User data judgment system fusing multivariate data

Abstract

The invention discloses a user data judgment system fusing multivariate data, which relates to the technical field of data processing and analysis, acquires data from multiple channels such as an e-commerce platform, social media, a search engine and an offline store through a multivariate data acquisition module, supports real-time and timed acquisition and has a data quality monitoring function. The data preprocessing module uses an advanced algorithm to clean noise, repeated and wrong data, correct abnormal values, fill missing values and standardize the data. The multivariate data fusion module comprehensively adopts a feature and model fusion method, processes high-dimensional sparse features by using an automatic encoder, and dynamically adjusts sub-model weights. And the user data judgment model module is optimized by means of a deep learning model in combination with a reinforcement learning algorithm. And the result output and application module presents the result in a visual form, is applied to the fields of precision marketing, customer relationship management, financial credit evaluation and the like, and can dynamically adjust the strategy according to the real-time behavior of the user.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field related to data processing and analysis, and specifically relates to a user data judgment system fusing multi-element data. BACKGROUND

[0002] In today's digital era, user data presents the characteristics of explosive growth, wide sources and various types. The traditional user data processing and judgment method often only analyzes a single or a few data sources, which is difficult to comprehensively and accurately reflect the real characteristics and behavior patterns of users. For example, relying only on the purchase records of users on an e-commerce platform to analyze user preferences cannot consider the user's interest expression on social media, search behavior in search engines and other important data.

[0003] Some existing data processing systems have poor data fusion effect, low data processing efficiency and low judgment accuracy when processing multi-element data. On the one hand, the great differences in data format and structure of different types make it difficult to effectively fuse them, resulting in low quality of fused data and affecting subsequent analysis. On the other hand, with the rapid increase in data volume, existing algorithms and system architectures are difficult to quickly process massive data, and cannot meet the real-time requirements. In addition, due to the lack of effective data mining and analysis models, the judgment of user data is often too simple and one-sided, and cannot deeply understand the potential needs and behavior trends of users. SUMMARY

[0004] The purpose of the present application is to provide a user data judgment system fusing multi-element data to solve the problems raised in the background.

[0005] To achieve the above purpose, the present application provides the following technical scheme:

[0006] The user data judgment system fusing multi-element data comprises:

[0007] The multi-element data acquisition module is configured to acquire user data from multiple data sources, including e-commerce platform transaction records, social media platform user dynamics and interaction information, search engine search logs, and offline store consumption records. For different data sources, appropriate data acquisition interfaces and protocols are adopted, and user order information, product browsing records and payment information are obtained regularly through the API interface of the e-commerce platform. Network crawler technology that complies with laws, regulations and platform rules is used to collect data such as user publication content, likes, comments and follow relationships on social media platforms to ensure accurate data acquisition. At the same time, the module has two modes of real-time acquisition and timed acquisition, which can be flexibly switched according to the characteristics of data sources and business needs to ensure the timeliness and integrity of data;

[0008] The data preprocessing module is used to perform cleaning operation on the collected raw data, including removing noise data by using statistical-based method, detecting and deleting duplicate data by using hash algorithm, correcting error format data by using data format checking rule, identifying and correcting abnormal values by using clustering analysis and box plot-based method, and reasonably inferring or supplementing missing values by using multiple imputation method, K-Nearest Neighbor algorithm, etc. In addition, the module performs standardization processing on the cleaned data, converts different magnitudes of data to the interval of [0, 1] by using normalization algorithm for numerical data, and converts different formats and magnitudes of data to unified standard format and magnitude by using one-hot encoding and label encoding for category data, so as to facilitate subsequent analysis and processing.

[0009] The multi-source data fusion module is used to fuse the preprocessed data by using the combination of feature fusion and model fusion. In the aspect of feature fusion, the weighted summation, principal component analysis (PCA) and linear discriminant analysis (LDA) are used to generate a comprehensive feature vector from the features extracted from different types of data. The PCA is used to reduce the dimension of high-dimensional features and remove redundant information to improve the data processing efficiency. In the aspect of model fusion, multiple sub-models are trained based on different types of data, a user consumption behavior model is trained based on transaction data, and a user interest preference model is trained based on social media data. Then, the prediction results of the multiple sub-models are fused by using voting method, stacking method and Bayesian model averaging method to fully utilize the advantages of different models and improve the accuracy and reliability of the prediction.

[0010] The user data judgment model module is used to construct a user data judgment model based on deep learning, which adopts recurrent neural network (RNN), long short-term memory network (LSTM) and gated recurrent unit network (GRU) to capture the dynamic changes and long-term dependencies of user behavior by using the powerful processing capability of time series data. The fused user data is modeled and analyzed, and reinforcement learning algorithms such as deep Q network (DQN) and policy gradient algorithm are used to interact and feedback with actual user behavior data, adjust model parameters according to reward mechanism, improve prediction accuracy and adaptability of the model, and make it better cope with complex and variable user behavior scenarios.

[0011] Result output and application module: output the user data judgment result in an intuitive and easy-to-understand form, including generating user portrait report and user behavior prediction chart; the user portrait report covers the user's basic attributes, interest preferences, and consumption habits, and is presented in the form of word cloud chart and column chart; the user behavior prediction chart shows the user's possible behavior trend in the future, the purchase probability and the activity change, and is presented in the form of line chart and area chart; and the judgment result is applied to related fields such as precision marketing, customer relationship management, and financial credit assessment. In precision marketing, personalized product recommendations and advertisements are pushed according to user interests and purchase intentions by using collaborative filtering algorithm and content-based recommendation algorithm; in customer relationship management, potential customer loss risks are discovered in advance by using early warning models based on user behavior prediction, and corresponding retention measures are taken.

[0012] Preferably, the multi-data collection module has a data quality monitoring function during the collection process, which monitors the integrity, accuracy and consistency of the data in real time, and when data quality problems are found, automatically triggers data re-collection or repair mechanism, and generates data quality report to record data problems and processing process.

[0013] Preferably, the data preprocessing module, when cleaning data, checks the price field in the transaction record for rationality, compares it with the price of the same type of goods, and analyzes the time series to determine whether the price is abnormal, and marks and corrects it. For missing user attribute information, in addition to supplementing it according to other related data, it also uses association rule mining algorithm to find the potential association between attributes from a large amount of user data, improving the accuracy of missing value filling.

[0014] Preferably, the multi-data fusion module, when fusing features, uses automatic encoder deep learning technology to compress and reconstruct high-dimensional sparse features, retaining key feature information and reducing the consumption of data dimension on computing resources. In model fusion, the performance of different sub-models is evaluated by cross-validation and bootstrap method, and the weight of sub-models in the fusion process is dynamically adjusted to adapt to different data distribution and business scenarios.

[0015] Preferably, the deep learning model in the user data judgment model module adopts adaptive learning rate adjustment strategy in the training process, including Adagrad, Adadelta, RMSProp and Adam algorithms, which automatically adjusts the learning rate according to the progress of model training, speeds up the convergence of the model, and avoids falling into local optimal solution; at the same time, regularization techniques such as L1 and L2 regularization and Dropout are used to prevent model overfitting and improve the generalization ability of the model.

[0016] Preferably, the result output is a user portrait report generated by the application module, which supports user-defined dimension and index filtering and display, meets the personalized needs of different business departments, in the precise marketing application, can update the recommendation strategy in real time according to the real-time behavior data of the user, and realizes dynamic and precise marketing push, and in the customer relationship management application, integrated with the customer service system, when detecting the potential customer loss risk, automatically sends a reminder to the customer service personnel, and provides a customer historical behavior analysis report, assisting the customer service personnel to formulate a personalized retention plan.

[0017] Compared with the prior art, the present application provides a user data judgment system that fuses multiple data, which has the following beneficial effects:

[0018] Improve data judgment accuracy: by fusing multiple data, fully mining the behavior and feature information of users in different scenarios, the judgment of users is more comprehensive and accurate, which can deeply understand the potential needs and behavior trends of users, and provide more reliable basis for enterprise decision-making.

[0019] Improve data processing efficiency: by using advanced data acquisition, preprocessing and fusion technology, and efficient deep learning and reinforcement learning algorithm, massive multi-data can be quickly processed and analyzed, meeting the real-time requirement and improving the overall performance of the system.

[0020] Enhance system adaptability: use reinforcement learning to continuously optimize the user data judgment model, so that the system can adjust the model parameters in time according to the dynamic changes of user behavior, adapt to the needs of different user groups and business scenarios, and has stronger universality and adaptability. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the present application will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0022] The present application provides a user data judgment system that fuses multiple data, including:

[0023] Multi-element data collection module: configured to collect user data from multiple data sources, including e-commerce platform transaction records, social media platform user dynamics and interaction information, search engine search logs, and offline store consumption records. For different data sources, adopt appropriate data collection interfaces and protocols, and regularly obtain user order information, product browsing records, and payment information through API interface docking with e-commerce platforms. Use network crawler technology that complies with laws, regulations, and platform rules to collect data on user posts, likes, comments, and follow relationships on social media platforms to ensure accurate data acquisition. At the same time, this module has both real-time collection and scheduled collection modes, which can be flexibly switched according to data source characteristics and business needs to ensure data timeliness and integrity.

[0024] Data preprocessing module: used for cleaning the collected raw data, including removing noise data using statistical-based methods, detecting and deleting duplicate data through hash algorithms, correcting error format data using data format verification rules, identifying and correcting outliers using clustering analysis and box plot-based methods, and for missing values, using multiple imputation, K-nearest neighbor algorithm, etc. for reasonable speculation or supplementation. In addition, the module standardizes the cleaned data, converts different magnitudes of numerical data to the [0, 1] interval using normalization algorithms, and converts categorical data using one-hot encoding and label encoding to convert different formats and magnitudes of data into a unified standard format and magnitude for subsequent analysis and processing.

[0025] Multi-element data fusion module: combines feature fusion and model fusion to fuse preprocessed data. In feature fusion, weighted summation, principal component analysis (PCA), and linear discriminant analysis (LDA) are used to generate comprehensive feature vectors from different types of data. PCA is used to reduce the dimensionality of high-dimensional features to remove redundant information and improve data processing efficiency. In model fusion, multiple sub-models are trained using different types of data, including a user consumption behavior model trained based on transaction data and a user interest preference model trained based on social media data. The prediction results of multiple sub-models are then fused using voting, stacking, and Bayesian model averaging methods to fully leverage the strengths of different models and improve prediction accuracy and reliability.

[0026] The user data judgment model module: constructs a user data judgment model based on deep learning, adopts recurrent neural network RNN, long short-term memory network LSTM, and gated recurrent unit network GRU, utilizes the powerful processing capability of the three networks for time series data, captures the dynamic changes and long-term dependencies of user behavior, models and analyzes the fused user data, meanwhile, utilizes reinforcement learning algorithms such as deep Q network DQN and policy gradient algorithm, interacts and feeds back with actual user behavior data, adjusts model parameters according to the reward mechanism, improves the prediction accuracy and adaptability of the model, and makes the model better cope with complex and changeable user behavior scenarios.

[0027] The result output and application module: outputs the user data judgment result in an intuitive and easy-to-understand form, including generating a user portrait report and a user behavior prediction chart; the user portrait report covers the basic attributes, interest preferences, and consumption habits of the user, and is presented in the form of word cloud and column chart; the user behavior prediction chart shows the possible behavior trend of the user in the future, the purchase probability and the change of activity level in the form of line chart and area chart; and the judgment result is applied to related fields such as precision marketing, customer relationship management, and financial credit assessment. In precision marketing, personalized product recommendations and advertisements are pushed according to user interests and purchase intentions by using collaborative filtering algorithm and content-based recommendation algorithm; in customer relationship management, potential customer churn risks are discovered in advance by using early warning models based on user behavior prediction, and corresponding retention measures are taken.

[0028] The multi-element data collection module has a data quality monitoring function in the collection process, and monitors the integrity, accuracy and consistency of the data in real time. When data quality problems are found, the data re-collection or repair mechanism is automatically triggered, and a data quality report is generated to record the data problems and processing process.

[0029] In the data cleaning process, the data preprocessing module checks the price field in the transaction record for rationality, compares it with the price of the same type of goods, and analyzes the time series to determine whether the price is abnormal, and marks and corrects it. For missing user attribute information, in addition to supplementing it according to other related data, the association rule mining algorithm is also used to find the potential association between attributes from a large amount of user data, improving the accuracy of missing value filling.

[0030] In feature fusion, the multi-element data fusion module uses automatic encoder deep learning technology to compress and reconstruct high-dimensional sparse features, retaining key feature information and reducing the consumption of computing resources by data dimension. In model fusion, the performance of different sub-models is evaluated by cross-validation and bootstrap method, and the weight of the sub-model in the fusion process is dynamically adjusted to adapt to different data distribution and business scenarios.

[0031] The deep learning model in the user data judgment model module adopts an adaptive learning rate adjustment strategy in the training process, including the algorithms of Adagrad, Adadelta, RMSProp, and Adam, to automatically adjust the learning rate according to the progress of model training, accelerate the convergence speed of the model, and avoid falling into a local optimal solution; at the same time, regularization techniques such as L1 and L2 regularization and Dropout are used to prevent model overfitting and improve the generalization ability of the model.

[0032] The result output and application module generates a user portrait report, supports user-defined dimension and index filtering and display, and meets the personalized needs of different business departments; in precision marketing applications, it can update the recommendation strategy in real time according to real-time user behavior data, and realize dynamic and precise marketing push; in customer relationship management applications, it is integrated with customer service systems, and when potential customer churn risk is detected, it automatically sends reminders to customer service personnel and provides customer historical behavior analysis reports to assist customer service personnel in developing personalized retention solutions.

[0033] Embodiment one: user purchase intention prediction in e-commerce field

[0034] Data collection: Collect user's historical order data from an e-commerce platform, including purchase time, purchase commodity category, purchase amount, etc.; at the same time, collect user's browsing records on the platform, such as browsing commodity category, browsing duration, etc. In addition, through cooperation with social media platforms, obtain user's discussion content related to e-commerce on social media, data such as liked commodity types, etc. In the collection process, according to the characteristics of the data source, the e-commerce platform data adopts a timed collection mode, and the data is updated every morning to ensure the timeliness of the data; the social media data adopts a real-time collection mode to obtain the latest user dynamics in time. Use the data quality monitoring function to monitor the integrity and accuracy of the data in real time, such as finding missing fields in the e-commerce platform order data, triggering the data re-collection mechanism immediately.

[0035] Data preprocessing: clean the collected order data, remove data with abnormal order status (such as canceled but recorded incorrectly not deleted), filter invalid browsing (such as very short browsing duration, which may be a mistake) in browsing records. For abnormal values in the price of order data, correct them through comparison with similar commodity prices and time series analysis; for missing values in user attribute information, use association rule mining algorithm to fill them in combination with other user data. Then standardize the order amount, browsing duration and other data, convert the order amount to the [0, 1] interval through normalization algorithm, and convert the commodity category and other category type data through one-hot encoding.

[0036] Data fusion: Extract features such as purchase frequency and average purchase amount from order data, and extract features such as category distribution of browsed goods and browsing depth from browsing records. Use weighted summation to fuse these features, set the order data feature weight to 0.6, and set the browsing record feature weight to 0.4. At the same time, train a purchase behavior prediction sub-model based on order data and an interest preference sub-model based on browsing records, and use voting to fuse the prediction results of the two sub-models. When fusing features, use PCA for dimensionality reduction processing for high-dimensional goods features; before model fusion, evaluate the performance of the two sub-models through cross-validation, and dynamically adjust the voting weight according to the performance.

[0037] User data judgment model: Construct an LSTM model and input the fused feature data into the model for training. Use reinforcement learning algorithm to adjust model parameters according to the difference between actual purchase behavior and model prediction results. In the training process, use Adam adaptive learning rate adjustment strategy to speed up model convergence, and use Dropout regularization technique to prevent model overfitting.

[0038] Result output and application: The model outputs the probability of a user purchasing a certain type of goods in the next week. Based on the prediction results, the e-commerce platform pushes related goods coupons and recommendation information to high-probability purchase users, improving marketing effectiveness. At the same time, e-commerce platform operators can customize filtering of user age, region, consumption preference and other dimension information according to user portrait report, deeply analyze the purchase intention of different user groups, and optimize marketing strategy.

[0039] (II) Example Two: User Credit Evaluation in the Financial Field

[0040] Data collection: Obtain user loan records, repayment records, credit card usage records, etc. from financial institutions; obtain user consumption behavior data (such as consumption amount and frequency in different consumption scenarios) and social network relationship data (such as social influence and social relationship stability) from third-party data platforms. For internal data of financial institutions, use a combination of real-time collection and scheduled collection to obtain new loan and repayment records in real time, and update credit card usage records daily; for third-party data platform data, set a reasonable collection period according to data update frequency to ensure data timeliness. Through data quality monitoring, ensure that the collected data is accurate and correct, such as strictly checking the overdue marks in loan records.

[0041] Data preprocessing: cleaning overdue data error labels in loan records, abnormal transaction records in credit card usage records, etc. Standardize data from different data sources, such as normalizing loan amounts, consumption amounts, etc. For abnormal overdue data in loan records, correct it by comparing with historical repayment data, industry overdue standard reference, etc. For missing social network relationship data, use K-Nearest Neighbor algorithm to fill in according to similar user social data. Normalize numerical data such as loan amount, consumption amount, etc. to [0, 1] interval, and use label encoding for categorical data in social network relationships (such as social relationship type).

[0042] Data fusion: extract credit history length, overdue times, etc. from loan records, and consumption stability, consumption diversity, etc. from consumption behavior data, and fuse these features through PCA method. Train a credit evaluation sub-model based on internal data of financial institutions, and train a risk assessment sub-model based on third-party data, and use stacking method to fuse the results of the two sub-models. When feature fusion, use auto-encoder for feature compression and reconstruction for high-dimensional consumption behavior features; when model fusion, evaluate sub-model performance multiple times through bootstrap method to determine the optimal stacking method and weight.

[0043] User data judgment model: use multilayer perceptron (MLP) to build user credit evaluation model and input fused feature data for model training. Through reinforcement learning, optimize the model according to user's actual credit performance (such as whether to repay on time, etc.). In the training process, use Adagrad adaptive learning rate adjustment strategy, combined with L1 and L2 regularization techniques, to improve the generalization ability and stability of the model. Result output and application: the model outputs the user's credit score and credit level. Financial institutions decide whether to provide loans to users, as well as loan amount and interest rate, based on the score and level, to reduce credit risk. At the same time, when communicating with users, if the user's credit risk rises, the system automatically provides customer service personnel with detailed information about the user's credit trend, recent consumption behavior, etc. based on user portrait reports and behavior prediction charts, to assist customer service personnel in communicating with users and developing appropriate risk response measures.

[0044] Finally, it should be noted that the above-described preferred embodiments of the present application are not intended to limit the present application, and although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements for some technical features, as long as they are within the spirit and principles of the present application. Any modifications, equivalent replacements, improvements, etc. made shall be included in the protection scope of the present application.

Claims

1. A user data judging system that fuses plural data, characterized by, Comprise: Multi-element data acquisition module: configured to collect user data from multiple data sources, covering e-commerce platform transaction records, social media platform user dynamics and interaction information, search engine search logs, offline store consumption records, for different data sources, adopt the appropriate data acquisition interface and protocol, through the API interface docking with e-commerce platform, regularly obtain user order information, commodity browsing record, payment information; use the network crawler technology conforming to laws and regulations and platform rules to collect the data of the user's published content, likes, comments, and attention relationship on the social media platform to ensure the accuracy of the data acquisition; at the same time, the module has real-time acquisition and timing acquisition two modes, which can be flexibly switched according to the characteristics of data sources and business needs, to ensure the timeliness and integrity of the data; Data preprocessing module: used for cleaning the collected raw data, including removing noise data by statistical methods, detecting and deleting duplicate data by hash algorithm, correcting error format data by data format verification rules, identifying and correcting abnormal values by clustering analysis and box plot method; for missing values, according to the data type and distribution, use multiple imputation method, K nearest neighbor algorithm for reasonable speculation or supplement; in addition, the module standardizes the cleaned data, for numerical data, use normalization algorithm to convert different orders of data to [0,1] interval; for category data, use one-hot encoding and label encoding to convert different formats and different orders of data to unified standard format and order, so as to facilitate subsequent analysis and processing; Multi-element data fusion module: adopts the method of combining feature fusion and model fusion to fuse the preprocessed data; In the aspect of feature fusion, use weighted sum, principal component analysis PCA, linear discriminant analysis (LDA) to generate comprehensive feature vectors from different types of data; through PCA, reduce the dimension of high-dimensional features, remove redundant information and improve data processing efficiency; in the aspect of model fusion, train multiple sub-models respectively using different types of data, train user consumption behavior model based on transaction data, train user interest preference model based on social media data, then fuse the prediction results of multiple sub-models by voting method, stacking method and Bayesian model averaging method, fully utilize the advantages of different models, improve the accuracy and reliability of prediction; User data judgment model module: construct a user data judgment model based on deep learning, use recurrent neural network RNN, long short-term memory network LSTM, gated recurrent unit network GRU, use its powerful processing ability for time series data to capture the dynamic changes and long-term dependence of user behavior, model analysis on the fused user data, at the same time, use reinforcement learning algorithm such as deep Q network DQN and policy gradient algorithm, through continuous interaction and feedback with user actual behavior data, adjust the model parameters according to the reward mechanism, improve the prediction accuracy and adaptability of the model, so that it can better cope with complex and variable user behavior scenarios; Result output and application module: output the user data judgment result in an intuitive and easy-to-understand form, including generating user portrait report and user behavior prediction chart; The user portrait report covers the basic attributes, interest preferences and consumption habits of the user, and is presented in the form of word cloud chart and column chart. The user behavior prediction chart shows the possible behavior trend of the user in the future period of time, the purchase probability and the activity change, and is presented in the form of line chart and area chart. The judgment result is applied to the related fields of precision marketing, customer relationship management and financial credit assessment. In precision marketing, personalized product recommendation and advertising are pushed according to the user interest and purchase intention by using collaborative filtering algorithm and content-based recommendation algorithm. In customer relationship management, potential customer loss risk is found in advance by using early warning model through user behavior prediction, and corresponding retention measures are taken.

2. The user data judging system of fused multi-element data according to claim 1, characterized by, The multi-data collection module has data quality monitoring function in the collection process, and monitors the integrity, accuracy and consistency of the data in real time. When data quality problems are found, the data re-collection or repair mechanism is automatically triggered, and a data quality report is generated to record the data problems and processing process.

3. The user data judging system of fused multi-dimensional data according to claim 1, characterized in that, In the data preprocessing module, in addition to checking the numerical rationality of the price field in the transaction record, the price is also judged to be abnormal by comparing it with the price of the same type of goods and analyzing the time sequence, and is marked and corrected. For missing user attribute information, in addition to supplementing it according to other related data, the association rule mining algorithm is also used to find the potential association between attributes from a large amount of user data, so as to improve the accuracy of missing value filling.

4. The user data judging system of fused multi-dimensional data according to claim 1, characterized in that, In the multi-data fusion module, for high-dimensional sparse features, automatic encoder deep learning technology is used for feature compression and reconstruction to retain key feature information and reduce the consumption of data dimension on computing resources. In model fusion, the performance of different sub-models is evaluated by cross-validation and bootstrap method, and the weight of the sub-model in the fusion process is dynamically adjusted to adapt to different data distribution and business scenarios.

5. The user data judging system of fused multi-dimensional data according to claim 1, characterized in that, In the deep learning model in the user data judgment model module, adaptive learning rate adjustment strategy is used in the training process, including Adagrad, Adadelta, RMSProp and Adam algorithms, to automatically adjust the learning rate according to the progress of model training, speed up the convergence speed of the model and avoid falling into local optimal solution. At the same time, regularization techniques such as L1 and L2 regularization and Dropout are used to prevent model overfitting and improve the generalization ability of the model.

6. The user data judging system of fused multi-dimensional data according to claim 1, characterized in that, The result output and the user portrait report generated by the application module support user-defined dimension and index filtering and display, and meet the personalized needs of different business departments; in the precise marketing application, the recommended strategy can be updated in real time according to the real-time behavior data of the user, and dynamic and precise marketing push can be realized; in the customer relationship management application, the customer service system is integrated, when the potential customer loss risk is detected, the customer service personnel are automatically sent a reminder, and a customer historical behavior analysis report is provided to assist the customer service personnel in formulating a personalized retention plan.