A method, device and electronic device for predicting user churn probability

Abstract

The present application relates to the field of data processing technology, and provides a method, device and electronic device for predicting the probability of user churn. The method obtains each current refresh time of a target user refreshing a recommended page on a target terminal and the historical behavior record of the recommended page on the target terminal in a historical time period; analyzes the behavioral characteristics of linear behaviors occurring in the historical time period to obtain the first activity score of the target user at the current refresh time; processes the behavioral characteristics of nonlinear behaviors to obtain the second activity score of the target user in the historical time period; determines the daily activity sequence of the target user based on the first activity score and the corresponding second activity score at each current refresh time; predicts the daily activity sequence based on a time series large language model to obtain the user churn probability of the target user in a preset future time period. The method improves the real-time and accuracy of predicting the probability of user churn.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method, device and electronic device for predicting user churn probability. Background Art

[0002] With the rapid development of network technology, various websites, APPs and other network products have emerged to meet the experience needs of different users.

[0003] At present, these network products are not necessarily in a stable operation and maintenance state after they go online. Users' interests and behaviors will change over time, and existing recommendation algorithms often cannot capture these changes in a timely manner, resulting in recommendations that do not match users' actual needs, thereby accelerating user churn.

[0004] However, existing solutions for analyzing user churn behavior cannot effectively predict user churn behavior and generally fail to fully consider the dynamic characteristics of user behavior and individual differences, which limits the accuracy of the prediction. Summary of the invention

[0005] The purpose of the embodiments of the present application is to provide a method, device and electronic device for predicting the probability of user churn, which is used to monitor multi-dimensional user behavior data such as user activity patterns and login status in real time, and use advanced large language models to predict the probability of user churn, so as to more finely capture the changing trends of user behavior and improve the real-time and accuracy of the prediction.

[0006] In a first aspect, a method for predicting user churn probability is provided, and the method may include:

[0007] Obtaining each current refresh time of the target user refreshing the recommended page on the target terminal and the historical behavior record of the target user on the recommended page on the target terminal in the historical time period; the historical behavior record includes linear behaviors that change linearly with time and non-linear behaviors that do not change linearly with time; each historical time period is a historical time period adjacent to the corresponding current refresh time;

[0008] For any current refresh time, analyze the behavior characteristics of linear behaviors that occurred in the historical time period to obtain the first activity score of the target user at the current refresh time;

[0009] Processing the behavior characteristics of the nonlinear behavior to obtain a second activity score of the target user in a historical time period;

[0010] Determine a daily activity sequence of the target user based on the first activity score and the corresponding second activity score at each current refresh time, wherein the daily activity sequence includes a correspondence between each natural day arranged in chronological order and the corresponding target activity score;

[0011] Based on the time series large language model, the daily level activity sequence is predicted to obtain the user churn probability of the target user in a preset future time period, and the preset future time period is a time period starting from the last current refresh time among the current refresh times.

[0012] In a possible implementation, the behavior characteristics of the linear behavior include the behavior occurrence time of the linear behavior;

[0013] Analyze the behavior characteristics of linear behaviors that occurred in the historical time period to obtain the first activity score of the target user at the current refresh time, including:

[0014] Using a preset behavior activity algorithm, the current refresh time and the behavior occurrence time of the linear behavior in the historical time period are processed to obtain an initial activity score corresponding to the linear behavior at the current refresh time;

[0015] Based on the configured time aggregation weights of each linear behavior, the initial activity scores of the corresponding linear behaviors are aggregated to obtain the first activity score of the target user at the current refresh time.

[0016] In a possible implementation, the linear behavior is a reading behavior; and the preset behavior activity algorithm is expressed as:

[0017]

[0018] in, The current refresh time when the target user refreshes the page. represents the first activity score of the i-th reading behavior at time T, They are the configuration parameters, is the time when the i-th reading behavior occurs, , n represents the number of reading behaviors in the historical period.

[0019] In a possible implementation, the behavior characteristics of the nonlinear behavior include the number of interactions, number of login days, and current refresh times of the target user;

[0020] The calculation formula of the second activity score is expressed as:

[0021]

[0022] in, Indicates the second activity score of the target user's j-th nonlinear behavior in the historical time period, For the configuration parameters, , It represents the number of the j-th nonlinear behavior, j=1 represents interactive nonlinear behavior, j=2 represents logged-in nonlinear behavior, and j=3 represents refresh nonlinear behavior.

[0023] In a possible implementation, determining the target user's daily level activity sequence based on the first activity score and the corresponding second activity score at each current refresh time includes:

[0024] Using the trained activity estimation model, the first activity score and the corresponding second activity score at each current refresh time are processed to obtain a real-time activity sequence of the target user, wherein the real-time activity sequence includes a corresponding relationship between each current refresh time and the real-time activity score;

[0025] According to the aggregation granularity of natural days, the real-time activity scores in the real-time activity sequence are aggregated and calculated to obtain a day-level activity sequence.

[0026] In a possible implementation, before predicting the day-level activity sequence based on the time series large language model, the method further includes:

[0027] According to a preset convolution algorithm, a preset time convolution kernel is used to perform a convolution operation on the target activity score in the day-level activity sequence to obtain a new day-level activity sequence; the new day-level activity sequence includes the correspondence between each natural day arranged in chronological order and the corresponding new activity score.

[0028] In a possible implementation, based on the time series large language model, the daily level activity sequence is predicted to obtain the user churn probability of the target user in a preset future time period, including:

[0029] Determining a scaling factor based on a non-empty target activity score in the day-level activity sequence;

[0030] Using the scaling factor, normalizing the non-empty target activity scores in the daily activity sequence to obtain a normalized activity sequence;

[0031] Use torch.bucketize function to map the normalized target activity scores in the normalized activity sequence into a set of predefined buckets; each predefined bucket has the same size and has its own bucket boundary value;

[0032] Determine a token sequence based on the bucket boundary values ​​of the predefined buckets and the number of configured sequence special tags;

[0033] The token sequence is input into the time series large language model to obtain the user churn probability of the target user in a preset future time period.

[0034] In a possible implementation, after obtaining the user churn probability of the target user in a preset future time period, the method further includes:

[0035] Based on the user churn probability, the control strategies in the recall stage, the rough ranking stage, the fine ranking stage and the re-ranking stage of the recommendation system are adjusted respectively to obtain new control strategies for the corresponding stages;

[0036] Among them, the new control strategy in the recall stage includes adjusting the recall quantity of the content that the target user is interested in; the new control strategy in the rough ranking stage includes adjusting the weight value of the target user's consumption behavior; the new control strategy in the fine ranking stage includes adjusting the weight value that characterizes the user value and content value of the target user; the new control strategy in the rearrangement stage includes adjusting the display order of the recommended content that the target user is interested in.

[0037] In a second aspect, a device for predicting user churn probability is provided, and the device may include:

[0038] An acquisition unit, used to acquire each current refresh time when a target user refreshes a recommended page on a target terminal and a historical behavior record of the target user on the recommended page on the target terminal in a historical time period; the historical behavior record includes a linear behavior that changes linearly with time and a non-linear behavior that does not change linearly with time; each historical time period is a historical time period adjacent to the corresponding current refresh time;

[0039] An analysis unit, configured to analyze the behavior characteristics of linear behaviors occurring in a historical time period at any current refresh time, and obtain a first activity score of the target user at the current refresh time;

[0040] A processing unit, configured to process the behavior characteristics of the nonlinear behavior to obtain a second activity score of the target user in a historical time period;

[0041] a determining unit, configured to determine a day-level activity sequence of the target user based on the first activity score and the corresponding second activity score at each current refresh time, wherein the day-level activity sequence includes a correspondence between each natural day arranged in chronological order and the corresponding target activity score;

[0042] A prediction unit is used to predict the daily level activity sequence based on a time series large language model to obtain the user churn probability of the target user in a preset future time period, where the preset future time period is a time period starting from the last current refresh time among the current refresh times.

[0043] In a third aspect, an electronic device is provided, the electronic device comprising a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus;

[0044] Memory, used to store computer programs;

[0045] The processor is used to implement any method step described in the first aspect when executing the program stored in the memory.

[0046] In a fourth aspect, a computer-readable storage medium is provided, wherein a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, any method step described in the first aspect is implemented.

[0047] The present application provides a method for predicting the probability of user churn, in which each current refresh time of a target user refreshing a recommended page on a target terminal and the historical behavior record of the target user on the recommended page on the target terminal in a historical time period are obtained; the historical behavior record includes linear behaviors that change linearly with time and nonlinear behaviors that do not change linearly with time; each historical time period is a historical time period adjacent to the corresponding current refresh time; for any current refresh time, the behavior characteristics of the linear behaviors occurring in the historical time period are analyzed to obtain the first activity score of the target user at the current refresh time; the behavior characteristics of the nonlinear behaviors are processed to obtain the second activity score of the target user in the historical time period; based on the first activity score of each current refresh time and the corresponding second activity score, the daily activity sequence of the target user is determined, and the daily activity sequence includes the corresponding relationship between each natural day arranged in chronological order and the corresponding target activity score; based on the time series large language model, the daily activity sequence is predicted to obtain the user churn probability of the target user in a preset future time period, and the preset future time period is the time period with the last current refresh time as the starting time in each current refresh time. The method improves the real-time and accuracy of predicting the probability of user churn. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0049] Figure 1 A flowchart of a method for predicting user churn probability provided in an embodiment of the present application;

[0050] Figure 2 A schematic diagram of the structure of a device for predicting user churn probability provided in an embodiment of the present application;

[0051] Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application. Unless otherwise defined, the technical terms or scientific terms used in this application should be the common meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in this application do not indicate any order, quantity or importance, but are only used to distinguish different components. "Including" or "including" and other similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connected", "coupled" or "connected" and other similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0053] In mobile apps, user retention is a crucial indicator that directly affects the long-term effectiveness and commercial value of the system. However, existing recommendation systems often focus on improving short-term user engagement and satisfaction, while ignoring the importance of long-term user retention. As a result, even recommendation systems that perform well in the short term may face the problem of gradual user loss, thus affecting the overall effectiveness and sustainable development of the system.

[0054] Existing technologies use a variety of different methods to predict the probability of user churn, including:

[0055] 1. Methods based on historical behavior analysis: This type of method mainly relies on users’ historical behavior data (such as browsing and consumption records) and predicts the probability of user churn through statistical analysis or simple machine learning models.

[0056] 2. User portrait-based method: This method builds user portraits to understand users' interest preferences, social attributes, etc., and combines this information to predict the probability of user churn.

[0057] 3. Methods based on time series analysis: This type of method uses time series analysis techniques (such as ARIMA, state space model, etc.) to capture the changing patterns of user behavior over time and predict future behavior patterns.

[0058] 4. Methods based on deep learning: This type of method uses deep learning technology (such as recurrent neural network RNN, long short-term memory network LSTM, etc.) to automatically extract the complex characteristics of user behavior and then predict the probability of user churn.

[0059] Although existing technologies can predict user churn to a certain extent, they usually fail to fully consider the dynamic characteristics of user behavior and individual differences, so there is still room for improvement in the accuracy of predicting user churn probability. In addition, many methods are limited by the amount of data and computing resources required, and also face certain challenges in practical applications.

[0060] In order to solve the above problems, the present application provides a method for predicting the probability of user churn, in which when the target user refreshes the recommendation page, the target terminal automatically collects the user behavior data and calculates the user activity score to evaluate the activity score of the target user at the current refresh time. After that, by analyzing the user's activity score data, the time series large language model is used to predict the possibility of future churn. The time series large language model combines time series analysis technology and natural language processing technology to better understand the changes in user behavior over time, especially using the powerful prediction ability of the large language model (LLM) to achieve accurate user churn prediction.

[0061] Furthermore, based on the estimated user churn probability, the probability is applied to each stage module of the recommendation system through a personalized control strategy to improve the user retention rate. Each stage module may include a recall control system, a rough sorting control system, a fine sorting control system, and a re-sorting control system.

[0062] The preferred embodiments of the present application are described below in conjunction with the drawings in the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application. In addition, the embodiments and features in the embodiments of the present application may be combined with each other if there is no conflict.

[0063] Figure 1 The following is a flow chart of a method for predicting the probability of user churn provided in an embodiment of the present application. Figure 1 As shown, the method may include:

[0064] Step S110: obtaining each current refresh time when the target user refreshes the recommended page on the target terminal and the historical behavior record of the target user on the recommended page on the target terminal within the historical time period corresponding to the current refresh time.

[0065] In the specific implementation, after obtaining the current refresh time T of the target user refreshing the recommended page on the target terminal, for any current refresh time T, the user portrait database is called to obtain the historical behavior records of the target user corresponding to each current refresh time on the recommended page on the target terminal within the historical time period.

[0066] The historical time period is a historical time period adjacent to the corresponding current refresh time. The historical behavior record may include linear behaviors that change linearly with time (such as reading behaviors of reading content) and non-linear behaviors that do not change linearly with time (such as refresh behaviors, user interaction behaviors, and login behaviors, etc.).

[0067] Step S120: Analyze the behavior characteristics of each linear behavior occurring in the historical time period to obtain the first activity score of the target user at the current refresh time of the corresponding linear behavior.

[0068] The behavior characteristics of the linear behavior may include the time when the linear behavior occurs.

[0069] In the specific implementation, for any current refresh time T, a preset behavior activity algorithm is used to process the current refresh time and the behavior occurrence time of the linear behavior in the corresponding historical time period to obtain the initial activity score corresponding to the linear behavior at the current refresh time. Afterwards, based on the configured time aggregation weight of each linear behavior, the initial activity score of the corresponding linear behavior is aggregated to obtain the first activity score of the target user at the current refresh time.

[0070] Specifically: First, obtain the time when each reading behavior occurred in the historical time period to obtain a historical reading behavior set. For example, assuming that the user exist Always read the content , then the user A collection of historical reading behaviors in the last month , is the reading content corresponding to the i-th reading behavior The time point of the moment.

[0071] Secondly, for any reading behavior, a preset behavior activity algorithm is used to process the current refresh time and the behavior occurrence time of the reading behavior in the corresponding historical time period to obtain the initial activity score corresponding to the reading behavior at the current refresh time. Among them, the behavior occurrence time of the reading behavior can be understood as the reading start time of the reading behavior.

[0072] Then, the product of the configured time aggregation weight of each linear behavior and the initial activity score of the corresponding linear behavior is accumulated to obtain the first activity score of the target user at the current refresh time. The first activity score G represents the activity score of the target user at the current refresh time after the historical time period.

[0073] In an example, when the linear behavior is reading behavior, the calculation formula of the first activity score G can be expressed as:

[0074]

[0075] in, represents the time aggregation weight of the i-th reading behavior, and its value can be set to , you can also set more complex exponential weights, or fixed values, such as , this application is not limited here.

[0076] In an example, when the linear behavior is reading behavior, the preset behavior activity algorithm is expressed as:

[0077]

[0078] in, The current refresh time when the target user refreshes the page. represents the first activity score of the i-th reading behavior at time T, They are the configuration parameters, is the time when the i-th reading behavior occurs, , n represents the number of reading behaviors in the historical period.

[0079] Step S130: Process the behavior characteristics of the nonlinear behavior to obtain a second activity score of the target user in a historical time period.

[0080] The behavior characteristics of non-linear behavior may include the number of interactions, login days, and current refresh times of the target user. Accordingly, the calculation formula for the second activity score may be expressed as:

[0081]

[0082] in, Indicates the second activity score of the target user's j-th nonlinear behavior in the historical time period, For the configuration parameters, , It represents the number of the j-th nonlinear behavior, j=1 represents interactive nonlinear behavior, j=2 represents logged-in nonlinear behavior, and j=3 represents refresh nonlinear behavior.

[0083] Step S140: Determine the daily activity sequence of the target user based on the first activity score and the corresponding second activity score at each current refresh time.

[0084] In the specific implementation, the trained activity estimation model is used to process the first activity score G and the corresponding second activity score Z at each current refresh time to obtain the real-time activity sequence of the target user. The real-time activity sequence may include the corresponding relationship between each current refresh time and the real-time activity score; the activity estimation model is a linear regression model.

[0085] Afterwards, the real-time activity scores at each current refresh time in the real-time activity sequence are aggregated and calculated according to the aggregation granularity of the natural day to obtain the daily level activity sequence. The daily level activity sequence may include the corresponding relationship between each natural day arranged in chronological order and the corresponding target activity score.

[0086] In one example, for any historical time period corresponding to the current refresh time T, the corresponding second activity score Z including the activity score of the interactive behavior can be obtained. , the activity score of login behavior and the activity score of the current refresh behavior When , the calculation formula of the real-time activity score s at the current refresh time T can be expressed as:

[0087]

[0088] in, and is the weight parameter trained by the model, The value range is .

[0089] After calculating the real-time activity score s at each current refresh time, the target user's ID and the corresponding real-time activity sequence can be updated in real time to the user activity score database through the Kafka message sending mechanism. That is, the user activity database stores: user ID, each current refresh time T and the corresponding real-time activity score s. As time goes by, the user Activity score Will be continuously recorded, thus forming a real-time activity sequence This step ensures the real-time and accuracy of the user activity score information, so that the recommendation strategy can be adjusted in time according to the latest activity score information, thereby improving the relevance of the recommendation and user satisfaction.

[0090] Furthermore, after sorting the real-time activity scores in the obtained real-time activity sequence in chronological order, the real-time activity scores are aggregated and calculated according to the aggregation granularity of natural days (days) to obtain the target activity scores corresponding to each natural day, thereby generating a day-level activity sequence S, that is, the day-level activity sequence may include the correspondence between each natural day arranged in chronological order and the corresponding target activity score.

[0091] In one example, to calculate the target activity score for day d The calculation formula is as follows:

[0092]

[0093] in, Indicates user In the Day's The real-time activity score s at the moment, Indicates user In the The number of activity scores for the day.

[0094] Step S150: Based on the time series large language model, the daily activity sequence is predicted to obtain the user churn probability of the target user in a preset future time period.

[0095] Since the data format of the target activity score in the daily activity sequence S is different from the data format of the input data of the time series large language model, the sequence needs to be tokenized, that is, the numerical sequence is converted into a data token form suitable for processing by the large language model. The conversion method includes three steps: calculating the scaling factor, scaling context, and bucketing.

[0096] (1) Calculate the scaling factor :

[0097] Determine the scaling factor based on the non-empty target activity scores in the daily activity sequence S. Specifically, obtain the cumulative sum of the absolute values ​​of all non-empty target activity scores in the daily activity sequence S, and then divide it by the quotient of the number of all non-empty target activity scores;

[0098] If the calculated quotient is greater than 0, the quotient is determined as the scaling factor. ; If the calculated quotient is not greater than 0, the scaling factor Set to 1.0 to ensure the validity of subsequent calculations, the scaling factor The calculation formula is as follows:

[0099]

[0100] in, is the number of all non-empty activity scores, is the kth non-empty target activity score in the daily activity sequence S.

[0101] (2) Scaling context:

[0102] The scaling factor is used to normalize the non-empty target activity scores in the daily activity sequence to obtain the corresponding standardized target activity scores, thereby obtaining a standardized activity sequence; specifically, the standardized target activity score is the corresponding non-empty target activity score and the scaling factor The quotient of the division.

[0103] (3) Bucketing:

[0104] The torch.bucketize function is used to map the standardized target activity scores in the standardized activity sequence into a set of predefined buckets; each predefined bucket has the same size and has its own bucket boundary value; the token sequence is determined based on the bucket boundary value of each predefined bucket and the number of configured target sequence tokens.

[0105] Specifically, torch.bucketize function is used to map the scaled sequence into a set of predefined buckets. The number of sequence-specific markers also needs to be considered during the mapping process. Sequence-specific markers refer to special markers that are characteristic of the input sequence of the time-series large language model, such as the sequence start marker [CLS], sequence end marker [SEP], etc.

[0106] First, determine the bucket boundary value of each predefined bucket (for example, after determining that the size of each bucket is approximately 0.0073, determine the bucket boundary value of each predefined bucket), which can be determined by uniform distribution or based on the characteristics of data distribution.

[0107] Second, using the torch.bucketize function, the scaled sequences are mapped into corresponding predefined buckets.

[0108] Afterwards, the bucket index determined by the bucket boundary value of each predefined bucket Convert to the corresponding token and add the number of special tags in the sequence to get the user Token sequence. Bucket index The calculation formula can be expressed as:

[0109]

[0110] in, is a list of bucket boundary values, is the number of special tokens in the sequence.

[0111] Finally, the token sequence is input into the time series large language model to obtain the user churn probability of the target user in a preset future time period. The preset future time period is the time period starting from the last current refresh time among the current refresh times.

[0112] In some embodiments, in order to better capture the changing pattern of the user activity score, before executing step S150, a time convolution kernel may be used according to a preset convolution algorithm. , performing a convolution operation on the target activity score in the daily activity sequence, this process can not only extract features, but also effectively smooth the noise in the daily activity sequence, and obtain a smoother new user activity sequence. Then, perform step S150.

[0113] Among them, the preset convolution algorithm can be expressed as:

[0114] Set the convolution kernel The length is , then the temporal convolution kernel Can write , then based on the new user activity sequence , the target user is the first in the sequence The new activity score for the day is , the calculation formula can be:

[0115]

[0116] in, The first in the day-level activity sequence The target activity score for the day.

[0117] In some embodiments, after obtaining the user churn probability of the target user within a preset future time period, the control strategies in the recall stage, rough ranking stage, fine ranking stage and re-ranking stage of the recommendation system can be adjusted based on the user churn probability to obtain new control strategies for the corresponding stages; wherein, the new control strategy for the recall stage may include adjusting the recall quantity of content that the target user is interested in; the new control strategy for the rough ranking stage may include adjusting the weight value of the target user's consumption behavior; the new control strategy for the fine ranking stage may include adjusting the weight value that characterizes the user value and content value of the target user; the new control strategy for the re-ranking stage may include adjusting the display order of recommended content that the target user is interested in.

[0118] Based on the probability of user churn, these probabilities are applied to each stage of the recommendation system through personalized control strategies to improve user retention. The adjustment systems corresponding to each stage can include: recall control system, rough sorting control system and fine sorting control system; the strategy for adjusting the display order of content types that target users are interested in can be applied in the re-sorting control system.

[0119] (1) Recall control system

[0120] Responsible for adjusting the recall quantity of recommended resources according to user preferences. By analyzing the user's historical behavior data, identify the user's preferred resource types, and increase the recall quantity of the corresponding type of resources accordingly.

[0121] Strategy: For users with a high probability of churn, increase the number of recalls based on the resource types that the user prefers. For resource types that the user is less interested in, appropriately reduce the number of recalls.

[0122] (2) Coarse discharge control system

[0123] It plays a role in the rough ranking stage, optimizing the recommended content by adjusting the multi-objective fusion weights in a personalized way. For different users, the multi-objective (CTR, likes, favorites, shares, creations, comments, and personal page access) fusion weights are adjusted to make the recommendation system more inclined to recommend content that users like.

[0124] Strategy: For users with a higher risk of churn (i.e., a higher probability of user churn), increase the weights corresponding to the CTR, likes, favorites, and sharing target scores in the multi-target set, and reduce the weights of the creation, comment, and entry to personal pages target scores, so that the system is more inclined to recommend content that users like.

[0125] (3) Precision control system

[0126] In the refinement stage, there are four comprehensive value goals: user value, creative value, content value, and commercial value. By increasing the weights of user value and content value goals and reducing the weights of commercial value and creative value goals, we can reduce the recommendations of content that users are not interested in or may cause loss.

[0127] Strategy: For users with a higher risk of churn (i.e., a higher probability of user churn), reduce the weight of ecological goals (such as recommendation diversity, novelty, etc.) and the weight of commercial value goals (promotion, advertising, etc.); strengthen the value of user-preferred content to improve user satisfaction.

[0128] (4) Rearrangement control system

[0129] Responsible for adjusting business strategies and commercial strategies in the final presentation stage of recommendation results to ensure that the recommended content can both meet user needs and reduce negative impacts.

[0130] Strategy: For users with a high probability of churn, adjust the order of recommended content display to prioritize the content that the user is more interested in. Reduce the display of advertisements or other commercial content that may cause user disgust.

[0131] Through the coordinated work of the above four control systems, personalized recommendations can be effectively adjusted based on the probability of user churn, thereby improving user satisfaction and retention rate, reducing user churn rate, and bringing better business results to the App.

[0132] Alternatively, for potential churn users with a higher probability of churn, additional incentives such as limited-time discounts, membership privileges, and point rewards can be provided to enhance user stickiness.

[0133] In some embodiments, in order to further improve the accuracy of the prediction, after obtaining the user churn probability of the target user in a preset future time period, the attribute information of the target user (such as basic information such as age, gender, occupation, etc.), the attribute information of the target user's friends in the social network, and the content that the target user and each friend are interested in can be obtained.

[0134] Based on the configured first cluster center and first cluster radius, the attribute information of the target user and the attribute information of the friends are clustered to obtain target friend information similar to the target user, and based on the configured second cluster center and second cluster radius, the content followed by the target user and the content followed by each friend are clustered to obtain target content followed similar to the content followed by the target user;

[0135] Then, find out whether the target content exists in the content of interest corresponding to the target friend information;

[0136] If so, the target user churn probability corresponding to the target friend information corresponding to the target content is obtained, and the mean of the target user churn probability is determined as the final user churn probability of the target user in a preset future time period.

[0137] If not, the user churn probability obtained in step S150 is determined as the final user churn probability of the target user in a preset future time period.

[0138] In some embodiments, after obtaining a new content distribution strategy, a quick response mechanism can be established to promptly collect feedback information from target users regarding the new content distribution strategy, and further adjust the new content distribution strategy accordingly, so that users feel that their opinions are valued, which can effectively improve satisfaction.

[0139] Corresponding to the above method, the embodiment of the present application also provides a device for predicting the probability of user churn, such as Figure 2 As shown, the device comprises:

[0140] The acquisition unit 210 is used to acquire each current refresh time when the target user refreshes the recommended page on the target terminal and the historical behavior record of the target user on the recommended page on the target terminal in the historical time period; the historical behavior record includes linear behaviors that change linearly with time and non-linear behaviors that do not change linearly with time; each historical time period is a historical time period adjacent to the corresponding current refresh time;

[0141] The analysis unit 220 is used to analyze the behavior characteristics of the linear behaviors occurring in the historical time period at any current refresh time to obtain the first activity score of the target user at the current refresh time;

[0142] A processing unit 230 is used to process the behavior characteristics of the nonlinear behavior to obtain a second activity score of the target user in a historical time period;

[0143] A determination unit 240 is configured to determine a day-level activity sequence of the target user based on the first activity score and the corresponding second activity score at each current refresh time, wherein the day-level activity sequence includes a correspondence between each natural day arranged in chronological order and the corresponding target activity score;

[0144] The prediction unit 250 is used to predict the daily level activity sequence based on the time series large language model to obtain the user churn probability of the target user in a preset future time period, and the preset future time period is a time period with the last current refresh time as the starting time among each current refresh time.

[0145] The functions of the various functional units of the user churn probability prediction device provided in the above-mentioned embodiments of the present application can be realized through the above-mentioned method steps. Therefore, the specific working process and beneficial effects of each unit in the user churn probability prediction device provided in the embodiments of the present application will not be repeated here.

[0146] The present application also provides an electronic device, such as Figure 3 As shown, it includes a processor 310 , a communication interface 320 , a memory 330 and a communication bus 340 , wherein the processor 310 , the communication interface 320 , and the memory 330 communicate with each other via the communication bus 340 .

[0147] Memory 330, for storing computer programs;

[0148] The processor 310 is used to execute the program stored in the memory 330 to implement the following steps:

[0149] Obtaining each current refresh time of the target user refreshing the recommended page on the target terminal and the historical behavior record of the target user on the recommended page on the target terminal in the historical time period; the historical behavior record includes linear behaviors that change linearly with time and non-linear behaviors that do not change linearly with time; each historical time period is a historical time period adjacent to the corresponding current refresh time;

[0150] For any current refresh time, analyze the behavior characteristics of linear behaviors that occurred in the historical time period to obtain the first activity score of the target user at the current refresh time;

[0151] Processing the behavior characteristics of the nonlinear behavior to obtain a second activity score of the target user in a historical time period;

[0152] Determine a daily activity sequence of the target user based on the first activity score and the corresponding second activity score at each current refresh time, wherein the daily activity sequence includes a correspondence between each natural day arranged in chronological order and the corresponding target activity score;

[0153] Based on the time series large language model, the daily level activity sequence is predicted to obtain the user churn probability of the target user in a preset future time period, and the preset future time period is a time period starting from the last current refresh time among the current refresh times.

[0154] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The communication bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0155] The communication interface is used for communication between the above electronic device and other devices.

[0156] The memory may include a random access memory (RAM) or a non-volatile memory (NVM), such as at least one disk memory. Optionally, the memory may also be at least one storage device located away from the aforementioned processor.

[0157] The above-mentioned processor 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, discrete hardware components.

[0158] The implementation methods and beneficial effects of the components of the electronic device in the above embodiments to solve the problems can be seen in Figure 1 The various steps in the illustrated embodiment are implemented, therefore, the specific working process and beneficial effects of the electronic device provided by the embodiment of the present application are not repeated here.

[0159] In another embodiment provided in the present application, a computer-readable storage medium is provided, in which instructions are stored. When the computer-readable storage medium is executed on a computer, the computer executes the method for predicting the user churn probability described in any of the above embodiments.

[0160] In another embodiment provided by the present application, a computer program product including instructions is also provided, which, when executed on a computer, enables the computer to execute the method for predicting the user churn probability described in any of the above embodiments.

[0161] Those skilled in the art will appreciate that the embodiments in the present application can be provided as methods, systems, or computer program products. Therefore, the embodiments in the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the embodiments in the present application can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0162] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0163] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0164] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0165] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present application.

[0166] Obviously, those skilled in the art can make various changes and modifications to the embodiments in the present application without departing from the spirit and scope of the embodiments in the present application. Thus, if these modifications and variations of the embodiments in the present application fall within the scope of the claims and their equivalents in the embodiments of the present application, the embodiments of the present application are also intended to include these modifications and variations.

Claims

1. A method for predicting user churn probability, characterized in that: The method comprises: Obtaining each current refresh time of the target user refreshing the recommended page on the target terminal and the historical behavior record of the target user on the recommended page on the target terminal in the historical time period; the historical behavior record includes linear behaviors that change linearly with time and non-linear behaviors that do not change linearly with time; each historical time period is a historical time period adjacent to the corresponding current refresh time; For any current refresh time, analyze the behavior characteristics of linear behaviors that occurred in the historical time period to obtain the first activity score of the target user at the current refresh time; Processing the behavior characteristics of the nonlinear behavior to obtain a second activity score of the target user in a historical time period; Determine a daily activity sequence of the target user based on the first activity score and the corresponding second activity score at each current refresh time, wherein the daily activity sequence includes a correspondence between each natural day arranged in chronological order and the corresponding target activity score; Based on the time series large language model, the daily level activity sequence is predicted to obtain the user churn probability of the target user in a preset future time period, where the preset future time period is a time period starting from the last current refresh time among the current refresh times; The daily level activity sequence is predicted based on the time series large language model to obtain the user churn probability of the target user in a preset future time period, including: Determining a scaling factor based on a non-empty target activity score in the day-level activity sequence; Using the scaling factor, normalizing the non-empty target activity scores in the daily activity sequence to obtain a normalized activity sequence; Use torch.bucketize function to map the normalized target activity scores in the normalized activity sequence into a set of predefined buckets; each predefined bucket has the same size and has its own bucket boundary value; Determine a token sequence based on the bucket boundary values ​​of the predefined buckets and the number of configured sequence special tags; The token sequence is input into the time series large language model to obtain the user churn probability of the target user in a preset future time period.

2. The method according to claim 1, characterized in that The behavior characteristics of the linear behavior include the time when the linear behavior occurs; Analyze the behavior characteristics of linear behaviors that occurred in the historical time period to obtain the first activity score of the target user at the current refresh time, including: Using a preset behavior activity algorithm, the current refresh time and the behavior occurrence time of the linear behavior in the historical time period are processed to obtain an initial activity score corresponding to the linear behavior at the current refresh time; Based on the configured time aggregation weights of each linear behavior, the initial activity scores of the corresponding linear behaviors are aggregated to obtain the first activity score of the target user at the current refresh time.

3. The method according to claim 2, characterized in that The linear behavior is reading behavior; the preset behavior activity algorithm is expressed as: in, The current refresh time when the target user refreshes the page. represents the first activity score of the i-th reading behavior at time T, They are the configuration parameters, is the time when the i-th reading behavior occurs, , n represents the number of reading behaviors in the historical period.

4. The method according to claim 1, characterized in that The behavior characteristics of the non-linear behavior include the number of interactions, number of login days and current refresh times of the target user; The calculation formula of the second activity score is expressed as: in, Indicates the second activity score of the target user's j-th nonlinear behavior in the historical time period, For the configuration parameters, , It represents the number of the j-th nonlinear behavior, j=1 represents interactive nonlinear behavior, j=2 represents logged-in nonlinear behavior, and j=3 represents refresh nonlinear behavior.

5. The method according to claim 1, characterized in that Determining a daily level activity sequence of the target user based on the first activity score and the corresponding second activity score at each current refresh time includes: Using the trained activity estimation model, the first activity score and the corresponding second activity score at each current refresh time are processed to obtain a real-time activity sequence of the target user, wherein the real-time activity sequence includes a corresponding relationship between each current refresh time and the real-time activity score; According to the aggregation granularity of natural days, the real-time activity scores in the real-time activity sequence are aggregated and calculated to obtain a day-level activity sequence.

6. The method according to claim 1, characterized in that Before predicting the day-level activity sequence based on the time series large language model, the method further includes: According to a preset convolution algorithm, a preset time convolution kernel is used to perform a convolution operation on the target activity score in the day-level activity sequence to obtain a new day-level activity sequence; the new day-level activity sequence includes the correspondence between each natural day arranged in chronological order and the corresponding new activity score.

7. The method according to claim 1, characterized in that After obtaining the user churn probability of the target user in a preset future time period, the method further includes: Based on the user churn probability, the control strategies in the recall stage, the rough ranking stage, the fine ranking stage and the re-ranking stage of the recommendation system are adjusted respectively to obtain new control strategies for the corresponding stages; Among them, the new control strategy in the recall stage includes adjusting the recall quantity of the content that the target user is interested in; the new control strategy in the rough ranking stage includes adjusting the weight value of the target user's consumption behavior; the new control strategy in the fine ranking stage includes adjusting the weight value that characterizes the user value and content value of the target user; the new control strategy in the rearrangement stage includes adjusting the display order of the recommended content that the target user is interested in.

8. A device for predicting user churn probability, characterized in that: The device comprises: An acquisition unit, used to acquire each current refresh time when a target user refreshes a recommended page on a target terminal and a historical behavior record of the target user on the recommended page on the target terminal in a historical time period; the historical behavior record includes a linear behavior that changes linearly with time and a non-linear behavior that does not change linearly with time; each historical time period is a historical time period adjacent to the corresponding current refresh time; An analysis unit, configured to analyze the behavior characteristics of linear behaviors occurring in a historical time period at any current refresh time, and obtain a first activity score of the target user at the current refresh time; A processing unit, configured to process the behavior characteristics of the nonlinear behavior to obtain a second activity score of the target user in a historical time period; a determining unit, configured to determine a day-level activity sequence of the target user based on the first activity score and the corresponding second activity score at each current refresh time, wherein the day-level activity sequence includes a correspondence between each natural day arranged in chronological order and the corresponding target activity score; A prediction unit, configured to predict the daily level activity sequence based on a time series large language model, and obtain a user churn probability of the target user in a preset future time period, wherein the preset future time period is a time period starting from the last current refresh time among the current refresh times; The prediction unit is specifically used for: Determining a scaling factor based on a non-empty target activity score in the day-level activity sequence; Using the scaling factor, normalizing the non-empty target activity scores in the daily activity sequence to obtain a normalized activity sequence; Use torch.bucketize function to map the normalized target activity scores in the normalized activity sequence into a set of predefined buckets; each predefined bucket has the same size and has its own bucket boundary value; Determine a token sequence based on the bucket boundary values ​​of the predefined buckets and the number of configured sequence special tags; The token sequence is input into the time series large language model to obtain the user churn probability of the target user in a preset future time period.

9. An electronic device, characterized in that: The electronic device comprises a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus; Memory, used to store computer programs; A processor, for implementing any of the methods described in claims 1-7 when executing a program stored in a memory.

Images