User behavior data acquisition method of APP

Abstract

The invention relates to the technical field of data acquisition methods, in particular to an APP user behavior data acquisition method, which specifically comprises the following steps: step 1, defining a behavior event model, and constructing a behavior event model based on ontology; step 2, establishing a dynamic point burying mechanism, realizing dynamic point burying by adopting a dynamic agent technology based on the behavior event model, and for each interactive component in the APP, automatically generating a point burying code through a dynamic agent during operation; 3, data acquisition and preprocessing: acquiring user behavior data in real time in the process of using the APP by the user, and preprocessing noise, missing values and needs in the acquired data; compared with the prior art, the method has remarkable advantages in all links of data collection, processing, analysis and application, comprehensive, accurate and real-time user behavior data can be efficiently collected, and the requirements of enterprises for user behavior analysis and business optimization can be met more comprehensively and deeply.

Description

Technical Field

[0001] The present invention relates to the technical field of data collection methods, and in particular to a method for collecting user behavior data of an APP. Background Art

[0002] In the mobile internet era, collecting app user behavior data is crucial for product optimization, improving user experience, and precision marketing. Currently, common methods for collecting app user behavior data mostly rely on traditional tracking technology and log-based collection methods. Traditional tracking technology requires manually adding a large amount of tracking code to the app code, which not only results in high development costs and low efficiency, but also makes subsequent maintenance difficult. Once the app functionality changes, the tracking code also needs to be modified accordingly. While log-based collection methods can record user behavior, they often struggle to accurately capture and analyze data in complex behavioral scenarios and those with high real-time requirements. This can lead to problems such as data loss and latency, making it unable to meet increasingly complex data analysis needs.

[0003] Existing methods for collecting app user behavior data suffer from low efficiency, insufficient accuracy, poor real-time performance, and weak scalability. This makes it difficult to efficiently collect comprehensive, accurate, and real-time user behavior data, making it difficult for businesses to conduct in-depth analysis and mining of user behavior. Summary of the Invention

[0004] In view of the technical problems that existing APP user behavior data collection methods have low collection efficiency, insufficient accuracy, poor real-time performance and weak scalability, and are unable to efficiently collect comprehensive, accurate and real-time user behavior data, and are difficult to support enterprises in conducting in-depth analysis and mining of user behavior, the present invention provides an APP user behavior data collection method.

[0005] The technical solution adopted by the present invention is: a method for collecting user behavior data of an APP, which specifically includes the following steps:

[0006] Step 1: Define the behavior event model and build a behavior event model based on ontology;

[0007] Step 2: Establish a dynamic tracking mechanism. Based on the above behavioral event model, use dynamic proxy technology to implement dynamic tracking. For each interactive component in the app, the dynamic proxy automatically generates tracking code at runtime.

[0008] Step 3: Data collection and preprocessing: When users use the app, user behavior data is collected in real time. Noise and missing values ​​in the collected data need to be preprocessed.

[0009] Step 4: Data transmission and storage, using asynchronous transmission based on message queue to transmit the collected data to the server;

[0010] Step 5: Real-time data analysis and mining: Use stream computing technology to perform real-time analysis and mining on the data transmitted to the server;

[0011] Step 6: Deep behavioral pattern mining: Introducing recurrent neural networks and their variants, long short-term memory networks, from deep learning to perform deep feature extraction and pattern mining on user behavior data.

[0012] Step 7: Data feedback and optimization. Feedback the results of analysis and mining to the APP development team and operation team. The development team optimizes the APP functions and interface design based on the data feedback, and the operation team formulates marketing strategies based on user behavior data.

[0013] In one embodiment, in step 1, the behavior event model is expressed as: E = (O, A, T, C);

[0014] Where: E represents the user behavior event set, that is, the set of all user behavior events generated in the APP; O represents the behavior object set, that is, the various elements involved in the user's operation in the APP, O = {o1, o2, ..., o m}, where o i Represents the i-th behavior object; A is a set of behavior actions, A={a1,a2,…,a k}, where a j represents the jth action; T is the timestamp set of the action, which is used to record the specific time when the action occurs, T = {t1, t2, ..., t n}, where t l Indicates the timestamp of the lth behavior event; C represents the context set of the behavior, C={c1,c2,…,c p}, where c q Represents the qth contextual factor.

[0015] In one embodiment, in step 2, let M be the set of buried points, M = {m1,m2,…,m n}, where m i Indicates the i-th buried point;

[0016] The process of dynamic proxy generating tracking code is expressed by the following formula: i =P(E i ,S);

[0017] Where: E i is the behavioral event related to the ith tracking point, belonging to the set E defined in step 1; S is the runtime state of the APP, S = {s1, s2, ..., s r}, where s tRepresents the tth runtime state parameter; the P function represents the mapping relationship between the behavior event and the runtime state to generate the tracking code, which is specifically expressed as: P(E i ,S)=f(E i )×g(S);

[0018] Among them, f(E i ) is about behavioral event E i is a function used to determine the basic type and content of the tracking point; g(S) is a function about the runtime state S, which is used to adjust the specific parameters of the tracking point according to the runtime state.

[0019] In one embodiment, in step 3, for numerical data, the mean interpolation method is used to handle missing values, and X = {x1, x2, ..., x n} is the original data set, is the data mean, For missing values ​​x j ,use Filling

[0020] For outliers, a density-based clustering algorithm is used to identify and process them;

[0021] Given a dataset D, a neighborhood radius ∈ and a minimum number of points MinPts, for a data point p, its ∈ neighborhood N ∈ (p) is defined as: N ∈ (p)={q∈D|distance(p,q)≤∈}, where distance(p,q) represents the distance between point p and point q; if |N ∈ (p)|≥MinPts, then point p is a core point; if point p is a core point and point q is in the ∈ neighborhood of p, then point q is said to be directly density-reachable from point p; if there exists a point chain p1,p2,…,p n ,p1=p,p n =q, and p i+1 From p i If point q is directly density reachable from point p, then point q is said to be density reachable from point p.

[0022] In one embodiment, in step 4, let Q be the message queue, D be the data set obtained in step 3, and the data transmission process is expressed as: Q.put(D);

[0023] The throughput of the message queue is calculated using the following formula:

[0024] Among them, Number of Messages represents the number of messages processed within the time interval TimeInterval. On the server side, a distributed storage system is used to store data, and the data is partitioned and stored according to dimensions such as time and user ID. The data partition rule is expressed as: Partition = Hash(userID) mod N;

[0025] Where userID is the unique identifier of the user, N is the number of partitions, and Hash is the hash function.

[0026] In one embodiment, in step 5, let D stream For real-time data streaming.

[0027] In one embodiment, in step 6, the input user behavior data sequence is X = [x1, x2, ..., x T ], where T is the sequence length, x t represents the behavioral data vector at time t, x t ∈R d , d is the data feature dimension;

[0028] For RNN, at time t, the hidden layer state h t The update formula is: t =σ(W xh x t +W hh h t-1 +b h );

[0029] Where: σ is the activation function; W xh is the weight matrix from the input layer to the hidden layer, with a dimension of h×d, where h is the number of neurons in the hidden layer, and W hh is the weight matrix from hidden layer to hidden layer, with dimension h×h; b h is the bias vector of the hidden layer, with dimension h×1;

[0030] LSTM is based on RNN and introduces a gating mechanism. The input gate i t 、Forget Gate t , output gate o t and memory unit c t ;

[0031] The formula is as follows:

[0032]

[0033] Where: ⊙ represents element-wise multiplication, It is the candidate memory unit state. By training the RNN or LSTM model, the complex patterns and long-term dependencies in the user behavior sequence are learned, and the output hidden layer state is used as the deep feature representation of the user behavior.

[0034] In one embodiment, in step 7, the degree of optimization of the data acquisition method is evaluated using the following formula:

[0035]

[0036] Among them, Post-optimizationPerformance represents the performance indicators of data collection and analysis after optimization, and Pre-optimizationPerformance represents the performance indicators before optimization.

[0037] The beneficial effects of the present invention are: compared with the existing technology, in the present invention, the ontology-based behavioral event model is used to perform a structured definition of user behavior, and the combination of the dynamic tracking mechanism improves the flexibility and scalability of tracking, and reduces the development and maintenance costs; the use of data preprocessing, message queue-based data transmission and distributed storage ensures the accuracy, stability and storage efficiency of the data; the use of stream computing technology to achieve real-time data analysis and mining can provide data support for corporate decision-making in a timely manner; the data feedback and optimization mechanism form a closed loop, continuously improving the quality of data collection and the accuracy of analysis; the newly added deep behavior pattern mining step mines the deep patterns of user behavior through deep learning models, providing enterprises with deeper user insights. Compared with the existing technology, it has significant advantages in all aspects of data collection, processing, analysis and application, and can efficiently collect comprehensive, accurate and real-time user behavior data, which can more comprehensively and deeply meet the needs of enterprises for user behavior analysis and business optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0039] In the description of the present invention, it should be noted that the terms "front", "up", "down", "left", "right", "vertical", "horizontal", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limiting the present invention.

[0040] Reference Attachment Figure 1 In order to solve the problems existing in the background technology, this application proposes the following technical solution: a method for collecting user behavior data of an APP, specifically comprising the following steps:

[0041] Step 1: Define the behavior event model and build a behavior event model based on ontology;

[0042] In step 1, the behavior event model is expressed as: E = (O, A, T, C);

[0043] Where: E represents the set of user behavior events, that is, the set of all user behavior events generated in the APP; O represents the set of behavior objects, that is, the various elements involved in user operations in the APP, such as buttons, pages, pictures, etc., O = {o1, o2, ..., o m}, where o i represents the i-th behavior object; A is a set of behavior actions, including click, slide, input, long press and other operations, A={a1,a2,…,a k}, where a j represents the jth action; T is the timestamp set of the action, which is used to record the specific time when the action occurs, T = {t1, t2, ..., t n}, where t l Indicates the timestamp of the lth behavior event; C represents the context set of the behavior, such as device model, network status, geographic location, APP version, etc. C = {c1, c2, ..., c p}, where c q Represents the qth contextual factor.

[0044] The above technical solution is explained as follows: Construct an ontology-based behavioral event model to define user behavior in a structured manner. By clarifying the behavioral event set, behavioral object set, behavioral action set, timestamp set, and context environment set, a clear framework is provided for subsequent data collection and analysis. This model enables data to have stronger semantic expression capabilities, making it easier to understand and process user behavior. In actual applications, it can more accurately capture various user behaviors in the APP, including operation content, operation time, and environmental information during the operation. This helps companies fully understand user behavior characteristics and provides a solid foundation for subsequent data analysis and business decisions. At the same time, the structured definition also enables data from different sources to be standardized, improves the integration and availability of data, and facilitates cross-platform and cross-application user behavior analysis.

[0045] Step 2: Establish a dynamic tracking mechanism. Based on the above behavioral event model, use dynamic proxy technology to implement dynamic tracking. For each interactive component in the app, the dynamic proxy automatically generates tracking code at runtime.

[0046] In step 2, let M be the set of buried points, M = {m1,m2,…,m n}, where m iIndicates the i-th buried point;

[0047] The process of dynamic proxy generating tracking code can be expressed by the following formula: i =P(E i ,S);

[0048] Where: E i is the behavioral event related to the i-th tracking point, belonging to the E set defined in step 1; S is the runtime state of the APP, including the current page, user login status, session ID, etc., S = {s1, s2, ..., s r}, where s t Represents the tth runtime state parameter; the P function represents the mapping relationship between the behavior event and the runtime state to generate the tracking code, which can be specifically expressed as: P(E i ,S)=f(E i )×g(S);

[0049] Among them, f(E i ) is about behavioral event E i is a function used to determine the basic type and content of the tracking point; g(S) is a function about the runtime state S, which is used to adjust the specific parameters of the tracking point according to the runtime state.

[0050] The above technical solution is explained as follows: Dynamic tracking is achieved using dynamic proxy technology, and there is no need to manually add a large amount of tracking code. Tracking is automatically generated according to behavioral events and runtime status, which greatly improves the flexibility and scalability of tracking. When the APP function changes, there is no need to rewrite a large amount of tracking code, and it can quickly adapt to changes, reducing development and maintenance costs. This dynamic tracking mechanism can automatically generate corresponding tracking points for different user behaviors and scenarios at runtime to ensure the comprehensiveness and accuracy of data collection. For example, for newly launched functional modules, the system can automatically generate tracking points for them and collect users' usage of new functions in a timely manner. In addition, dynamic tracking can also collect data in a targeted manner based on users' personalized characteristics and behavioral habits, improve the value density of data, and provide strong support for precision marketing and personalized recommendations.

[0051] Step 3: Data collection and preprocessing: When users use the app, user behavior data is collected in real time. Noise and missing values ​​in the collected data need to be preprocessed.

[0052] Among them, in step 3, for numerical data, the mean interpolation method is used to deal with missing values, and X={x1,x2,…,x n} is the original data set, is the data mean, For missing values ​​x j ,use Filling

[0053] For outliers, a density-based clustering algorithm is used to identify and process them;

[0054] Given a dataset D, a neighborhood radius ∈ and a minimum number of points MinPts, for a data point p, its ∈ neighborhood N ∈ (p) is defined as: N ∈ (p)={q∈D|distance(p,q)≤∈}, where distance(p,q) represents the distance between point p and point q; if |N ∈ (p)|≥MinPts, then point p is a core point; if point p is a core point and point q is in the ∈ neighborhood of p, then point q is said to be directly density-reachable from point p; if there exists a point chain p1,p2,…,p n ,p1=p,p n =q, and p i+1 From p i If point q is directly density reachable from point p, then point q is said to be density reachable from point p.

[0055] The above technical solution is explained as follows: data is collected in real time while the user is using the APP, and the collected data is preprocessed. Missing values ​​are processed by the mean interpolation method, and outliers are identified and processed by a density-based clustering algorithm to ensure the accuracy and completeness of the data. The clean data after preprocessing provides a high-quality data foundation for subsequent analysis and mining. In the actual data collection process, data may be missing or abnormal due to various reasons. If these problems are not resolved, they will seriously affect the results of data analysis. Through data preprocessing, these problems can be effectively solved and the quality of the data can be improved. For example, for missing user age data, the mean interpolation method can be used to fill it in to make the data more complete; for abnormal user behavior data, such as a large number of click operations in a short period of time, it can be identified and processed by a clustering algorithm to eliminate interference factors and ensure the accuracy of data analysis.

[0056] Step 4: Data transmission and storage, using asynchronous transmission based on message queue to transmit the collected data to the server;

[0057] In step 4, let Q be the message queue and D be the data set obtained in step 3. The data transmission process is expressed as: Q.put(D);

[0058] The throughput of the message queue can be calculated using the following formula:

[0059] Among them, Number of Messages represents the number of messages processed within the time interval TimeInterval. On the server side, a distributed storage system is used to store data, and the data is partitioned and stored according to dimensions such as time and user ID. The data partition rule is expressed as: Partition = Hash(userID) mod N;

[0060] Where userID is the unique identifier of the user, N is the number of partitions, and Hash is the hash function.

[0061] The above technical solution is explained as follows: Data is transmitted to the server using an asynchronous transmission method based on a message queue, which improves the stability and efficiency of data transmission. Message queues have the characteristics of high throughput and asynchronous processing, which can effectively avoid data loss caused by network fluctuations and other problems. On the server side, data is partitioned and stored through a distributed storage system, which improves the efficiency of data query and processing, and can effectively meet the storage and management needs of large amounts of user behavior data. The distributed storage system can partition and store data according to dimensions such as time and user ID, making data storage more reasonable and queries faster. For example, when conducting user behavior analysis, the behavior data of a specific user within a specific time period can be quickly located, which improves the efficiency of data analysis. At the same time, the distributed storage system also has high availability and scalability, and can flexibly adjust storage resources as the number of users and the amount of data grow to ensure stable operation of the system.

[0062] Step 5: Real-time data analysis and mining: Use stream computing technology to perform real-time analysis and mining on the data transmitted to the server;

[0063] In step 5, let D stream For real-time data stream transmission, a series of computing operators (such as mapping operators, filtering operators, aggregation operators, etc.) are defined to process the data stream.

[0064] The above technical solution is explained as follows: Stream computing technology is used to perform real-time analysis and mining of data transmitted to the server. By defining a series of computing operators to process the data stream, user behavior patterns and abnormal behaviors can be discovered in a timely manner, providing real-time data support for corporate decision-making. In today's rapidly changing market environment, companies need to understand user behavior and market dynamics in a timely manner in order to respond quickly. Real-time data analysis and mining technology can meet this demand. Through real-time processing and analysis of user behavior data, companies can promptly discover changes in user preferences and needs and adjust product and marketing strategies in a timely manner. For example, when a large number of users are found to frequently visit a specific page within a certain period of time, the company can promptly optimize the content and layout of the page to improve the user experience; when abnormal user behavior is detected, such as a sudden large number of unsubscribe operations, the company can take timely measures to avoid losses.

[0065] Step 6: Deep behavioral pattern mining: Introducing recurrent neural networks and their variants, long short-term memory networks, from deep learning to perform deep feature extraction and pattern mining on user behavior data.

[0066] In step 6, let the input user behavior data sequence be X = [x1, x2, ..., x T ], where T is the sequence length, x t represents the behavioral data vector at time t, x t ∈R d , d is the data feature dimension;

[0067] For RNN, at time t, the hidden layer state h t The update formula is: t =σ(W xh x t +W hh h t-1 +b h );

[0068] Where: σ is the activation function, such as the sigmoid function or the tanh function, which is used to map the input to the appropriate interval; W xh is the weight matrix from the input layer to the hidden layer, with a dimension of h×d, where h is the number of neurons in the hidden layer, and W hh is the weight matrix from hidden layer to hidden layer, with dimension h×h. h is the bias vector of the hidden layer, with dimension h×1;

[0069] LSTM is based on RNN and introduces a gating mechanism. The input gate i t 、Forget Gate t , output gate o t and memory unit c t, dealing with long-term dependency problems in long sequence data;

[0070] The formula is as follows:

[0071]

[0072] Among them: the meaning of each weight matrix and bias vector is similar to that of RNN, which is used to calculate the gating signal and memory unit state respectively. ⊙ represents element-level multiplication. It is the candidate memory unit state. By training the RNN or LSTM model, the complex patterns and long-term dependencies in the user behavior sequence are learned. The output hidden layer state can be used as the deep feature representation of user behavior for subsequent tasks such as user classification and behavior prediction.

[0073] The above technical solution is explained as follows: Recurrent neural networks (RNNs) and their variant, long short-term memory (LSTM) networks, from deep learning, are introduced to perform deep feature extraction and pattern mining on user behavior data. These models are capable of learning complex patterns and long-term dependencies within user behavior sequences. The output hidden layer states serve as deep feature representations of user behavior, providing richer and more valuable information for subsequent tasks such as user classification and behavior prediction. Through deep behavioral pattern mining, companies can gain a deeper understanding of user behavior and needs, identifying potential market opportunities. For example, by analyzing historical user behavior data, future purchase intentions and needs can be predicted, providing a more accurate basis for personalized recommendations. Market segments within user groups can also be identified, allowing differentiated marketing strategies to be developed for different market groups. Furthermore, deep behavioral pattern mining can help companies identify abnormal patterns in user behavior, promptly detect potential risks and issues, and take appropriate measures to prevent and resolve them.

[0074] Step 7: Data feedback and optimization. Feedback the results of analysis and mining to the APP development team and operation team. The development team optimizes the APP functions and interface design based on the data feedback, and the operation team formulates marketing strategies based on user behavior data.

[0075] In step 7, the degree of optimization of the data acquisition method is evaluated using the following formula:

[0076]

[0077] Among them, Post-optimizationPerformance represents the performance indicators of data collection and analysis after optimization (such as collection efficiency, accuracy, etc.), and Pre-optimizationPerformance represents the performance indicators before optimization.

[0078] The above technical solution is explained as follows: The results of the analysis and mining are fed back to the app development team and the operations team. The development team optimizes the app's functionality and interface design based on this data feedback, and the operations team develops more precise marketing strategies. By continuously adjusting and optimizing data collection methods and analysis models, a closed-loop optimization mechanism is formed to continuously improve the quality of data collection and the accuracy of analysis. This data feedback and optimization mechanism enables companies to continuously improve their products and services based on user behavior data, thereby increasing user satisfaction and loyalty. For example, the development team can optimize the app's interface layout and interaction methods based on user operating habits and feedback, making the user experience smoother; the operations team can develop more precise marketing strategies based on user purchasing behavior and preferences to improve marketing effectiveness. At the same time, through the continuous optimization of data collection methods and analysis models, the quality of data and the accuracy of analysis can be further improved, forming a virtuous cycle and driving the continuous development of the company.

[0079] In summary, in the present invention, the behavioral event model based on ontology is used to perform a structured definition of user behavior, and the combination with the dynamic tracking mechanism improves the flexibility and scalability of tracking, and reduces the development and maintenance costs; data preprocessing, message queue-based data transmission and distributed storage are used to ensure the accuracy, stability and storage efficiency of the data; stream computing technology is used to achieve real-time data analysis and mining, which can provide data support for corporate decision-making in a timely manner; data feedback and optimization mechanisms form a closed loop, continuously improving the quality of data collection and the accuracy of analysis; the newly added deep behavior pattern mining step mines the deep patterns of user behavior through deep learning models, providing enterprises with deeper user insights. Compared with the existing technology, it has significant advantages in all aspects of data collection, processing, analysis and application, and can efficiently collect comprehensive, accurate and real-time user behavior data, which can more comprehensively and deeply meet the needs of enterprises for user behavior analysis and business optimization.

[0080] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for collecting user behavior data of an APP, characterized in that: The specific steps include: Step 1: Define the behavior event model and build a behavior event model based on ontology; Step 2: Establish a dynamic tracking mechanism. Based on the above behavioral event model, use dynamic proxy technology to implement dynamic tracking. For each interactive component in the app, the dynamic proxy automatically generates tracking code at runtime. Step 3: Data collection and preprocessing: When users use the app, user behavior data is collected in real time. Noise and missing values ​​in the collected data need to be preprocessed. Step 4: Data transmission and storage, using asynchronous transmission based on message queue to transmit the collected data to the server; Step 5: Real-time data analysis and mining: Use stream computing technology to perform real-time analysis and mining on the data transmitted to the server; Step 6: Deep behavioral pattern mining: Introducing recurrent neural networks and their variants, long short-term memory networks, from deep learning to perform deep feature extraction and pattern mining on user behavior data. Step 7: Data feedback and optimization. Feedback the results of analysis and mining to the APP development team and operation team. The development team optimizes the APP functions and interface design based on the data feedback, and the operation team formulates marketing strategies based on user behavior data.

2. The method for collecting user behavior data of an APP according to claim 1, characterized in that: In step 1, the behavioral event model is expressed as: E = (O, A, T, C); Where: E represents the user behavior event set, that is, the set of all user behavior events generated in the APP; O represents the behavior object set, that is, the various elements involved in the user's operation in the APP, O = {o1, o2, ..., o m }, where o i Represents the i-th behavior object; A is a set of behavior actions, A={a1,a2,…,a k }, where a j represents the jth action; T is the timestamp set of the action, which is used to record the specific time when the action occurs, T = {t1, t2, ..., t n }, where t l Indicates the timestamp of the lth behavior event; C represents the context set of the behavior, C={c1,c2,…,c p }, where c q Represents the qth contextual factor.

3. The method for collecting user behavior data of an APP according to claim 2, characterized in that: In step 2, let M be the set of buried points, M={m1,m2,…,m n }, where m i Indicates the i-th buried point; The process of dynamic proxy generating tracking code is expressed by the following formula: i =P(E i ,S); Where: E i is the behavioral event related to the ith tracking point, belonging to the set E defined in step 1; S is the runtime state of the APP, S = {s1, s2, ..., s r }, where s t Represents the tth runtime state parameter; the P function represents the mapping relationship between the behavior event and the runtime state to generate the tracking code, which is specifically expressed as: P(E i ,S)=f(E i )×g(S); Among them, f(E i ) is about behavioral event E i is a function used to determine the basic type and content of the tracking point; g(S) is a function about the runtime state S, which is used to adjust the specific parameters of the tracking point according to the runtime state.

4. The method for collecting user behavior data of an APP according to claim 3, characterized in that: In step 3, for numerical data, the mean interpolation method is used to deal with missing values. Let X = {x1, x2, ..., x n } is the original data set, is the data mean, For missing values ​​x j ,use Filling For outliers, a density-based clustering algorithm is used to identify and process them; Given a dataset D, a neighborhood radius ∈ and a minimum number of points MinPts, for a data point p, its ∈ neighborhood N ∈ (p) is defined as: N ∈ (p)={q∈D|distance(p,q)≤∈}, where distance(p,q) represents the distance between point p and point q; if |N ∈ (p)|≥MinPts, then point p is the core point; If point p is a core point and point q is in the ∈ neighborhood of p, then point q is said to be directly density-reachable from point p; If there exists a point chain p1,p2,…,p n ,p1=p,p n =q, and p i+1 From p i If point q is directly density reachable from point p, then point q is said to be density reachable from point p.

5. The method for collecting user behavior data of an APP according to claim 4, characterized in that: In step 4, let Q be the message queue and D be the data set obtained in step 3. The data transmission process is expressed as: Q.put(D); The throughput of the message queue is calculated using the following formula: Among them, Number of Messages represents the number of messages processed within the time interval TimeInterval. On the server side, a distributed storage system is used to store data, and the data is partitioned and stored according to dimensions such as time and user ID. The data partition rule is expressed as: Partition = Hash(userID) mod N; Where userID is the unique identifier of the user, N is the number of partitions, and Hash is the hash function.

6. The method for collecting user behavior data of an APP according to claim 5, characterized in that: In step 5, let D stream For real-time transmission of data streams.

7. The method for collecting user behavior data of an APP according to claim 6, characterized in that: In step 6, let the input user behavior data sequence be X = [x1, x2, ..., x T ], where T is the sequence length, x t represents the behavioral data vector at time t, x t ∈R d , d is the data feature dimension; For RNN, at time t, the hidden layer state h t The update formula is: t =σ(W xh x t +W hh h t-1 +b h ); Where: σ is the activation function; W xh is the weight matrix from the input layer to the hidden layer, with a dimension of h×d, where h is the number of neurons in the hidden layer, and W hh is the weight matrix from hidden layer to hidden layer, with dimension h×h; b h is the bias vector of the hidden layer, with dimension h×1; LSTM is based on RNN and introduces a gating mechanism. The input gate i t 、Forget Gate t , output gate o t and memory unit c t ; The formula is as follows: i t =σ(W xi x t +W hi h t-1 +b i ); f t =σ(W xf x t +W hf h t-1 +b f ); o t =σ(W xo x t +W ho h t-1 +b o ); h t =o t ⊙tanh(c t ); Among them: ⊙ represents element-wise multiplication, It is the candidate memory unit state. By training the RNN or LSTM model, the complex patterns and long-term dependencies in the user behavior sequence are learned, and the output hidden layer state is used as the deep feature representation of the user behavior.

8. The method for collecting user behavior data of an APP according to claim 7, characterized in that: In step 7, the degree of optimization of the data acquisition method is evaluated using the following formula: Among them, Post-optimizationPerformance represents the performance indicators of data collection and analysis after optimization, and Pre-optimizationPerformance represents the performance indicators before optimization.

Images