A Multi-Dimensional Mining-Based Intelligent Analysis Method for Carrier Logs

Abstract

The present invention provides an intelligent analysis method of operator logs based on multi-dimensional mining in the field of log analysis technology, including: step S10, obtaining communication logs; step S20, constructing location data based on grids and cells; step S30, obtaining user consumption data; step S40, pre-processing communication logs, location data, and user consumption data to construct a log data set; step S50, filtering the log data set to divide it into residential state data and working state data; step S60, clustering the residential state data and working state data to obtain spatial features; step S70, creating a behavior feature extraction model, using the log data set to train the behavior feature extraction model, and using the trained behavior feature extraction model to extract behavior features; step S80, using spatial features and behavior features to optimize resource allocation and analyze user behavior. The advantages of the present invention are: accurate extraction of user spatial features and behavior features is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of log analysis, and in particular to an intelligent analysis method of operator logs based on multi-dimensional mining. Background Art

[0002] With the widespread use of mobile networks and mobile payments, and the emergence of various social apps, the access traffic brought by mobile Internet is growing explosively at an unprecedented rate. However, in the context of such rapid development of mobile Internet, the value generated by operators' voice, SMS and value-added services is gradually becoming saturated. At the same time, with multiple rounds of rate cuts, the cancellation of roaming fees for traffic services, and the promotion of unlimited traffic packages, the profits created by traffic have also shown signs of sluggish growth.

[0003] The reasons why operators have invested a lot of manpower, material and financial resources but have only seen weak revenue growth can be summarized as follows: 1. Compared with the traffic value created by Internet products such as WeChat, Weibo, and Douyin short videos, operators can only obtain a small amount of profit from it, and thus gradually tend to the edge of the industrial chain, and gradually become the "plumbers" between Internet companies and users; 2. Due to the explosive growth in traffic brought about by the rapid development of the Internet, telecom operators are also facing the problem of increased costs; 3. There are many problems in the marketing and service methods of operators facing users, and they cannot carry out multi-faceted and precise marketing to users.

[0004] For the above reasons, operators are in urgent need of optimizing resource allocation and conducting precision marketing to users to ensure revenue. Therefore, how to provide an intelligent analysis method for operator logs based on multi-dimensional mining to accurately extract user spatial features and behavioral features has become a technical problem that needs to be solved. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent analysis method of operator logs based on multi-dimensional mining, so as to accurately extract user space features and behavior features.

[0006] The present invention is implemented as follows: a method for intelligent analysis of operator logs based on multi-dimensional mining, comprising the following steps:

[0007] Step S10: Obtain the communication logs of each user including call logs and Internet access logs;

[0008] Step S20, setting a plurality of grids of preset sizes on the map based on the location distribution of the base stations, and constructing location data based on the grids and the locations of the cells carried on the map;

[0009] Step S30: Obtain user consumption data;

[0010] Step S40: pre-processing the communication log, location data and user consumption data to construct a log data set;

[0011] Step S50, filtering the log data set by status, and further dividing it into residential status data and working status data;

[0012] Step S60: clustering the residential status data and the work status data respectively using the DBSCAN clustering algorithm to obtain spatial features including residential locations and work locations;

[0013] Step S70: creating a behavior feature extraction model based on deep learning technology, using the log data set to train the behavior feature extraction model, and using the trained behavior feature extraction model to extract the user's behavior features;

[0014] Step S80: Utilize the spatial characteristics and behavioral characteristics to optimize resource allocation and analyze user behavior.

[0015] Furthermore, the step S10 is specifically as follows:

[0016] Obtaining the communication logs of each user including call logs and Internet access logs, and aggregating the communication logs in hours;

[0017] The communication log carries at least user ID, date, time, call record, LAC, CI, longitude, latitude and connected base station; the CI is the base station sector ID.

[0018] Furthermore, in step S20, the grid carries the land use; the land use is residential, commercial or industrial;

[0019] The relationship between the grid and the cells is many-to-many.

[0020] Furthermore, in step S30, the user consumption data includes at least user ID, VIP level, network access year, user age, user gender and consumption characteristics.

[0021] Furthermore, the step S40 specifically includes:

[0022] Step S41, removing irrelevant data items from the communication log, location data and user consumption data;

[0023] Step S42: removing noise data from the communication log, location data, and user consumption data;

[0024] Step S43: removing duplicate data from the communication log, location data, and user consumption data;

[0025] Step S44: feature-encoding the communication log, location data, and user consumption data;

[0026] Step S45: completing missing values ​​in the communication log, location data, and user consumption data based on a linear interpolation method;

[0027] Step S46: construct a log data set based on the communication log, location data and user consumption data.

[0028] Furthermore, the step S50 specifically includes:

[0029] Step S51, determining the cell where the user is located through the LAC, CI, longitude and latitude carried in the communication log;

[0030] Step S52: matching the corresponding grid from the location data based on the cell, thereby obtaining the land use corresponding to the cell;

[0031] Step S53: based on the date and time carried in the communication log, counting the user's visit frequency and residence time in the corresponding cell;

[0032] Step S54: combining the land use, visit frequency and residence time, dividing the log data set into residential status data and work status data.

[0033] Furthermore, the step S60 specifically includes:

[0034] Step S61: clustering the residential status data and the work status data respectively based on longitude and latitude using a DBSCAN clustering algorithm in units of user ID to obtain a residential location cluster and a work location cluster;

[0035] Step S62: Determine the residential location and the work location through the number of days and duration that the base station is connected in the residential location cluster and the work location cluster, and construct spatial features based on the residential location and the work location.

[0036] Furthermore, in step S70, the convolution layer of the behavior feature extraction model uses a 3D convolution kernel, the activation function uses a nonlinear function, and the pooling method uses average pooling.

[0037] Furthermore, in step S70, the use of the log data set to train the behavior feature extraction model is specifically as follows:

[0038] Converting the log data set into a three-dimensional single-channel grayscale feature map, using the three-dimensional single-channel grayscale feature map as input of a behavior feature extraction model, and using the behavior feature as output of the behavior feature extraction model, so as to train the behavior feature extraction model;

[0039] The three-dimensional single-channel grayscale feature map includes an X-axis feature map, a Y-axis feature map and a Z-axis feature map; the X-axis feature map is composed of user ID and call records; the Y-axis feature map is segmented data with a granularity of 5 minutes; the Z-axis feature map is segmented data with a time slice as the granularity; the time slice is 5 hours;

[0040] The behavioral characteristics include at least: the frequency and standard deviation of calls between the calling and called users; the mean, variance, skewness and kurtosis of the calling user's call duration; the mean number of calls made by the calling user; the mode, skewness and standard deviation of the calling user's province; the mean, standard deviation, skewness and kurtosis of the calling user's call charges; the category, rate package and real-name attributes of the calling and called users; the account opening time of the calling user; the calling user's downtime status and downtime duration; and the mean monthly consumption amount of the user.

[0041] Furthermore, in the step S80, the resource configuration optimization is specifically optimizing the layout of base stations; and the user behavior analysis is specifically analyzing abnormal behaviors of users.

[0042] The advantages of the present invention are:

[0043] 1. By combining the time dimension and the space dimension, the log data set is divided into residential status data and working status data, and the DBSCAN clustering algorithm is used to cluster the residential status data and the working status data respectively to obtain spatial features, which greatly improves the accuracy of user spatial feature extraction; by preprocessing the communication logs, location data and user consumption data and constructing a log data set, and then using the log data set to train the behavior feature extraction model, the preprocessing eliminates interference data, and the integration of communication logs, location data and user consumption data is closer to the real business scenario, so that the trained behavior feature extraction model has strong robustness, which greatly improves the accuracy of user behavior feature extraction, and finally realizes the accurate extraction of user spatial features and behavior features, which is convenient for operators to reasonably optimize resource allocation and conduct precision marketing to users.

[0044] 2. By using 3D convolution kernels to explore the potential relationship between behavioral features, since 3D convolution kernels are very suitable for spatiotemporal feature learning, compared with 2D convolution kernels, 3D convolution and 3D pooling operations can better model temporal information, further improving the extraction accuracy of the behavioral feature extraction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The present invention will be further described below in conjunction with embodiments with reference to the accompanying drawings.

[0046] Figure 1 It is a flow chart of an operator log intelligent analysis method based on multi-dimensional mining of the present invention.

[0047] Figure 2 It is a schematic diagram of the log data set construction process of the present invention. DETAILED DESCRIPTION

[0048] The technical solution in the embodiments of the present application has the following overall idea: by combining the time dimension and the space dimension, the log data set is divided into residential status data and work status data, and the residential status data and the work status data are clustered respectively using the DBSCAN clustering algorithm to obtain spatial features, so as to improve the accuracy of user spatial feature extraction; by preprocessing the communication log, location data and user consumption data and constructing a log data set, and then using the log data set to train the behavior feature extraction model, since the preprocessing eliminates the interference data, the integration of the communication log, location data and user consumption data is closer to the real business scenario, so as to improve the accuracy of user behavior feature extraction, and finally realize the accurate extraction of user spatial features and behavior features.

[0049] Please refer to Figure 1 to Figure 2 As shown, a preferred embodiment of the operator log intelligent analysis method based on multi-dimensional mining of the present invention includes the following steps:

[0050] Step S10: Obtain the communication logs of each user including call logs and Internet access logs;

[0051] Step S20, setting a plurality of grids of preset sizes on the map based on the location distribution of the base stations, and constructing location data based on the grids and the locations of the cells carried on the map;

[0052] Step S30: Obtain user consumption data;

[0053] Step S40: pre-processing the communication log, location data and user consumption data to construct a log data set;

[0054] Preprocessing is the process of discovering and resolving data quality issues, including cleaning similar data, duplicate data, abnormal data, inconsistent data, etc., which can provide guarantees for data mining and analysis.

[0055] Step S50, filtering the log data set by status, and further dividing it into residential status data and working status data;

[0056] Step S60: clustering the residential status data and the work status data respectively using the DBSCAN clustering algorithm to obtain spatial features including residential locations and work locations;

[0057] Step S70: creating a behavior feature extraction model based on deep learning technology, using the log data set to train the behavior feature extraction model, and using the trained behavior feature extraction model to extract the user's behavior features;

[0058] Since convolutional neural networks (deep learning technology) can extract the potential relationship between features, more semantic information can be obtained through deep feature extraction, which is beneficial to the final classification. Therefore, deep learning technology is used to create a behavioral feature extraction model.

[0059] Step S80: Utilize the spatial characteristics and behavioral characteristics to optimize resource allocation and analyze user behavior.

[0060] The step S10 is specifically as follows:

[0061] Obtaining the communication logs of each user including call logs and Internet access logs, and aggregating the communication logs in hours;

[0062] The communication log carries at least user ID, date, time, call record, LAC (location area code), CI, longitude, latitude and connected base station; the CI is the base station sector ID.

[0063] When users use operator services such as voice and traffic data, their mobile phones will connect to adjacent base stations and generate communication logs. These communication logs are a type of mobile phone trajectory data, which have the characteristics of large data volume, large user base, and wide coverage of the population. The use of communication logs can not only provide a basis for operator billing, but also can be used to study more abundant user behaviors.

[0064] Every time a user makes or receives a call, the operator generates a call log; every time a user uses the mobile network, the operator generates an Internet access log. In order to reduce the space for data storage, communication logs are aggregated in units of hours. For example, user x makes a call in cell A for m seconds at a certain time between 0:00 and 1:00, then makes a call in cell A for n seconds at a certain time between 0:00 and 1:00, and finally makes a call in cell B for t seconds at a certain time between 0:00 and 1:00. Then, if the call logs are aggregated in units of hours, two call logs will be obtained: user x makes a call in cell A for m+n seconds from 0:00 to 1:00; user x makes a call in cell B for t seconds from 0:00 to 1:00.

[0065] In the step S20, the grid carries the land use; the land use is residential, commercial or industrial;

[0066] The grid and the cells are in a many-to-many relationship, that is, one cell may belong to multiple grids, and one grid may also contain multiple cells.

[0067] In step S30, the user consumption data at least includes user ID, VIP level, network access year, user age, user gender and consumption characteristics.

[0068] The step S40 specifically includes:

[0069] Step S41, removing irrelevant data items from the communication log, location data and user consumption data;

[0070] The irrelevant data items include at least log ID, IP address, MAC address and staff;

[0071] Step S42: removing noise data from the communication log, location data, and user consumption data;

[0072] The noise data includes at least URL addresses, special symbols, garbled text and emoticons;

[0073] Step S43: removing duplicate data from the communication log, location data, and user consumption data;

[0074] Step S44: feature-encoding the communication log, location data, and user consumption data;

[0075] Since some features in the communication log, location data and user consumption data are unordered categorical variables, such as base station type, VIP level, etc., feature encoding should be performed and converted into a matrix; One-Hot encoding is used for features with less than 10 categories; frequency encoding is used for features with more than 10 categories; features that need to be encoded include at least fault level, error type, etc.

[0076] Step S45: completing missing values ​​in the communication log, location data, and user consumption data based on a linear interpolation method;

[0077] Step S46: construct a log data set based on the communication log, location data and user consumption data.

[0078] The step S50 specifically includes:

[0079] Step S51, determining the cell where the user is located through the LAC, CI, longitude and latitude carried in the communication log;

[0080] Step S52: matching the corresponding grid from the location data based on the cell, thereby obtaining the land use corresponding to the cell;

[0081] Step S53: based on the date and time carried in the communication log, counting the user's visit frequency and residence time in the corresponding cell;

[0082] Step S54: combining the land use, visit frequency and residence time, dividing the log data set into residential status data and work status data.

[0083] Since operators set up base stations based on the number of users in the area, and users' behavioral activities tend to show regularity, that is, users always make periodic changes in their locations around their homes and work locations and spend most of their time in these two locations, the users' home locations and work locations can be identified through their communication logs and location data.

[0084] The state of the log data set is defined as a combination of land use and time period. The land use is divided into residential, commercial, and industrial, and the time period is divided into early morning period and working period. The following four states can be obtained through permutations and combinations: 1. Residential state: the time of connecting to the base station is between 1:00 and 6:00 in the morning, and the location of the base station is in the residential area; 2. Working state: the time of connecting to the base station is between 9:00 and 11:00 in the morning and 14:00 to 16:00 in the afternoon on weekdays, and the location of the base station is in the commercial area or industrial area; 3. Early morning working area state: the time of connecting to the base station is between 1:00 and 6:00 in the morning, and the location of the base station is in the commercial area or industrial area; 4. Working period residential area state: the time of connecting to the base station is between 9:00 and 11:00 in the morning and 14:00 to 16:00 in the afternoon on weekdays, and the location of the base station is in the residential area. Through statistics, it is found that in the user's communication log, there is less data in state 3 and state 4, so this application only considers state 1 and state 2, and divides the log data set into residential state data and working state data through state 1 and state 2.

[0085] The step S60 specifically includes:

[0086] Step S61: clustering the residential status data and the work status data respectively based on longitude and latitude using a DBSCAN clustering algorithm in units of user ID to obtain a residential location cluster and a work location cluster;

[0087] Step S62: Determine the residential location and the work location through the number of days and duration that the base station is connected in the residential location cluster and the work location cluster, and construct spatial features based on the residential location and the work location.

[0088] The DBSCAN clustering algorithm is a classic density-based clustering algorithm that can identify clusters of any number and shape in a noisy data set. The DBSCAN clustering algorithm includes the following parameters:

[0089] ε-neighborhood: The area within the ε radius of an object p is called the ε-neighborhood of object p;

[0090] Density: the number of data points contained in the ε-nearest neighborhood;

[0091] Core point: Set X t For any point in the set, when the number of points in its ε-neighborhood exceeds the density threshold Minpts, point X t As the core point;

[0092] Direct density reachability: used to determine whether two points belong to the same cluster; point p2 is said to be directly density reachable from point p1 if the following two conditions are met: (a) the two points are close enough to satisfy distance(p1,p2)<ε; (b) there are enough points in the ε-neighborhood of p1 to satisfy p1 num >Minpts, that is, p1 is the core point;

[0093] Density reachable: In a set D, if there exists an object chain p1,p2,…,p n , let p1 = q, p n =p,p t ∈D, 1≤i≤n, p i It is from p i-1 If p is directly density-reachable with respect to ε and Minpts, then object p is density-reachable from object q with respect to ε and Minpts;

[0094] Density connectability: Assume that there are points O, A and B in the sample set D. When points O and A satisfy density reachability, and points O and B also satisfy density reachability, then points A and B satisfy density connectability.

[0095] Based on the above parameters, the basic idea of ​​the DBSCAN clustering algorithm is to determine whether each point in the entire clustered data set belongs to the same class according to whether the number of sample points within the threshold radius is greater than the density threshold; if the current point p is a core point, then other points in its immediate neighborhood are also added to the set, and the cluster where the current data point is located is continuously expanded until all data points that meet the current cluster are traversed to form a complete cluster. If point p is a boundary point, that is, the number of other points in its immediate neighborhood is less than the threshold, then skip this point and continue to select the next point in the data set for processing. The remaining points in the data set are selected as noise (abnormal) points.

[0096] Compared with other clustering algorithms, the DBSCAN clustering algorithm has greater advantages in location identification: first, the DBSCAN clustering algorithm does not need to input the number of clusters k, which meets the identification requirements of important user locations, that is, the number of important user locations is unknown; second, unlike other algorithms such as k-means that are very sensitive to noise and isolated point data, the DBSCAN clustering algorithm can identify noise points while clustering and is insensitive to abnormal points in the data set, which is very consistent with the scenario of this application, because in addition to the user's stay information at important locations, the communication log is also mixed with some records of stay at other non-important locations, that is, noise data; finally, the initial values ​​of other clustering algorithms have a great influence on the clustering results, while the DBSCAN clustering algorithm is not sensitive to the order of samples in the data set, and the clustering results are relatively unbiased.

[0097] The DBSCAN clustering algorithm is used to cluster each user in the residential status data and work status data with "longitude" and "latitude", respectively, and the actual geographical distance is used as the clustering distance function to obtain the alternative clusters of the user's residential location and work location; important locations generally refer to locations with the highest frequency of occurrence or the longest stay time. Based on this feature and combined with the characteristics of the communication log itself, the locations that meet the following conditions are identified as residential locations and work locations from the alternative clusters: 1. At least one base station in the cluster is connected for more than seven days; 2. The residential location and work location have the most total connected hours for all base stations in the cluster.

[0098] In step S70, the convolution layer of the behavior feature extraction model uses a 3D convolution kernel, the activation function uses a nonlinear function, and the pooling method uses average pooling.

[0099] Since 3D convolutional neural networks are very suitable for spatiotemporal feature learning, and can better model temporal information through 3D convolution and 3D pooling operations than 2D convolutional neural networks, 3D convolution kernels are used; a nonlinear function is introduced as an activation function after each convolution layer, so that the convolutional neural network set can fit any function and still perform well in complex sample spaces; pooling is essentially downsampling, and a reasonable choice of pooling method can allow the model to focus on important features, reduce dimensions, significantly reduce the number of hyperparameters, and reduce the degree of model overfitting, so this application adopts an average pooling method; the last fully connected layer of the feature extraction model outputs the output of two neurons, which is used as the final network classification, representing the confidence that the user's behavior is abnormal behavior.

[0100] In step S70, the use of the log data set to train the behavior feature extraction model is specifically as follows:

[0101] Converting the log data set into a three-dimensional single-channel grayscale feature map, using the three-dimensional single-channel grayscale feature map as input of a behavior feature extraction model, and using the behavior feature as output of the behavior feature extraction model, so as to train the behavior feature extraction model;

[0102] The three-dimensional single-channel grayscale feature map includes an X-axis feature map, a Y-axis feature map and a Z-axis feature map; the X-axis feature map is composed of user ID and call records; the Y-axis feature map is segmented data with a granularity of 5 minutes; the Z-axis feature map is segmented data with a time slice as the granularity; the time slice is 5 hours;

[0103] Since the call behavior of ordinary users is periodic within a day, the call peaks are mainly distributed around 9:00, 14:00, and 19:00. Therefore, 5 hours is taken as a time slice, and 7:00-12:00, 12:00-17:00, and 17:00-22:00 are taken as observation windows of the day. Each observation window is a time slice of user behavior.

[0104] The behavioral characteristics include at least: the frequency and standard deviation of calls between the calling and called users; the mean, variance, skewness and kurtosis of the calling user's call duration; the mean number of calls made by the calling user; the mode, skewness and standard deviation of the calling user's province; the mean, standard deviation, skewness and kurtosis of the calling user's call charges; the category, rate package and real-name attributes of the calling and called users; the account opening time of the calling user; the calling user's downtime status and downtime duration; and the mean monthly consumption amount of the user.

[0105] In the step S80, the resource configuration optimization is specifically to optimize the layout of the base stations; and the user behavior analysis is specifically to analyze the abnormal behavior of the users.

[0106] Since users' abnormal behavior will not only affect the behavior patterns of normal users in the system, but also affect the stable operation of the entire system, users' abnormal behavior may also affect their work quality and efficiency, causing large economic losses and threatening the stability, security and development of operators, it is necessary to analyze users' abnormal behavior.

[0107] In summary, the advantages of the present invention are:

[0108] 1. By combining the time dimension and the space dimension, the log data set is divided into residential status data and working status data, and the DBSCAN clustering algorithm is used to cluster the residential status data and the working status data respectively to obtain spatial features, which greatly improves the accuracy of user spatial feature extraction; by preprocessing the communication logs, location data and user consumption data and constructing a log data set, and then using the log data set to train the behavior feature extraction model, the preprocessing eliminates interference data, and the integration of communication logs, location data and user consumption data is closer to the real business scenario, so that the trained behavior feature extraction model has strong robustness, which greatly improves the accuracy of user behavior feature extraction, and finally realizes the accurate extraction of user spatial features and behavior features, which is convenient for operators to reasonably optimize resource allocation and conduct precision marketing to users.

[0109] 2. By using 3D convolution kernels to explore the potential relationship between behavioral features, since 3D convolution kernels are very suitable for spatiotemporal feature learning, compared with 2D convolution kernels, 3D convolution and 3D pooling operations can better model temporal information, further improving the extraction accuracy of the behavioral feature extraction model.

[0110] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. An intelligent analysis method of operator logs based on multi-dimensional mining, characterized by: The steps include: Step S10: Obtain the communication logs of each user including call logs and Internet access logs; Step S20, setting a plurality of grids of preset sizes on the map based on the location distribution of the base stations, and constructing location data based on the grids and the locations of the cells carried on the map; Step S30: Obtain user consumption data; Step S40: pre-processing the communication log, location data and user consumption data to construct a log data set; Step S50, filtering the log data set by status, and further dividing it into residential status data and working status data; Step S60: clustering the residential status data and the work status data respectively using the DBSCAN clustering algorithm to obtain spatial features including residential locations and work locations; Step S70: creating a behavior feature extraction model based on deep learning technology, using the log data set to train the behavior feature extraction model, and using the trained behavior feature extraction model to extract the user's behavior features; Step S80: Utilizing the spatial characteristics and behavioral characteristics to optimize resource allocation and analyze user behavior; The step S60 specifically includes: Step S61: clustering the residential status data and the work status data respectively based on longitude and latitude using a DBSCAN clustering algorithm in units of user ID to obtain a residential location cluster and a work location cluster; Step S62: determining the residential location and the work location by the number of days and duration that the base station is connected in the residential location cluster and the work location cluster, and constructing a spatial feature based on the residential location and the work location; In step S70, the use of the log data set to train the behavior feature extraction model is specifically as follows: Converting the log data set into a three-dimensional single-channel grayscale feature map, using the three-dimensional single-channel grayscale feature map as input of a behavior feature extraction model, and using the behavior feature as output of the behavior feature extraction model, so as to train the behavior feature extraction model; The three-dimensional single-channel grayscale feature map includes an X-axis feature map, a Y-axis feature map and a Z-axis feature map; the X-axis feature map is composed of user ID and call records; the Y-axis feature map is segmented data with a granularity of 5 minutes; the Z-axis feature map is segmented data with a time slice as the granularity; the time slice is 5 hours; The behavioral characteristics include at least: the frequency and standard deviation of calls between the calling and called users; the mean, variance, skewness and kurtosis of the calling user's call duration; the mean number of calls made by the calling user; the mode, skewness and standard deviation of the calling user's province; the mean, standard deviation, skewness and kurtosis of the calling user's call charges; the category, rate package and real-name attributes of the calling and called users; the account opening time of the calling user; the calling user's downtime status and downtime duration; and the mean monthly consumption amount of the user.

2. The method for intelligent analysis of operator logs based on multi-dimensional mining according to claim 1, characterized in that: The step S10 is specifically as follows: Obtaining the communication logs of each user including call logs and Internet access logs, and aggregating the communication logs in hours; The communication log carries at least user ID, date, time, call record, LAC, CI, longitude, latitude and connected base station; the CI is the base station sector ID.

3. The method for intelligent analysis of operator logs based on multi-dimensional mining according to claim 1, characterized in that: In the step S20, the grid carries the land use; the land use is residential, commercial or industrial; The relationship between the grid and the cells is many-to-many.

4. The method for intelligent analysis of operator logs based on multi-dimensional mining according to claim 1, characterized in that: In step S30, the user consumption data at least includes user ID, VIP level, network access year, user age, user gender and consumption characteristics.

5. The method for intelligent analysis of operator logs based on multi-dimensional mining according to claim 1, characterized in that: The step S40 specifically includes: Step S41, removing irrelevant data items from the communication log, location data and user consumption data; Step S42: removing noise data from the communication log, location data, and user consumption data; Step S43: removing duplicate data from the communication log, location data, and user consumption data; Step S44: feature-encoding the communication log, location data, and user consumption data; Step S45: completing missing values ​​in the communication log, location data, and user consumption data based on a linear interpolation method; Step S46: construct a log data set based on the communication log, location data and user consumption data.

6. The method for intelligent analysis of operator logs based on multi-dimensional mining according to claim 2, characterized in that: The step S50 specifically includes: Step S51, determining the cell where the user is located through the LAC, CI, longitude and latitude carried in the communication log; Step S52, matching the corresponding grid from the location data based on the cell, and then obtaining the land use corresponding to the cell; Step S53: based on the date and time carried in the communication log, counting the user's visit frequency and residence time in the corresponding cell; Step S54: combining the land use, visit frequency and residence time, dividing the log data set into residential status data and work status data.

7. The method for intelligent analysis of operator logs based on multi-dimensional mining according to claim 1, characterized in that: In step S70, the convolution layer of the behavior feature extraction model uses a 3D convolution kernel, the activation function uses a nonlinear function, and the pooling method uses average pooling.

8. The method for intelligent analysis of operator logs based on multi-dimensional mining according to claim 1, characterized in that: In the step S80, the resource configuration optimization is specifically to optimize the layout of the base stations; and the user behavior analysis is specifically to analyze the abnormal behavior of the users.

Images