Model training method, push switch state prediction method, device and electronic equipment
By training a classification model on the application's multidimensional feature data, the system can predict whether users will turn off push notifications, thus solving the problem of rising rates of users turning off push notifications and improving push notification operation effectiveness and user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING QIYI CENTURY SCI & TECH CO LTD
- Filing Date
- 2022-12-30
- Publication Date
- 2026-05-19
AI Technical Summary
In existing technologies, as push notifications become more frequent, the rate at which users turn off push notifications is constantly increasing, leading to a decline in the effectiveness of push notification operations, a decrease in user activity and retention rates, and the possibility that users may miss important content.
By acquiring multidimensional feature data of the application, including push interaction data, user profile data, and application usage records, training data and labels are generated. The classification model is then trained using this data until it converges, resulting in a classification model for predicting push on/off states.
It improved the effectiveness of push notification operations, increased user activity and retention rates, and prevented users from missing important content.
Smart Images
Figure CN116010865B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of computer technology, and in particular to a model training method, a push switch state prediction method, an apparatus, and an electronic device. Background Technology
[0002] Sending push notifications has become an increasingly important operational tactic for major internet companies, used to launch apps and thus increase user activity and retention. As push notifications become more frequent and the volume of push notifications increases, the rate at which push notifications are turned off is also rising.
[0003] If users turn off push notifications, the effectiveness of push notification operations will be greatly reduced, leading to lower user activity and retention rates. Furthermore, users may miss content they want to view. Therefore, predicting the state of push notifications is a pressing issue that needs to be addressed. Summary of the Invention
[0004] To solve the above-mentioned technical problems, or at least partially solve them, this application provides a model training method, a push switch state prediction method, an apparatus, and an electronic device.
[0005] Firstly, this application provides a model training method, including:
[0006] Obtain multidimensional feature data of an application, the application having a push switch for controlling whether to push push messages to the user, the multidimensional feature data containing multiple feature data related to the user's behavior of turning off the push switch;
[0007] Obtain the on / off state of the push switch after the application pushes the push message;
[0008] Based on the multidimensional feature data, the training data corresponding to the user is determined, and the switch state is determined as the label corresponding to the training data.
[0009] Using the training data and the labels, a pre-established classification model is trained until the classification model converges, resulting in a trained classification model used to classify the on / off state of the push switch.
[0010] Optionally, the multidimensional feature data includes: push interaction data, user profile data, and application usage records. The push interaction data includes interaction records of users interacting with push messages pushed by the application. The application usage records include records of users using the application. The user profile data includes at least one attribute of the user.
[0011] Based on the multidimensional feature data, the training data corresponding to the user is determined, including:
[0012] The training data for each user is determined based on the push interaction data, user profile data, and application usage records in the multidimensional feature data.
[0013] Optionally, training data corresponding to a user is determined based on push interaction data, user profile data, and application usage records in the multidimensional feature data, including:
[0014] For each of the multidimensional feature data, the message content of the push message triggered by the user and the first trigger number of the user for each message content are determined based on the push interaction data.
[0015] Determine the content category to which the message content belongs based on the message content;
[0016] The sum of the first trigger counts of the message content included in the content category is determined as the second trigger count corresponding to the content category;
[0017] The content categories are sorted, and the second trigger counts corresponding to the content categories are sorted according to the order of the content categories. A push preference vector is generated based on the sorted second trigger counts.
[0018] The push preference vector is determined as the training data.
[0019] Optionally, training data corresponding to a user is determined based on push interaction data, user profile data, and application usage records in the multidimensional feature data, including:
[0020] For each application usage record in the multidimensional feature data, the viewing time information and viewing content information are determined based on the application usage record.
[0021] A time preference vector is generated based on the viewing time information, and a content preference vector is generated based on the viewing time information and the viewing content information.
[0022] The time preference vector and the content preference vector are determined as the training data.
[0023] Optionally, determining the training data corresponding to a user based on push interaction data, user profile data, and application usage records from the multiple multidimensional feature data further includes:
[0024] A viewing time sequence is determined based on the viewing time information, and a viewing content sequence is generated based on the viewing content information.
[0025] Based on the viewing time sequence and the viewing content sequence, users with similar viewing time and viewing content are aggregated, and the viewing habit category corresponding to the user is determined.
[0026] The category feature vector is generated based on the movie viewing habit category, and the category feature vector is determined as the training data.
[0027] Optionally, training data corresponding to a user is determined based on push interaction data, user profile data, and application usage records in the multidimensional feature data, including: 0. Determining the user's age information, geographic information, and gender information based on the user profile data;
[0028] An age feature vector is generated based on the age information, a regional feature vector is generated based on the regional information, and a gender feature vector is generated based on the gender information. The age feature vector, the regional feature vector, and the gender feature vector are then used as the training data.
[0029] Optionally, using the training data and the labels, the pre-established classification model is trained, including:
[0030] The training data is input into the classification model so that the classification model outputs the predicted state of the push switch;
[0031] If the predicted state is different from the label, adjust the model parameters of the classification model and repeat the above steps until the classification model converges.
[0032] 0. In a second aspect, this application provides a method for predicting the state of a push switch, comprising:
[0033] Obtain multidimensional user characteristic data of the applications used by the target users;
[0034] The user's multidimensional feature data is input into a classification model obtained by any of the model training methods described in the first aspect, and the predicted switch state output by the classification model is obtained.
[0035] Thirdly, this application provides a model training apparatus, comprising: 5 a first acquisition module, used to acquire multidimensional feature data of an application, the application having
[0036] A push switch used to control whether to push push messages to users, the multidimensional feature data includes multiple feature data that are associated with the user's behavior of turning off the push switch;
[0037] The second acquisition module is used to acquire the on / off state of the push switch after the application pushes the push message;
[0038] The first determining module is used to determine the training data corresponding to the user based on the multidimensional feature data, and to determine the switch state as the label corresponding to the training data.
[0039] The training module is used to train a pre-established classification model using the training data and the labels until the classification model converges, thus obtaining a trained classification model. The classification model is used to classify the on / off state of the push switch.
[0040] Optionally, the multidimensional feature data includes: push interaction data, user profile data, and application usage records. The push interaction data includes interaction records of users interacting with push messages pushed by the application. The application usage records include records of users using the application. The user profile data includes at least one attribute of the user.
[0041] The first determining module includes:
[0042] The construction unit is used to determine the training data corresponding to each user based on the push interaction data, user profile data and application usage records in the multidimensional feature data.
[0043] Optionally, the building unit includes:
[0044] The first determining subunit is used to determine the message content of the push message triggered by the user and the first number of times the user triggers each message content, based on the push interaction data in each of the multidimensional feature data.
[0045] The second determining subunit is used to determine the content category to which the message content belongs based on the message content;
[0046] The third determining subunit is used to determine the sum of the first trigger counts of the message content included in the content category as the second trigger count corresponding to the content category;
[0047] The first generation subunit is used to sort the content categories, sort the second trigger counts corresponding to the content categories according to the order of the content categories, and generate a push preference vector based on the sorted second trigger counts;
[0048] The fourth determining subunit is used to determine the push preference vector as the training data.
[0049] Optionally, the building unit includes:
[0050] The fifth determining subunit is used to determine the viewing time information and viewing content information based on the application usage record in each of the multidimensional feature data.
[0051] The second generation subunit is used to generate a time preference vector based on the movie viewing time information, and to generate a content preference vector based on the movie viewing time information and the movie viewing content information.
[0052] The sixth determining subunit is used to determine the time preference vector and the content preference vector as the training data.
[0053] Optionally, the building unit further includes:
[0054] The sorting subunit is used to determine the viewing time sequence based on the viewing time information and to generate the viewing content sequence based on the viewing content information.
[0055] The clustering subunit is used to aggregate users with similar viewing times and viewing content based on the viewing time sequence and the viewing content sequence, and to determine the viewing habit category corresponding to the user.
[0056] The seventh determining subunit is used to generate the category feature vector according to the movie viewing habit category, and determine the category feature vector as the training data.
[0057] Optionally, the building unit includes:
[0058] The eighth determining subunit is used to determine the user's age information, geographic information, and gender information based on the user profile data;
[0059] The eighth determining subunit is used to generate an age feature vector based on the age information, a regional feature vector based on the regional information, and a gender feature vector based on the gender information, and to determine the age feature vector, the regional feature vector, and the gender feature vector as the training data.
[0060] Optionally, the training module includes:
[0061] An input subunit is used to input the training data into the classification model so that the classification model outputs the predicted state of the push switch;
[0062] The adjustment subunit is used to adjust the model parameters of the classification model if the predicted state is different from the label. The above steps are repeated until the classification model converges.
[0063] Fourthly, this application provides a push switch state prediction device, comprising:
[0064] The third acquisition module is used to acquire multidimensional user feature data of the applications used by the target user.
[0065] The inference module is used to input the user's multidimensional feature data into a classification model obtained by any of the model training devices described in the third aspect, and to obtain the predicted switch state output by the classification model.
[0066] Fifthly, this application provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;
[0067] Memory, used to store computer programs;
[0068] When a processor executes a program stored in memory, it implements either the model training method described in the first aspect or the push switch state prediction method described in the second aspect.
[0069] In a sixth aspect, this application provides a computer-readable storage medium storing a program for a model training method or a push switch state prediction method, wherein when the program for the model training method or the push switch state prediction method is executed by a processor, it implements the steps of the model training method described in the first aspect or the push switch state prediction method described in the second aspect.
[0070] The technical solutions provided in this application have the following advantages compared with the prior art:
[0071] This application embodiment can automatically acquire multi-dimensional feature data and on / off status of the application, determine the training data corresponding to the user based on the multi-dimensional feature data, determine the on / off status as the label corresponding to each training data, and then use the training data and labels to train the classification model to obtain the trained classification model, which can be used to predict whether the push switch will be turned off by the user after a push message is pushed. This facilitates the formulation of a push message push strategy suitable for the user based on the predicted on / off status, improves the effectiveness of push operation, increases user activity and retention rate, and prevents users from missing the content they want to watch. Attached Figure Description
[0072] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.
[0073] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0074] Figure 1 A flowchart illustrating a model training method provided in this application embodiment;
[0075] Figure 2 A flowchart illustrating a push switch state prediction method provided in this application embodiment;
[0076] Figure 3 A structural diagram of a model training device provided in an embodiment of this application;
[0077] Figure 4 A structural diagram of a push switch state prediction device provided in an embodiment of this application;
[0078] Figure 5 This is a structural diagram of an electronic device provided in an embodiment of this application. Detailed Implementation
[0079] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0080] Since turning off the push notification switch in an application significantly reduces the effectiveness of push notification operations, lowering user activity and retention rates, and potentially causing users to miss content they want to view, this application provides a model training method, a push notification switch state prediction method, apparatus, and electronic device. These methods automatically acquire multi-dimensional feature data and switch data from various applications, generate training data and labels based on the multi-dimensional feature data, and then use the training data and labels to train a classification model. The trained classification model is used to predict whether the push notification switch will be turned off by the user after a push message is sent. This allows for the specification of appropriate push notification strategies based on the predicted switch state, improving push notification effectiveness, increasing user activity and retention rates, and preventing users from missing content they want to view.
[0081] like Figure 1 As shown, the model training method of this application embodiment can be applied to the server, which communicates with multiple applications set in the user terminal. The method may include the following steps:
[0082] Step S101: Obtain the multidimensional feature data of the application;
[0083] In this embodiment of the application, the application has a push switch for controlling whether to push push messages to the user, and the multidimensional feature data includes multiple feature data that are associated with the user's behavior of turning off the push switch.
[0084] Push notifications, also known as pop-up notifications, are a technology where the server sends real-time notifications to the client. Common use cases include message alerts, event notifications, and content push notifications. Applications have a push switch to control whether push notifications are allowed. When the push switch is on, push notifications are sent to the user. When the screen is locked, push notifications can be displayed on the lock screen; when the screen is unlocked, push notification banners can pop up from the top. When the push switch is off, the application cannot send push notifications to the user.
[0085] The multidimensional feature data includes: push interaction data, user profile data, and application usage records. The push interaction data includes interaction records of users interacting with push messages pushed by the application. The application usage records include records of users using the application. The user profile data includes at least one user attribute.
[0086] In this embodiment of the application, the application can automatically record multidimensional feature data. Every certain period of time, each application can automatically synchronize multidimensional feature data to the server; alternatively, the server can obtain multidimensional feature data from each application at certain intervals.
[0087] Step S102: Obtain the on / off state of the push switch after the application pushes the push message;
[0088] Because push notifications can be disruptive to users, they may choose to turn them off after seeing a push notification. The decision to turn off push notifications is influenced by various factors. Users of different ages and in different cities have different thresholds for the number of messages they can accept. Furthermore, different users have different content preferences; for example, some prefer light entertainment while others are interested in negative news. If a push notification contains content that users find offensive, they are highly likely to immediately turn off the notification switch. To predict whether users will turn off push notifications after seeing them, the on / off status of the push notification switches for each application can be obtained within a preset time period after push notifications are sent, serving as labels for model training.
[0089] In this embodiment of the application, the application can collect the switch status within a preset time period after the user sees the push message through automatic pingback notification. Every once in a while, each application can automatically synchronize the switch status to the server; alternatively, the server can obtain the switch status from each application at regular intervals.
[0090] Step S103: Determine the training data corresponding to the user based on the multidimensional feature data, and determine the switch state as the label corresponding to the training data;
[0091] In this step, multidimensional feature data from at least one application can be processed to obtain training data. Specifically, one or more corresponding feature vectors can be generated based on the feature data of each dimension in the multidimensional feature data (see the embodiments below for details), and the generated feature vectors are determined as training data.
[0092] Based on the on / off state of each application, labels corresponding to the training data are generated. Furthermore, when the application is on, its corresponding label is "0"; when the application is off, its corresponding label is "1".
[0093] In one embodiment of this application, training data for each user can be determined based on push interaction data, user profile data, and application usage records in the multidimensional feature data.
[0094] Step S104: Using the training data and the labels, train the pre-established classification model until the classification model converges to obtain a trained classification model. The classification model is used to classify the on / off state of the push switch.
[0095] In this step, the training data can be input into the classification model so that the classification model outputs the predicted switch state of the push switch. If the predicted switch state is different from the label, the model parameters of the classification model are adjusted, and the above steps are repeated until the classification model converges.
[0096] This application embodiment can automatically acquire multi-dimensional feature data and on / off status of the application, determine the training data corresponding to the user based on the multi-dimensional feature data, determine the on / off status as the label corresponding to each training data, and then use the training data and labels to train the classification model to obtain the trained classification model, which can be used to predict whether the push switch will be turned off by the user after a push message is pushed. This facilitates the formulation of a push message push strategy suitable for the user based on the predicted on / off status, improves the effectiveness of push operation, increases user activity and retention rate, and prevents users from missing the content they want to watch.
[0097] In another embodiment of this application, training data corresponding to a user is determined based on push interaction data, user profile data, and application usage records in the multidimensional feature data, including:
[0098] Step S201: For each of the multidimensional feature data, determine the message content of the push message triggered by the user and the first trigger number of the user for each message content based on the push interaction data.
[0099] In this embodiment of the application, the push interaction data includes the message content of the push message triggered by the user and the first trigger count of each message content by the user.
[0100] In practical applications, automatic pingback notifications can be used to collect the message content triggered after a user sees a push message, and the number of times each message content is triggered can be counted to obtain the first trigger count.
[0101] Step S202: Determine the content category to which the message content belongs based on the message content;
[0102] In this embodiment of the application, each message content can be processed based on the application's natural language, and each message content can be classified based on its meaning to obtain the content category to which the message content triggered by each user belongs.
[0103] Step S203: The sum of the first trigger counts of the message content included in the content category is determined as the second trigger count corresponding to the content category;
[0104] The second trigger count is obtained by counting the total number of times multiple message contents within each content category are triggered.
[0105] Step S204: Sort the content categories, sort the second trigger counts corresponding to the content categories according to the order of the content categories, and generate a push preference vector based on the sorted second trigger counts.
[0106] This application embodiment can generate a push preference vector according to a preset format based on each content category and its corresponding second trigger count. Furthermore, different bits of the vector can be associated with different content categories, and the value corresponding to each bit represents the second trigger count. The values of each bit are standardized, for example: [0.3, 0, 0.1...]. The corresponding interpretation is that the first bit represents celebrity entertainment news, which the user has clicked 3 times; the second bit represents technology news, which the user has not clicked; and the third bit represents sports news, which the user has clicked once.
[0107] Step S205: The push preference vector is determined as the training data.
[0108] This application embodiment can generate a push preference vector based on push interaction data in multidimensional feature data and determine it as training data for model training. By using push interaction data containing interaction records of user interaction with push messages pushed by the application to generate a push preference vector and use it to train the model, the trained model can be sensitive to the user's interaction records with push messages, understand what kind of push messages the user is interested in, and thus more accurately predict the state of push on / off.
[0109] In another embodiment of this application, training data corresponding to a user is determined based on push interaction data, user profile data, and application usage records in the multidimensional feature data, including:
[0110] Step S301: For each application usage record in the multidimensional feature data, determine the viewing time information and viewing content information based on the application usage record;
[0111] In this embodiment of the application, the viewing time information refers to the time information of the user watching videos through the application in the past period, and the viewing content information refers to the user's viewing records in the past period, such as the title of the movie watched.
[0112] In this step, you can either directly extract the viewing time information from the application usage record, or you can parse the application usage record, such as converting it to a corresponding format, and then extract the viewing time information from the format-converted application usage record. Similarly, you can directly extract the viewing content information from the application usage record, or you can parse the application usage record, such as converting it to a corresponding format, and then extract the viewing content information from the format-converted application usage record.
[0113] Step S302: Generate a time preference vector based on the viewing time information, and generate a content preference vector based on the viewing time information and the viewing content information;
[0114] In this step, statistical analysis can be performed on the user's movie-watching time information, and a time preference vector `time_prefer` can be generated based on the analysis results. For example, the time preference vector can contain multiple bits, each bit corresponding to a time period. The value of each bit represents the weight of the user's movie-watching within that time period. For example, [0.05, 0.02, 0.03...], which means that the first bit corresponds to the time period 00:00-01:00, and 0.05 represents the weight of the user's probability of watching a movie during that time period; the second bit corresponds to the time period 01:01-02:00, and 0.02 represents the weight of the user's probability of watching a movie during that time period; the third bit corresponds to the time period 02:01-03:00, and 0.03 represents the weight of the user's probability of watching a movie during that time period.
[0115] Statistical analysis can be performed on users' viewing content information, and a content preference vector (content_prefer) can be generated based on the analysis results. For example, the content preference vector can contain multiple bits, each...
[0116] Each digit corresponds to a film category. Different films belong to different film categories. The 5-digit value for each digit represents the user's viewing time for that film, for example: [8, 5, 2…]. The corresponding explanation is…
[0117] The first digit corresponds to film category A, and 8 indicates that the user's total viewing time for film content belonging to film category A is 8 hours. The second digit corresponds to film category B, and 5 indicates that the user's total viewing time for film content belonging to film category A is 5 hours. The third digit corresponds to film category C, and 2 indicates that the user's total viewing time for film content belonging to film category A is 2 hours.
[0118] Step S303: The time preference vector and the content preference vector are determined as the training data.
[0119] The embodiments of this application enable the trained model to be sensitive to the user's usage time of the application and the content viewed, thereby understanding when it is convenient for the user to view push messages and which messages contain which information.
[0120] The system can detect content that is interesting to the user without turning off push notifications, thus more accurately predicting whether push notifications are on or off.
[0121] In another embodiment of this application, training data corresponding to a user is determined based on push interaction data, user profile data, and application usage records in the multidimensional feature data, including:
[0122] Step S401: Determine the viewing time sequence based on the viewing time information, and generate the viewing content sequence based on the viewing content information;
[0123] In the above embodiments, the viewing time information corresponding to each viewing content information can be determined based on the application usage records, the viewing time sequence can be determined based on the viewing time information, and the viewing content sequence can be generated based on the viewing content information.
[0124] Step S402: Based on the viewing time sequence and the viewing content sequence, aggregate users with similar viewing times and viewing content, and determine the viewing habits corresponding to the users.
[0125] category;
[0126] In this step, users with similar viewing content and viewing time can be grouped into one category. Through clustering, at least one viewing habit category can be obtained. Each cluster category corresponds to a user's viewing habit. The viewing habit category to which the user belongs is determined as the user's corresponding viewing habit category.
[0127] Step S403: Generate the category feature vector according to the movie viewing habit category, and determine the category feature vector as the training data.
[0128] This application embodiment can cluster the viewing content sequences of multiple users, aggregating users with similar viewing habits into one category and generating a higher-dimensional category feature vector. Since the viewing behavior of a single user is relatively simple, aggregation will enrich the viewing behavior of users belonging to the same viewing habit category. Furthermore, by using richer viewing behavior to train the model, the generalization ability of the trained model can be enhanced, and the accuracy of the model in predicting the state of push notification switches can be improved.
[0129] In another embodiment of this application, training data corresponding to a user is determined based on push interaction data, user profile data, and application usage records in the multidimensional feature data, including:
[0130] Step S501: Determine the user's age, geographic location, and gender information based on the user profile data;
[0131] Step S502: Generate an age feature vector based on the age information, generate a regional feature vector based on the regional information, and generate a gender feature vector based on the gender information, and determine the age feature vector, the regional feature vector, and the gender feature vector as the training data.
[0132] In this embodiment of the application, each bit in the age feature vector corresponds to an age group, such as: minor, youth, middle-aged, etc., and the value corresponding to each bit represents the user's actual situation belonging to this category. Each bit in the geographic feature vector corresponds to a city level, such as: first-tier city, second-tier city, etc., and the value corresponding to each bit represents the user's actual situation belonging to this category. Each bit in the gender feature vector corresponds to a gender, such as: male, female, and the value corresponding to each bit represents the user's actual situation belonging to this category.
[0133] For example, users' age groups can be pre-divided into minors: [0-18], youth [18-30], middle-aged [30-50], etc. Each bit of the age feature vector corresponds to an age group, with the first bit corresponding to minors, the second bit to youth, and the third bit to middle-aged. If the user's actual age is 20 years old, the age feature vector age can be [0, 1, 0].
[0134] Each bit of the gender feature vector corresponds to a gender, with the first bit corresponding to male and the second bit corresponding to female. If the user is actually female, the gender feature sex is [0, 1].
[0135] Each bit of the regional feature vector corresponds to a city level. The first bit corresponds to a first-tier city, the second bit corresponds to a second-tier city, and the third bit corresponds to a third-tier city. If the user's location is Beijing, which is a first-tier city, then the regional feature vector city is [1, 0, 0].
[0136] In yet another embodiment of this application, as Figure 2 As shown, a method for predicting the state of a push switch is also provided, including:
[0137] Step S601: Obtain multidimensional user feature data of the application used by the target user;
[0138] The user multidimensional feature data is similar to the multidimensional feature data in the aforementioned embodiments, including: push interaction data, user profile data, and application usage records. The push interaction data includes interaction records of users interacting with push messages pushed by the application, the application usage records include records of users using the application, and the user profile data includes at least one user attribute.
[0139] In this step, in order to predict whether the target user will turn off the push notification after receiving the push message, multidimensional user feature data can be obtained from the applications used by the target user.
[0140] Step S602: Input the user's multidimensional feature data into the classification model obtained by the model training method described in any of the aforementioned method embodiments, and obtain the predicted switch state output by the classification model.
[0141] Since the classification model has been trained in the aforementioned embodiments, after inputting the user's multidimensional feature data into the classification model, the classification model can output a predicted switch state.
[0142] This application embodiment can use a classification model to automatically predict whether the push switch will be turned off by the user after a push message is sent. This makes it easier to formulate a push message push strategy suitable for the user based on the predicted switch status, thereby improving the effectiveness of push operations, increasing user activity and retention rates, and preventing users from missing the content they want to watch.
[0143] In yet another embodiment of this application, as Figure 3 As shown, a model training device is also provided, comprising:
[0144] The first acquisition module 11 is used to acquire multidimensional feature data of an application, wherein the application has a push switch for controlling whether to push push messages to the user, and the multidimensional feature data includes multiple feature data that are associated with the user's behavior of turning off the push switch;
[0145] The second acquisition module 12 is used to acquire the on / off state of the push switch after the application pushes the push message;
[0146] The first determining module 13 is used to determine the training data corresponding to the user based on the multidimensional feature data, and to determine the switch state as the label corresponding to the training data.
[0147] Training module 14 is used to train a pre-established classification model using the training data and the labels until the classification model converges, thereby obtaining a trained classification model. The classification model is used to classify the on / off state of the push switch.
[0148] Optionally, the multidimensional feature data includes: push interaction data, user profile data, and application usage records. The push interaction data includes interaction records of users interacting with push messages pushed by the application. The application usage records include records of users using the application. The user profile data includes at least one attribute of the user.
[0149] The first determining module includes:
[0150] The construction unit is used to determine the training data corresponding to each user based on the push interaction data, user profile data and application usage records in the multidimensional feature data.
[0151] Optionally, the building unit includes:
[0152] The first determining subunit is used to determine the message content of the push message triggered by the user and the first number of times the user triggers each message content, based on the push interaction data in each of the multidimensional feature data.
[0153] The second determining subunit is used to determine the content category to which the message content belongs based on the message content;
[0154] The third determining subunit is used to determine the sum of the first trigger counts of the message content included in the content category as the second trigger count corresponding to the content category;
[0155] The first generation subunit is used to sort the content categories, sort the second trigger counts corresponding to the content categories according to the order of the content categories, and generate a push preference vector based on the sorted second trigger counts;
[0156] The fourth determining subunit is used to determine the push preference vector as the training data.
[0157] Optionally, the building unit includes:
[0158] The fifth determining subunit is used to determine the viewing time information and viewing content information based on the application usage record in each of the multidimensional feature data.
[0159] The second generation subunit is used to generate a time preference vector based on the movie viewing time information, and to generate a content preference vector based on the movie viewing time information and the movie viewing content information.
[0160] The sixth determining subunit is used to determine the time preference vector and the content preference vector as the training data.
[0161] Optionally, the building unit further includes:
[0162] The sorting subunit is used to determine the viewing time sequence based on the viewing time information and to generate the viewing content sequence based on the viewing content information.
[0163] The clustering subunit is used to aggregate users with similar viewing times and viewing content based on the viewing time sequence and the viewing content sequence, and to determine the viewing habit category corresponding to the user.
[0164] The seventh determining subunit is used to generate the category feature vector according to the movie viewing habit category, and determine the category feature vector as the training data.
[0165] Optionally, the building unit includes:
[0166] The eighth determining subunit is used to determine the user's age information, geographic information, and gender information based on the user profile data;
[0167] The eighth determining subunit is used to generate an age feature vector based on the age information, a regional feature vector based on the regional information, and a gender feature vector based on the gender information, and to determine the age feature vector, the regional feature vector, and the gender feature vector as the training data.
[0168] Optionally, the training module includes:
[0169] An input subunit is used to input the training data into the classification model so that the classification model outputs the predicted state of the push switch;
[0170] The adjustment subunit is used to adjust the model parameters of the classification model if the predicted state is different from the label. The above steps are repeated until the classification model converges.
[0171] In yet another embodiment of this application, as Figure 4 As shown, a push switch state prediction device is also provided, comprising:
[0172] The third acquisition module 21 is used to acquire multidimensional user feature data of the application used by the target user.
[0173] The inference module 22 is used to input the user's multidimensional feature data into the classification model obtained by the model training device as described in the previous embodiment, and obtain the predicted switch state output by the classification model.
[0174] In another embodiment of this application, an electronic device is also provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.
[0175] Memory, used to store computer programs;
[0176] When the processor executes the program stored in the memory, it implements the model training method described in any of the foregoing embodiments, or the push switch state prediction method described in the foregoing embodiments.
[0177] The electronic device provided in this invention allows the processor to automatically acquire multi-dimensional feature data and on / off states of an application by executing a program stored in the memory. Based on the multi-dimensional feature data, the processor determines the training data corresponding to the user and identifies the on / off states as labels corresponding to each training data point. The processor then uses the training data and labels to train a classification model, resulting in a trained classification model. This model is used to predict whether the push notification switch will be turned off by the user after a push message is sent. This facilitates the development of a push message strategy suitable for the user based on the predicted on / off states, improving push operation effectiveness, increasing user activity and retention rates, and preventing users from missing content they want to view.
[0178] Furthermore, the embodiments of this application can use a classification model to automatically predict whether the push switch will be turned off by the user after the push message is pushed. This makes it easier to formulate a push message push strategy suitable for the user based on the predicted switch status, improve the effectiveness of push operations, increase user activity and retention rates, and prevent users from missing the content they want to watch.
[0179] The communication bus 1140 mentioned in the above-mentioned electronic device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 1140 can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.
[0180] The communication interface 1120 is used for communication between the above-mentioned electronic device and other devices.
[0181] The memory 1130 may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.
[0182] The processor 1110 mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.
[0183] In another embodiment of this application, a computer-readable storage medium is provided, on which a program for a model training method or a push switch state prediction method is stored. When the program for the model training method or the push switch state prediction method is executed by a processor, it implements the steps of the model training method or the push switch state prediction method described in any of the foregoing method embodiments.
[0184] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0185] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.
Claims
1. A model training method, characterized in that, include: Obtain multidimensional feature data of an application, the application having a push switch for controlling whether to push push messages to the user, the multidimensional feature data containing multiple feature data related to the user's behavior of turning off the push switch; Obtain the on / off state of the push switch after the application pushes the push message; Based on the multidimensional feature data, the training data corresponding to the user is determined, and the switch state is determined as the label corresponding to the training data. The multidimensional feature data includes: push interaction data, user profile data, and application usage records. The push interaction data includes interaction records of users interacting with push messages pushed by the application. The application usage records include records of users using the application. The user profile data includes at least one attribute of the user. Based on the multidimensional feature data, the training data corresponding to the user is determined, including: Based on the push interaction data, user profile data and application usage records in the multidimensional feature data, the training data corresponding to each user is determined. Based on the push interaction data, user profile data, and application usage records in the multidimensional feature data, the training data corresponding to the user is determined, including: For each of the multidimensional feature data, the message content of the push message triggered by the user and the first trigger number of the user for each message content are determined based on the push interaction data. Determine the content category to which the message content belongs based on the message content; The sum of the first trigger counts of the message content included in the content category is determined as the second trigger count corresponding to the content category; The content categories are sorted, and the second trigger counts corresponding to the content categories are sorted according to the order of the content categories. A push preference vector is generated based on the sorted second trigger counts. The push preference vector is determined as the training data; Using the training data and the labels, a pre-established classification model is trained until the classification model converges, resulting in a trained classification model used to classify the on / off state of the push switch.
2. The model training method according to claim 1, characterized in that, Based on the push interaction data, user profile data, and application usage records in the multidimensional feature data, the training data corresponding to the user is determined, including: For each application usage record in the multidimensional feature data, the viewing time information and viewing content information are determined based on the application usage record. A time preference vector is generated based on the viewing time information, and a content preference vector is generated based on the viewing time information and the viewing content information. The time preference vector and the content preference vector are determined as the training data.
3. The model training method according to claim 2, characterized in that, The training data for determining the user based on push interaction data, user profile data, and application usage records from multiple multidimensional feature data sources also includes: A viewing time sequence is determined based on the viewing time information, and a viewing content sequence is generated based on the viewing content information. Based on the viewing time sequence and the viewing content sequence, users with similar viewing times and viewing content are aggregated, and the viewing habit category corresponding to the user is determined. Generate category feature vectors based on movie viewing habits, and use these category feature vectors as the training data.
4. The model training method according to claim 1, characterized in that, Based on the push interaction data, user profile data, and application usage records in the multidimensional feature data, the training data corresponding to the user is determined, including: Based on the user profile data, determine the user's age, geographic location, and gender information; An age feature vector is generated based on the age information, a regional feature vector is generated based on the regional information, and a gender feature vector is generated based on the gender information. The age feature vector, the regional feature vector, and the gender feature vector are then used as the training data.
5. The model training method according to claim 1, characterized in that, Using the training data and the labels, the pre-established classification model is trained, including: The training data is input into the classification model so that the classification model outputs the predicted state of the push switch; If the predicted state is different from the label, adjust the model parameters of the classification model and repeat the training steps of the classification model until the classification model converges.
6. A method for predicting the state of a push switch, characterized in that, include: Obtain multidimensional user characteristic data of the applications used by the target users; The user's multidimensional feature data is input into the classification model obtained by any of the model training methods described in claims 1 to 5, and the predicted switch state output by the classification model is obtained.
7. A model training device, characterized in that, include: The first acquisition module is used to acquire multidimensional feature data of an application, wherein the application has a push switch for controlling whether to push push messages to the user, and the multidimensional feature data includes multiple feature data that are associated with the user's behavior of turning off the push switch; The second acquisition module is used to acquire the on / off state of the push switch after the application pushes the push message; The first determining module is used to determine the training data corresponding to the user based on the multidimensional feature data, and to determine the switch state as the label corresponding to the training data. The multidimensional feature data includes: push interaction data, user profile data, and application usage records. The push interaction data includes interaction records of users interacting with push messages pushed by the application. The application usage records include records of users using the application. The user profile data includes at least one attribute of the user. Determining training data corresponding to a user based on the multidimensional feature data includes: determining training data corresponding to each user based on the push interaction data, user profile data, and application usage records in the multidimensional feature data; determining training data corresponding to a user based on the push interaction data, user profile data, and application usage records in the multidimensional feature data includes: for each push interaction data in the multidimensional feature data, determining the message content of the push message triggered by the user and the first trigger count of each message content; determining the content category to which the message content belongs based on the message content; determining the second trigger count corresponding to the content category by summing the first trigger counts of the message content included in the content category; sorting the content categories; sorting the second trigger counts corresponding to the content categories according to the order of the content categories; generating a push preference vector based on the sorted second trigger counts; and determining the push preference vector as the training data. The training module is used to train a pre-established classification model using the training data and the labels until the classification model converges, thus obtaining a trained classification model. The classification model is used to classify the on / off state of the push switch.
8. A push switch state prediction device, characterized in that, include: The third acquisition module is used to acquire multidimensional user feature data of the applications used by the target user. The inference module is used to input the user's multidimensional feature data into the classification model obtained by the model training device as described in claim 7, and obtain the predicted switch state output by the classification model.
9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the model training method according to any one of claims 1 to 5, or the push switch state prediction method according to claim 6.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for a model training method or a push switch state prediction method. When the program for the model training method or the push switch state prediction method is executed by a processor, it implements the steps of the model training method according to any one of claims 1-5 or the push switch state prediction method according to claim 6.