User behavior big data analysis method and system for cloud office resource push

Abstract

The embodiments of the present invention relate to the technical fields of cloud office and big data push, and specifically provide a user behavior big data analysis method and system applied to cloud office resource push, which can accurately locate the non-directional push situation in the historical office resource push event by obtaining the collaborative office session behavior information and prior push feedback data of historical cloud office users. The knowledge of session scene jumps and the characteristics of collaborative interaction status are obtained from the collaborative office session behavior information, which enriches the analysis dimension and makes the analysis of push events more comprehensive. The original big data push analysis algorithm is used in combination with these characteristics to obtain the initial push analysis viewpoint, which helps to accurately judge the non-directional push results. The algorithm is debugged based on the prior push feedback data to obtain the target algorithm, which improves the accuracy of the algorithm and makes the target push analysis viewpoint more consistent with the prior data, thereby providing a reliable basis for the directional push of cloud office resources and improving the accuracy of office resource push and user office efficiency.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of cloud office and big data push, and in particular, to a user behavior big data analysis method and system applied to cloud office resource push. Background Art

[0002] As cloud office becomes increasingly popular, effective push of office resources has become the key to improving office efficiency. However, traditional office resource push lacks precision and there are a lot of non-directional pushes. For example, users receive office resources that do not match their needs, which not only wastes resources but also interferes with the user's normal office process. Existing push analysis methods often do not fully consider the various behavioral characteristics of users in cloud office, making it difficult to accurately determine whether the push is targeted. Summary of the invention

[0003] The embodiments of the present invention at least provide a method and system for analyzing user behavior big data applied to cloud office resource push.

[0004] The embodiment of the present invention provides a user behavior big data analysis method applied to cloud office resource push, which is applied to a user behavior big data analysis system. The method includes: obtaining collaborative office session behavior information of historical cloud office users, and prior push feedback data of historical office resource push events corresponding to the collaborative office session behavior information, wherein the prior push feedback data is prior authentication data used to characterize that the historical office resource push events have non-directional push results; obtaining the session scene jump knowledge of the historical cloud office users and the collaborative interaction state characteristics of the historical collaborative interaction terminals corresponding to the historical cloud office users from the collaborative office session behavior information of the historical cloud office users; using the original big data push analysis algorithm, based on the session scene jump, According to the change knowledge and the collaborative interaction state characteristics, the historical office resource push event is pushed and analyzed to obtain the initial push analysis viewpoint of the historical office resource push event; the initial push analysis viewpoint is a discriminant analysis result used to characterize that the historical office resource push event has a non-directional push result; based on the prior push feedback data and the initial push analysis viewpoint, the original big data push analysis algorithm is debugged to obtain a target big data push analysis algorithm; the consistency evaluation weight between the target push analysis viewpoint of the historical office resource push event generated by the target big data push analysis algorithm and the prior push feedback data is greater than the consistency evaluation weight between the initial push analysis viewpoint and the prior push feedback data.

[0005] An embodiment of the present invention also provides a user behavior big data analysis system, including a processor and a memory; the processor and the memory are communicatively connected, and the processor is used to read a computer program from the memory and execute it to implement the above method.

[0006] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. The computer program implements the above method when running.

[0007] The technical solution provided by the embodiment of the embodiment of the present invention may include the following beneficial effects: by obtaining the collaborative office session behavior information and prior push feedback data of historical cloud office users, the non-directional push situation in the historical office resource push event can be accurately located. The knowledge of session scene jumps and the characteristics of collaborative interaction status are obtained from the collaborative office session behavior information, which enriches the analysis dimension and makes the analysis of push events more comprehensive. The original big data push analysis algorithm is used in combination with these characteristics to obtain the initial push analysis viewpoint, which helps to accurately judge the non-directional push results. The algorithm is debugged based on the prior push feedback data to obtain the target algorithm, which improves the accuracy of the algorithm and makes the target push analysis viewpoint more consistent with the prior data, thereby providing a reliable basis for the targeted push of cloud office resources and improving the accuracy of office resource push and user office efficiency.

[0008] For a description of the effects of the above-mentioned user behavior big data analysis system and computer-readable storage medium, please refer to the description of the above-mentioned method.

[0009] In order to make the above-mentioned purposes, features and advantages of the embodiments of the present invention more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the embodiments of the present invention and are used together with the specification to illustrate the technical solutions of the embodiments of the present invention. It should be understood that the following drawings only illustrate certain embodiments of the embodiments of the present invention and should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 It is a block diagram of a user behavior big data analysis system shown in an embodiment of the present invention.

[0012] Figure 2 It is a flow chart of a method for analyzing user behavior big data applied to cloud office resource push, shown in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] Here, exemplary embodiments will be described in detail, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the embodiments of the present invention. Instead, they are only examples of devices and methods consistent with some aspects of the embodiments of the present invention.

[0014] Figure 1 The structural diagram of the user behavior big data analysis system 10 provided in the embodiment of the present invention includes a processor 102, a memory 104, and a bus 106. Among them, the memory 104 is used to store execution instructions, including a memory and an external memory. The memory can also be understood as an internal memory, which is used to temporarily store the operation data in the processor 102 and the data exchanged with the external memory such as the hard disk. The processor 102 exchanges data with the external memory through the memory. When the user behavior big data analysis system 10 is running, the processor 102 communicates with the memory 104 through the bus 106, so that the processor 102 executes the user behavior big data analysis method applied to the cloud office resource push according to the embodiment of the present invention.

[0015] Please combine Figure 2 , Figure 2 It is a flow chart of a user behavior big data analysis method for cloud office resource push provided by an embodiment of the present invention, which is applied to a user behavior big data analysis system. The method may exemplarily include the following steps 100 to 160.

[0016] Step 100: The user behavior big data analysis system obtains the collaborative office session behavior information of historical cloud office users, and the prior push feedback data of the historical office resource push events corresponding to the collaborative office session behavior information. The prior push feedback data is the prior authentication data used to characterize the existence of non-directional push results in the historical office resource push events.

[0017] Step 120: The user behavior big data analysis system obtains the session scene transition knowledge of the historical cloud office user and the collaborative interaction state characteristics of the historical collaborative interaction terminal corresponding to the historical cloud office user from the collaborative office session behavior information of the historical cloud office user.

[0018] Step 140: The user behavior big data analysis system utilizes the original big data push analysis algorithm to perform a push analysis on the historical office resource push event based on the session scene jump knowledge and the collaborative interaction state characteristics, and obtains an initial push analysis view of the historical office resource push event; the initial push analysis view is a discriminant analysis result used to characterize the existence of non-directional push results in the historical office resource push event.

[0019] Step 160: The user behavior big data analysis system debugs the original big data push analysis algorithm based on the prior push feedback data and the initial push analysis viewpoint to obtain a target big data push analysis algorithm; the consistency evaluation weight between the target push analysis viewpoint of the historical office resource push event generated by the target big data push analysis algorithm and the prior push feedback data is greater than the consistency evaluation weight between the initial push analysis viewpoint and the prior push feedback data.

[0020] Based on the above steps 100 to 160, the user behavior big data analysis system plays a core execution subject role in the user behavior big data analysis method pushed by cloud office resources.

[0021] First, in step 100, the system obtains the collaborative office session behavior information of historical cloud office users and the prior push feedback data of the corresponding historical office resource push events. The collaborative office session behavior information in the embodiment of the present invention includes many aspects, such as rich information such as the communication content, communication frequency, and communication objects of users in the cloud office environment. The prior push feedback data is a key prior authentication data, which is specifically used to characterize the existence of non-directional push results in historical office resource push events. For example, there may be a numerical indicator to indicate the degree of this non-directional push, such as a value of 0.6. When this value is higher than a certain threshold, it indicates that there is a more obvious non-directional push situation. These data provide a basis for subsequent analysis.

[0022] Next, in step 120, the system further mines the session scene jump knowledge of historical cloud office users and the collaborative interaction state characteristics of the historical collaborative interaction terminals corresponding to the historical cloud office users from the collaborative office session behavior information of the historical cloud office users that has been obtained. Session scene jump knowledge may refer to relevant information about a user switching from one work scene to another during a collaborative office session, such as suddenly switching from a document editing scene to a video conferencing scene. The collaborative interaction state characteristics may include the network connection stability of the interactive terminal, the performance status of the interactive device, and the like. Taking network connection stability as an example, if the network connection is frequently interrupted within a certain period of time, this is a manifestation of a collaborative interaction state characteristic.

[0023] Then, entering step 140, the user behavior big data analysis system uses the original big data push analysis algorithm to carry out push analysis on the historical office resource push events based on the previously acquired knowledge of session scene jumps and collaborative interaction state characteristics. The original big data push analysis algorithm in the embodiment of the present invention can be an algorithm based on rules or machine learning models. For example, it can be an algorithm based on a decision tree, which makes branch judgments based on different session scene jump situations and collaborative interaction state characteristics. Through the analysis of this algorithm, the initial push analysis viewpoint of the historical office resource push event is obtained. This initial push analysis viewpoint is a discriminant analysis result of whether there is a non-directional push result in the historical office resource push event. For example, the probability of a non-directional push result given by this initial push analysis viewpoint is 0.4, which indicates that according to the preliminary analysis of the original algorithm, there is a certain degree of possibility of non-directional push.

[0024] Finally, in step 160, the system debugs the original big data push analysis algorithm based on the prior push feedback data and the initial push analysis viewpoint. The prior push feedback data is the accurate data that has been determined to characterize the non-directional push result, and the initial push analysis viewpoint is the result initially obtained by the algorithm. By comparing and analyzing the two, the original algorithm is adjusted to obtain the target big data push analysis algorithm. This target big data push analysis algorithm has an important characteristic, that is, the consistency evaluation weight between the target push analysis viewpoint of the historical office resource push event generated by it and the prior push feedback data is greater than the consistency evaluation weight between the initial push analysis viewpoint and the prior push feedback data. For example, if an evaluation indicator is used to measure this consistency evaluation weight, for the initial push analysis viewpoint, this weight can be 0.3, and the consistency evaluation weight of the target push analysis viewpoint and the prior push feedback data obtained by the debugged target big data push analysis algorithm may be increased to 0.8, which indicates that the target algorithm is more accurate in judging whether there is a non-directional push result in the historical office resource push event. This accurate target algorithm can provide better guidance for subsequent cloud office resource push, especially in directional push. Through in-depth analysis of historical data and optimization of algorithms, we can avoid non-directional push, improve the accuracy of office resource push, and enhance user office efficiency.

[0025] In more detail, when obtaining the collaborative office session behavior information of historical cloud office users, this is a very complex and comprehensive data collection process. The system can collect from multiple data sources, such as the server log of the cloud office platform, the user operation record database, etc. These data sources contain a huge amount of information, and the system needs to effectively filter and organize it. For example, in the server log, it may contain information such as the user's login time, login location, and the functional module of the operation. This information needs to be integrated into the collaborative office session behavior information. As for the prior push feedback data of historical office resource push events, its source can be the user's feedback survey results, the actual usage statistics of the pushed resources, etc. If in the user feedback survey, a large number of users indicate that the office resources they received are not related to their needs, this can be used as prior authentication data for the existence of non-directional push results.

[0026] When mining the knowledge of session scene jumps, the system can use data mining techniques, such as association rule mining. By analyzing a large number of collaborative office session records, the association between certain operations or events is discovered, thereby determining the jump of the session scene. For example, if it is found that the user frequently initiates video call requests after continuously performing document editing operations, this indicates a jump from the document editing scene to the video conferencing scene. For the collaborative interaction status characteristics of the collaborative interaction end, the system can use performance monitoring tools to obtain relevant data. For example, the performance status of the interactive device can be reflected by monitoring indicators such as the CPU usage and memory occupancy of the device. If the CPU usage is high for a long time, this can affect the push effect of office resources, and may also be related to non-directional push.

[0027] The original big data push analysis algorithm can build an analysis model when performing push analysis based on the knowledge of session scene jumps and the characteristics of collaborative interaction states. This model can use the knowledge of session scene jumps and the characteristics of collaborative interaction states as input variables, and output the initial push analysis viewpoint after a series of internal processing logic. In this process, if a rule-based algorithm is used, it can be determined whether there is a non-directional push situation based on pre-set rules, such as the frequency of session scene jumps and the quality of collaborative interaction state characteristics. If it is an algorithm based on machine learning, historical data can be used for model training so that the model can learn the relationship between different input variables and non-directional push results.

[0028] When debugging the original big data push analysis algorithm to obtain the target big data push analysis algorithm, the system can use an optimization algorithm, such as a gradient descent algorithm. By continuously adjusting the parameters in the algorithm, the consistency evaluation weight between the target push analysis viewpoint and the prior push feedback data is continuously improved. This process may require multiple iterations, and each iteration will fine-tune the algorithm according to the current results until a satisfactory consistency evaluation weight is achieved. Moreover, during the debugging process, the system may also introduce new data for verification to ensure the generalization ability of the target algorithm. For example, a part of the data that did not participate in the training is used as a verification set to observe the performance of the target algorithm on these data. If the performance is good, it means that the target algorithm has good generalization ability and can be applied to the actual cloud office resource push analysis, thereby effectively identifying the non-directional push situation in the office resource push process, and providing accurate guidance for subsequent directional push.

[0029] To further explain in detail, the collection of collaborative office session behavior information of historical cloud office users, in addition to the basic information mentioned above, may also include the user's specific operating habits in the collaborative office software. For example, some users are accustomed to processing emails in a specific time period, while some users prefer to check emails after the task is completed. These operating habits are also part of the collaborative office session behavior information. For a priori push feedback data, if it is calculated from the actual usage of the pushed resources, it may involve indicators such as the number of times the resource is opened, the duration of use, and whether it is shared. If an office resource is pushed in large quantities but rarely opened or used for a short time, this strongly suggests the existence of non-directional push.

[0030] In terms of mining knowledge of session scene jumps, in addition to association rule mining, sequence pattern mining can also be used. For example, by analyzing the operation sequence of users in collaborative office sessions, some typical scene jump patterns are discovered. For example, first perform file sharing operations, then conduct instant messaging discussions, and finally enter the project management module. This is a typical operation sequence pattern that reflects the jump of session scenes. For the collaborative interaction state characteristics of the collaborative interaction end, the state of the input and output devices of the device can also be considered, such as the movement speed of the mouse, the key frequency of the keyboard, etc. If the movement speed of the mouse is abnormally slow or the key frequency of the keyboard suddenly increases, this may be related to the user's work status or the acceptance of office resource push.

[0031] When constructing the analysis model of the original big data push analysis algorithm, if it is an algorithm based on machine learning, a neural network model can be used. In this model, the conversation scene jump knowledge and the characteristics of the collaborative interaction state are used as the input of the neurons in the input layer, and are processed by the neurons in the hidden layer to finally obtain the initial push analysis viewpoint in the output layer. The neurons in the hidden layer can build a complex relationship between input and output by learning the patterns in the historical data. When debugging the original algorithm, in addition to using the gradient descent algorithm, other optimization algorithms such as genetic algorithms can also be used. Genetic algorithms optimize the parameters in the algorithm by simulating the biological evolution process. For example, the parameters in the algorithm are regarded as the genes of biological individuals. Through operations such as selection, crossover and mutation, better parameter combinations are continuously evolved, thereby improving the consistency evaluation weight between the target push analysis viewpoint and the prior push feedback data.

[0032] Throughout the entire process, the system needs to effectively manage and store data. For the massive amount of historical cloud office users' collaborative office session behavior information and prior push feedback data, a distributed storage system such as Hadoop Distributed File System (HDFS) can be used. This storage system can effectively handle large-scale data storage and management issues to ensure data security and availability. At the same time, during the operation of the algorithm, the system can use the computing resources of the cloud computing platform to improve the operation efficiency of the algorithm. For example, when using a neural network model for analysis, the cloud computing platform can provide powerful computing power to accelerate the model training and analysis process.

[0033] In summary, the user behavior big data analysis system conducts a comprehensive and in-depth analysis of the non-directional push of cloud office resources through a series of operations, from data acquisition, mining, analysis to algorithm debugging. By continuously optimizing the algorithm and improving the accuracy of the analysis, it ultimately provides a reliable guide for the directional push of cloud office resources, which is of great significance for improving the efficiency of cloud office and user experience.

[0034] In addition, in the user behavior big data analysis method of cloud office resource push, session scene jump knowledge and collaborative interaction state characteristics are core technical features and can be represented in the form of feature vectors. These two core technical features are further explained below.

[0035] In the user behavior big data analysis method of cloud office resource push, the knowledge of session scene jumps and the characteristics of collaborative interaction states are represented by feature vectors, which are of great significance in many aspects and have a comprehensive impact.

[0036] For the feature vector representation of conversation scene jump knowledge, in terms of scene classification and vector dimension definition, multi-dimensional scene classification covers a variety of cloud office scenes, such as document editing, video conferencing, etc., and constructs 10-dimensional feature vectors, each dimension corresponding to a scene jump situation. The jump frequency quantification takes the actual number of jumps as the dimension value. For example, if the jump from document editing to video conferencing occurs 5 times a day, the dimension value is 5. In the representation of jump time intervals and sequences, in addition to the jump frequency, the time interval is included in a relative manner, such as short, medium, and long classifications, and the jump sequence is reflected through special coding, such as coding by scene sequence number. These play an important role in the analysis and debugging of the original algorithm. The original algorithm, such as the decision tree, will make decisions based on the dimensions of the feature vector, and the neural network will use it as the input layer data to learn the relationship. When debugging the algorithm, it can be adjusted according to the specific dimensions to improve accuracy.

[0037] The feature vector representation of the collaborative interaction state feature contains multiple parts. Among the network connection related features, the network connection stability can be refined into three sub-indicators composed of bandwidth, latency and packet loss rate, and the network connection type and quality assessment criteria are associated at the same time. Device performance related features include comprehensive evaluation of CPU and memory, such as weighted calculation of comprehensive values ​​and classification, and also involve disk I / O and GPU performance (if relevant). Input and output device status related features refine the mouse and keyboard status, and also cover the device status representation of microphones, cameras, etc. In the original algorithm analysis, the rule-based algorithm judges resource push adjustments based on the feature vector, and the neural network algorithm learns the relationship between input features and push results. The feature vector provides an optimization reference when debugging the algorithm.

[0038] The feature vector representation has a comprehensive impact on the overall analysis. In terms of data integration and analysis efficiency, it integrates data from different data sources, such as session scene jump knowledge from operation logs, and collaborative interaction status features from monitoring tools, which improves data management efficiency, facilitates storage, query and retrieval, and enables various analysis algorithms to process data efficiently. In terms of improving analysis accuracy, it comprehensively considers the impact of various factors on office resource push, such as the impact of session scene jump and collaborative interaction status association on push demand. The original algorithm analysis and debugging can make full use of multi-factor information and optimize the basis for push decision-making. By comprehensively judging the session scene jump knowledge and collaborative interaction status features, the accuracy of non-directional push judgment is improved, providing a reliable basis for directional push.

[0039] In an optional embodiment, the session scenario jump knowledge includes the session interaction characteristics of the historical cloud office users in X session nodes, the X office resource usage events in the X session nodes, and the push decision change characteristics of the historical office resource push events, where X is an integer greater than 1. On this basis, the use of the original big data push analysis algorithm described in step 140 is based on the session scene jump knowledge and the collaborative interaction state characteristics to perform push analysis on the historical office resource push event to obtain an initial push analysis view of the historical office resource push event, including: performing a full connection operation on the session interaction characteristics of the X session nodes to obtain a scene interaction full connection vector of the historical cloud office user; based on the push decision change characteristics of the X office resource usage events and the historical office resource push event, generating an office resource unique hot semantics for reflecting the X office resource usage events and the historical office resource push event; using the original big data push analysis algorithm, based on the scene interaction full connection vector, the office resource unique hot semantics, the session scene jump knowledge and the collaborative interaction state characteristics, to perform push analysis on the historical office resource push event to obtain an initial push analysis view of the historical office resource push event.

[0040] In this embodiment, the session scene transition knowledge includes the session interaction features of historical cloud office users at X session nodes, X office resource usage events, and push decision change features of historical office resource push events, where X is an integer greater than 1. This structure provides a detailed information basis for in-depth analysis of user behavior in cloud office.

[0041] From the perspective of the conversation interaction features of the conversation nodes, each conversation node contains rich information. For example, if X is 10, the conversation interaction features of each of the 10 conversation nodes are multi-dimensional. For example, in the first conversation node, the user's interaction frequency may be 50 times / minute, the interaction duration is 15 minutes, and it is also accompanied by a high frequency of specific operation types (such as frequent file search operations). The second conversation node may present a different situation, with the interaction frequency dropping to 30 times / minute and the interaction duration shortened to 8 minutes, but more collaborative operations with others (such as shared document editing) appear. When the conversation interaction features of these 10 conversation nodes are fully connected, this is a way to deeply integrate information. The full connection operation associates each feature dimension of each conversation node to form a fully connected vector of scene interactions. This vector can comprehensively display the comprehensive situation of the user's interaction status on multiple conversation nodes. For example, after the full connection operation, it may be found that there is a potential correlation between certain features, such as conversation nodes with high interaction frequencies are often accompanied by specific types of operations, while conversation nodes with low interaction frequencies are related to other types of operations.

[0042] Generating office resource unique hot semantics based on the push decision change characteristics of X office resource usage events and historical office resource push events is a key link. Take office resource usage events as an example. For example, there are 5 main types of office resources, namely document editing resources, video conferencing resources, project management resources, email processing resources and data storage resources. The usage of office resources varies in different session nodes. For example, in the third session node, document editing resources are used a lot, so in the office resource unique hot semantics, the dimension corresponding to document editing resources will be marked as 1, and other resource dimensions will be marked as 0. If in the fifth session node, video conferencing resources begin to be used, then in the office resource unique hot semantics, the dimension of video conferencing resources will be changed to 1, and the document editing resource dimension previously marked as 1 will be restored to 0 (for example, one office resource is mainly used at the same time). For the push decision change characteristics, if in a certain period of time, the push decision of the historical office resource push event changes from pushing document editing resources to pushing project management resources, this change will be reflected in the office resource unique hot semantics through specific encoding or marking. This office resource unique hot semantics can clearly reflect the dynamic changes of office resource usage and push decisions.

[0043] Next, the original big data push analysis algorithm is used to push and analyze historical office resource push events based on scene interaction fully connected vectors, office resource unique hot semantics, session scene jump knowledge, and collaborative interaction state characteristics to obtain initial push analysis views. The original big data push analysis algorithm plays the role of a core analysis engine here. For example, this algorithm is an algorithm based on a complex neural network architecture, which is similar to a highly complex information processing network. The scene interaction fully connected vector, office resource unique hot semantics, session scene jump knowledge, and collaborative interaction state characteristics are used as input data for the input layer.

[0044] In the input layer of the neural network, each element in the fully connected vector of the scene interaction represents a specific information point in the user's interaction with the conversation node. Taking the interaction frequency and interaction duration mentioned earlier as an example, if the element with a high interaction frequency enters the neural network in the input layer, it will establish a connection with the neurons in the hidden layer. For example, the hidden layer has 200 neurons, and these neurons have different weighted treatments on the input elements. For elements with a high interaction frequency, the neurons associated with them may be given a larger weight. This is because in the analysis, a high interaction frequency may be closely related to the demand for office resources and push decisions. For example, if the user interaction frequency is high and concentrated on document editing operations, then the push demand for document editing resources may be higher, and the corresponding neuron weights will reflect this correlation.

[0045] Office resource unique semantics also plays an important role in neural networks. When the mark in the office resource unique semantics shows that a specific office resource is being used, the neural network can identify this situation and perform correlation analysis with historical office resource push events. For example, if the office resource unique semantics shows that video conferencing resources are being used, and the email processing resources are pushed in the historical office resource push event, this may be a signal of non-directional push. The neural network will perform complex calculations internally based on this signal, and may adjust the neuron connection weights related to the two resources in order to more accurately judge the rationality of the push in subsequent analysis.

[0046] The push decision change characteristics in the session scenario jump knowledge also have a key impact on the analysis. For example, when jumping from the document editing scenario to the video conferencing scenario, the push decision should be changed from pushing document editing resources to pushing video conferencing resources. If the push is not made according to this reasonable transition in the actual historical office resource push event, such as still pushing document editing resources, then this will be regarded as an abnormal situation by the neural network, and in the subsequent calculation, it will be analyzed in the direction of increasing the possibility of non-directional push. The collaborative interaction state characteristics, such as device performance status and network connection status, will also affect the analysis results. For example, if the device's CPU usage is 80% and the memory usage is 60%, this indicates that the device performance is under a high load state. In this case, if the historical office resource push event still pushes a large number of complex resources, such as high-definition video conferencing resources, this may be regarded as an unreasonable push because the device may not be able to handle these resources well. Similarly, if the network connection status is unstable, such as the packet loss rate reaches 10%, then the rationality of the office resource push that relies on the network, such as the push of online collaborative editing resources, may need to be re-evaluated.

[0047] Through such a comprehensive and detailed analysis process, we finally get the initial push analysis view of the historical office resource push event. This initial push analysis view is the discriminant analysis result of whether there is a non-directional push result in the historical office resource push event. For example, if the probability of non-directional push given by the initial push analysis view is 0.3, it means that according to the current analysis, there is a certain degree of possibility of non-directional push.

[0048] This design, by carefully incorporating multi-faceted conversation scene jump knowledge, covers conversation interaction characteristics of conversation nodes, office resource usage events, and push decision change characteristics, and adopts full-connection operations and generates office resource unique hot semantics, etc., to describe user behavior and resource push status in cloud office in an extremely detailed and comprehensive way. The original big data push analysis algorithm can significantly improve the accuracy of historical office resource push event analysis when combined with this rich information for push analysis. This comprehensiveness and accuracy helps to more effectively identify non-directional push situations, thereby providing a more reliable basis for subsequent directional pushes, greatly improving the accuracy of cloud office resource pushes, and thus improving user office efficiency. In the context of increasingly complex cloud office environments and increasingly high requirements for resource push accuracy, this technical solution can better meet user needs, optimize the allocation and utilization of office resources, and reduce unnecessary resource waste and user operation troubles.

[0049] On the basis of the above optional embodiment, the use of the original big data push analysis algorithm to perform push analysis on the historical office resource push events based on the scene interaction fully connected vector, the office resource unique hot semantics, the session scene jump knowledge and the collaborative interaction state characteristics to obtain the initial push analysis viewpoint of the historical office resource push events further includes: using the user behavior preference mining branch of the original big data push analysis algorithm to perform user behavior preference mining on the session scene jump knowledge to obtain a resource operation behavior preference vector for identifying non-directional push results; using the convolution of the original big data push analysis algorithm A processing branch is used to perform convolution processing on the scene interaction fully connected vector and the collaborative interaction state feature to obtain a collaborative scene interaction convolution vector; a residual connection branch of the original big data push analysis algorithm is used to perform residual connection on the collaborative scene interaction convolution vector to obtain a collaborative scene residual vector in the collaborative scene interaction convolution vector; a push analysis branch of the original big data push analysis algorithm is used to perform push analysis on the historical office resource push events based on the resource operation behavior preference vector, the collaborative scene residual vector and the office resource unique hot semantics to obtain an initial push analysis viewpoint of the historical office resource push events.

[0050] Under the above-mentioned technical solution framework, the original big data push analysis algorithm has a more detailed and complex operation process when it pushes and analyzes historical office resource push events based on the fully connected vector of scene interaction, the unique semantics of office resources, the knowledge of session scene jumps, and the characteristics of collaborative interaction status to obtain the initial push analysis point of view.

[0051] First, the user behavior preference mining branch of the original big data push analysis algorithm mines user behavior preferences for session scene jump knowledge, thereby obtaining a resource operation behavior preference vector for identifying non-directional push results. Session scene jump knowledge contains rich user behavior information, such as operation characteristics, office resource usage, and push decision changes in multiple session nodes. When mining user behavior preferences, it is a process of in-depth analysis of these complex information. For example, the office resource usage events in session scene jump knowledge involve multiple types, such as document editing resources, video conferencing resources, project management resources, etc. For document editing resources, in a specific sequence of session nodes, if it is found that the user uses the document editing resource frequently and the duration of each use is long, and the document editing resource is used with strong consistency when jumping from one session node to another, then in the resource operation behavior preference vector, the dimension related to the document editing resource will reflect the preference characteristics for the resource. For example, a higher value, such as 80 (indicating a higher preference tendency on this dimension), may be marked on this dimension to reflect the user's operation behavior preference for document editing resources.

[0052] Next, the convolution processing branch of the original big data push analysis algorithm convolves the scene interaction fully connected vector and the collaborative interaction state feature to obtain the collaborative scene interaction convolution vector. The scene interaction fully connected vector is a vector that integrates the conversation interaction features of multiple conversation nodes, while the collaborative interaction state feature contains information such as device performance and network connection status. For example, the dimension of the scene interaction fully connected vector is 100, and the dimension of the collaborative interaction state feature vector is 50. Convolution processing is an effective way to extract key features. By sliding the convolution kernel on these two vectors for convolution operations, hidden patterns and relationships can be discovered. For example, the convolution kernel may capture the association between certain features related to high interaction frequency in the scene interaction fully connected vector and the high performance state of the device in the collaborative interaction state feature. The collaborative scene interaction convolution vector obtained after convolution processing combines the key information in these two vectors and presents it in a new feature representation form, which is more conducive to subsequent analysis.

[0053] Then, the residual connection branch of the original big data push analysis algorithm performs residual connection on the collaborative scene interaction convolution vector to obtain the collaborative scene residual vector in the collaborative scene interaction convolution vector. The residual connection is to alleviate the gradient vanishing or gradient exploding problems that may occur in deep neural networks, and it also helps to better transmit information. In the collaborative scene interaction convolution vector, there may be different levels of feature representation. Through the residual connection, the deeper feature information is supplemented to the overall feature representation in the form of a residual. For example, the feature representation of the collaborative scene interaction convolution vector in a certain intermediate layer is [10, 20, 30]. After the residual connection, the obtained collaborative scene residual vector may add some additional information on the basis of the original features, such as [12, 22, 32]. These supplementary values ​​reflect the adjustment of the features brought about by the residual connection, so that the vector can more comprehensively and accurately represent the information in the collaborative scene.

[0054] Finally, the push analysis branch of the original big data push analysis algorithm performs push analysis on historical office resource push events based on the resource operation behavior preference vector, the collaborative scene residual vector, and the office resource unique hot semantics to obtain the initial push analysis view. The resource operation behavior preference vector reflects the user's operation preference for different office resources, the collaborative scene residual vector contains the key information of the collaborative scene interaction convolution vector after residual connection, and the office resource unique hot semantics clearly represents the usage of office resources and the change of push decision. For example, if the resource operation behavior preference vector shows that the user has a high preference for video conferencing resources (such as the relevant dimension value is 70), and the collaborative scene residual vector indicates that the current collaborative scene device has good performance and stable network connection (this state can be represented by a specific value range in the vector), the office resource unique hot semantics shows that the video conferencing resource is currently being used, but the document editing resource is pushed in the historical office resource push event, then in the push analysis branch, it will be judged based on this information that this may be a non-directional push situation, thereby reflecting a high possibility of non-directional push in the initial push analysis view, such as the non-directional push probability given by the initial push analysis view is 0.6.

[0055] It can be seen that through different processing operations performed by multiple branches of the original big data push analysis algorithm, various information related to cloud office resource push is comprehensively and deeply mined and analyzed. The user behavior preference mining branch can accurately obtain the user's operational behavior preferences for office resources, which helps to identify push situations that do not match user preferences. The convolution processing branch effectively integrates the scene interaction fully connected vector and the collaborative interaction state features through convolution operations to extract key collaborative scene information. The residual connection branch further optimizes the information representation of the collaborative scene interaction convolution vector and enhances the accuracy of information transmission. The push analysis branch analyzes the information of multiple vectors and improves the accuracy of judging whether the historical office resource push event is a non-directional push, thereby providing a more reliable basis for subsequent directional pushes and improving the accuracy and effectiveness of cloud office resource pushes.

[0056] In the next step, the user behavior preference mining branch of the original big data push analysis algorithm is used to perform user behavior preference mining on the session scene jump knowledge to obtain a resource operation behavior preference vector for identifying non-directional push results, including: using the user behavior preference mining branch of the original big data push analysis algorithm to generate behavior preference change knowledge reflecting the session scene jump knowledge; integrating the event activation label features of the X office resource usage events into the behavior preference change knowledge to obtain behavior preference fusion knowledge; using the first knowledge reinforcement coefficient, the second knowledge reinforcement coefficient and the third knowledge reinforcement coefficient of the user behavior preference mining branch to respectively reinforce the resource features of the behavior preference fusion knowledge to obtain first resource preference reinforcement knowledge, second resource preference reinforcement knowledge and third resource preference reinforcement knowledge; based on the first resource preference reinforcement knowledge, the second resource preference reinforcement knowledge and the third resource preference reinforcement knowledge, determining a resource operation behavior preference vector for identifying non-directional push results.

[0057] Furthermore, the resource operation behavior preference vector for identifying non-directional push results is determined based on the first resource preference reinforcement knowledge, the second resource preference reinforcement knowledge and the third resource preference reinforcement knowledge, including: performing knowledge feature collision on the first resource preference reinforcement knowledge and the second resource preference reinforcement knowledge to obtain a push matching discrimination feature reflecting the existence of non-directional push results in the historical office resource push event; based on the push matching discrimination feature and the third resource preference reinforcement knowledge, determining the resource operation behavior preference vector for identifying non-directional push results.

[0058] In this technical solution, the user behavior preference mining branch of the original big data push analysis algorithm has a series of complex and orderly operation processes when processing the session scene jump knowledge to obtain the resource operation behavior preference vector used to identify non-directional push results.

[0059] First, the user behavior preference mining branch of the original big data push analysis algorithm is used to generate behavior preference change knowledge that reflects the knowledge of session scene jumps. The knowledge of session scene jumps covers rich information of historical cloud office users in multiple session nodes, including session interaction characteristics, office resource usage events, and push decision change characteristics of historical office resource push events. The knowledge of behavior preference change is a deep refinement of these session scene jump knowledge. It summarizes the change pattern of user behavior preferences in different session scenarios from the perspective of user behavior. For example, in the knowledge of session scene jumps, if there is a frequent jump from the document editing scene to the video conferencing scene, the knowledge of behavior preference change may reflect the change of user behavior preference during the transition between these two scenes, such as the change from frequent use of document editing tools to the change of demand for video conferencing related functions (such as screen sharing, multi-person calls, etc.).

[0060] Next, the event activation label features of X office resource usage events are fused into the behavior preference change knowledge to obtain the behavior preference fusion knowledge. The event activation label features of office resource usage events are a way to mark the usage of office resources. For example, there are 5 types of office resources, such as document editing resources, video conferencing resources, project management resources, mail processing resources, and data storage resources. Each resource will have a corresponding event activation label feature when it is used. Taking video conferencing resources as an example, when a user starts a video conference, there may be a specific event activation label feature, such as [1, 0, 0, 0, 0] (here the first value is 1, indicating that the video conferencing resource is activated, and the others are 0, indicating that other resources are not activated). Fusion of these event activation label features into the behavior preference change knowledge can enable the behavior preference fusion knowledge to more comprehensively reflect the user's behavior preference in the process of office resource use. For example, if the behavior preference change knowledge shows that the user has a demand for video conferencing related functions after a certain session scene jumps, and the fused event activation label features indicate that the video conferencing resource is indeed activated in this scene, then the behavior preference fusion knowledge will more accurately reflect the association between this behavior preference and resource usage.

[0061] Then, the first knowledge reinforcement coefficient, the second knowledge reinforcement coefficient and the third knowledge reinforcement coefficient of the user behavior preference mining branch are used to reinforce the resource characteristics of the behavior preference fusion knowledge, respectively, to obtain the first resource preference reinforcement knowledge, the second resource preference reinforcement knowledge and the third resource preference reinforcement knowledge. These knowledge reinforcement coefficients are parameters set within the algorithm, which are used to adjust the resource feature weights in the behavior preference fusion knowledge. For example, the first knowledge reinforcement coefficient is 0.8, the second knowledge reinforcement coefficient is 0.6, and the third knowledge reinforcement coefficient is 0.4. If there is a dimension in the behavior preference fusion knowledge that represents the user's preference for document editing resources, and its initial value is 50 (indicating a degree of preference on this dimension), when it is reinforced with the first knowledge reinforcement coefficient of 0.8, the value of this dimension in the first resource preference reinforcement knowledge may become 50*0.8=40 (an exemplary adjustment based on coefficients, and the actual calculation may be more complicated). Similarly, similar adjustments will be made to other resources in the second resource preference reinforcement knowledge and the third resource preference reinforcement knowledge according to the corresponding coefficients.

[0062] Based on the first resource preference reinforcement knowledge, the second resource preference reinforcement knowledge, and the third resource preference reinforcement knowledge, a resource operation behavior preference vector for identifying non-directional push results is determined. This process includes performing knowledge feature collision on the first resource preference reinforcement knowledge and the second resource preference reinforcement knowledge to obtain a push matching discrimination feature that reflects the existence of non-directional push results in historical office resource push events. Knowledge feature collision is an operation that compares and analyzes feature differences and similarities in two knowledge structures. For example, the value of the dimension representing the user's preference for video conferencing resources in the first resource preference reinforcement knowledge is 40, and the value of this dimension in the second resource preference reinforcement knowledge is 30. This difference may be identified in the knowledge feature collision. If this difference exceeds a certain threshold, it may be regarded as a signal that a non-directional push result may exist, and thus reflected in the push matching discrimination feature.

[0063] Then, based on the push matching discriminant feature and the third resource preference reinforcement knowledge, the resource operation behavior preference vector for identifying non-directional push results is determined. For example, if the push matching discriminant feature shows that there may be problems in the push of video conferencing resources, and the preference for video conferencing resources in the third resource preference reinforcement knowledge also provides a certain reference, the combination of these two aspects of information can determine the dimension value related to the video conferencing resource in the resource operation behavior preference vector. For example, the push matching discriminant feature strongly suggests that there is a possibility of non-directional push of video conferencing resources, and the preference for video conferencing resources in the third resource preference reinforcement knowledge is low, then the dimension related to the video conferencing resource in the resource operation behavior preference vector may be marked with a lower value, such as 20, to indicate the situation of this resource in identifying non-directional push results.

[0064] Therefore, by generating behavioral preference change knowledge and integrating the event activation label features of office resource usage events, the relationship between the user's behavioral preference and office resource usage during the session scene jump process can be fully and deeply reflected. Different knowledge reinforcement coefficients are used to reinforce the resource characteristics of the behavioral preference fusion knowledge, and the weights of the resource characteristics are further refined and adjusted to make the reinforced knowledge more targeted. The collision of knowledge features and the determination of the resource operation behavior preference vector based on the push matching discriminant features and the third resource preference reinforcement knowledge improve the accuracy of the recognition of non-directional push results. This series of operations helps to more accurately identify non-directional push situations, thereby providing a more reliable basis for subsequent directional pushes, optimizing the effect of cloud office resource pushes, and improving users' office efficiency.

[0065] Under another design idea, the push analysis branch of the original big data push analysis algorithm is used to perform push analysis on the historical office resource push event based on the resource operation behavior preference vector, the collaborative scenario residual vector and the office resource unique hot semantics to obtain an initial push analysis viewpoint of the historical office resource push event, including: using the push analysis branch of the original big data push analysis algorithm to decompose the office resource unique hot semantics to obtain Y office resource unique hot semantic blocks; at least one of the Y office resource unique hot semantic blocks includes an office resource usage event and does not include a section of the historical office resource push event, and Y is an integer greater than 1; feature integration is performed on the resource operation behavior preference vector, the collaborative scenario residual vector and the Y office resource unique hot semantic blocks to obtain a push analysis integration vector; through the push analysis branch, based on the push analysis integration vector, push analysis is performed on the historical office resource push event to obtain an initial push analysis viewpoint of the historical office resource push event.

[0066] Under this design idea, the push analysis branch of the original big data push analysis algorithm has a complex and highly correlated operation process when it performs push analysis on historical office resource push events based on resource operation behavior preference vectors, collaborative scenario residual vectors, and office resource unique hot semantics to obtain initial push analysis views.

[0067] The push analysis branch of the original big data push analysis algorithm decomposes the unique semantics of office resources to obtain Y unique semantic blocks of office resources. The unique semantics of office resources is a semantic structure that is specially constructed to represent information related to office resource usage events and historical office resource push events. Consider a situation where, for example, the unique semantics of office resources covers comprehensive information related to 10 types of office resources. These 10 resource types are document editing resources, video conferencing resources, project management resources, email processing resources, data storage resources, team collaboration tool resources, schedule management resources, report making resources, office software update resources, and security protection resources. When it is decomposed into Y unique semantic blocks of office resources, for example, Y is 5, this means that the originally complex unique semantics of office resources must be carefully segmented.

[0068] In these five office resource unique semantic blocks, each semantic block contains specific office resource related information. At least one of the office resource unique semantic blocks includes office resource usage events and does not include the sections of historical office resource push events. For example, one of the office resource unique semantic blocks may focus on the relevant semantics of document editing resource usage events. This semantic block may contain various detailed information when the document editing resource is used, such as the creation time of the document, the frequency of editing, the type of document edited (whether it is a normal document, report document or project document, etc.), and some specific operations in the editing process (such as frequent format adjustments, content copy and paste, etc.). However, this semantic block does not contain the sections about document editing resource push in historical office resource push events, which is a very targeted information splitting. This splitting helps to analyze the different situations of office resource usage and push links more deeply, and effectively separate the information that may have been mixed together, thus laying the foundation for subsequent precise analysis.

[0069] Next, the resource operation behavior preference vector, the collaborative scenario residual vector, and the Y office resource unique hot semantic blocks are feature integrated to obtain the push analysis integration vector. The resource operation behavior preference vector is a vector that can reflect the user's preference for office resource operation behavior. For example, the dimension of this vector is 200, and each dimension corresponds to a feature related to office resource operation preference. For example, in this vector, the dimension related to video conferencing resources may indicate the user's preference for video conferencing resources. If the value of this dimension is 90, it means that the user has a higher preference for video conferencing resources in terms of operation behavior. For another example, the dimension value related to project management resources is 70, indicating that the user's preference for project management resources is slightly lower than that for video conferencing resources. These values ​​are obtained through a series of complex analyses and calculations based on the user's behavior data in the actual office process, and they can accurately reflect the user's operation tendency on different office resources.

[0070] The collaborative scene residual vector contains the key information in the collaborative scene interaction convolution vector after a series of complex processing. For example, the dimension of the collaborative scene residual vector is 100, and the values ​​in this vector reflect the various feature information in the collaborative scene. For example, some of these dimensions may be related to device performance. If the value of a dimension related to the device CPU utilization is 80, this means that the CPU utilization of the device is at a relatively high level. In addition, some dimensions may be related to the network connection status, such as the value of the network delay dimension is 30, indicating that the network delay is at a relatively low level. These values ​​together constitute the collaborative scene residual vector, which reflects the status of the collaborative scene in terms of device performance and network connection, and this information is closely related to the push and use of office resources.

[0071] When performing feature integration, the resource operation behavior preference vector, the collaborative scenario residual vector, and the information in the Y office resource hot semantic blocks should be comprehensively integrated. For the information in the office resource hot semantic block, the features representing the type of office resource, usage, and relationship with other office resources should be deeply associated and integrated with the preference features of the resource in the resource operation behavior preference vector. For example, for the office resource hot semantic block containing information related to document editing resources, the information about the frequency of use of document editing resources, document type, etc. should be associated with the degree of preference for document editing resources in the resource operation behavior preference vector. If the frequency of use of document editing resources is high in the office resource hot semantic block, and the degree of preference for document editing resources in the resource operation behavior preference vector is also high, then this association will be strengthened during the feature integration process.

[0072] At the same time, it is also necessary to combine the relevant collaborative scene information in the collaborative scene residual vector. For example, if the collaborative scene residual vector shows good device performance (such as low CPU usage, sufficient memory, etc.), then when integrating features, for those office resources that have high requirements for device performance (such as large project management software or high-definition video conferencing resources), this good device performance information will be integrated with the relevant information of these office resources in other vectors. The push analysis integration vector obtained in this way is a vector that integrates multi-faceted information. Each of its dimensions integrates relevant features from different information sources, so that it can more comprehensively and accurately reflect the various factors related to historical office resource push events.

[0073] Finally, through the push analysis branch, the historical office resource push events are analyzed based on the push analysis integration vector to obtain the initial push analysis view of the historical office resource push events. This push analysis branch will conduct in-depth analysis and judgment based on the information in the push analysis integration vector. For example, if in the push analysis integration vector, a dimension shows that users have a very high preference for a certain office resource (such as report making resources) (the relevant dimension value is 95), and in the office resource hot semantic block, it is shown that the report making resource has a specific operation mode in the office resource use event, such as frequent data import and export, complex formula application, etc., while the email processing resource is pushed in the historical office resource push event, this forms an obvious difference. This difference will be regarded as a signal of possible non-directional push in the push analysis process, resulting in the initial push analysis view giving a higher probability of non-directional push. For example, in this case, the non-directional push probability given by the initial push analysis view is 0.8, which indicates that according to the current analysis, the historical office resource push event has a high possibility of non-directional push.

[0074] It can be seen that by disassembling the unique semantics of office resources to obtain multiple semantic blocks, this method can analyze the information related to office resources in detail, and effectively separate and refine the information related to usage events and push events. This helps to more accurately explore the characteristics and relationships of office resources in different links, avoiding confusion and general processing of information. The push analysis integration vector is obtained by integrating the resource operation behavior preference vector, the collaborative scene residual vector, and the unique semantic block of office resources. This process comprehensively integrates various key information from different aspects that are closely related to push analysis. It organically combines multiple aspects of information such as user operation preferences for office resources, the status of collaborative scenes, and the specific situation of office resource use and push, avoiding the limitation of relying solely on a single information source for analysis. Push analysis based on the push analysis integration vector can improve the accuracy of judging whether the historical office resource push event is a non-directional push. By comprehensively considering various factors, it can more accurately identify possible non-directional push situations, thereby providing a more reliable basis for subsequent directional push. This helps to optimize the strategy of cloud office resource push, improve the accuracy and efficiency of office resource push, and thus improve the work efficiency and experience of users in the cloud office environment.

[0075] In a further technical solution of the above optional embodiment, the generating of office resource unique hot semantics reflecting the X office resource usage events and the historical office resource push events based on the push decision change characteristics of the X office resource usage events and the historical office resource push events includes: generating an initial cloud office session log including the X office resource usage events and the historical office resource push events based on the push decision change characteristics of the X office resource usage events and the historical office resource push events; if the feature granularity of the initial cloud office session log is lower than the preset feature granularity, performing text optimization on the initial cloud office session log to obtain a session optimization log; performing semantic mining on the X office resource usage events in the session optimization log to obtain office resource unique hot semantics reflecting the X office resource usage events and the historical office resource push events.

[0076] In this further technical solution of the above optional embodiment, office resource unique hot semantics reflecting these events are generated based on the push decision change characteristics of X office resource usage events and historical office resource push events, involving a series of orderly and closely related operation processes.

[0077] First, based on the push decision change characteristics of X office resource usage events and historical office resource push events, an initial cloud office session log including X office resource usage events and historical office resource push events is generated. This initial cloud office session log is a way to record the usage and push decision of relevant office resources. For example, if X is 5, these 5 office resource usage events may involve document editing resources, video conferencing resources, project management resources, email processing resources, and data storage resources, respectively. In the initial cloud office session log, relevant information of these office resource usage events will be recorded, such as the time sequence of use, usage duration, frequency of use, etc. At the same time, the push decision change characteristics of historical office resource push events will also be recorded, such as the decision change point from pushing document editing resources to pushing video conferencing resources and the corresponding timestamp and other information.

[0078] Next, we need to consider the feature granularity of the initial cloud office session log. If the feature granularity of the initial cloud office session log is lower than the preset feature granularity, we need to perform text optimization on the initial cloud office session log to obtain a session optimization log. Feature granularity is an indicator to measure the level of detail of the cloud office session log. For example, in the initial cloud office session log, if the usage event of the document editing resource only records the usage time of 30 minutes, and the preset feature granularity requires recording more details, such as the type of editing operation within these 30 minutes (whether it is simple text input, format adjustment or picture insertion, etc.), the size change of the document, etc. If the initial cloud office session log does not reach such a detailed level, text optimization is required. In the text optimization process, data mining technology, natural language processing technology and other means may be used to supplement the missing information to obtain a session optimization log.

[0079] Then, semantic mining is performed on the X office resource usage events in the session optimization log to obtain the unique office resource semantics used to reflect the X office resource usage events and historical office resource push events. In the session optimization log, the information of the office resource usage events is more detailed and comprehensive. Semantic mining is a technical means to extract meaningful information from text data. For example, for the document editing resource usage event, in the semantic mining process, it may be identified that in a specific time period, the document editing resource is mainly used to write project reports (this is a semantic information), and in this process there is an association with the project management resource (for example, the project progress information is obtained from the project management resource for document editing). For the video conference resource usage event, semantic mining may find that the main purpose of the meeting is to discuss the technical difficulties in the project, and the participants are mainly technical personnel in the project team. Through such a semantic mining process, the unique office resource semantics can be constructed for each office resource usage event and historical office resource push event. This semantics can accurately reflect the various relationships and characteristics of office resources in the use and push process.

[0080] With this design, by generating the initial cloud office session log, it is possible to preliminarily record the relevant information of office resource usage events and historical office resource push events, providing basic data for subsequent analysis. The consideration of fine-grained features and text optimization when necessary ensure that the cloud office session log has sufficient detail and avoids inaccurate analysis due to missing or insufficient information. Semantic mining is performed on office resource usage events in the session optimization log to obtain the unique semantics of office resources. This method can deeply mine the semantic information in the process of office resource use and push, thereby more accurately reflecting the relationship between office resources and the characteristics of use and push. This helps to more accurately analyze whether there is non-directional push in historical office resource push events in the subsequent big data push analysis algorithm, thereby providing a more reliable basis for directional push and improving the accuracy and efficiency of cloud office resource push.

[0081] In a preferred embodiment, the debugging of the original big data push analysis algorithm based on the prior push feedback data and the initial push analysis viewpoint in step 160 to obtain a target big data push analysis algorithm includes: determining the non-directional push analysis loss of the original big data push analysis algorithm based on the prior push feedback data and the initial push analysis viewpoint; determining the stability quantitative characteristics of the original big data push analysis algorithm based on the non-directional push analysis loss; if the stability quantitative characteristics of the original big data push analysis algorithm do not meet the loss compliance conditions, then correcting the original big data push analysis algorithm based on the non-directional push analysis loss to obtain a target big data push analysis algorithm.

[0082] Further, if the stability quantification feature of the original big data push analysis algorithm does not meet the loss standard condition, then based on the non-directional push analysis loss, the original big data push analysis algorithm is corrected to obtain a target big data push analysis algorithm, including: if the stability quantification feature of the original big data push analysis algorithm does not meet the loss standard condition, then based on the non-directional push analysis loss, the algorithm parameters of the original big data push analysis algorithm are corrected to obtain a corrected original big data push analysis algorithm; through the corrected original big data push analysis algorithm, based on the session scene jump knowledge and the collaborative interaction state characteristics, the historical office resource push event is pushed and analyzed to obtain a target push analysis viewpoint of the historical office resource push event; the target push analysis viewpoint is a detection result for characterizing that the historical office resource push event has a non-directional push result; based on the target push analysis viewpoint and the prior push feedback data, the stability quantification feature of the corrected original big data push analysis algorithm is determined; if the stability quantification feature of the corrected original big data push analysis algorithm is a stability quantification feature, the corrected original big data push analysis algorithm is determined as the target big data push analysis algorithm.

[0083] In this preferred embodiment, the original big data push analysis algorithm is debugged based on the prior push feedback data and the initial push analysis viewpoint to obtain the target big data push analysis algorithm. This process is a complex and progressive technical solution, covering multiple key links and there is a close logical connection between each link.

[0084] First, the non-directional push analysis loss of the original big data push analysis algorithm is determined based on the prior push feedback data and the initial push analysis viewpoint. As an authoritative prior authentication data that characterizes the existence of non-directional push results in historical office resource push events, the prior push feedback data has high accuracy and reference value. For example, the prior push feedback data may come from the precise statistics and in-depth analysis of resource push situations in a large number of actual office scenarios. For example, in a batch of historical office resource push events, for a specific type of office resource push (such as an update push for a certain professional software), the prior push feedback data shows that the actual proportion of its non-directional push is 0.7 (0.7 in the embodiment of the present invention is a numerical example reflecting the severity of the non-directional push situation in the office resource push, and the higher the value, the more serious the non-directional push situation).

[0085] The initial push analysis viewpoint is the initial judgment and analysis result of the original big data push analysis algorithm on whether there are non-directional push results in the historical office resource push events. This initial analysis result is obtained based on the logic of the original algorithm itself and the data processed. For example, when the original algorithm analyzes the above-mentioned specific office resource push, the initial push analysis viewpoint gives a non-directional push ratio of 0.5. Then, the difference between the two (0.7-0.5=0.2) reflects a certain degree of non-directional push analysis loss. This analysis loss is not just a simple numerical difference, it reflects the degree of deviation between the original algorithm in the current state and the accurate judgment of non-directional push situations. This deviation may be caused by a variety of factors such as the algorithm's own model structure, parameter settings, or the data characteristics used.

[0086] Next, the stability quantitative characteristics of the original big data push analysis algorithm are determined based on the non-directional push analysis loss. The stability quantitative characteristics are an indicator that comprehensively considers the stability performance of the original algorithm in many aspects when processing non-directional push analysis tasks. It not only focuses on the non-directional push analysis loss in a single analysis, but also considers the fluctuation of this non-directional push analysis loss in the analysis of multiple different historical office resource push events.

[0087] For example, 10 different historical office resource push events are analyzed, and each analysis obtains a corresponding non-directional push analysis loss value. If these loss values ​​fluctuate greatly, such as the loss value is 0.1 in the first analysis, 0.3 in the second, and 0.05 in the third, this indicates that the original algorithm is unstable when processing different data. This instability may imply that the algorithm has poor adaptability to factors such as different types of office resource push events, different session scenario jump knowledge, or collaborative interaction state characteristics. On the contrary, if these loss values ​​are relatively stable and have a small fluctuation range, such as between 0.1-0.2, then the stability quantitative characteristics of the original algorithm are relatively high.

[0088] When determining the quantitative characteristics of stability, other relevant factors may also be considered. For example, the performance of the algorithm when faced with data of different sizes. If the loss of non-directional push analysis is small and stable when processing a small amount of historical office resource push event data, but as the amount of data increases, the loss value fluctuates more, this will also affect the evaluation of the quantitative characteristics of stability. In addition, different types of office resource push events (such as document resource push, conference resource push, etc.) may have different effects on algorithm stability. These factors need to be comprehensively considered in the process of determining the quantitative characteristics of stability.

[0089] If the stability quantitative characteristics of the original big data push analysis algorithm do not meet the loss compliance conditions, it is necessary to modify the original big data push analysis algorithm based on the non-directional push analysis loss to obtain the target big data push analysis algorithm. The loss compliance condition is a strict pre-set standard based on the accuracy requirements for cloud office resource push and the needs of actual application scenarios. This standard is set to ensure that the algorithm can stably and accurately judge non-directional push situations in actual applications.

[0090] For example, the loss compliance condition is set as the stability quantification feature must reach a certain threshold, such as 0.8 (a numerical example indicating the degree of stability quantification, the higher the value, the better the stability). If the stability quantification feature of the original algorithm is evaluated to be 0.6, which is lower than this threshold, then the loss compliance condition is not met and the algorithm needs to be modified.

[0091] Specifically, if the stability quantitative characteristics of the original big data push analysis algorithm do not meet the loss criteria, the algorithm parameters of the original big data push analysis algorithm are modified based on the non-directional push analysis loss to obtain the modified original big data push analysis algorithm. Algorithm parameters are the core components of the original big data push analysis algorithm, which directly determine the algorithm's operating logic and analysis results. In different types of algorithms, algorithm parameters have different manifestations.

[0092] Taking the original big data push analysis algorithm based on neural networks as an example, the algorithm parameters may include the connection weights between neurons, bias values, the number of network layers, the number of neurons in each layer, etc. If it is found in the previous analysis that the non-directional push analysis loss is large and the stability quantitative characteristics do not meet the requirements, it may be because the connection weights between some neurons are set unreasonably. For example, the neuron connection weights related to a specific office resource push event are too high or too low, causing the algorithm to overemphasize or ignore certain key factors during analysis. At this time, these connection weights need to be adjusted.

[0093] For example, in the original algorithm, the weight of the neuron connection related to the frequency of use of a certain office resource is 0.8. After analysis, it is found that this weight setting causes the algorithm to have a large deviation when judging non-directional push, and this weight may be adjusted to 0.6. At the same time, similar adjustments may be made to the bias value. If the original bias value is 0.1, it may be adjusted to 0.05 based on the loss of non-directional push analysis. In addition, for the number of network layers and the number of neurons in each layer, if it is found that the algorithm has difficulties in processing complex conversation scenario jump knowledge and collaborative interaction state characteristics, the number of network layers may be increased or the number of neurons in each layer may be adjusted to improve the analysis ability and accuracy of the algorithm.

[0094] Then, through the revised original big data push analysis algorithm, the historical office resource push events are pushed and analyzed based on the session scene jump knowledge and collaborative interaction state characteristics, and the target push analysis view of the historical office resource push events is obtained. The session scene jump knowledge contains rich information about the user's behavior in the cloud office process, such as the user's conversion between different office tasks (such as document editing, video conferencing, project management, etc.), conversion frequency, and operation mode in different scenarios. The collaborative interaction state characteristics reflect the performance status of office equipment (such as CPU usage, memory occupancy, etc.), network connection status (such as bandwidth, latency, packet loss rate, etc.) and the interaction between users and devices (such as the frequency of use of input and output devices, operation speed, etc.).

[0095] When performing push analysis, the revised algorithm will take these factors into consideration. For example, for a historical office resource push event, such as pushing project management tool resources to a team. If the session scene jump knowledge shows that the team members have just switched from the video conferencing scene to the project management scene, and the switching frequency is high, this may imply that the team currently has a high demand for project management resources. At the same time, if the collaborative interaction status characteristics indicate that the device performance is good (CPU utilization is 30%, memory occupancy is 40%), and the network connection is stable (bandwidth is 100Mbps, latency is 10ms, and packet loss rate is 0.1%), then the revised algorithm will make a comprehensive judgment based on this information when analyzing this push event.

[0096] For example, the target push analysis viewpoint obtained by the revised algorithm gives a non-directional push degree of 0.65. This value is closer to 0.7 in the prior push feedback data than the initial push analysis viewpoint of the original algorithm (such as 0.5), indicating that the revised algorithm is more accurate in judging the results of non-directional push.

[0097] Determine the stability quantitative characteristics of the revised original big data push analysis algorithm based on the target push analysis viewpoint and the prior push feedback data. This process is similar to the previous process of determining the stability quantitative characteristics of the original algorithm, but this time the evaluation is based on the target push analysis viewpoint obtained by the revised algorithm.

[0098] For example, we analyze multiple groups of different historical office resource push events again to obtain the corresponding target push analysis viewpoints and the corresponding non-target push analysis loss values. If we find that the fluctuation range of these loss values ​​is significantly smaller than that of the original algorithm and is closer to the prior push feedback data, then the stability quantitative characteristics of the revised original big data push analysis algorithm are improved. For example, in the original algorithm, the stability quantitative characteristic is 0.6. After correction, in the new analysis process, the stability quantitative characteristic is improved to 0.85.

[0099] If the stability quantification feature of the revised original big data push analysis algorithm is a stable quantification feature, that is, it meets the pre-set requirements for stability quantification features, the revised original big data push analysis algorithm is determined as the target big data push analysis algorithm. This target big data push analysis algorithm has higher accuracy and stability, can better adapt to the complex situation in the cloud office resource push process, and accurately judge whether there is a non-directional push result in the historical office resource push event.

[0100] In summary, by accurately determining the non-directional push analysis loss and stability quantitative characteristics, the performance of the original big data push analysis algorithm in processing non-directional push analysis tasks can be comprehensively and deeply evaluated. When the stability quantitative characteristics do not meet the requirements, the algorithm parameters are targeted and corrected according to the non-directional push analysis loss. This correction process can gradually optimize the algorithm and make it more accurate in judging the non-directional push results. Based on the target push analysis point of view and prior push feedback data, the stability quantitative characteristics are determined again and judged to ensure that the final target big data push analysis algorithm reaches a high level in terms of stability and accuracy. This helps to significantly improve the reliability of office resource push analysis, thereby providing a more accurate basis for the targeted push of cloud office resources, effectively improving the effectiveness and user experience of cloud office resource push, reducing resource waste and user troubles caused by non-directional push, and improving the overall efficiency of cloud office.

[0101] Based on the above content, in an extended embodiment, the method also includes: determining the current push analysis point of view of the cloud office user to be analyzed through the target big data push analysis algorithm; based on the current non-directional push results in the current push analysis point of view, generating an office resource directional push optimization strategy for the cloud office user to be analyzed.

[0102] In detail, first, the current push analysis viewpoint of the cloud office user to be analyzed is determined by the target big data push analysis algorithm. The target big data push analysis algorithm is an algorithm obtained through a series of previous operation optimizations, and its accuracy and stability are high. This current push analysis viewpoint includes multiple judgments on the office resource push situation of the cloud office user to be analyzed. For example, there may be a value indicating the degree of non-directional push in the current push analysis viewpoint, such as 0.6 (indicating that there is a certain degree of non-directional push in the office resource push, and the higher the value, the more serious the non-directional push situation). This value may be based on a combination of multiple factors, including the user's operating behavior in different office scenarios, the use of office resources, and the collaborative interaction status. At the same time, the current push analysis viewpoint may also include a judgment on the rationality of the push of different office resources, such as some office resources are judged to have a higher possibility of non-directional push, while others are considered to be more reasonable to push.

[0103] Based on the current non-directional push results in this current push analysis viewpoint, we start to generate a targeted push optimization strategy for office resources for the cloud office users to be analyzed. This process is a comprehensive and in-depth analysis process involving multiple aspects of consideration.

[0104] From the perspective of the types of office resources, it is necessary to classify and analyze the office resources in the cloud office environment in detail. For example, office resources can be divided into document processing resources, communication and collaboration resources, project management resources, etc. For document processing resources, it is necessary to further analyze their performance in the current push analysis point of view. If in the current non-directional push results, the non-directional push value of document processing resources is high, such as 0.7 (indicating that document processing resources have serious non-directional problems in push), then when generating a targeted push optimization strategy, it is necessary to focus on how to adjust the push for such resources. This may involve in-depth exploration of the functional characteristics of document processing resources, user demand characteristics, etc. For example, some users may need document processing software with advanced format editing functions, but the current push may not take this demand into account, resulting in non-directional push.

[0105] In terms of user behavior habits, it is necessary to deeply analyze the operation behavior patterns of the cloud office users to be analyzed. For example, by analyzing the frequency of users' use of office resources in different time periods. For example, the data shows that users use communication and collaboration resources most frequently between 9 am and 11 am, reaching 30 operations per hour (30 operations in the embodiment of the present invention is a numerical example representing the frequency of use), while the frequency of using project management resources between 2 pm and 4 pm is higher, with 20 operations per hour. At the same time, analyze the operation sequence and preferences of users in different office scenarios. For example, in the project execution scenario, users usually use project management resources for task allocation first, and then use communication and collaboration resources for team communication. If it is found in the current non-directional push result that the push sequence does not match the user's behavior habits, such as pushing communication and collaboration resources first in the project execution scenario, and then pushing project management resources, this may be a factor leading to non-directional push. Therefore, when generating an optimization strategy for directional push of office resources, the push sequence and time should be adjusted according to the user's behavior habits.

[0106] The collaborative interaction state is also a key factor. The collaborative interaction state includes the device performance state and the network connection state. For example, the CPU usage rate of the device is 80% (here 80% means that the CPU of the device is in a high usage state), and the memory occupancy rate is 70%. In this case, the device performance may affect the push effect of office resources. If the current push analysis view shows that some office resources with high requirements for device performance, such as large project management software, are pushed under this device performance state, and non-directional push results appear, then when generating a directional push optimization strategy, it is necessary to consider adjusting the type of office resources pushed according to the device performance state. For the network connection state, if the network bandwidth is 50Mbps (50Mbps in the embodiment of the present invention is a numerical example representing the network bandwidth) and the delay is 100ms, under this network condition, if a high-definition video conferencing resource (this resource has high network requirements) is pushed and a non-directional push result appears, then in the optimization strategy, it is necessary to consider avoiding pushing such resources with high network requirements when the network conditions are poor, or adjusting the parameters of the pushed resources to adapt to the network conditions, such as reducing the image quality requirements of the video conference.

[0107] The historical push data of office resources should not be ignored either. Analyzing the historical push data of office resources can reveal a lot of useful information. For example, the frequency, time and user feedback of pushing a certain office resource in the past. For example, in the past month, office software update resources were pushed once every Monday morning, but user feedback showed that few users updated at this time, and the update rate was only 10% (here 10% is a numerical example representing the update rate), which indicates that there may be problems with this push method. In the current non-directional push results, if it is found that the push of office software update resources is also unreasonable, then when generating a directional push optimization strategy, it is necessary to refer to the historical push data and adjust the time or method of push. For example, you can change to push office software update resources during the user's free time or according to the user's usage habits.

[0108] The association and complementarity between office resources are also important factors to consider when generating targeted push optimization strategies. For example, document processing resources and data storage resources are associated because users usually need to store documents after processing them. If it is found in the current non-targeted push results that the push of these two types of resources is not well combined, for example, document processing resources are often pushed separately, but the related data storage resources are not pushed in time, this may affect the user's office efficiency, resulting in non-targeted push. Therefore, when generating optimization strategies, it is necessary to consider the reasonable combination and push of related and complementary office resources. For example, when pushing document processing resources, you can also recommend related data storage resources or provide an integrated solution for the two.

[0109] By comprehensively considering the above factors, based on the current non-directional push results in the current push analysis viewpoint, an optimization strategy for the directional push of office resources for the cloud office users to be analyzed can be generated. This optimization strategy can accurately adjust the push of office resources based on the actual needs, behavioral habits, collaborative interaction status, and historical push of office resources of users, improve the directional nature of office resource push, reduce the situation of non-directional push, and thus improve the office efficiency and satisfaction of users.

[0110] With this design, by comprehensively considering multiple factors to generate an optimization strategy for targeted push of office resources, users' office resource needs can be met more accurately. Considering the non-targeted push results in the current push analysis point of view, problems in the push can be directly optimized. Detailed analysis of office resource types helps to push reasonably according to the characteristics of different resources. Analyzing user behavior habits can match the push with the user's operation mode and improve the user's acceptance of pushed resources. Considering the collaborative interaction state can ensure that the pushed office resources are adapted to the device performance and network conditions, avoiding resource waste and push failure. Referring to the historical push data of office resources can avoid repeating past erroneous push methods and improve the success rate of push. Paying attention to the correlation and complementarity between office resources can further improve users' office efficiency, make office resource push more scientific and reasonable, and improve the effectiveness and user experience of cloud office resource push as a whole.

[0111] In summary, by obtaining the collaborative office session behavior information and prior push feedback data of historical cloud office users, the non-directional push situation in the historical office resource push events can be accurately located. The knowledge of session scene jumps and the characteristics of collaborative interaction status are obtained from the collaborative office session behavior information, which enriches the analysis dimension and makes the analysis of push events more comprehensive. The original big data push analysis algorithm is combined with these characteristics to obtain the initial push analysis point of view, which helps to accurately judge the non-directional push results. The algorithm is debugged based on the prior push feedback data to obtain the target algorithm, which improves the accuracy of the algorithm and makes the target push analysis point of view more consistent with the prior data, thereby providing a reliable basis for the targeted push of cloud office resources and improving the accuracy of office resource push and user office efficiency.

[0112] Furthermore, a readable storage medium is provided, on which a program is stored, and the program implements the above method when executed by a processor.

[0113] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described in the above exemplary embodiment can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the embodiments of the present invention, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

Claims

1. A user behavior big data analysis method applied to cloud office resource push, characterized in that: The method is implemented by a user behavior big data analysis system, and the method includes: Acquire the collaborative office session behavior information of historical cloud office users, and the prior push feedback data of the historical office resource push event corresponding to the collaborative office session behavior information, wherein the prior push feedback data is the prior authentication data used to characterize that the historical office resource push event has a non-directional push result; From the collaborative office session behavior information of the historical cloud office user, the session scene jump knowledge of the historical cloud office user and the collaborative interaction state characteristics of the historical collaborative interaction terminal corresponding to the historical cloud office user are obtained, wherein the session scene jump knowledge refers to the relevant information of the user switching from one work scene to another work scene during the collaborative office session, and the collaborative interaction state characteristics include the network connection stability of the interaction terminal and the performance status of the interactive device; Using the original big data push analysis algorithm, based on the session scene jump knowledge and the collaborative interaction state characteristics, the historical office resource push event is pushed and analyzed to obtain an initial push analysis view of the historical office resource push event; the initial push analysis view is a discriminant analysis result used to characterize that the historical office resource push event has a non-directional push result; Based on the prior push feedback data and the initial push analysis viewpoint, the original big data push analysis algorithm is debugged to obtain a target big data push analysis algorithm; the consistency evaluation weight between the target push analysis viewpoint of the historical office resource push event generated by the target big data push analysis algorithm and the prior push feedback data is greater than the consistency evaluation weight between the initial push analysis viewpoint and the prior push feedback data.

2. The method according to claim 1, characterized in that The session scenario jump knowledge includes the session interaction features of the historical cloud office user at X session nodes, the X office resource usage events at the X session nodes, and the push decision change features of the historical office resource push events, where X is an integer greater than 1; The method of using the original big data push analysis algorithm to perform push analysis on the historical office resource push events based on the session scene transition knowledge and the collaborative interaction state characteristics to obtain an initial push analysis viewpoint of the historical office resource push events includes: Performing a full connection operation on the session interaction features of the X session nodes to obtain a scene interaction full connection vector of the historical cloud office user; Based on the push decision change characteristics of the X office resource usage events and the historical office resource push events, generating office resource unique hot semantics for reflecting the X office resource usage events and the historical office resource push events; By utilizing the original big data push analysis algorithm, based on the scene interaction fully connected vector, the office resource unique hot semantics, the session scene jump knowledge and the collaborative interaction state characteristics, the historical office resource push event is pushed and analyzed to obtain an initial push analysis view of the historical office resource push event.

3. The method according to claim 2, characterized in that The method of using the original big data push analysis algorithm to perform push analysis on the historical office resource push events based on the scene interaction fully connected vector, the office resource unique hot semantics, the session scene jump knowledge, and the collaborative interaction state characteristics, and obtaining an initial push analysis viewpoint of the historical office resource push events includes: Utilizing the user behavior preference mining branch of the original big data push analysis algorithm, user behavior preference mining is performed on the session scene transition knowledge to obtain a resource operation behavior preference vector for identifying non-directional push results; Using the convolution processing branch of the original big data push analysis algorithm, the scene interaction fully connected vector and the collaborative interaction state feature are convolved to obtain a collaborative scene interaction convolution vector; Utilizing the residual connection branch of the original big data push analysis algorithm, residual connection is performed on the collaborative scene interaction convolution vector to obtain a collaborative scene residual vector in the collaborative scene interaction convolution vector; By utilizing the push analysis branch of the original big data push analysis algorithm, based on the resource operation behavior preference vector, the collaborative scenario residual vector and the office resource unique hot semantics, a push analysis is performed on the historical office resource push events to obtain an initial push analysis view of the historical office resource push events.

4. The method according to claim 3, characterized in that The user behavior preference mining branch of the original big data push analysis algorithm is used to perform user behavior preference mining on the session scene transition knowledge to obtain a resource operation behavior preference vector for identifying non-directional push results, including: Utilize the user behavior preference mining branch of the original big data push analysis algorithm to generate behavior preference change knowledge reflecting the session scene jump knowledge; Integrating the event activation label features of the X office resource usage events into the behavior preference change knowledge to obtain behavior preference fusion knowledge; The first knowledge reinforcement coefficient, the second knowledge reinforcement coefficient and the third knowledge reinforcement coefficient of the user behavior preference mining branch are used to respectively reinforce the resource characteristics of the behavior preference fusion knowledge to obtain the first resource preference reinforcement knowledge, the second resource preference reinforcement knowledge and the third resource preference reinforcement knowledge; Based on the first resource preference reinforcement knowledge, the second resource preference reinforcement knowledge, and the third resource preference reinforcement knowledge, a resource operation behavior preference vector for identifying non-directional push results is determined.

5. The method according to claim 4, characterized in that The determining, based on the first resource preference enhancement knowledge, the second resource preference enhancement knowledge, and the third resource preference enhancement knowledge, a resource operation behavior preference vector for identifying a non-directional push result includes: Performing knowledge feature collision on the first resource preference reinforcement knowledge and the second resource preference reinforcement knowledge to obtain a push matching discrimination feature reflecting that the historical office resource push event has a non-directional push result; Based on the push matching discrimination feature and the third resource preference reinforcement knowledge, a resource operation behavior preference vector for identifying non-directional push results is determined.

6. The method according to claim 3, characterized in that The push analysis branch using the original big data push analysis algorithm performs push analysis on the historical office resource push events based on the resource operation behavior preference vector, the collaborative scenario residual vector, and the office resource unique hot semantics to obtain an initial push analysis view of the historical office resource push events, including: The push analysis branch of the original big data push analysis algorithm is used to decompose the office resource unique hot semantics to obtain Y office resource unique hot semantic blocks; at least one of the Y office resource unique hot semantic blocks includes an office resource usage event and does not include a section of the historical office resource push event, and Y is an integer greater than 1; Performing feature integration on the resource operation behavior preference vector, the collaborative scenario residual vector, and the Y office resource unique hot semantic blocks to obtain a push analysis integration vector; Through the push analysis branch, based on the push analysis integration vector, the historical office resource push event is subjected to push analysis to obtain an initial push analysis viewpoint of the historical office resource push event.

7. The method according to claim 2, characterized in that The generating, based on the push decision change characteristics of the X office resource usage events and the historical office resource push events, office resource unique hot semantics for reflecting the X office resource usage events and the historical office resource push events includes: Based on the push decision change characteristics of the X office resource usage events and the historical office resource push events, generating an initial cloud office session log including the X office resource usage events and the historical office resource push events; If the feature granularity of the initial cloud office session log is lower than the preset feature granularity, text optimization is performed on the initial cloud office session log to obtain a session optimization log; Semantic mining is performed on the X office resource usage events in the session optimization log to obtain office resource unique hot semantics that reflect the X office resource usage events and the historical office resource push events.

8. The method according to claim 1, characterized in that The debugging of the original big data push analysis algorithm based on the prior push feedback data and the initial push analysis viewpoint to obtain the target big data push analysis algorithm includes: Determining a non-directional push analysis loss of the original big data push analysis algorithm based on the prior push feedback data and the initial push analysis viewpoint; Determining the stability quantitative characteristics of the original big data push analysis algorithm based on the non-directional push analysis loss; If the stability quantitative characteristics of the original big data push analysis algorithm do not meet the loss standard condition, then based on the non-directional push analysis loss, the original big data push analysis algorithm is modified to obtain the target big data push analysis algorithm; If the stability quantification feature of the original big data push analysis algorithm does not meet the loss standard condition, the original big data push analysis algorithm is corrected based on the non-directional push analysis loss to obtain a target big data push analysis algorithm, including: If the stability quantitative characteristics of the original big data push analysis algorithm do not meet the loss compliance condition, then based on the non-directional push analysis loss, the algorithm parameters of the original big data push analysis algorithm are corrected to obtain a corrected original big data push analysis algorithm; By using the modified original big data push analysis algorithm, based on the session scene jump knowledge and the collaborative interaction state characteristics, the historical office resource push event is pushed and analyzed to obtain a target push analysis viewpoint of the historical office resource push event; the target push analysis viewpoint is a detection result for characterizing that the historical office resource push event has a non-directional push result; Determining the stability quantitative characteristics of the modified original big data push analysis algorithm based on the target push analysis viewpoint and the prior push feedback data; If the stability quantification feature of the modified original big data push analysis algorithm is a stability quantification feature, the modified original big data push analysis algorithm is determined as the target big data push analysis algorithm.

9. A user behavior big data analysis system, characterized in that: It comprises a processor and a memory; the processor and the memory are communicatively connected, and the processor is used to read a computer program from the memory and execute it to implement the method described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program implements the method according to any one of claims 1 to 8 when running.

Images