0023] The specific implementation of the present invention is described in detail below with reference to the accompanying drawings.

[0024] Figure 1 FIG. 1 is a flow chart of a click feedback method for a push message according to an embodiment of the present invention. Figure 1 As shown, the click feedback method of the push message includes:

[0025] Step S101: collecting user click feedback information on a push message, wherein the click feedback information includes: click time information, click content type information, and feature information associated with the click time and / or click content type;

[0026] Step S103: using the click feedback information to generate training data, and training the prediction system according to the generated training data;

[0027] Step S105: extract multiple users from the database, obtain user feature data corresponding to the multiple users, and input the user feature data as test data into the prediction system to obtain prediction results, wherein the prediction results include: the probability of click time enumeration and the probability of click content type enumeration.

[0028] The embodiment of the present invention provides a feedback mechanism, which collects the user's click feedback information on the pushed message, generates training data using the above click feedback information, and trains the prediction system based on the generated training data; inputs the user feature data corresponding to multiple users retrieved from the database into the above prediction system to obtain the prediction result, wherein the above prediction result includes: the probability of click time enumeration, the probability of click content type enumeration. Considering the user's usage habits, usage experience, and the time and occasion of opening the message, according to the prediction result, the message that the user really needs is pushed to the user at the appropriate time, so as to achieve the purpose of accurately locating the user for the push message and allowing the user to receive the message within the time period when the user hopes to receive the message.

[0029] Preferably, the above-mentioned characteristic information associated with the click time and / or click content type may include but is not limited to at least one of the following: user region information, device screen information, browser type information, operating system type information, device manufacturer information, device type information, device name information, and push channel information.

[0030] First, let's describe feature engineering. Feature engineering refers to using a series of engineering methods to filter out better data features from raw data to improve the training effect of the model. Data and features determine the upper limit of machine learning, while models and algorithms are approaching this upper limit. It can be seen that good data and features are the prerequisites for models and algorithms to play a greater role. Feature engineering usually includes data preprocessing, feature selection, dimensionality reduction and other links. The quality of the features extracted from the data will directly affect the effect of the model.

[0031] Preferably, using the above-mentioned click feedback information to generate training data includes: converting the click time in the log data into a click time enumeration, and then encoding the click time enumeration using a fixed-position one-hot encoding method; converting the click content type in the log data into a click content type enumeration, and then encoding the click content type enumeration using a fixed-position one-hot encoding method; converting the feature information associated with the click time and / or click content type into feature data, and then using a dynamic one-hot encoding method to reduce the dimension of the above-mentioned user feature data and compress it into a dynamic one-hot interval.

[0032] Combine the following Figure 2 Describe feature engineering.

[0033] like Figure 2 As shown, firstly, user fields are selected from the user profile database, that is, feature information associated with the click time and / or click content type (e.g., user region information, device screen information, browser type information, operating system type information, device manufacturer information, device type information, device name information, push channel information, etc.) is converted into feature data; then feature processing is performed, that is, feature representation based on Embedding; Embedding representation: also known as Distributed representation, compared to the high-dimensional sparse one-hot encoding representation, embedding learns a low-dimensional dense real number vector, that is, compressing sparse data with more bits into a space with fewer bits. In fact, it is to process the one-hot encoding into an embedded vector representation of a fixed dimension.

[0034] The above Embedding representation includes: dynamic one-hot encoding and fixed position one-hot encoding. The following describes these two encoding methods in detail:

[0035] The first encoding method: dynamic one-hot encoding

[0036] The field name + field value are concatenated into a string, the int64 hash value of the string is taken, and then the hash value is modulo the length of the dynamic one-hot area. The formula is as follows:

[0037] Int64_Hash(field name_field value)% dynamic one-hot area length

[0038] The advantage of dynamic one-hot is that it can reduce the dimensionality of sparse features to a fixed-length feature interval, and both the feature field and the feature field value can be dynamically increased. As long as the feature interval length is selected appropriately, the occurrence of collisions can be ignored.

[0039] The second encoding method: fixed position one-hot encoding method

[0040] Mapped to a fixed one-hot position by enumeration number.

[0041] The advantage of fixed one-hot is that the feature interval length is fixed and the feature position can be known in advance.

[0042] In a preferred implementation, the features may further include: item features and user features (ie: feature information associated with click time and / or click content type).

[0043] The first type of feature data: item feature data

[0044] (1) Enumeration of click time of pushed content

[0045] That is, the click time of the pushed content. For example, there are 24 enumerations (0 to 23), representing 24-hour time periods. Processing method: fixed position one-hot

[0046] for example:

[0047] 0 o'clock = 000000000000000000000001

[0048] 1 o'clock=000000000000000000000010 ...

[0050] 23:00 = 10000000000000000000000

[0051] Convert the click time enumeration of the pushed content into a time one-hot, see Figure 3 .

[0052] (2) Enumeration of push content types

[0053] That is, the type of content to be pushed. For example, (a=0, t=1, v=2) has 3 enumerations, representing 3 media types, where a=article, t=question-and-answer post, v=video. Processing method: fixed position one-hot

[0054] for example:

[0055] a=0000000001

[0056] t=0000000010

[0057] v=0000000100 ...

[0059] The type enumeration of pushed content is converted into content type one-hot, please refer to Figure 4 .

[0060] The second type of characteristic data: user characteristic data (i.e., characteristic information associated with click time and / or click content type)

[0061] Since user features are usually numerous and have many enumerations, too many enumerations will lead to feature sparseness and a large model. In addition, user features are uncertain (more enumerations may be added in the future), so it is not suitable to use a fixed one-hot method to extract features. You can use a dynamic one-hot method to extract features. For example:

[0062] City (ip_area), processing method: dynamic one-hot

[0063] Device screen (device_screen_width, device_screen_height), processing method: dynamic one-hot

[0064] Browser type (browser_name), processing method: dynamic one-hot

[0065] System type (os_name), processing method: dynamic one-hot

[0066] Device vendor (device_vendor), processing method: dynamic one-hot

[0067] Device type (device_type), processing method: dynamic one-hot

[0068] Device name (device_model), processing method: dynamic one-hot

[0069] Push channel enumeration (Huawei, vivo, OPPO, Xiaomi), processing method: dynamic one-hot ...

[0071] The remaining user features are reduced and compressed into a dynamic one-hot range, see Figure 5 .

[0072] Preferably, in the above step S101, collecting the user's click feedback information on the push message may further include: obtaining a click behavior log of the user clicking on the push message; extracting log data in the above user click behavior log, wherein the above log data includes: click time, click content type, and feature information associated with the click time and / or click content type (for example, user id, push channel, etc.).

[0073] In the preferred implementation process, click feedback information can be obtained by collecting information in the user's click behavior log. For example, the user's click behavior log is a log of the user's behavior of clicking on the pushed content after receiving the pushed content on the terminal. After the log is processed by the message queue and streaming, it is aggregated and generated into training data.

[0074] The log format is as follows:

[0075] Click time, user ID (identification information), content type, push channel, whether clicked

[0076] Figure 6 1 is a flow chart of a method for generating training data training prediction system according to a preferred embodiment of the present invention. Figure 6 As shown in , first extract the log data from the user click behavior log stream. The click time needs to be converted into a click time enumeration, and then converted into a time enumeration one-hot, such as Figure 3 As shown; the content type needs to be converted into a content type enumeration one-hot, such as Figure 4 As shown in the figure, user ID and push channel are converted into user features through feature engineering, and then compressed into dynamic one-hot intervals through dynamic one-hot method, and training data is generated, as shown in the figure. Figure 5 As shown. Then the prediction system is trained based on the generated training data.

[0077] Preferably, the above prediction system may further include: two serially connected logistic regression classification models, wherein the first logistic regression classification model is a time prediction model, and the second logistic regression classification model is a content type prediction model.

[0078] Preferably, in the above step S103, training the prediction system according to the generated training data may further include: inputting the encoded data for encoding the user characteristic data in the training data and the encoded data for encoding the click time enumeration in the training data into the time prediction model to train the time prediction model; inputting the encoded data for encoding the user characteristic data in the training data, the encoded data for encoding the click content type enumeration in the training data, and the encoded data for encoding the click time enumeration number in the training data into the content type prediction model to train the content type prediction model.

[0079] Figure 7 1 is a diagram showing the training working principle of the prediction system according to the preferred embodiment of the present invention. Figure 7 As shown, the prediction system uses two logistic regression classification models in series, named LR-Time (time prediction model) and LR-Type (content type prediction model).

[0080] Among them, the logistic regression classification model, also known as the LR regression classification model, is based on the linear regression model. It uses the sigmoid function to compress the result of the linear model w^Tx to between [0,1] to make it have probabilistic meaning. Its essence is still a linear model, and its implementation is relatively simple. Through logistic regression analysis, the weights of independent variables can be obtained. The LR model is the basic component of deep learning. The formula is: y=f(x)=logistics(linear(x)). The logistic regression model continuously corrects the weights through deep learning. When the weight is corrected to an error less than the predetermined value, the logistic regression classification model completes the training.

[0081] like Figure 7 As shown, the input of LR-Time includes: the encoded data for encoding user feature data in the training data, and the encoded data for encoding click time enumeration in the training data. The output is the weight of LR-Time. The weight is continuously corrected through deep learning. When the weight is corrected to a very small error and meets the predetermined requirements, the LR-Time model is trained.

[0082] The formula is as follows:

[0083] Click time enumeration = LR-Time(x) = logistics(linear(user feature))

[0084] The input of LR-Type includes: the encoded data of the user feature data in the training data + the output of LR-Time (which needs to be converted into click time features first, that is, the time enumeration number is encoded using a fixed position one-hot encoding method to generate encoded data, such as Figure 3 and Figure 7 The LR-Time model is trained by encoding the LR-Type weights in the training data (as shown in the figure) + encoding the click content type enumeration in the training data. The weights are continuously corrected through deep learning. When the weights are corrected to a very small error and meet the predetermined requirements, the LR-Time model is trained.

[0085] The formula is as follows:

[0086] Content type enumeration = LR-Type(x) = logistics(linear(user feature + click time feature))

[0087] Therefore, with sufficient data support, a prediction system that meets the needs can be trained. It should be noted that in order to train a prediction system with small errors and high precision, all the acquired training data are usually used to train the prediction system. However, if the amount of training data is huge, in order to reduce the amount of calculation, some training data can be randomly extracted from the overall training data according to a certain proportion (for example, 30%) to train the prediction system.

[0088] Preferably, in step S105, retrieving a plurality of users from a database may further include: retrieving the plurality of users ranked top in terms of activity from the database.

[0089] Preferably, in step S105, the user feature data is input into the above-mentioned prediction system as the data to be tested, and obtaining the prediction result may further include: using a dynamic one-hot encoding method to reduce the dimension of the above-mentioned user feature data and compress it into a dynamic one-hot interval to generate the above-mentioned data to be tested; inputting the above-mentioned data to be tested into the above-mentioned time prediction system to obtain the probability of the above-mentioned click time enumeration; using a fixed position one-hot encoding method to encode at least one time enumeration number in the above-mentioned time enumeration number to generate encoded data; inputting the above-mentioned generated encoded data and the above-mentioned data to be tested into the above-mentioned content type prediction model to obtain the probability of the above-mentioned content type prediction model.

[0090] Preferably, at least one of the above-mentioned time enumeration numbers can be determined in the following manner: sorting the above-mentioned time enumeration numbers according to the order of the probability of the above-mentioned click time enumeration; determining the top X large probabilities among the above-mentioned click time enumeration probabilities, and taking the time enumeration numbers corresponding to the top X large probabilities as the above-mentioned at least one time enumeration number, wherein X is an integer greater than or equal to 1 and less than the total amount of the above-mentioned time enumeration numbers.

[0091] In the preferred implementation process, a model optimization scheme can be adopted, that is, for the serial logistic regression classification model, the input of the content type prediction model is not the encoded data for encoding all the click time enumeration numbers, but the encoded data of the time enumeration numbers corresponding to the top X most probable click time enumeration probabilities are input into the content type prediction model.

[0092] For example, the number of users is 1 million, there are 24 time enumerations, and 3 content types. If all are calculated once using the model, it will take 1 million * 24 * 3 = 72 million times. Since the time prediction and content type prediction are separated into two different models, the probability of clicking the time enumeration is obtained first, and it is the input of the content type prediction model. Therefore, the probability of time enumeration can be sorted from large to small, and only the top X probabilities of clicking the time enumeration (for example, the top 3 time enumerations with high probabilities) can be extracted as the input of the content type prediction model, which can greatly reduce the amount of calculation. The amount of calculation is 1 million * 3 * 3 = 9 million times, and the performance is improved by 87.5%. That is: the output of the time prediction model is the top three click time priorities (Top3) of the pushed content, and the output of the content type prediction model is the type priority of the content to be pushed.

[0093] The following combination Figure 8 and Fig. 9 The above preferred embodiments are further described.

[0094] like Figure 8 As shown in FIG. 1 , firstly, active and / or sub-active users are extracted from the user database, and then the extracted users and push channels are converted into user feature data through the above feature engineering, and then the dynamic one-hot method is used to reduce the dimension and compress it into the dynamic one-hot interval, and the data to be predicted is generated, as shown in FIG. Figure 5 shown.

[0095] like Fig. 9As shown, the user data to be predicted (user feature data in the figure) is input into LR-Time (time prediction model) to obtain the probability of the predicted result click time enumeration. After that, the time enumeration number output by the first-level LR-Time (time prediction model) can be encoded and converted into click time feature data, and then input into LR-Type (content type prediction model) together with the user data to be predicted to obtain the probability of the predicted result content type enumeration.

[0096] Preferably, for the click time enumeration numbers output by the first-level LR-Time (time prediction model), these time enumeration numbers can be sorted according to the order of the probability of the above click time enumeration, and the top X probabilities (for example, the top 3) among the above click time enumeration probabilities are determined, where X is an integer greater than or equal to 1 and less than the total amount of the above time enumeration numbers, and the time enumeration numbers corresponding to the top X probabilities are used as the input of the second-level content type prediction model. That is, only the time enumeration numbers corresponding to the top X probabilities can be encoded using a fixed position one-hot encoding method and converted into click time feature data, such as Figure 3 As shown. Then, it is input into LR-Type (content type prediction model) together with the user's data to be predicted to obtain the probability of the content type enumeration of the prediction result. As mentioned above, this can greatly reduce the amount of calculation, improve the calculation efficiency and system performance.

[0097] Preferably, after inputting the above-mentioned user characteristic data as the test data into the above-mentioned prediction system and obtaining the prediction result in step S105, it can also include: receiving the prediction result from the above-mentioned prediction system, wherein the above-mentioned prediction result includes: user identification information corresponding to the above-mentioned multiple users, the probability of click time enumeration corresponding to each of the above-mentioned user identification information, and the probability of click content type enumeration; for each of the above-mentioned user identification information, determine the top M probabilities from the probability of click time enumeration corresponding to the user identification information, randomly select a probability from the above-mentioned top M probabilities, and use the click time enumeration number corresponding to the selected probability as the message push time of the current user, wherein M is greater than or equal to 1, and M is less than or equal to the total number of click time enumeration numbers corresponding to the user identification information; for each of the above-mentioned user identification information, determine the top N probabilities from the probability of click content type enumeration corresponding to the user identification information, randomly select a probability from the above-mentioned top N probabilities, and use the click content type enumeration number corresponding to the selected probability as the message push content type of the current user, wherein N is greater than or equal to 1, and N is less than or equal to the total number of click content type enumeration numbers corresponding to the user identification information.

[0098] In the preferred implementation process, M and N can usually be selected as numbers greater than or equal to 2. This can avoid pushing information according to the click time enumeration number and click content type enumeration number with the highest probability every time, which lacks diversity. It randomly selects one from multiple click time enumeration numbers and multiple click content type enumeration numbers with higher probabilities for message push, which has high flexibility and diversity and can effectively provide user experience.

[0099] Preferably, the above method may also include: creating a task bucket timer, wherein, for each click time enumeration number of all click time enumeration numbers, a task bucket corresponding to the click time enumeration number is set; then after using the click time enumeration number corresponding to the selected probability as the message push time of the current user, it may also include: placing the identification information of the current user into the task bucket of the click time enumeration number corresponding to the above selected probability.

[0100] Preferably, after placing the identification information of the current user into the task bucket of the click time enumeration number corresponding to the above-selected probability, it can also include: configuring a message push table according to the created task bucket, the click content type enumeration number corresponding to the probability selected for each user, and the push channel information corresponding to each user; according to the click time enumeration number and the click content type enumeration number in the above-mentioned message push table, the push engine pushes messages to each user.

[0101] During the optimization implementation process, the Push task engine obtains the top 3 click time priorities of the pushed content and the type priority of the content to be pushed from the prediction system. The first parameter determines the time period in which the content is pushed to the user, and the second parameter determines the type of content pushed to the user. Since push is pushed by time period, the Push task engine divides the users to be pushed into buckets by time. The Push task engine needs to do three steps:

[0102] Step 1: Create a task bucket timer.

[0103] The result obtained from the prediction system is a user list data containing the following:

[0104] User ID, Top 3 push times, push content type, push channel

[0105] Divide 24 hours into a task bucket for each hour, randomly select the user list according to the probability of the three push times, and put the user into the corresponding time bucket:

[0106] For example, the three push times of user a are as follows:

[0107] When push time 1=18, the probability is 0.5

[0108] When push time 2 = 21, the probability is 0.3

[0109] When push time 3 = 22, the probability is 0.2

[0110] According to the probability random function Random (0.5, 0.3, 0.2), the randomly selected push time is 18:00, so user a will be put into the time bucket of 18:00 to be sent:

[0111] 0 time = [user 1, user 2, user 3] ...

[0113] 18:00 = [User 4, User 5, User A] ...

[0115] 23:00 = = [...]

[0116] Step 2: Match the task content in the bucket and obtain the content pushed to the user from the content engine (call the batch interface) according to the type priority of the content.

[0117] For example, the probability of user a’s push content is as follows:

[0118] The probability of type a is 0.5

[0119] Type t probability is 0.4

[0120] Type v probability is 0.1

[0121] According to the probability random function Random (0.5, 0.4, 0.1), the push content randomly selected is of type t, then the content of type t will be found in the content pool and sent to user a.

[0122] Step 3: Use the terminal manufacturer's push channel to push the content to the terminal user.

[0123] After the above steps, a large schedule table is obtained (that is, the above configuration message push table):

[0124] Time 0 = [{id: user 1, content: a, channel: 1}, {id: user 2, content: t, channel: 1}, {id: user 3, content: v, channel: 2}] ...

[0126] 18:00 = [{id: user 4, content: a, channel: 1}, {id: user 5, content: a, channel: 3}, {id: user a, content: t, channel: 2}] ...

[0128] 23:00 = = [...]

[0129] The Push engine pushes the messages that users really need at the right time according to this schedule, thereby reducing the user's sense of being harassed and achieving the purpose of optimizing the message push system.

[0130] Fig.10 FIG. 1 is a general flow chart of a click feedback method for a Push message pushed by a Push system according to a preferred embodiment of the present invention. Fig.10 As shown, the click feedback method of the Push system push message mainly includes the following parts:

[0131] S1001: Use feature engineering on user data, select features to implement user characterization, and output user feature data. Through feature engineering processing, subsequent training data and test data can be obtained.

[0132] S1003: By collecting the user's click feedback on the push message, these behavior records are converted into training data for the neural network in combination with the user's characteristic data, and the constructed prediction system is trained. For example, the click behavior log of the user clicking on the push message can be obtained; the log data in the above user click behavior log is extracted, wherein the above log data includes: click time, click content type, and characteristic information associated with the click time and / or click content type. The characteristic information associated with the click time and / or click content type is converted into user characteristic data, and then the above user characteristic data is reduced in dimension and compressed into a dynamic one-hot interval using a dynamic one-hot encoding method.

[0133] S1005: Users are obtained according to the user's active attributes, and these users are converted into data to be predicted in combination with user feature data, and input into the prediction system to obtain the parameters (time, content type) of the push message. For example, the above-mentioned multiple users with the highest activity ranking are obtained from the user profile database, and then the user feature data is obtained, and the above-mentioned user feature data is compressed into a dynamic one-hot interval by dynamic one-hot encoding to generate the above-mentioned data to be tested; the above-mentioned data to be tested is input into the first-level time prediction model of the prediction system to obtain the probability of the above-mentioned click time enumeration; the above-mentioned time enumeration numbers are sorted according to the order of the probability of the above-mentioned click time enumeration; the top X probabilities of the above-mentioned click time enumeration are determined, and the time enumeration numbers corresponding to the top X probabilities are encoded by fixed position one-hot encoding to generate encoded data; the above-mentioned generated encoded data and the above-mentioned data to be tested are input into the above-mentioned content type prediction model to obtain the probability of the above-mentioned content type prediction model.

[0134] S1007: Submit the parameters of the push message to the recommendation system, and the recommendation system extracts appropriate content from the content library and pushes personalized messages to the user through the message channel. For example, the Push engine receives the prediction results from the above prediction system, and for each of the above user identification information, determines the top M probabilities from the probability of the click time enumeration corresponding to the user identification information, randomly selects a probability from the above top M probabilities, and uses the click time enumeration number corresponding to the selected probability as the message push time of the current user, where M is greater than or equal to 1, and M is less than or equal to the total number of click time enumeration numbers corresponding to the user identification information; for each of the above user identification information, determines the top N probabilities from the probability of the click content type enumeration corresponding to the user identification information, randomly selects a probability from the above top N probabilities, and uses the click content type enumeration number corresponding to the selected probability as the message push content type of the current user, where N is greater than or equal to 1, and N is less than or equal to the total number of click content type enumeration numbers corresponding to the user identification information.

[0135] Fig.11 is a structural block diagram of a click feedback device for a push message according to an embodiment of the present invention. Fig.11 As shown, the click feedback device for the push message includes: an information collection module 10, which is used to collect the user's click feedback information on the push message, wherein the above-mentioned click feedback information includes: click time information, click content type information, and feature information associated with the click time and / or click content type; a system training module 12, which is used to generate training data using the above-mentioned click feedback information, and build a prediction system based on the generated above-mentioned training data; a result acquisition module 14, which is used to retrieve multiple users from a database, obtain user feature data corresponding to the above-mentioned multiple users, and input the above-mentioned user feature data as test data into the above-mentioned prediction system to obtain prediction results, wherein the above-mentioned prediction results include: the probability of click time enumeration, the probability of click content type enumeration.

[0136] The embodiment of the present invention provides a click feedback device for push messages, wherein the information collection module 10 collects the user's click feedback information on the push messages, the model building module 12 uses the click feedback information to generate training data, and trains the prediction system according to the generated training data; the result acquisition module 14 inputs the user feature data corresponding to multiple users retrieved from the database into the prediction system to obtain the prediction results, wherein the prediction results include: the probability of click time enumeration, the probability of click content type enumeration. Considering the user's usage habits, usage experience, and the time and occasion of opening the message, according to the prediction results, the message that the user really needs is pushed to the user at the appropriate time, so as to achieve the purpose of accurately locating the user for the push message and allowing the user to receive the message within the time period when the user hopes to receive the message.

[0137] Preferably, the above prediction system may further include: two serially connected logistic regression classification models, wherein the first logistic regression classification model is a time prediction model, and the second logistic regression classification model is a content type prediction model.

[0138] In the preferred implementation process, user feature data is input as test data into the prediction system to obtain prediction results. Specifically, the user feature data is compressed into a dynamic one-hot interval using a dynamic one-hot encoding method to generate the test data; the test data is input into the time prediction system to obtain the probability of the click time enumeration; at least one of the time enumeration numbers is encoded using a fixed position one-hot encoding method to generate encoded data; the generated encoded data and the test data are input into the content type prediction model to obtain the probability of the content type prediction model.

[0139] At least one of the above-mentioned time enumeration numbers is determined in the following manner: the above-mentioned time enumeration numbers are sorted in the order of the probabilities of the above-mentioned click time enumerations, the top X probabilities among the above-mentioned click time enumerations are determined, and the time enumeration numbers corresponding to the top X probabilities are used as the above-mentioned at least one time enumeration number, wherein X is an integer greater than or equal to 1 and less than the total number of the above-mentioned time enumeration numbers.

[0140] Preferably, if Fig.11 As shown, the above-mentioned device may also include: a result receiving module 16, which is used to receive the prediction results from the above-mentioned prediction system, wherein the above-mentioned prediction results include: user identification information corresponding to the above-mentioned multiple users, the probability of click time enumeration corresponding to each of the above-mentioned user identification information, and the probability of click content type enumeration; a first selection module 18, which is used to determine the top M probabilities from the probability of click time enumeration corresponding to the user identification information for each of the above-mentioned user identification information, randomly select a probability from the above-mentioned top M probabilities, and use the click time enumeration number corresponding to the selected probability as the message push time of the current user, wherein M is greater than or equal to 1, and M is less than or equal to the total number of click time enumeration numbers corresponding to the user identification information; a second selection module 20, which is used to determine the top N probabilities from the probability of click content type enumeration corresponding to the user identification information for each of the above-mentioned user identification information, randomly select a probability from the above-mentioned top N probabilities, and use the click content type enumeration number corresponding to the selected probability as the message push content type of the current user, wherein N is greater than or equal to 1, and N is less than or equal to the total number of click content type enumeration numbers corresponding to the user identification information.

[0141] It should be noted that the preferred implementation of the combination of the modules in the above push message click feedback device can be found in Figures 1 to 10 The description is not repeated here.

[0142] In summary, with the above embodiments provided by the present invention, a click feedback mechanism for push messages is provided, the idea and implementation method of the push channel are optimized, and the key parameters of the push message content are estimated, thereby realizing the implementation and strategy of user personalized push message content. In addition, a two-level series prediction system is constructed, which is more flexible. The coded data of the time enumeration numbers corresponding to the first few high probabilities in the probability of click time enumeration are input into the content type prediction model, which greatly simplifies the calculation amount of the prediction and improves the processing capacity.

[0143] The above disclosures are only several specific embodiments of the present invention; however, the present invention is not limited thereto, and any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.

Claims

1. A click feedback method for push messages, characterized in that: include: Collecting user click feedback information on the push message, wherein the click feedback information includes: click time information, click content type information, and feature information associated with the click time and / or click content type; Using the click feedback information to generate training data, and training a prediction system according to the generated training data; Extract multiple users from the database, obtain user feature data corresponding to the multiple users, and input the user feature data as test data into the prediction system to obtain prediction results, wherein the prediction results include: the probability of click time enumeration and the probability of click content type enumeration; The prediction system comprises: Two serially connected logistic regression classification models, wherein the first logistic regression classification model is a time prediction model and the second logistic regression classification model is a content type prediction model; Inputting the user characteristic data as the test data into the prediction system to obtain the prediction result; including: using a dynamic one-hot encoding method to reduce the dimension of the user characteristic data to a dynamic one-hot interval to generate the test data; Inputting the test data into the time prediction model to obtain the probability of the click time enumeration; Encoding at least one of the time enumeration numbers using a fixed position one-hot encoding method to generate encoded data; Inputting the generated encoded data and the data to be tested into the content type prediction model to obtain the probability of the content type prediction model; At least one of the time enumeration numbers is determined by: sorting the time enumeration numbers in order of the probability of clicking the time enumeration; Determine the top X probabilities of the click time enumeration, and use the time enumeration numbers corresponding to the top X probabilities as the at least one time enumeration number, where X is an integer greater than or equal to 1 and less than the total amount of the time enumeration numbers; After inputting the user characteristic data as the test data into the prediction system and obtaining the prediction result, the method further includes: receiving the prediction result from the prediction system, wherein the prediction result includes: user identification information corresponding to the plurality of users, the probability of click time enumeration corresponding to each of the user identification information, and the probability of click content type enumeration; For each user identification information, determine the top M probabilities from the click time enumeration probabilities corresponding to the user identification information, randomly select a probability from the top M probabilities, and use the click time enumeration number corresponding to the selected probability as the message push time of the current user, where M is greater than or equal to 1, and M is less than or equal to the total number of click time enumeration numbers corresponding to the user identification information; For each of the user identification information, determine the top N probabilities from the enumerated probabilities of the click content type corresponding to the user identification information, randomly select a probability from the top N probabilities, and use the enumeration number of the click content type corresponding to the selected probability as the message push content type of the current user, where N is greater than or equal to 1, and N is less than or equal to the total number of the enumeration numbers of the click content type corresponding to the user identification information; The method further includes: creating a task bucket timer, wherein for each click time enumeration number of all click time enumeration numbers, a task bucket corresponding to the click time enumeration number is set; After using the click time enumeration number corresponding to the selected probability as the message push time of the current user, the method further includes: placing the identification information of the current user into the task bucket of the click time enumeration number corresponding to the selected probability.

2. The method according to claim 1, characterized in that The collecting of user click feedback information on the push message includes: obtaining a click behavior log of the user clicking on the push message; Extract log data from the user click behavior log, wherein the log data includes: click time, click content type, and feature information associated with the click time and / or click content type.

3. The method according to claim 2, characterized in that Generating training data using the click feedback information includes: converting the click time in the log data into a click time enumeration, and then encoding the click time enumeration using a fixed position one-hot encoding method; Converting the click content type in the log data into a click content type enumeration, and then encoding the click content type enumeration using a fixed position one-hot encoding method; The feature information associated with the click time and / or the click content type is converted into user feature data, and then the user feature data is reduced in dimension and compressed into a dynamic one-hot interval using a dynamic one-hot encoding method.

4. The method according to claim 1, characterized in that: Training the prediction system according to the generated training data includes: inputting the coded data for encoding the user feature data in the training data and the coded data for encoding the click time enumeration in the training data into the time prediction model to train the time prediction model; The encoded data for encoding the user feature data in the training data, the encoded data for encoding the click content type enumeration in the training data, and the encoded data for encoding the click time enumeration number in the training data are input into the content type prediction model to train the content type prediction model.

5. The method according to claim 1, characterized in that After placing the identification information of the current user into the task bucket of the click time enumeration number corresponding to the selected probability, the method further includes: configuring a message push table according to the created task bucket, the click content type enumeration number corresponding to the probability selected for each user, and the push channel information corresponding to each user; According to the click time enumeration number and the click content type enumeration number in the message push table, the push engine pushes a message to each user.

6. The method according to any one of claims 1 to 5, characterized in that The retrieving a plurality of users from the database includes: retrieving the plurality of users ranked top in terms of activity from the database.

7. The method according to any one of claims 1 to 5, characterized in that The characteristic information associated with the click time and / or click content type includes at least one of the following: user region information, device screen information, browser type information, operating system type information, device manufacturer information, device type information, device name information, and push channel information.

8. A click feedback device for push messages, characterized in that: include: An information collection module, used to collect user click feedback information on the push message, wherein the click feedback information includes: click time information, click content type information, and feature information associated with the click time and / or click content type; A system training module, used to generate training data using the click feedback information, and train a prediction system based on the generated training data; comprising: two serially connected logistic regression classification models, wherein the first logistic regression classification model is a time prediction model, and the second logistic regression classification model is a content type prediction model; wherein the training prediction system comprises: Inputting the user characteristic data as the test data into the prediction system to obtain the prediction result; including: using a dynamic one-hot encoding method to reduce the dimension of the user characteristic data to a dynamic one-hot interval to generate the test data; Inputting the test data into the time prediction model to obtain the probability of the click time enumeration; Encoding at least one of the time enumeration numbers using a fixed position one-hot encoding method to generate encoded data; Inputting the generated encoded data and the data to be tested into the content type prediction model to obtain the probability of the content type prediction model; At least one of the time enumeration numbers is determined by: sorting the time enumeration numbers in order of the probability of clicking the time enumeration; Determine the top X probabilities of the click time enumeration, and use the time enumeration numbers corresponding to the top X probabilities as the at least one time enumeration number, where X is an integer greater than or equal to 1 and less than the total amount of the time enumeration numbers; After inputting the user characteristic data as the test data into the prediction system and obtaining the prediction result, the method further includes: receiving the prediction result from the prediction system, wherein the prediction result includes: user identification information corresponding to the plurality of users, the probability of click time enumeration corresponding to each of the user identification information, and the probability of click content type enumeration; For each user identification information, determine the top M probabilities from the click time enumeration probabilities corresponding to the user identification information, randomly select a probability from the top M probabilities, and use the click time enumeration number corresponding to the selected probability as the message push time of the current user, where M is greater than or equal to 1, and M is less than or equal to the total number of click time enumeration numbers corresponding to the user identification information; For each of the user identification information, determine the top N probabilities from the enumerated probabilities of the click content type corresponding to the user identification information, randomly select a probability from the top N probabilities, and use the enumeration number of the click content type corresponding to the selected probability as the message push content type of the current user, where N is greater than or equal to 1, and N is less than or equal to the total number of the enumeration numbers of the click content type corresponding to the user identification information; The method further includes: creating a task bucket timer, wherein for each click time enumeration number of all click time enumeration numbers, a task bucket corresponding to the click time enumeration number is set; After using the click time enumeration number corresponding to the selected probability as the message push time of the current user, the method further includes: placing the identification information of the current user into the task bucket of the click time enumeration number corresponding to the selected probability; a result acquisition module for extracting multiple users from a database, acquiring user feature data corresponding to the multiple users, and inputting the user feature data as test data into the prediction system to acquire prediction results, wherein the prediction results include: the probability of click time enumeration and the probability of click content type enumeration; A result receiving module, used for receiving the prediction result from the prediction system, wherein the prediction result includes: user identification information corresponding to a plurality of users, the probability of click time enumeration corresponding to each of the user identification information, and the probability of click content type enumeration; A first selection module is used to determine the top M probabilities from the click time enumeration probabilities corresponding to each user identification information, randomly select a probability from the top M probabilities, and use the click time enumeration number corresponding to the selected probability as the message push time of the current user, wherein M is greater than or equal to 1, and M is less than or equal to the total number of click time enumeration numbers corresponding to the user identification information; The second selection module is used to determine the top N probabilities from the click content type enumeration probabilities corresponding to each user identification information, randomly select a probability from the top N probabilities, and use the click content type enumeration number corresponding to the selected probability as the message push content type of the current user, wherein N is greater than or equal to 1, and N is less than or equal to the total number of click content type enumeration numbers corresponding to the user identification information.