User communication behavior analysis method based on big data

Abstract

The invention relates to the technical field of big data analysis, and discloses a user communication behavior analysis method based on big data, and the method comprises the following steps: S1, collecting user multi-source communication behavior data, generating a basic communication behavior data set and a derivative behavior feature data set, s2, performing behavior index quantification processing on the basic communication behavior data set to generate a user communication behavior quantification index set, and S3, performing dynamic weight distribution processing on the derived behavior characteristic data set based on a time decay model to generate a user behavior aging characteristic matrix; through an innovatively designed time decay weight mechanism, a modeling capability of dynamic evolution of behavior characteristics is endowed, and the technology captures an aging change rule of user communication preferences, identifies dynamic characteristics of sudden increase of video traffic in a commuting period and conversation mode conversion in holidays and festivals, enables an analysis model to have a self-adaptive updating capability, and improves user experience. The problem that response to user behavior changes is lagged in a traditional scheme is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data analysis, and particularly to a user communication behavior analysis method based on big data. BACKGROUND

[0002] Communication refers to the transmission of information from one party to another party accurately and safely without violating the will of each party in the broad sense. Data communication is a new communication mode generated by the combination of communication technology and computer technology. Transmission channels are required for transmitting information between two places. According to the difference of transmission media, there are wired data communication and wireless data communication. However, they are all connected by transmission channels to link data terminals and computers, so that different data terminals at different places can share software, hardware and information resources.

[0003] The current user communication behavior analysis technology mainly has the following defects: 1. Static modeling limitations: The existing user communication behavior analysis method only relies on historical communication data for clustering analysis, does not consider the time decay effect of behavior characteristics, and equates early communication records with recent behaviors, which leads to a lag in the response of the model to user preference changes, and cannot capture time-sensitive behavior patterns, resulting in a decrease in the prediction accuracy of the model over time.

[0004] 2. Lack of group association: The existing individual communication behavior portrait construction technology analyzes the behavior characteristics of a single user through a classification algorithm, ignores the influence mechanism of group behavior on individual decision-making, and does not establish a user-group association model, which leads to the inability to quantify the transmission effect of social circles on consumption behavior, and causes the service recommendation system to ignore the transmission influence in the social network, resulting in a disconnection between the recommended strategy and the actual demand.

[0005] 3. Coarse granularity of service matching: The existing communication service matching technology adopts a service matching mechanism based on simple rule logic, which has the following defects: first, the rules completely rely on manual experience setting and lack self-adaptive adjustment ability, second, the quantitative mapping relationship between portrait characteristics and service strategies is not established, and finally, the cross-dimension feature coordination mechanism is missing, especially the correlation between social influence and use preference is not utilized, which leads to insufficient accuracy of service recommendation.

[0006] Therefore, in view of the above problems, the present application provides a user communication behavior analysis method based on big data, which can integrate a new analysis framework of behavior time correction, group association quantization and service accurate mapping. SUMMARY

[0007] (I) Technical problems solved

[0008] In view of the deficiencies of the prior art, the present application provides a user communication behavior analysis method based on big data, which solves the problems in the background art.

[0009] (II) Technical solutions

[0010] To achieve the above object, the present application provides the following technical solutions: a user communication behavior analysis method based on big data, comprising the following steps: S1, collecting user multi-source communication behavior data to generate a basic communication behavior data set and a derived behavior feature data set; S2, performing behavior index quantization processing on the basic communication behavior data set to generate a user communication behavior quantization index set; S3, performing dynamic weight distribution processing on the derived behavior feature data set based on a time decay model to generate a user behavior time effectiveness feature matrix; S4, constructing a user behavior feature vector space according to the user communication behavior quantization index set and the user behavior time effectiveness feature matrix; S5, performing group behavior feature matching processing on the user behavior feature vector space through a group behavior calibration algorithm to generate user group behavior correlation data; S6, generating user multi-dimensional communication portrait data based on the user group behavior correlation data and combining a behavior portrait layering model; S7, performing communication service optimization matching processing according to the user multi-dimensional communication portrait data and a service adaptation matrix to generate target communication service strategy data.

[0011] Preferably, the S1 comprises: S11, collecting user basic communication data through a mobile terminal, including call frequency, call duration, traffic consumption type and APP use period, to generate a basic communication behavior data set U; S12, extracting derived behavior feature data based on the basic communication behavior data set U, including communication stability coefficient, behavior preference entropy value and social network density, to generate a derived behavior feature data set V, wherein category data adopts embedding coding Embedding instead of one-hot coding One-Hot, and original one-hot coding will cause high-dimensional sparse space, and embedding coding compresses it into a low-dimensional continuous vector: wherein is a discrete category feature, is an embedding function, is a continuous vector after embedding; The embedding layer is trained by a shallow neural network to map the category features to a continuous space and smooth the non-convex region. S13, generate user space trajectory feature data W by fusing base station positioning data and WiFi probe data, and incorporate the derived behavior feature data set V.

[0012] Preferably, the S2 includes: S21, perform index normalization processing on the base communication behavior data set U, and the calculation formula is: wherein is the original value, , are the maximum and minimum values of the same index, respectively; Use normalization + integer quantization processing and add integer truncation mechanism: wherein is the normalized feature value, is the discretized feature value, is the discretization factor, which is set based on data distribution; S22, calculate the user communication entropy value based on the normalization result wherein is the proportion of the ith behavior index, is the communication behavior entropy value; S23, construct a user communication behavior quantization index set wherein is the behavior time fluctuation coefficient.

[0013] Preferably, the S3 includes: S31, set a time decay factor wherein is the decay rate parameter, is the time length from the behavior occurrence, is the time decay factor, and a piecewise linearization processing is added in the time decay model to convert the time period discrete value into a continuous probability distribution: wherein is a Gaussian kernel function based on time t, is a reference time point of behavior occurrence, is a parameter for controlling the speed of time decay, is the behavior feature value after time weighting, is the original behavior feature value, is the current time point; S32, calculate the decay factor of each feature in the derived behavior feature data set V Weighted by decay factor: wherein is the original behavior feature value, is the time-effect weighted feature value, is the time decay factor.

[0014] Preferably, the S4 comprises: S41, mapping the user communication behavior quantitative indicator set to a dimension behavior vector ; S42, expanding the user behavior time-effect feature matrix to a feature vector ; S43, merging and to construct a user behavior feature vector space .

[0015] 7. Preferably, the group behavior calibration algorithm execution step in the S5 comprises: S51, position initialization supports discrete sampling, changes continuous uniform distribution to integer or mixed sampling: wherein is the lower bound of the feature space, is the upper bound of the feature space, type is the feature type identifier, is the initialization population position, is the discrete-continuous mixed sampler, is the population individual number, is the population size, is the integer set, is the population size; DiscreteRand samples from integer uniform distribution for discrete features, and still uses uniform distribution for continuous features; S52, exploration stage: replace continuous position update with discrete random walk to avoid gradient failure: wherein is the position vector of the i-th fish eagle individual at time t, is the position vector of the i-th fish eagle individual at time t+1, is the current optimal group position vector, is the fish eagle individual index, t is the iteration number, is the discrete disturbance term, d is the maximum step length, This is the learning rate coefficient. For random disturbance coefficients, For random number generator, Centralized random numbers; S53. During the development phase, to handle non-convex spaces, multiple group collaboration and multiple restart strategies were added: in For the Lévy flight function, This is the step size control coefficient. The dimension of the feature space; S54. Introduce penalty functions to handle discrete constraints and avoid invalid solutions: in The final fitness function is defined as fitness, and the basic fit function is defined as fitness. Let be the feature vector to be evaluated. For the current user's feature vector, For group feature template vectors, The penalty coefficient is... For the penalty term function; S55. Add a validity check for group association data: Feasibility indicators for calculating the solution: in The function for calculating the Pearson correlation coefficient. For relevance, when If the value is less than 0.7, the algorithm will be restarted. Before outputting, cluster-based filtering will be applied to remove outlier solutions.

[0016] Preferably, S6 includes: S61. Construct a hierarchical profile model: Base layer: Communication pattern labels are generated based on Q-clustering; Preference layer: Behavioral tendency weights are based on T-weighting; Social layer: The group influence index is calculated based on G; S62. Generate multi-dimensional user communication profile data .

[0017] Preferably, S7 includes: S71. Constructing a service adaptability matrix This represents the matching weight between service i and profile dimension j; S72, Computational Service Strategy Score As the image layer weight, For service-profile adaptation, is an image layer feature value, is a service policy score; S73, generating target communication service policy data .

[0018] Preferably, the step S1 of collecting user multi-source communication behavior data comprises: S14, performing correlation analysis on the basic communication behavior data set and external social network behavior data through a cross-domain data fusion engine to generate a cross-platform behavior correlation matrix; S15, performing abnormal behavior filtering processing on the cross-platform behavior correlation matrix based on a space-time constraint condition, and the calculation formula is: wherein is a social network behavior vector, is a communication behavior vector, is a behavior similarity threshold, is a behavior time interval, is a time window threshold; S16, injecting the filtered correlation data into the derived behavior feature data set V, and updating the social network density feature weight: wherein is an updated social network density weight, is an original social network density weight, is a social enhancement coefficient, is a correlation event counter.

[0019] Preferably, it comprises: The population calibration algorithm engine is configured with a parallel computing architecture, and a MapReduce framework is used to realize distributed computing of E-OOA population position updating, and the MapReduce framework supports discrete computing: Mapper processes feature encoding, and Reducer performs discrete position updating of E-OOA; The service matching matrix library updates the policy weight in real time, and connects the communication service provider database through an API interface.

[0020] (Three) beneficial effects

[0021] Compared with the prior art, the present application provides an analysis method of user communication behavior based on big data, which has the following beneficial effects: 1. Overcome the limitation of static modeling, realize dynamic behavior evolution tracking: Through the time decay weight mechanism designed by innovation, the modeling ability of dynamic evolution of behavior characteristics is given, the time change law of user communication preference is captured, the dynamic characteristics of video traffic surge in commuting period and holiday call mode change are identified, the adaptive updating ability of the analysis model is obtained, and the problem of slow response of traditional scheme to user behavior change is solved.

[0022] 2. Break through the bottleneck of missing group association, and establish a social transmission quantization model: By developing a group behavior calibration algorithm, a mathematical association network of users-group is constructed in the behavior characteristic space. The technology realizes the quantitative analysis of the influence of social circle, accurately describes the group behavior transmission mechanism of suite following and application propagation effect, injects social dynamics elements into the service recommendation system, and solves the problem of low utilization rate of group influence factor in the traditional scheme.

[0023] 3. Reconstruct service matching logic to realize accurate quantization of strategy output: By the collaborative design of hierarchical portrait and service adaptation matrix, a mathematical mapping system of characteristics-strategy is established. The technology upgrades the fuzzy experience rule to a calculable quantization matching model, ensures the interpretability of the strategy, realizes the matching of service supply and user demand, and solves the pain point problem of lack of quantization basis in the traditional service recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0026] Please refer to Figure 1 The user communication behavior analysis method based on big data comprises the following steps: S1, collecting user multi-source communication behavior data to generate basic communication behavior data set and derived behavior characteristic data set; S2, performing behavior index quantization processing on the basic communication behavior data set to generate a user communication behavior quantization index set; S3, performing dynamic weight distribution processing on the derived behavior characteristic data set based on a time decay model to generate a user behavior time-effect characteristic matrix; S4, constructing a user behavior characteristic vector space according to the user communication behavior quantization index set and the user behavior time-effect characteristic matrix; S5, performing group behavior feature matching processing on the user behavior feature vector space through a group behavior calibration algorithm to generate user group behavior correlation degree data; S6, generating user multi-dimensional communication portrait data based on the user group behavior correlation degree data and in combination with a behavior portrait hierarchical model; S7, performing communication service optimization matching processing according to the user multi-dimensional communication portrait data and a service adaptation degree matrix to generate target communication service strategy data.

[0027] S1 includes: S11, collecting user basic communication data through a mobile terminal, including call frequency, call duration, traffic consumption type and APP use period, to generate a basic communication behavior data set U; S12, based on the basic communication behavior data set U, extracting derived behavior feature data, including communication stability coefficient, behavior preference entropy value and social network density, to generate a derived behavior feature data set V, wherein category data uses embedding coding Embedding instead of one-hot coding One-Hot, and original one-hot coding will cause high-dimensional sparse space, and embedding coding compresses it into a low-dimensional continuous vector: wherein is a discrete category feature, is an embedding function, is a continuous vector after embedding; The embedding layer maps the category feature to a continuous space through a shallow neural network training, and smooths the non-convex region; S13, generating user spatial trajectory feature data W through fusion processing of base station positioning data and WiFi probe data, and incorporating the derived behavior feature data set V.

[0028] S2 includes: S21, performing index normalization processing on the basic communication behavior data set U, and the calculation formula is: wherein is an original value, , are maximum and minimum values of the same index, respectively; Use normalization + integer quantization processing to add an integer truncation mechanism: wherein is a normalized feature value, is a discretized feature value, is a discretization factor, which is set based on data distribution; S22, calculating user communication entropy value based on the normalization result wherein is the proportion of the i-th behavior indicator, is the communication behavior entropy value; S23, constructing a user communication behavior quantification indicator set wherein is the behavior time fluctuation coefficient.

[0029] S3 includes: S31, setting a time decay factor wherein is the decay rate parameter, is the time length from the occurrence of the behavior, is the time decay factor, and a piecewise linearization process is added to the time decay model to convert the discrete time period value into a continuous probability distribution: wherein is the Gaussian kernel function based on time t, is the reference time point of the behavior occurrence, is the parameter for controlling the speed of time decay, is the behavior feature value weighted by time, is the original behavior feature value, is the current time point; S32, weighting each feature in the derived behavior feature data set V according to the decay factor: wherein is the original behavior feature value, is the time-effective weighted feature value, is the time decay factor.

[0030] S4 includes: S41, mapping the user communication behavior quantification indicator set to a dimensional behavior vector ; S42, expanding the user behavior time-effective feature matrix into a feature vector ; S43, merging and to construct a user behavior feature vector space .

[0031] The group behavior calibration algorithm execution steps in S5 include: S51, position initialization supports discrete sampling, which changes the continuous uniform distribution to integer or mixed sampling: ​ where is the feature space lower bound, is the feature space upper bound, type is the feature type identifier, is the initialization population position, is the discrete-continuous hybrid sampler, is the population individual number, is the population size, is the integer set, is the population size; DiscreteRand samples from the uniform distribution of integers for discrete features, and still uses the uniform distribution for continuous features ; S52, Exploration phase: replace continuous position update with discrete random walk to avoid gradient failure: where is the position vector of the i-th fish eagle individual at time t, is the position vector of the i-th fish eagle individual at time t+1, is the current optimal group position vector, is the fish eagle individual index, t is the iteration number, is the discrete disturbance term, d is the maximum step length, is the learning rate coefficient, is the random disturbance coefficient, is the random number generator, is the centralized random number; S53, Development phase is to deal with non-convex space, add multi-population cooperation and multiple restart strategy: where is the Levy flight function, is the step length control coefficient, is the feature space dimension; S54, introduce penalty function to deal with discrete constraints, avoid invalid solution: where is the final fitness function, fitness is the basic matching degree function, is the feature vector to be evaluated, is the current user feature vector, is the group feature template vector, is the penalty coefficient, is the penalty term function; S55, add validity check of group correlation degree data: A feasibility index of the solution is calculated: wherein is the Pearson correlation coefficient calculation function, is the correlation degree, when is less than 0.7, the algorithm is triggered to restart, and before output, a screening based on clustering is applied to remove outlier solutions.

[0032] S6 includes: S61, constructing a hierarchical portrait model: The basic layer: the communication mode label is generated based on Q clustering; The preference layer: the behavior tendency weight is based on T weighting; The social layer: the group influence index is calculated based on G; S62, generating user multi-dimensional communication portrait data .

[0033] S7 includes: S71, constructing a service adaptation degree matrix represents the matching weight of service i and portrait dimension j; S72, calculating the service strategy score is the portrait layer weight, is the service-portrait adaptation degree, is the portrait layer feature value, is the service strategy score; S73, generating target communication service strategy data .

[0034] The user multi-source communication behavior data collected in step S1 includes: S14, performing correlation analysis on the basic communication behavior data set and external social network behavior data through a cross-domain data fusion engine to generate a cross-platform behavior correlation matrix; S15, performing abnormal behavior filtering processing on the cross-platform behavior correlation matrix based on a space-time constraint condition, and the calculation formula is: wherein is the social network behavior vector, is the communication behavior vector, is the behavior similarity threshold, is the behavior time interval, is the time window threshold; S16, injecting the filtered correlation data into the derived behavior feature data set V, and updating the social network density feature weight: wherein a social network density weight for the updated, a social network density weight for the original, a social enhancement coefficient, a correlation event counter.

[0035] comprising: The population calibration algorithm engine is configured with a parallel computing architecture, and the distributed computing of the E-OOA population position update is realized by using the MapReduce framework. The MapReduce framework is extended to support discrete computing: the Mapper processes the feature encoding, and the Reducer performs the discrete position update of the E-OOA; The service matching matrix library updates the strategy weight in real time, and is connected with the communication service provider database through an API interface.

[0036] comprising: The data acquisition module includes a terminal data acquisition unit, a base station fusion unit, and a behavior feature extraction unit. The behavior analysis module includes a quantitative processing unit, a time correction unit, and a population calibration algorithm engine. The service optimization module includes a portrait generation unit, a service matching matrix library, and a strategy execution unit.

[0037] The fish eagle optimization algorithm in step S5 includes a discrete position update and a penalty function mechanism, wherein the position update adopts a hybrid discrete-continuous sampling, and the formula is: wherein is the position at the current time, is the position at the next time, i is the individual index, and t is the iteration time, is a discrete disturbance term, this formula is used for discrete characteristics, and a discrete constraint penalty term is added to the fitness function.

[0038] Embodiment 1: Dynamic behavior modeling and time correction: The embodiment specifically implements a user behavior dynamic modeling process, collects basic communication data with accurate time stamps through a mobile terminal SDK, including call frequency, traffic consumption type and APP use period, simultaneously fuses base station positioning and WiFi probe data to capture user space movement trajectory, performs sliding window verification on original data, filters abnormal data points caused by signal jitter, in the time weight distribution stage, innovatively establishes a dual-mode attenuation model: the basic communication features adopt a linear attenuation algorithm, and the complex behavior features adopt an exponential attenuation algorithm, and constructs a three-layer time correction engine: the input layer receives the original behavior vector, the attenuation calculation layer automatically matches the algorithm according to the feature type, and the output layer generates a time-weighted feature matrix, finally inputs the corrected feature matrix into a hierarchical clustering model, the first layer of clustering identifies the high-frequency call type of the basic communication mode, the second layer of clustering captures the dynamic evolution law of the traffic surge feature in the commuting period, and finally outputs the user behavior portrait coding with a time label, realizing the transition from static analysis to dynamic tracking.

[0039] Embodiment 2: Group behavior calibration and social association The embodiment implements a group feature matching process, first constructs a five-dimensional behavior feature space: maps the user behavior quantitative index set to a two-dimensional plane coordinate, X-axis: behavior diversity index, Y-axis: service preference intensity, simultaneously projects the time feature matrix to a three-dimensional space to add Z-axis: behavior stability coefficient, unifies the dimension through a space transformation algorithm, then executes the fish eagle optimization algorithm: initializes to randomly generate 20 groups of feature vectors representing typical group modes; in the exploration stage, each fish eagle detects in the feature space along three random directions, calculates the feature distance with the target user and retains the top five positions; in the development stage, a fine search grid is established at the optimal positions, the local optimal solution is located through the gradient descent method, and the most matched group feature vector is output, finally analyzes the social association factor in the vector, accurately quantifies the social circle layer penetration rate and behavior transmission intensity, and generates an association degree report labeled with influence dimension and action intensity.

[0040] Embodiment 3: Hierarchical portrait and service matching: The embodiment completes the service strategy generation whole process. In the hierarchical portrait construction stage: the basic layer labels the dominant type through communication behavior clustering and adds time limit label, the preference layer establishes behavior decision tree model, automatically assigns weight coefficient according to feature importance, the social layer quantifies group influence factor and calculates propagation effect value, the service adaptation matrix construction stage: the strategy library containing basic communication service package and value-added service is established, the three-dimensional evaluation system is innovatively designed, X dimension evaluates basic layer matching degree, Y dimension measures preference layer fit degree, Z dimension quantifies social layer gain degree, when the strategy is generated, portrait data is input into the adaptation matrix, three-dimensional score is calculated in parallel, strategy sorting queue is established to automatically filter low-fit options, three strategies and matching degree analysis report are output, and real-time feedback mechanism is used to monitor user response behavior, and the matrix weight parameter is dynamically adjusted to form a closed-loop optimization.

[0041] It should be noted that the relational terms herein, such as first and second, are used solely to distinguish one from another entity or action without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises... a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0042] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, alternatives, and variations can be made in the embodiments without departing from the spirit and scope of the present application as defined by the appended claims and their equivalents.

Claims

1. A method for analyzing user communication behavior based on big data, characterized in that, The method comprises the following steps: S1, collecting user multi-source communication behavior data to generate a basic communication behavior data set and a derived behavior feature data set; S2, performing behavior index quantization processing on the basic communication behavior data set to generate a user communication behavior quantization index set; S3, performing dynamic weight allocation processing on the derived behavior feature data set based on a time decay model to generate a user behavior time effectiveness feature matrix; S4, constructing a user behavior feature vector space according to the user communication behavior quantization index set and the user behavior time effectiveness feature matrix; S5, performing group behavior feature matching processing on the user behavior feature vector space through a group behavior calibration algorithm to generate user group behavior correlation data; S6, generating user multi-dimensional communication portrait data based on the user group behavior correlation data and combining a behavior portrait layering model; S7, performing communication service optimization matching processing on the user multi-dimensional communication portrait data and a service adaptation matrix to generate target communication service strategy data.

2. The big data based analysis method of user communication behavior according to claim 1, wherein, The S1 comprises: S11, collecting user basic communication data through a mobile terminal, including call frequency, call duration, traffic consumption type and APP use period, to generate a basic communication behavior data set U; S12, extracting derived behavior feature data based on the basic communication behavior data set U, including communication stability coefficient, behavior preference entropy value and social network density, to generate a derived behavior feature data set V, wherein category data adopts embedding coding Embedding instead of one-hot coding One-Hot, and original one-hot coding will cause high-dimensional sparse space, and embedding coding compresses it into a low-dimensional continuous vector: wherein is a discrete categorical feature, is an embedding function, is an embedded continuous vector; An embedding layer maps category features to a continuous space through a shallow neural network training, and smooths non-convex regions; S13, performing fusion processing on base station positioning data and WiFi probe data to generate user spatial trajectory feature data W, and incorporating the derived behavior feature data set V. 3.The big data based analysis method of user communication behavior according to claim 1, wherein, The S2 comprises: S21, performing index normalization processing on the basic communication behavior data set U, and the calculation formula is: wherein is the original value, , are the maximum and minimum values of the same index, respectively; Using normalization + integer quantization processing, adding an integer truncation mechanism: wherein is a normalized feature value, is a discretized feature value, is a discretization factor, set based on data distribution; S22, based on the normalization result, calculate the user communication entropy value wherein is the proportion of the i-th type of behavior index, is the communication behavior entropy value; S23, constructing a user communication behavior quantification index set wherein is a behavior time fluctuation coefficient. 4.The big data based analysis method of user communication behavior according to claim 1, wherein, The S3 comprises: S31, set time decay factor wherein is a decay rate parameter, is a time-to-occurrence length, is a time decay factor, in which a piecewise linearization process is added to the time decay model to convert the discrete values of time periods into a continuous probability distribution: wherein is a Gaussian kernel function based on time t, is a reference time point of the behavior occurrence, is a parameter to control the speed of time decay, is the behavior feature value after time weighting, is the original behavior feature value, is the current time point; S32, derive a feature data set V from the feature data set U Weighted by a decay factor: wherein is the original behavior characteristic value, is the time-dependent weighted characteristic value, is a time decay factor. 5.The big data based analysis method of user communication behavior according to claim 1, wherein, The S4 comprises: S41, quantifying the set of user communication behavior indicators mapping to dimensional behavior vector ; S42, the user behavior time-effect feature matrix is expanded into a feature vector ; S43, merge with constructing a user behavior feature vector space . 6.The big data based analysis method of user communication behavior according to claim 1, wherein, The group behavior calibration algorithm execution steps in the S5 comprise: S51, supporting discrete sampling for position initialization, changing continuous uniform distribution to integer or hybrid sampling: wherein is a feature space lower bound, is a feature space upper bound, type is a feature type identifier, is an initialization population position, is a discrete-continuous hybrid sampler, is a population individual number, is a population size, is an integer set, is a population size; For discrete features, DiscreteRand samples from an integer uniform distribution, for continuous features, still a uniform distribution ; S52, exploration stage: replacing continuous position update with discrete random walk to avoid gradient failure: wherein is the position vector of the i-th fish eagle individual at time t, is the position vector of the i-th fish eagle individual at time t+1, is the current best swarm position vector, is the fish eagle individual index, t is the iteration number, is the discrete perturbation term, d is the maximum step size, is the learning rate coefficient, is the random perturbation coefficient, is the random number generator, is the centralizing random number; S53, development stage for processing non-convex space, adding multi-population cooperation and multiple restart strategies: wherein is the Levy flight function, is a step control coefficient, is the characteristic space dimension; S54, introducing a penalty function to process discrete constraints to avoid invalid solutions: wherein is the final fitness function, fitness is the base matching function, is the feature vector to be evaluated, is the current user feature vector, is the population feature template vector, is the penalty coefficient, is the penalty term function; S55, adding effectiveness checking of group correlation data: Calculating the feasibility index of the solution: wherein is the Pearson correlation coefficient calculation function, is the degree of association, when less than 0.7, the algorithm is triggered to restart, and before output, a cluster-based screening is applied to remove outliers. 7.The big data based analysis method of user communication behavior according to claim 1, wherein, The S6 comprises: S61, constructing a layered portrait model: The basic layer: communication mode labels are generated based on Q clustering; The preference layer: behavior tendency weights are based on T weighting; The social layer: group influence index is based on G calculation; S62, generating user multi-dimensional communication portrait data . 8.The big data based analysis method of user communication behavior according to claim 1, wherein, The S7 comprises: S71, constructing a service adaptation degree matrix represents the matching weight of service i and image dimension j; S72, compute service policy score for image layer weight, for service-image fit, for image layer feature value, for service policy score; S73, generating target communication service policy data 9.The big data based analysis method of user communication behavior according to claim 1, wherein, The step S1 of collecting user multi-source communication behavior data comprises: S14, through the cross-domain data fusion engine, the basic communication behavior data set is associated with external social network behavior data, and cross-platform behavior correlation matrix is generated; S15, based on the space-time constraint condition, the cross-platform behavior correlation matrix is filtered and handled, and the calculation formula is: wherein is a social network behavior vector, is a communication behavior vector, is a behavior similarity threshold, is a behavior time interval, is a time window threshold; S16, the filtered correlation data is injected into the derived behavior feature data set V, and the social network density feature weight is updated: wherein is the updated social network density weight, is the original social network density weight, is the social enhancement factor, is the associated event counter.

10. The method of claim 1, wherein, It includes: The group calibration algorithm engine is configured with a parallel computing architecture, and a MapReduce framework is used to realize distributed calculation of E-OOA population position update, and the MapReduce framework supports discrete calculation: Mapper processes feature encoding, and Reducer executes discrete position update of E-OOA; The service matching matrix library updates the strategy weight in real time, and is connected with the communication service provider database through an API interface.

Images