A method and system for analyzing project data records based on CRM customer data analysis

Abstract

This invention discloses a project data record analysis method and system based on CRM customer data analysis, belonging to the field of data processing technology. It solves the technical problems of existing technologies where CRM customer data and PM project data are isolated, and existing project success prediction models rely on traditional features and fail to mine the deep temporal features of core indicators. The method includes: extracting customer historical behavior data from CRM and project execution data from the PM system, then calculating project-customer correlation strength data, constructing and training a dynamic correlation model between customers and projects, calculating a multi-dimensional project health index based on project execution data, constructing a machine learning classification model, using the correlation strength index and health index as core features, extracting deep temporal features through corresponding subnets, combining temporal attention layer fusion features, outputting the project success probability, and finally constructing a management dashboard to display key indicators in real time.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method and system for analyzing project data records based on CRM customer data analysis. Background Technology

[0002] In the field of project management, the success of a project is influenced by both client behavior and the project execution status.

[0003] However, existing technologies suffer from the following technical problems: First, customer data stored in CRM systems and project data stored in PM systems are often isolated, making deep integration difficult and hindering the accurate capture of the correlation between customer behavior and project progress. This results in a lack of customer-centric decision-making support for project management. Second, project health assessments often focus on a single dimension (e.g., only progress or budget), failing to construct a multi-dimensional assessment system encompassing progress, resources, budget, and customers. Furthermore, the corresponding threshold values ​​are often fixed, unable to adapt to different project characteristics, leading to low accuracy in early warnings. Third, project success prediction models rely heavily on traditional business characteristics (e.g., customer industry and project scale), failing to explore the deep temporal characteristics of correlation strength indices and health indices. Moreover, feature fusion lacks a dynamic weight adjustment mechanism, making prediction accuracy insufficient to meet actual needs. Fourth, key project indicators are displayed in a fragmented manner, lacking a unified real-time visualization platform, making it difficult for managers to quickly obtain the core status of projects. These problems result in the difficulty in timely identification of project risks and significant deviations between predicted and actual project success probabilities, hindering the improvement of project management efficiency and quality.

[0004] Therefore, a project data record analysis method and system based on CRM customer data analysis has emerged. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes a project data record analysis method and system based on CRM customer data analysis, to solve the following technical problem:

[0006] First, customer data in storage and CRM systems and project data in storage and PM systems are often isolated, making deep integration difficult and hindering the accurate capture of the correlation between customer behavior and project progress. This results in a lack of customer-centric decision-making support for project management. Second, project health assessments often focus on a single dimension (e.g., only schedule or budget), failing to build a multi-dimensional assessment system encompassing schedule, resources, budget, and customers. Furthermore, the corresponding thresholds are often fixed, unable to adapt to different project characteristics, leading to low accuracy in early warnings. Third, project success prediction models rely heavily on traditional business characteristics (e.g., customer industry and project size), failing to explore the deep temporal characteristics of correlation strength and health indices. The feature fusion lacks a dynamic weight adjustment mechanism, making prediction accuracy insufficient for practical needs. Fourth, key project indicators are displayed in a fragmented manner, lacking a unified real-time visualization platform, making it difficult for managers to quickly grasp the core status of projects. These problems result in the difficulty in timely identification of project risks and significant discrepancies between predicted and actual project success probabilities, hindering the improvement of project management efficiency and quality.

[0007] To address the aforementioned problems, a first aspect of the present invention provides a method for analyzing project data records based on CRM customer data analysis, comprising the following steps:

[0008] S1: Extract historical customer behavior data from the CRM system, and simultaneously extract project execution data from the PM system;

[0009] S2: Based on the customer's historical behavior data and project execution data, calculate the project-customer association strength index, construct a customer-project dynamic association model, and train the constructed customer-project dynamic association model;

[0010] S3: Based on the project execution data, calculate the multidimensional project health index, and set a health index threshold based on historical data. When the multidimensional project health index value is lower than the threshold, trigger an early warning.

[0011] S4: Construct a machine learning classification model, taking the project-customer association strength index and the multidimensional project health index as core features, and inputting them into the machine learning classification model along with traditional features. The model extracts deep temporal features of customer association and project health through the association strength analysis subnet and the health analysis subnet, respectively, and fuses all features through the temporal attention layer to learn the dynamic weights of each feature under different project stages, and finally outputs the project success probability.

[0012] S5: Build a management dashboard to display the project-customer relationship strength index, multi-dimensional project health index, and project success probability for each project in real time.

[0013] Preferably, step S1 includes the following steps:

[0014] Extract customer historical behavior data from CRM, including: historical transaction records, customer support data, and customer interaction data;

[0015] The historical transaction records include transaction amount, frequency, and product lines; the customer support data includes the number of work orders and average resolution time; and the customer interaction data includes the frequency of communication with key personnel and the number of meetings.

[0016] Extract project execution data from the PM system, including: task completion progress, milestone achievement time, budget consumption rate, and resource investment hours.

[0017] Preferably, the calculation of the customer association strength index includes the following steps:

[0018] Define the project-customer association strength index as follows:

[0019]

[0020] in, The project-customer relationship strength index. , , , , , , and These are the weighting coefficients for the corresponding parameters. The proportion of transactions for core products, For transaction frequency density, For transaction growth rate, The percentage of work orders specific to the project. To resolve timeliness discrepancies in work orders, For the frequency of interaction at the decision-making level, This is an index of interaction depth. To ensure timely information synchronization;

[0021] The transaction ratio of core products, transaction frequency density, and transaction growth rate are classified as transaction relevance stratification indicators; the proportion of project-specific work orders and the deviation in work order resolution timeliness are classified as support relevance stratification indicators; and the interaction frequency, interaction depth index, and information synchronization timeliness of the decision-making level are classified as interaction relevance stratification indicators.

[0022] The weighting coefficients are determined using a two-factor iterative optimization method, specifically:

[0023] The initial weights are set by industry experts based on the project type. Based on historical project data, the weights are iteratively adjusted using a particle swarm optimization algorithm with the goal of maximizing the Pearson correlation coefficient between the project-customer association strength index and the project success rate. In addition, the weights are re-optimized every quarter based on the data of newly added projects.

[0024] Preferably, the construction of the customer-project dynamic association model includes the following steps:

[0025] The project lifecycle is divided into a continuous sequence of time windows, using weeks as the time window. ,in, For the current window, This is the number of the time window;

[0026] Real-time updates of various layer indicators for each window;

[0027] The moving average algorithm is used to smooth the indicators of each stratum.

[0028] A time decay factor is introduced to apply exponential weight decay to historical data exceeding 3 months. The calculation of the project-customer association strength index for the current window is as follows:

[0029]

[0030] in, The current time window, i.e., the [number]th [time window]. Weekly Project - Customer Relationship Strength Index , and These are the hierarchical indicator numbers for transaction correlation, support correlation, and interaction correlation, respectively. , and These are the basic weights for the indicators at each level: transaction, support, and interaction. This is the historical window offset. The current window, i.e., the [number]th window. week, The previous window, i.e. week, To obtain The smaller value between 1 and 12 is used to limit the inclusion of data to only the most recent 12 windows. Historical data older than 3 months is excluded from the calculation because it decays to a negligible threshold. In other words, no more than 12 windows, corresponding to 3 months, are included. The time decay factor, After smoothing by moving average, the first The first window Individual transaction stratification indicator values, After smoothing by moving average, the first The first window One supporting hierarchical indicator value, After smoothing by moving average, the first The first window Each interaction layered indicator value.

[0031] Preferably, training the constructed customer-project dynamic association model includes the following steps:

[0032] Projects completed in the past three years were selected as training samples. Each sample contains raw hierarchical indicator data for each time window within the project's entire life cycle, as well as manually annotated 5-point project-related quality labels.

[0033] The system employs a 1D-CNN temporal feature extraction module and a GRU dynamic weight learning module.

[0034] The loss function is weighted mean squared error, and the optimizer is the Adam adaptive optimization algorithm. The initial learning rate is set to 0.001, and it decays by 50% every 50 training epochs.

[0035] The model outputs the customer-project correlation strength index and the contribution of hierarchical indicators for each window.

[0036] Preferably, the multidimensional health index includes the following steps:

[0037] The multidimensional project health index is obtained by weighted addition of schedule health, resource health, budget health, and customer health, with the weights determined by random forest importance analysis.

[0038] The progress health status is calculated based on the task completion rate and the number of milestone delay days, specifically:

[0039]

[0040] in, For progress health status, For task completion rate, The number of days for the milestone extension The maximum allowed extension period;

[0041] The resource health status is calculated using the resource input deviation rate and the key resource idle rate, specifically:

[0042]

[0043] in, For resource health, For resource input deviation rate, Key resource idle rate;

[0044] The budget health status is calculated using the budget consumption rate and cost deviation rate, specifically:

[0045]

[0046] in, For budget health, For budget consumption rate, Cost deviation rate;

[0047] The customer health score is calculated based on customer satisfaction scores and the frequency of changes in requirements, specifically as follows:

[0048]

[0049] in, For the health of our customers, Rate customer satisfaction. To accommodate the frequency of demand changes, Maximum number of changes;

[0050] The health index threshold is the 70th percentile of the multidimensional project health index of all successful projects in the past 3 years.

[0051] Preferably, the machine learning classification model includes the following steps:

[0052] Successful and unsuccessful samples were selected from historical project data and divided into training and test sets;

[0053] The samples are preprocessed, and the imbalance of samples is handled by a time decay-based weighted sampling strategy. Specifically, a weight coefficient is assigned to each training sample. The weight coefficient is inversely proportional to the time distance between the sample's current time point and the current time point. Weighted sampling is performed based on this weight, and then the SMOTE algorithm is used to generate synthetic samples from the sampled minority class samples.

[0054] Input core features and traditional features. The core features are the project-customer association strength index and the multidimensional project health index. The traditional features are customer industry type, project scale, team experience and project complexity.

[0055] A machine learning classification model is constructed using a dynamic feature weighted fusion network model, and the model is trained.

[0056] Preferably, the dynamic feature weighted fusion network model includes the following steps:

[0057] This model includes:

[0058] The association strength analysis subnet takes historical time-series data of the project-customer association strength index as input, uses a gated recurrent unit network to capture the dynamic change pattern of customer association, and outputs the first feature vector.

[0059] The health analysis subnet takes historical time-series data of multi-dimensional health index as input, uses a one-dimensional convolutional neural network to extract the local dependencies and models of health index, and outputs a second feature vector.

[0060] The feature concatenation layer is used to concatenate the first feature vector, the second feature vector, and traditional features to obtain a fused feature vector;

[0061] A temporal attention layer is used to receive the fused feature vector and automatically learn the dynamic weights of various features in the fused feature vector on the success probability at different lifecycle stages of the project.

[0062] The output layer, connected to the temporal attention layer, is used to output the project success probability.

[0063] Preferably, training the model includes the following steps:

[0064] The dynamic feature weighted fusion network model is trained using a custom weighted cross-entropy loss function, which is specifically as follows:

[0065]

[0066] in, For loss function, The true label is 1 for successful samples and 0 for failed samples. To predict probabilities, This is a hyperparameter, and its value is positively correlated with the project size.

[0067] Using AUC as the core metric, the hyperparameters of the model are tuned through Bayesian optimization, and the success probability of the project is finally output. The weight distribution output by the time-series attention layer is then visualized through a management dashboard.

[0068] A second aspect of the present invention provides a project data record analysis system based on CRM customer data analysis, comprising the following modules:

[0069] Multi-source data acquisition module: Extracts historical customer behavior data from CRM, and simultaneously extracts project execution data from PM system;

[0070] Customer-Project Association Analysis and Modeling Module: Based on the customer's historical behavior data and project execution data, calculate the project-customer association strength index, construct a dynamic customer-project association model, and train the constructed dynamic customer-project association model;

[0071] Project health assessment and early warning module: Based on the project execution data, calculate a multidimensional project health index and set a health index threshold based on historical data. When the multidimensional project health index value is lower than the threshold, an early warning is triggered.

[0072] Project success probability prediction module: Construct a machine learning classification model, take the project-customer association strength index and the multidimensional project health index as core features, and input them into the machine learning classification model along with traditional features. The model extracts deep temporal features of customer association and project health through association strength analysis subnet and health analysis subnet respectively, and fuses all features through temporal attention layer to learn the dynamic weight of each feature under different project stages, and finally outputs the project success probability.

[0073] Project data visualization management module: Constructs a management dashboard to display the project-customer relationship strength index, multi-dimensional project health index, and project success probability for each project in real time.

[0074] The beneficial effects of this invention are:

[0075] This invention constructs a dynamic customer-project association model, optimizes the calculation of the customer-project association strength index by using time window segmentation, moving average smoothing, and time decay factor, determines the weights by combining a two-factor iterative optimization method, and trains the model with a 1D-CNN combined with a GRU dual model. This breaks down the data barriers between CRM and PM, and achieves deep integration of customer behavior and project execution data. The customer-project association strength index can dynamically reflect the binding relationship between customers and projects. Compared with isolated data applications, it improves the accuracy of customer-project association analysis and provides a customer-dimensional reference for project management.

[0076] This invention constructs a multi-dimensional project health index by weighting and adding up progress, resources, budget, and customer health. The weights are determined by random forest importance analysis, and the threshold is set by the 70th percentile of the health index of successful projects in the past 3 years. This overcomes the shortcomings of traditional single-dimensional assessment, and the threshold is dynamically set based on historical data to adapt to different project characteristics, thus achieving accurate identification of project risks.

[0077] This invention constructs a dynamic feature weighted fusion network model, extracts deep temporal features using a correlation strength analysis subnet and a health analysis subnet, learns dynamic weights of features at different project stages by combining a temporal attention layer, and uses a custom weighted cross-entropy loss function. This technique breaks through the limitations of traditional models that rely on static features and lack dynamic weights, and greatly improves the AUC value for predicting the probability of project success. Attached Figure Description

[0078] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0079] Figure 2 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0080] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0081] Please see Figure 1 As shown, this invention is a project data record analysis method based on CRM customer data analysis, including the following steps:

[0082] S1: Extract historical customer behavior data from the CRM system, and simultaneously extract project execution data from the PM system;

[0083] S2: Based on the customer's historical behavior data and project execution data, calculate the project-customer association strength index, construct a customer-project dynamic association model, and train the constructed customer-project dynamic association model;

[0084] S3: Based on the project execution data, calculate the multidimensional project health index, and set a health index threshold based on historical data. When the multidimensional project health index value is lower than the threshold, trigger an early warning.

[0085] S4: Construct a machine learning classification model, taking the project-customer association strength index and the multidimensional project health index as core features, and inputting them into the machine learning classification model along with traditional features. The model extracts deep temporal features of customer association and project health through the association strength analysis subnet and the health analysis subnet, respectively, and fuses all features through the temporal attention layer to learn the dynamic weights of each feature under different project stages, and finally outputs the project success probability.

[0086] S5: Build a management dashboard to display the project-customer relationship strength index, multi-dimensional project health index, and project success probability for each project in real time.

[0087] Specifically, an ETL tool is used to connect the CRM and PM systems to periodically (hourly) extract historical customer behavior data (historical transaction records: transaction amount, frequency, product line; customer support data: number of work orders, average resolution time; customer interaction data: frequency of communication with key personnel, number of meetings) and project execution data (task completion progress, milestone achievement time, budget consumption rate, resource input hours). Data cleaning and format standardization are then performed. Based on the collected data, a hierarchical index of the correlation between transactions, support, and interactions is calculated. Time windows are divided by week, and after smoothing with a moving average, a time decay factor is introduced to calculate the correlation strength index. A 1D-CNN+GRU dual-module model is built, using data from the past three years... The project data is used for training, and the model is optimized using weighted mean squared error as the loss function. The model outputs the correlation strength index for each window and calculates progress, resources, budget, and customer health according to the formula. The weights of the multidimensional health index are determined by random forest. The 70th percentile of the health index of successful projects in the past 3 years is used as the threshold. The index is monitored in real time to see if it exceeds the threshold. If it exceeds the threshold, a red warning is triggered. Otherwise, monitoring continues. Historical successful / failed project samples are selected. Successful samples are defined as projects that are delivered on time and accepted by the customer. The samples are processed by time decay weighted sampling combined with SMOTE. The core features (correlation strength, health index) and traditional features are input. A model containing a GRU correlation subnet, a 1D-CNN health subnet, and a temporal attention layer is built. The model is trained with a custom weighted cross-entropy loss function and outputs the project success probability. Finally, a management dashboard is built and pushes real-time data via WebSocket to display correlation strength, health index, and success probability.

[0088] In one embodiment of the present invention, step S1 includes the following steps:

[0089] Extract customer historical behavior data from CRM, including: historical transaction records, customer support data, and customer interaction data;

[0090] The historical transaction records include transaction amount, frequency, and product lines; the customer support data includes the number of work orders and average resolution time; and the customer interaction data includes the frequency of communication with key personnel and the number of meetings.

[0091] Extract project execution data from the PM system, including: task completion progress, milestone achievement time, budget consumption rate, and resource investment hours.

[0092] Specifically, customer historical behavior data is defined as a dataset extracted from the CRM system that reflects the customer's past interactions with the enterprise. It includes three main categories: historical transaction records, customer support data, and customer interaction data. It is one of the core input data for calculating the project-customer association strength index and is used to quantify the basis of the relationship between the customer and the project. Project execution data is defined as a dataset extracted from the PM system that reflects the real-time execution status of the project. It is the core input data for calculating the multidimensional project health index and also provides project-side data support for the customer-project association strength index.

[0093] In one embodiment of the present invention, the calculation of the project-customer association strength index includes the following steps:

[0094] Define the project-customer association strength index as follows:

[0095]

[0096] in, The project-customer relationship strength index. , , , , , , and These are the weighting coefficients for the corresponding parameters. The proportion of transactions for core products, For transaction frequency density, For transaction growth rate, The percentage of work orders specific to the project. To resolve timeliness discrepancies in work orders, For the frequency of interaction at the decision-making level, This is an index of interaction depth. To ensure timely information synchronization;

[0097] The transaction ratio of core products, transaction frequency density, and transaction growth rate are classified as transaction relevance stratification indicators; the proportion of project-specific work orders and the deviation in work order resolution timeliness are classified as support relevance stratification indicators; and the interaction frequency, interaction depth index, and information synchronization timeliness of the decision-making level are classified as interaction relevance stratification indicators.

[0098] The weighting coefficients are determined using a two-factor iterative optimization method, specifically:

[0099] The initial weights are set by industry experts based on the project type. Based on historical project data, the weights are iteratively adjusted using a particle swarm optimization algorithm with the goal of maximizing the Pearson correlation coefficient between the project-customer association strength index and the project success rate. In addition, the weights are re-optimized every quarter based on the data of newly added projects.

[0100] Specifically, three tiered indicators for correlation are defined: transaction correlation, support correlation, and interaction correlation. Transaction correlation includes the proportion of core product transactions, transaction frequency density, and transaction growth rate. Support correlation includes the proportion of project-specific work orders and the timeliness deviation of work order resolution. Interaction correlation includes the frequency of interaction at the decision-making level, the interaction depth index, and the timeliness of information synchronization. Historical customer behavior data is extracted from CRM and project execution data is extracted from the PM system. The original values ​​are calculated according to the formulas for each indicator and standardized to the [0, 1] range. The weight coefficients are first determined by industry experts according to project type (implementation, consulting, and R&D). Then, based on historical project data from the past three years, the weights are iteratively adjusted using a particle swarm optimization algorithm with the goal of maximizing the Pearson correlation coefficient between the correlation strength index and project success rate. New project data is incorporated every quarter for re-optimization to ensure that the weights adapt to business changes. , , and These represent the total weights for the three tiered indicators, with corresponding values ​​of 0.3, 0.4, and 0.3 respectively. , and The corresponding values ​​are all 0.1. and The corresponding values ​​are all 0.2. , and The corresponding values ​​are all 0.1; among them, the project-customer association strength index is a comprehensive indicator used to quantify the closeness of the current project and customer association, with a value range of [0, 1]. The higher the value, the closer the association, which is used to support the construction of the customer-project dynamic association model and the prediction of project success probability; the core product transaction ratio is calculated by the ratio of the customer's historical transaction amount for the project's corresponding core product line to the customer's total transaction amount during the same period. If the customer has no other product line transactions, it is taken as 1; the transaction frequency density is calculated by the ratio of the customer's average annual transaction number in the project-related field to the average annual transaction number of similar customers in the industry. If it exceeds 1, it is calculated as 1; the transaction growth rate ,in, This represents the client's transaction volume in project-related fields over the past two years. For the client's transaction volume in the relevant fields of the project in the previous two years, the transaction growth rate is standardized to the range of [0,1], with negative growth being 0; the proportion of project-specific work orders is calculated by the ratio of the number of client-specific support work orders for the project during the project cycle to the total number of work orders for the client during the same period; the deviation in work order resolution time is calculated by subtracting the average resolution time of project-related work orders from 1 to the average resolution time of all client work orders; the frequency of interaction at the decision-making level is calculated by the ratio of the average number of monthly communications between the client's decision-making level and the project team to the average number of monthly communications for similar projects in the industry; the interaction depth index... ,in, For the number of strategic meetings, For the number of technical docking sessions, For the number of daily communications, The total number of weeks in the project cycle; the timeliness of information synchronization is calculated by subtracting the ratio of the average time for the client to respond to the project team's information to the industry standard response time from 1.

[0101] The method for determining the weighting coefficients, the "two-factor" approach, is defined as combining expert experience (subjective factors) with historical data-driven factors (objective factors). The two types of factors are iteratively balanced using a particle swarm optimization algorithm to optimize the correlation strength index with the project success rate. The results are updated quarterly to adapt to business changes.

[0102] In one embodiment of the present invention, the construction of the customer-project dynamic association model includes the following steps:

[0103] The project lifecycle is divided into a continuous sequence of time windows, using weeks as the time window. ,in, For the current window, This is the number of the time window;

[0104] Real-time updates of various layer indicators for each window;

[0105] The moving average algorithm is used to smooth the indicators of each stratum.

[0106] A time decay factor is introduced to apply exponential weight decay to historical data exceeding 3 months. The calculation of the project-customer association strength index for the current window is as follows:

[0107]

[0108] in, The current time window, i.e., the [number]th [time window]. Weekly Project - Customer Relationship Strength Index , and These are the hierarchical indicator numbers for transaction correlation, support correlation, and interaction correlation, respectively. , and These are the basic weights for the indicators at each level: transaction, support, and interaction. This is the historical window offset. The current window, i.e., the [number]th window. week, The previous window, i.e. week, To obtain The smaller value between 1 and 12 is used to limit the inclusion of data to only the most recent 12 windows. Historical data older than 3 months is excluded from the calculation because it decays to a negligible threshold. In other words, no more than 12 windows, corresponding to 3 months, are included. The time decay factor, After smoothing by moving average, the first The first window Individual transaction stratification indicator values, After smoothing by moving average, the first The first window One supporting hierarchical indicator value, After smoothing by moving average, the first The first window Each interaction layered indicator value.

[0109] Specifically, the entire project lifecycle is divided into a continuous sequence of time windows, using natural weeks as the unit, and marked as follows: ,in, For the first week of the project launch, The current week, i.e., the current window, corresponds to a 7-day data collection period for indicators. At the end of each time window t, the raw data within that window is automatically extracted from the CRM and PM systems to calculate the raw values ​​of the transaction correlation layered indicator, the support correlation layered indicator, and the interaction correlation layered indicator, ensuring that each indicator is automatically updated to the latest window every Monday. The time-series data of each layered indicator is smoothed using the following formula: ,in, After smoothing, the first The first time window The values ​​for each transaction correlation stratification indicator are consistent with the original indicator. The sequence number represents the stratification index of transaction correlation. These correspond to the core product transaction share, transaction frequency density, and transaction growth rate, respectively. The time window number, , and The first , and Within the first time window, the first The original calculated values ​​of the first transaction correlation stratification index were used; the other two stratification indices were calculated using the same method to obtain smoothed stratification index values. A time decay factor was preset according to project type (0.9 for implementation projects, 0.85 for consulting projects, and 0.95 for R&D projects) to make the weight of historical data decay exponentially over time (with recent data having a greater impact). The project-customer correlation strength index was then calculated accordingly. , and The corresponding values ​​are 0.3, 0.4 and 0.3 respectively. The calculated customer-project association strength index and the corresponding window information (timestamp, original index value, smoothed value) are stored in the database as input to the customer-project dynamic association model for subsequent model training and project success probability prediction.

[0110] In one embodiment of the present invention, training the constructed customer-project dynamic association model includes the following steps:

[0111] Projects completed in the past three years were selected as training samples. Each sample contains raw hierarchical indicator data for each time window within the project's entire life cycle, as well as manually annotated 5-point project-related quality labels.

[0112] The system employs a 1D-CNN temporal feature extraction module and a GRU dynamic weight learning module.

[0113] The loss function is weighted mean squared error, and the optimizer is the Adam adaptive optimization algorithm. The initial learning rate is set to 0.001, and it decays by 50% every 50 training epochs.

[0114] The model outputs the customer-project correlation strength index and the contribution of hierarchical indicators for each window.

[0115] Specifically, projects completed within the past three years are selected as training samples. Each sample contains raw hierarchical indicator data for each time window within the project's entire lifecycle, as well as manually labeled 5-point project-related quality tags. The raw indicator data is preprocessed and divided into training and validation sets. A dual-module architecture of temporal feature extraction and dynamic weight learning is adopted, including a temporal feature extraction module and a dynamic weight learning module. The temporal feature extraction module consists of a three-layer one-dimensional convolutional neural network with kernel sizes of 3, 5, and 7, corresponding to capturing temporal association patterns over 3, 5, and 7 weeks. Each convolutional layer is followed by a ReLU activation function and a max-pooling layer (with a pooling kernel size of 2). The input is a smoothed hierarchical indicator time series, and the output is a multi-scale temporal feature vector for each hierarchical indicator, used to extract different time scales from the indicator sequence after moving average. The model employs a two-layer gated recurrent unit (64 hidden layer neurons) to input the temporal feature vector output by a 1D-CNN and output the real-time weights of each layer's indicators. The loss function uses weighted mean square error, assigning a 1.5x weight to high-quality related samples, which are the 4-point and 5-point labels in a 5-point item association quality scale. The optimizer uses the Adam adaptive optimization algorithm with an initial learning rate of 0.001, decaying by 50% every 50 training epochs until the validation set loss shows no decrease for 10 consecutive epochs. A time decay factor is embedded as a learnable parameter in the model, and iterative optimization through backpropagation allows the model to automatically adapt the decay rate according to the item type. For each time window, the model outputs a dynamic customer-item association strength index and the real-time contribution of each layer's indicators.

[0116] In one embodiment of the present invention, the multidimensional health index includes the following steps:

[0117] The multidimensional project health index is obtained by weighted addition of schedule health, resource health, budget health, and customer health, with the weights determined by random forest importance analysis.

[0118] The progress health status is calculated based on the task completion rate and the number of milestone delay days, specifically:

[0119]

[0120] in, For progress health status, For task completion rate, The number of days for the milestone extension The maximum allowed extension period;

[0121] The resource health status is calculated using the resource input deviation rate and the key resource idle rate, specifically:

[0122]

[0123] in, For resource health, For resource input deviation rate, Key resource idle rate;

[0124] The budget health status is calculated using the budget consumption rate and cost deviation rate, specifically:

[0125]

[0126] in, For budget health, For budget consumption rate, Cost deviation rate;

[0127] The customer health score is calculated based on customer satisfaction scores and the frequency of changes in requirements, specifically as follows:

[0128]

[0129] in, For the health of our customers, Rate customer satisfaction. To accommodate the frequency of demand changes, Maximum number of changes;

[0130] The health index threshold is the 70th percentile of the multidimensional project health index of all successful projects in the past 3 years.

[0131] Specifically, the formula for calculating the multidimensional health index is as follows:

[0132]

[0133] in, For the multidimensional health index, For progress health status, For resource health, For budget health, For the health of our customers, , , and The weighting coefficients for the four health indicators are 0.3, 0.2, 0.3, and 0.2, respectively. The resource input deviation rate is obtained by dividing the planned resource input amount in the current project window by the actual resource input amount extracted from the PM system. The critical resource idle rate is obtained by dividing the total available time of critical resources in the current window by the actual usage time extracted from the PM system. The cost deviation rate is obtained by dividing the actual cost in the current project window by the planned cost extracted from the PM system. The requirement change frequency is obtained by extracting the total number of formally submitted requirement change requests from customers in the current project window extracted from the PM system.

[0134] In one embodiment of the present invention, the machine learning classification model includes the following steps:

[0135] Successful and unsuccessful samples were selected from historical project data and divided into training and test sets;

[0136] The samples are preprocessed, and the imbalance of samples is handled by a time decay-based weighted sampling strategy. Specifically, a weight coefficient is assigned to each training sample. The weight coefficient is inversely proportional to the time distance between the sample's current time point and the current time point. Weighted sampling is performed based on this weight, and then the SMOTE algorithm is used to generate synthetic samples from the sampled minority class samples.

[0137] Input core features and traditional features. The core features are the project-customer association strength index and the multidimensional project health index. The traditional features are customer industry type, project scale, team experience and project complexity.

[0138] A machine learning classification model is constructed using a dynamic feature weighted fusion network model, and the model is trained.

[0139] Specifically, data from projects that have completed closed-loop cycles within the past 5 years will be collected as a sample pool, meeting the following conditions: each project must include a full lifecycle customer-project correlation strength index sequence and a multi-dimensional project health index sequence; each project must be clearly labeled with success / failure tags; and the sample pool must cover multiple project types (implementation, consulting, R&D) and customer industries (finance, manufacturing, internet) to ensure sample diversity. A time-stratified partitioning method will be used to avoid data leakage. Specifically, the past 5 years' data will be divided into "historical segment (first 4 years)" and "recent segment (5th year)" based on project closure time. Historical segment data will be used as the training set, and recent segment data as the test set, with a partitioning ratio of 4:1. Time decay weights will be assigned to each training set sample to highlight the reference value of recent samples. Specifically, the time distance will be defined as the number of months between the sample project's closure time and the current time. For example, if the current time is October 2024, then projects that closed in October 2023 will have a weight of 12. The weight coefficient calculation formula is as follows: ,in, The time distance is 0.05, and the decay coefficient is 0.05 to ensure that the weight decreases as time goes on. Based on the weight of the training set samples, a weighted random sampling method is used to extract samples (with the same size as the original training set) so that the probability of recent high-weight samples being extracted is higher, which strengthens the model's learning of recent business patterns. For the minority class (failure samples), the SMOTE algorithm is used to synthesize samples.

[0140] In one embodiment of the present invention, the dynamic feature weighted fusion network model includes the following steps:

[0141] This model includes:

[0142] The association strength analysis subnet takes historical time-series data of the project-customer association strength index as input, uses a gated recurrent unit network to capture the dynamic change pattern of customer association, and outputs the first feature vector.

[0143] The health analysis subnet takes historical time-series data of multi-dimensional health index as input, uses a one-dimensional convolutional neural network to extract the local dependencies and models of health index, and outputs a second feature vector.

[0144] The feature concatenation layer is used to concatenate the first feature vector, the second feature vector, and traditional features to obtain a fused feature vector;

[0145] A temporal attention layer is used to receive the fused feature vector and automatically learn the dynamic weights of various features in the fused feature vector on the success probability at different lifecycle stages of the project.

[0146] The output layer, connected to the temporal attention layer, is used to output the project success probability.

[0147] Specifically, dynamic weights are assigned to the fused feature vectors, as follows:

[0148]

[0149]

[0150] in, This is a dynamic weight vector, the feature dynamic weight vector output by the temporal attention layer, used to represent the importance of each dimension of the fused feature vector to the successful prediction of the project. For normalization function, Here, the learnable weight matrix for the temporal attention layer is a core parameter that is continuously optimized through backpropagation during model training. The fused feature vector is the input to the temporal attention layer. For bias vectors, This is the weighted fusion feature vector.

[0151] In one embodiment of the present invention, training the model includes the following steps:

[0152] The dynamic feature weighted fusion network model is trained using a custom weighted cross-entropy loss function, which is specifically as follows:

[0153]

[0154] in, For loss function, The true label is 1 for successful samples and 0 for failed samples. To predict probabilities, This is a hyperparameter, and its value is positively correlated with the project size.

[0155] Using AUC as the core metric, Bayesian optimization is used to tune the model's hyperparameters.

[0156] The final output is the project success probability, and the weight distribution output by the time-series attention layer is visualized through the management dashboard.

[0157] Specifically, This is a hyperparameter, and its value is positively correlated with the project size; specifically: for projects with a size ≤ 1 million, The value is 0.4; 1 million < scale ≤ 5 million. Value is 0.5; size > 5 million. The value is 0.6.

[0158] Please see Figure 2 As shown, this invention is a project data record analysis system based on CRM customer data analysis, comprising the following modules:

[0159] Multi-source data acquisition module: Extracts historical customer behavior data from CRM, and simultaneously extracts project execution data from PM system;

[0160] Customer-Project Association Analysis and Modeling Module: Based on the customer's historical behavior data and project execution data, calculate the project-customer association strength index, construct a dynamic customer-project association model, and train the constructed dynamic customer-project association model;

[0161] Project health assessment and early warning module: Based on the project execution data, calculate a multidimensional project health index and set a health index threshold based on historical data. When the multidimensional project health index value is lower than the threshold, an early warning is triggered.

[0162] Project success probability prediction module: Construct a machine learning classification model, take the project-customer association strength index and the multidimensional project health index as core features, and input them into the machine learning classification model along with traditional features. The model extracts deep temporal features of customer association and project health through association strength analysis subnet and health analysis subnet respectively, and fuses all features through temporal attention layer to learn the dynamic weight of each feature under different project stages, and finally outputs the project success probability.

[0163] Project data visualization management module: Constructs a management dashboard to display the project-customer relationship strength index, multi-dimensional project health index, and project success probability for each project in real time.

[0164] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for analyzing project data records based on CRM customer data analysis, characterized by, The method comprises the following steps: S1: extracting customer historical behavior data from CRM, and extracting project execution data from PM system; S2: calculating project-customer correlation strength index based on the customer historical behavior data and project execution data, and constructing a customer-project dynamic correlation model, and training the constructed customer-project dynamic correlation model; S3: calculating a multi-dimensional project health index based on the project execution data, and setting a health index threshold based on historical data, and triggering an early warning when the multi-dimensional project health index value is lower than the threshold; S4: constructing a machine learning classification model, taking the project-customer correlation strength index and the multi-dimensional project health index as core features, and inputting them into the machine learning classification model together with traditional features, the model extracting deep time sequence features of customer correlation and project health through a correlation strength analysis subnetwork and a health degree analysis subnetwork respectively, and fusing all features through a time sequence attention layer to learn dynamic weights of each feature in different project stages, and finally outputting a project success probability; S5: constructing a management dashboard for real-time display of the project-customer correlation strength index, the multi-dimensional project health index and the project success probability of each project; The step S1 comprises the following steps: extracting customer historical behavior data from CRM, including historical transaction records, customer support data and customer interaction data; the historical transaction records include transaction amount, frequency and product line, the customer support data include the number of work orders and the average resolution time, and the customer interaction data include the communication frequency of key persons and the number of meetings; extracting project execution data from PM system, including task completion progress, milestone achievement time, budget consumption rate and resource input working hours; The step of constructing a customer-project dynamic correlation model comprises the following steps: dividing the project life cycle into a sequence of consecutive time windows wherein, is the current window, is the number of the time window; updating each hierarchical index of each window in real time; adopting a moving average algorithm to smooth each hierarchical index; introducing a time decay factor to implement exponential weight decay on historical data over 3 months, and the calculation of the project-customer correlation strength index of the current window is as follows: wherein, is the current time window, i.e., the first week of the project-customer association strength index, , and are the hierarchical index numbers of the transaction association, support association and interaction association, respectively, , and are the basic weights of the transaction, support and interaction hierarchical indexes, respectively, is the historical window offset, is the smaller value of and 12, is the time decay factor, is the th transaction hierarchical index value of the th week after being smoothed by the moving average, is the th support hierarchical index value of the th week after being smoothed by the moving average, is the th interaction hierarchical index value of the th week after being smoothed by the moving average; The step of training the constructed customer-project dynamic correlation model comprises the following steps: selecting completed projects in the past 3 years as training samples, each sample containing original hierarchical index data of each time window in the project life cycle and a 5-point project correlation quality label manually labeled; adopting a 1D-CNN time sequence feature extraction module and a GRU dynamic weight learning module; the loss function adopts a weighted mean square error, the optimizer adopts an Adam adaptive optimization algorithm, the initial learning rate is set to 0.001, and the learning rate is attenuated by 50% every 50 training periods; the model outputs the customer-project correlation strength index and the hierarchical index contribution degree of each window.

2. The method of claim 1, wherein the CRM customer data analysis based project data record analysis method is characterized by, The step of calculating the project-customer correlation strength index comprises the following steps: defining the project-customer correlation strength index as follows: wherein, is the project-customer association strength index, , , , , , , and is the weight coefficient of the corresponding parameter, is the core product transaction proportion, is the transaction frequency density, is the transaction growth rate, is the project-specific work order proportion, is the work order solution timeliness deviation, is the decision layer interaction frequency, is the interaction depth index, is the information synchronization and timeliness; The core product transaction proportion, transaction frequency density and transaction growth rate are divided into transaction correlation degree hierarchical indexes; the project exclusive work order proportion and work order solution time deviation are divided into support correlation degree hierarchical indexes; and the decision layer interaction frequency, interaction depth index and information synchronization timeliness are divided into interaction correlation degree hierarchical indexes; The weight coefficient is determined by a double-factor iterative optimization method, specifically as follows: The initial weight is set by an industry expert according to the project type, and based on historical project data, the weight is iteratively adjusted by a particle swarm optimization algorithm to maximize the Pearson correlation coefficient of the project-customer correlation strength index and the project success rate, and at the same time, the weight is re-optimized once every quarter according to the newly added project data.

3. The method of claim 1, wherein the CRM customer data analysis based project data record analysis method is characterized by, The multi-dimensional project health index comprises the following steps: The multi-dimensional project health index is obtained by weighting and adding the progress health degree, the resource health degree, the budget health degree and the customer health degree, and the weight is determined by random forest importance analysis; The progress health degree is calculated by the task completion rate and the milestone delay days, specifically as follows: wherein, is a progress health, is a task completion rate, is a milestone delay days, is a delay allowed maximum days; The resource health degree is calculated by the resource input deviation rate and the key resource idle rate, specifically as follows: wherein, is a resource health, is a resource input bias rate, is a key resource idle rate; The budget health degree is calculated by the budget consumption rate and the cost deviation rate, specifically as follows: wherein, is a budget health, is a budget consumption rate, is a cost variance rate; The customer health degree is calculated by the customer satisfaction score and the demand change frequency, specifically as follows: wherein, is a customer health score, is a customer satisfaction score, is a demand change frequency, is a maximum number of changes; The health index threshold is the 70th percentile of the multi-dimensional project health index of all successful projects in the past three years.

4. The method of claim 1, wherein the CRM customer data analysis based project data record analysis method is characterized by, The machine learning classification model comprises the following steps: Select successful and failed samples from historical project data and divide them into a training set and a test set; Preprocess the samples, process sample imbalance based on a time decay weighted sampling strategy, specifically as follows: assign a weight coefficient to each training sample, the weight coefficient is inversely proportional to the time distance from the time point of the sample to the current time, and based on the weight, perform weighted sampling, and then generate synthetic samples for the minority class samples after sampling by using the SMOTE algorithm; Input the core features and traditional features, the core features are the project-customer correlation strength index and the multi-dimensional project health index, and the traditional features are the customer industry type, the project scale, the team experience and the project complexity; A dynamic feature weighted fusion network model is used to construct the machine learning classification model, and the model is trained.

5. The method of claim 4, wherein the CRM customer data analysis based project data record analysis method is characterized by, The dynamic feature weighted fusion network model comprises the following steps: The model comprises: A correlation strength analysis subnet, which takes historical time series data of the project-customer correlation strength index as input, uses a gated recurrent unit network to capture the dynamic change pattern of customer correlation degree, and outputs a first feature vector; A health degree analysis subnet, which takes historical time series data of the multi-dimensional project health index as input, uses a one-dimensional convolutional neural network to extract the local dependency relationship and model of the health degree index, and outputs a second feature vector; A feature splicing layer, which is used to splice the first feature vector, the second feature vector and the traditional features to obtain a fusion feature vector; A time series attention layer, which is used to receive the fusion feature vector and automatically learn the dynamic weight of each type of feature in the fusion feature vector on the success probability at different project life cycle stages. An output layer connected with the time-series attention layer, configured to output a project success probability.

6. The project data record analysis method based on CRM customer data analysis of claim 4, wherein, The model is trained, including the following steps: The dynamic feature weighted fusion network model is trained by using a self-defined weighted cross-entropy loss function, and the loss function is specifically: wherein, is a loss function, is a true label, the value of the true label corresponding to a successful sample is 1, and the value of the true label corresponding to a failed sample is 0, is a predicted probability, is a hyperparameter, and the value is positively correlated with the size of the project; Taking AUC as the core index, the model hyperparameters are tuned by Bayesian optimization, and finally the project success probability is output, and the weight distribution output by the time-series attention layer is visualized and displayed through a management dashboard.

7. A project data record analysis system based on CRM customer data analysis, for implementing a project data record analysis method based on CRM customer data analysis according to any one of claims 1 to 6, characterized in that, The method comprises the following modules: A multi-source data acquisition module: extracting customer historical behavior data from CRM, and simultaneously extracting project execution data from a PM system; A customer-project correlation analysis and modeling module: based on the customer historical behavior data and the project execution data, calculating a project-customer correlation strength index, and constructing a customer-project dynamic correlation model, and training the constructed customer-project dynamic correlation model; A project health assessment and early warning module: based on the project execution data, calculating a multi-dimensional project health index, and setting a health index threshold based on historical data, and triggering an early warning when the multi-dimensional project health index value is lower than the threshold; A project success probability prediction module: constructing a machine learning classification model, taking the project-customer correlation strength index and the multi-dimensional project health index as core features, and inputting them into the machine learning classification model together with traditional features, the model extracts deep time-series features of customer correlation and project health through a correlation strength analysis subnetwork and a health degree analysis subnetwork respectively, and fuses all features through a time-series attention layer to learn the dynamic weights of each feature in different project stages, and finally outputs a project success probability; A project data visualization management module: constructing a management dashboard for real-time display of the project-customer correlation strength index, the multi-dimensional project health index and the project success probability of each project.

Images