Property collection monitoring and early warning method and system based on artificial intelligence

Abstract

The invention provides a property collection monitoring and early warning method and system based on artificial intelligence. The method comprises the following steps: acquiring historical time series data of a user, preprocessing the historical time series data of the user, and extracting features from the preprocessed data; establishing a time weighted attention mechanism model based on features extracted from the divided training set; measuring the difference between the prediction probability of the time-weighted attention mechanism and the actual risk category by using a cross entropy loss function, and adjusting a time-weighted attention mechanism model; the method comprises the following steps: acquiring user time sequence feature data, inputting the user time sequence feature data into an adjusted time weighted attention mechanism model, and adopting a collection strategy according to a model prediction result. The time weighted attention mechanism and the weighted cross entropy loss function can be combined to improve the prediction accuracy of the model for the customer payment risk, so that the collection strategy is adopted in a targeted manner, the collection efficiency is improved, and the resource allocation is optimized.

Description

Technical Field

[0001] The present invention proposes a property collection monitoring and early warning method and system based on artificial intelligence, belonging to the field of data processing technology. Background Art

[0002] Traditional debt collection work often relies on manual operations, including telephone collection, sending notifications, etc. These methods are not only time-consuming and labor-intensive, but also inefficient, especially when dealing with a large number of overdue accounts. Without a systematic monitoring and early warning mechanism, staff may find it difficult to prioritize the highest-risk cases, thereby slowing down the speed of problem resolution.

[0003] Moreover, in the case of debt collection, the customer's payment history often has time series characteristics. Although the traditional LSTM model can process time series data, it ignores the changing weights of data at different time points. Or after the model is trained, the traditional cross entropy loss function treats all categories equally and cannot deal with the problem of category imbalance by assigning different weights to different categories. Summary of the invention

[0004] The present invention provides a property debt collection monitoring and early warning method and system based on artificial intelligence to solve the above-mentioned problems.

[0005] The present invention proposes a property collection monitoring and early warning method based on artificial intelligence, the method comprising: Acquire user historical time series data, preprocess the user historical time data, and extract features from the preprocessed data; A temporal weighted attention mechanism model is established based on the features extracted from the partitioned training set; Using a cross entropy loss function to measure the difference between the predicted probability of the time-weighted attention mechanism and the actual risk category, and adjusting the time-weighted attention mechanism model; Obtain user time series feature data, input the user time series feature data into the adjusted time-weighted attention mechanism model, and adopt a collection strategy based on the model prediction results.

[0006] Further, obtaining user historical time series data, preprocessing the user historical time series data, and extracting features from the preprocessed data include: Acquire historical time series data of the user, the historical time series data including: payment records, property service requests and property fee adjustment records, the payment records including: the date and amount of each payment, the payment method, whether there is overdue payment, the overdue amount and the number of overdue days, the property service request including: the date and type of each service request and the date and result of service completion; the property fee adjustment record including: the date of adjustment, the amount of adjustment and the reason; Preprocessing the user historical data includes: for missing payment records, using forward filling to process missing values; converting the payment date into a unified time format; standardizing the payment amount and the overdue amount to have the same scale; aligning the time series data of each user to ensure that all sequences have the same length; Divide the preprocessed data into training set and validation set; Extract payment record features from the pre-processed payment records, wherein the payment record features include payment frequency, average payment amount, unique hot code of payment method, number of overdue payments, average overdue days, maximum days, and overdue payment amount ratio; Extracting property service request features from the pre-processed property service request records, wherein the property service request features include the total number of service requests, a unique hot encoding of the service request type, a response time, and a service satisfaction score; The property fee adjustment record features are extracted from the pre-processed property fee adjustment records, wherein the property fee adjustment record features include the number of fee adjustments, the average adjustment amount, the fee increase ratio and the adjustment reason unique hot coding.

[0007] Furthermore, a time-weighted attention mechanism model is established based on the features extracted from the divided training set, including: The features extracted from the training set are used to construct the feature vector x according to time. t , the feature vectors constitute the time series feature data {x1,x2,x3,…,x T}, input the time series feature data into the LSTM layer; The LSTM layer processes the time series feature data through the following formula to generate a hidden state vector for each time point t , the hidden state vector {h1,h2,h3,…,h T};

[0008] in, is the input feature vector at time t, It's the Gate of Forgetfulness. is the input gate, is the output gate, is a candidate memory cell, is the weight matrix of the forget gate, is the weight matrix of the input gate, is the weight matrix of the output gate, is the weight matrix of the candidate memory, σ is the Sigmoid function, whose output range is [0, 1], , , and They are the forget gate, input gate, output gate, and bias term of candidate memory; Calculate the weight of each time point, the calculation formula is:

[0009] in, is the attention weight, Represents the hidden state vector at time point t, the hidden state of the current step, which is used to capture all historical information before this time point. Represents the hidden state vector at time point j, which is the hidden state of each historical time point j, used to compare and calculate with the state at the current time point t. Represents the risk rating vector at time point j, which contains risk-related features and enters the attention mechanism to help the model better understand the risk status at each time point. For conversion , making it compatible with other vectors, For conversion , making it compatible with other vectors, Used to transform the risk rating vector , allowing it to be efficiently incorporated into the model's computation, is the bias vector, is the weight vector of the attention mechanism; The hidden states are weighted summed using the attention weights to obtain the context vector c:

[0010] The context vector c is input into the fully connected layer. The calculation formula of the fully connected layer is:

[0011] in, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer, and z is the linear transformation result without the activation function; The linear transformation result z is input into the softmax activation function to obtain the probability of the category to which the user payment risk belongs. The user payment risk is divided into 5 levels.

[0012] in, represents the probability that a given input x belongs to category k, represents the linear transformation result of category k, l represents the category index traversed by the current loop, and A, B, C, D and E represent different risk categories.

[0013] Furthermore, the cross entropy loss function is used to measure the difference between the predicted probability of the time-weighted attention mechanism and the actual risk category, including: Determine the weight of each category, wherein the weight of each category can be set based on the degree of sample imbalance, including: if category A, i.e., high-reputation customers, rarely have serious consequences of prediction errors, a lower weight is given, while category E, i.e., high-risk customers, are given a higher weight due to serious consequences; In each batch of model training, there are multiple samples. For each sample, the model outputs the predicted probability that it belongs to each category. These predicted probabilities and the actual category labels are used to calculate the loss function value through the weighted cross entropy loss function. Specifically, the cross entropy loss function is:

[0014] Among them, L represents the loss function, m represents the category index, from 1 to 5, corresponding to risk rating categories A to E, is the weight associated with category m, which is used to adjust the penalty for misclassification of different categories, including: misclassification of high-risk customers may bring greater costs, so this category can be given a higher weight, N is the total number of samples, is the binary representation of the true label. If sample i belongs to category m, then is 1, otherwise it is 0. is the probability that the model predicts that the sample belongs to category m; The calculated loss value is used to update the model weights through the back-propagation algorithm to minimize the loss function; Repeat the above process until the performance of the model on the validation set no longer improves.

[0015] Furthermore, based on the predicted risk probability, the collection strategy is determined, including: Input the user's current time series feature data into the model to obtain the predicted probability; For users with a predicted probability of A, obtain the user's historical payment method and payment status on the payment date. If the user has not paid on the payment date, make a reminder on the payment date based on the user's historical payment method. If the user's historical payment method is bank transfer, remind the user via SMS. If the user's historical payment method is online payment, remind the user via APP notification. If the user's historical payment method is cash, remind the user by calling and provide value-added services. For users with a predicted probability of B, obtain the user's historical payment method and payment status three days before the payment date. If the user has not paid three days before the payment date, make reminders three days before the payment date and on the payment date based on the user's historical payment method. For users with a predicted probability of C, obtain the user's historical payment method and payment status five days before the payment date. If the user has not paid five days before the payment date, make reminders five days, three days, and one day before the payment date, respectively, based on the user's historical payment method; For users with a predicted probability of D, obtain the user's historical payment method and payment status seven days before the payment date. If the user has not paid seven days before the payment date, make reminders seven days, five days, and three days before the payment date based on the user's historical payment method. If the user has not paid on the payment date, post a reminder at the user's door. For users with a predicted probability of E, the user's historical payment method and payment status ten days before the payment date are obtained. If the user has not paid ten days before the payment date, reminders will be made ten days, seven days, and five days before the payment date respectively based on the user's historical payment method. If the user has not paid three days before the payment date, a reminder will be posted at the user's door. If the user has not paid on the payment date, a message will be sent to the property staff that door-to-door collection is required.

[0016] The present invention proposes a property debt collection monitoring and early warning system based on artificial intelligence, the system comprising: A data preprocessing module is used to obtain user historical time series data, preprocess the user historical time data, and extract features from the preprocessed data; Establish a preliminary model module for building a temporal weighted attention mechanism model based on the features extracted from the partitioned training set; A loss function-based model adjustment module is used to measure the difference between the predicted probability of the time-weighted attention mechanism and the actual risk category using a cross-entropy loss function, and adjust the time-weighted attention mechanism model; A collection suggestion module is provided to obtain user time series feature data, input the user time series feature data into the adjusted time-weighted attention mechanism model, and adopt a collection strategy according to the model prediction results.

[0017] Furthermore, the data preprocessing module includes: A module for obtaining user historical time series data is used to obtain user historical time series data, wherein the historical time series data includes payment records, property service requests and property fee adjustment records. The payment records include the date and amount of each payment, the payment method, whether there is overdue payment, the amount of overdue payment and the number of overdue days. The property service request includes the date and type of each service request and the date and result of service completion. The property fee adjustment record includes the date of adjustment, the amount of adjustment and the reason. A preprocessing module is used to preprocess the user's historical data, including: for missing payment records, using forward filling to process missing values; converting the payment date into a unified time format; standardizing the payment amount and the overdue amount to have the same scale; aligning the time series data of each user to ensure that all sequences have the same length; The data set partitioning module is used to divide the preprocessed data into a training set and a validation set; A payment feature extraction module is used to extract payment record features from the preprocessed payment records, wherein the payment record features include payment frequency, average payment amount, unique hot code of payment method, number of overdue payments, average overdue days, maximum days and overdue payment amount ratio; A property service request feature extraction module is used to extract property service request features from pre-processed property service request records, wherein the property service request features include the total number of service requests, a unique hot code of the service request type, a response time, and a service satisfaction score; Extracting property expense adjustment record features is used to extract property expense adjustment record features from preprocessed property expense adjustment records, wherein the property expense adjustment record features include the number of expense adjustments, the average adjustment amount, the expense increase ratio, and the unique hot coding of the adjustment reason.

[0018] Furthermore, the building of the preliminary model module includes: The vector construction module is used to construct the feature vector x based on time from the features extracted from the training set t , the feature vectors constitute the time series feature data {x1,x2,x3,…,x T}, input the time series feature data into the LSTM layer; Generate hidden state vector module, the LSTM layer processes the time series feature data through the following formula to generate a hidden state vector for each time point t , the hidden state vector {h1,h2,h3,…,h T};

[0019] in, is the input feature vector at time t, It's the Gate of Forgetfulness. is the input gate, is the output gate, is a candidate memory cell, is the weight matrix of the forget gate, is the weight matrix of the input gate, is the weight matrix of the output gate, is the weight matrix of the candidate memory, σ is the Sigmoid function, whose output range is [0, 1], , , and They are the forget gate, input gate, output gate, and bias term of candidate memory; The time point weight calculation module is used to calculate the weight of each time point. The calculation formula is:

[0020] in, is the attention weight, Represents the hidden state vector at time point t, the hidden state of the current step, which is used to capture all historical information before this time point. Represents the hidden state vector at time point j, which is the hidden state of each historical time point j, used to compare and calculate with the state at the current time point t. Represents the risk rating vector at time point j, which contains risk-related features and enters the attention mechanism to help the model better understand the risk status at each time point. For conversion , making it compatible with other vectors, For conversion , making it compatible with other vectors, Used to transform the risk rating vector , allowing it to be efficiently incorporated into the model's computation, is the bias vector, is the weight vector of the attention mechanism; The hidden state weighting module is used to weight the hidden state using the attention weights to obtain the context vector c:

[0021] The input fully connected layer module is used to input the context vector c into the fully connected layer. The calculation formula of the fully connected layer is:

[0022] in, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer, and z is the linear transformation result without the activation function; The payment risk probability module is used to input the linear transformation result z into the softmax activation function to obtain the probability of the category to which the user payment risk belongs. The user payment risk is divided into 5 levels.

[0023] in, represents the probability that a given input x belongs to category k, represents the linear transformation result of category k, l represents the category index traversed by the current loop, and A, B, C, D and E represent different risk categories.

[0024] Furthermore, the loss function-based model adjustment module includes: A weight setting module is used to determine the weight of each category, and the weight of each category can be set based on the degree of sample imbalance, including: if category A, i.e., high-reputation customers, rarely have serious consequences of prediction errors, a lower weight is given, while category E, i.e., high-risk customers, are given a higher weight due to serious consequences; The loss calculation module is used for each batch of model training. There are multiple samples. For each sample, the model outputs the predicted probability that it belongs to each category. These predicted probabilities and the actual category labels are used to calculate the loss function value through the weighted cross entropy loss function. Specifically, the cross entropy loss function is:

[0025] Among them, L represents the loss function, m represents the category index, from 1 to 5, corresponding to risk rating categories A to E, is the weight associated with category m, which is used to adjust the penalty for misclassification of different categories, including: misclassification of high-risk customers may bring greater costs, so this category can be given a higher weight, N is the total number of samples, is the binary representation of the true label. If sample i belongs to category m, then is 1, otherwise it is 0. is the probability that the model predicts that the sample belongs to category m; Back propagation module, the calculated loss value is used to update the model weights through the back propagation algorithm to minimize the loss function; The iterative optimization module is used to repeat the above process until the performance of the model on the validation set no longer improves.

[0026] Furthermore, the module for providing collection suggestions includes: The prediction probability acquisition module is used to input the user's current time series feature data into the model to obtain the prediction probability; The module with predicted probability A is used to obtain the user's historical payment method and payment status on the payment date for the user with predicted probability A. If the user has not paid on the payment date, a reminder will be made on the payment date based on the user's historical payment method. If the user's historical payment method is bank transfer, the user will be reminded via SMS. If the user's historical payment method is online payment, the user will be reminded via APP notification. If the user's historical payment method is cash, the user will be reminded by phone and value-added services will be provided. The module with predicted probability B is used to obtain the user's historical payment method and payment status three days before the payment date for the user with predicted probability B. If the user has not paid three days before the payment date, a reminder will be made three days before the payment date and on the payment date according to the user's historical payment method; The module with a predicted probability of C is used to obtain the user's historical payment method and payment status five days before the payment date for the user with a predicted probability of C. If the user has not paid five days before the payment date, a reminder will be given five days, three days and one day before the payment date respectively according to the user's historical payment method; The module with a predicted probability of D is used to obtain the user's historical payment method and payment status seven days before the payment date for the user with a predicted probability of D. If the user has not paid seven days before the payment date, reminders will be made seven days, five days and three days before the payment date respectively according to the user's historical payment method. If the user has not paid on the payment date, a warm reminder will be posted at the user's door; The module with predicted probability E is used to obtain the user's historical payment method and payment status ten days before the payment date for the user with predicted probability E. If the user has not paid ten days before the payment date, reminders will be made ten days, seven days and five days before the payment date respectively according to the user's historical payment method. If the user has not paid three days before the payment date, a reminder will be posted at the user's door. If the user has not paid on the payment date, a message will be sent to the property staff that door-to-door collection is required.

[0027] The beneficial effects of the present invention are as follows: improved accuracy: through the time-weighted attention mechanism, the model can more accurately identify and utilize the historical data points that are most critical to the prediction results, thereby improving the accuracy of the prediction; real-time and adaptability: the model can be quickly adjusted according to the data collected in real time, better adapt to changes in user behavior, and timely update risk assessment and collection strategies; the use of automated and intelligent collection prediction models can reduce manual collection activities, reduce management costs, and improve work efficiency; customer relationship management: through accurate risk assessment and collection strategies, collection behaviors can be arranged more reasonably, avoiding unnecessary damage to customer relationships, and enhancing user satisfaction and loyalty; this AI model based on the time-weighted attention mechanism provides property management companies with an efficient, accurate, and cost-effective collection solution, which helps to improve cash flow management and customer service quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a schematic diagram of a property collection early warning method based on artificial intelligence according to the present invention; Figure 2 This is a block diagram of an artificial intelligence-based property debt collection early warning system described in the present invention. DETAILED DESCRIPTION

[0029] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0030] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. The embodiments described are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0032] One embodiment of the present invention is a property collection monitoring and early warning method based on artificial intelligence, such as Figure 1 As shown, the method includes: Acquire user historical time series data, preprocess the user historical time data, and extract features from the preprocessed data; A temporal weighted attention mechanism model is established based on the features extracted from the partitioned training set; Using a cross entropy loss function to measure the difference between the predicted probability of the time-weighted attention mechanism and the actual risk category, and adjusting the time-weighted attention mechanism model; Obtain user time series feature data, input the user time series feature data into the adjusted time-weighted attention mechanism model, and adopt a collection strategy based on the model prediction results.

[0033] The working principle and effect of the above technical solution are as follows: data preprocessing and feature extraction, first, obtain the user's historical time series data, such as payment history, transaction records, etc.; preprocess these data, including cleaning data (processing missing values, outliers), standardizing or normalizing data for easy processing; from the processed data, use statistical methods or machine learning techniques to extract key features that affect the user's payment behavior, such as payment periodicity, average payment delay, etc.; use the extracted features to establish a model based on the time-weighted attention mechanism, which can assign different weights according to the importance of data at each time point, and more accurately capture the key information in the time series data; the time-weighted attention mechanism improves the model's prediction accuracy for future behavior by focusing on the historical data points that have the greatest impact on the prediction results; model training and optimization, using the cross entropy loss function to measure the difference between the probability distribution predicted by the model and the actual risk category. This loss function is suitable for classification problems and can effectively optimize model parameters; adjust model parameters through methods such as backpropagation algorithm and gradient descent to minimize the loss function value, thereby improving the prediction performance of the model. The model is applied to the collection strategy. After the model training is completed, the newly collected user time series feature data is input into the adjusted and optimized model. According to the prediction results of the model, such as the user's possible future payment behavior and risk level, the corresponding collection strategy is formulated. High-risk users may require more frequent or more direct collection intervention. Improve accuracy. Through the time-weighted attention mechanism, the model can more accurately identify and utilize the most critical historical data points for the prediction results, thereby improving the accuracy of the prediction; real-time and adaptability. The model can quickly adjust according to the real-time collected data, better adapt to changes in user behavior, and update risk assessment and collection strategies in a timely manner; using automated and intelligent collection prediction models, it can reduce the reliance on manual collection activities, reduce management costs, and improve work efficiency; customer relationship management, through accurate risk assessment and collection strategies, can more reasonably arrange collection behavior, avoid unnecessary damage to customer relationships, and enhance user satisfaction and loyalty; this AI model based on the time-weighted attention mechanism provides property management companies with an efficient, accurate, and cost-effective collection solution, which helps to improve cash flow management and customer service quality.

[0034] One embodiment of the present invention is a property collection monitoring and early warning method based on artificial intelligence, which obtains user historical time series data, preprocesses the user historical time data, and extracts features from the preprocessed data, including: Acquire historical time series data of the user, the historical time series data including: payment records, property service requests and property fee adjustment records, the payment records including: the date and amount of each payment, the payment method, whether there is overdue payment, the overdue amount and the number of overdue days, the property service request including: the date and type of each service request and the date and result of service completion; the property fee adjustment record including: the date of adjustment, the amount of adjustment and the reason; Preprocessing the user historical data includes: for missing payment records, using forward filling to process missing values; converting the payment date into a unified time format; standardizing the payment amount and the overdue amount to have the same scale; aligning the time series data of each user to ensure that all sequences have the same length; Divide the preprocessed data into training set and validation set; Extract payment record features from the pre-processed payment records, wherein the payment record features include payment frequency, average payment amount, unique hot code of payment method, number of overdue payments, average overdue days, maximum days, and overdue payment amount ratio; Extracting property service request features from the pre-processed property service request records, wherein the property service request features include the total number of service requests, a unique hot encoding of the service request type, a response time, and a service satisfaction score; The property fee adjustment record features are extracted from the pre-processed property fee adjustment records, wherein the property fee adjustment record features include the number of fee adjustments, the average adjustment amount, the fee increase ratio and the adjustment reason unique hot coding.

[0035] The working principle and effect of the above technical solution are as follows: it involves systematic processing of historical time series data of users, through a series of preprocessing, feature extraction and data analysis steps, aiming to provide accurate input data for further risk assessment and decision analysis, data collection, collecting users' payment records, property service requests and property fee adjustment records, these data are the basis for building a portrait of users' financial behavior and service needs; data preprocessing, cleaning and normalizing the data of the collection, including filling in missing values, unifying time formats, standardizing numerical data, and ensuring data length consistency, so that the output of the data preparation stage is more suitable for machine learning modeling; feature extraction, payment The payment record characteristics are used to extract key indicators such as payment frequency and average payment amount to reflect the user's payment ability and payment habits. The property service request characteristics collect the number of service requests and response time, and use this to reflect the user's dependence on and satisfaction with property services. The property fee adjustment record characteristics analyze the frequency and scale of fee adjustments to evaluate the impact of external factors on the user's financial status. The preprocessed data is divided into a training set and a validation set to ensure that the model can be fairly evaluated and optimized during the training process. Model training and validation use the extracted feature data to train the prediction model, and use the validation set to adjust and optimize the model parameters to ensure that the model has good universality and prediction accuracy. Risk management optimization: through in-depth analysis of users’ payment behaviors and property service interaction records, property management can more accurately identify potential high-risk users and formulate corresponding collection strategies or adjust service plans; service personalization: insights mined from property service request records can help property management formulate more personalized service strategies, thereby improving user satisfaction and loyalty; transparency of financial budgets and adjustments: through analysis of property fee adjustment records, property management can better understand the necessity and frequency of fee adjustments, thereby making more informed decisions in financial budget formulation and adjustment; improving the efficiency of decision support systems: incorporating data-driven insights into daily management decisions can reduce intuition-based decisions and improve management efficiency and accuracy; enhancing predictive capabilities: through in-depth analysis and pattern recognition of historical data, the ability to predict future user behavior is enhanced, especially in predicting overdue payments and changes in service demand; the above technical solutions provide powerful support tools for property management by systematically processing and analyzing user data, helping them make more accurate and effective decisions in a complex and changing management environment.

[0036] An embodiment of the present invention is an artificial intelligence-based property collection monitoring and early warning method, which establishes a time-weighted attention mechanism model based on features extracted from a divided training set, including: The features extracted from the training set are used to construct the feature vector x according to time. t , the feature vectors constitute the time series feature data {x1,x2,x3,…,xT}, input the time series feature data into the LSTM layer; The LSTM layer processes the time series feature data through the following formula to generate a hidden state vector for each time point t , the hidden state vector {h1,h2,h3,…,h T};

[0037] in, is the input feature vector at time t, It's the Gate of Forgetfulness. is the input gate, is the output gate, is a candidate memory cell, is the weight matrix of the forget gate, is the weight matrix of the input gate, is the weight matrix of the output gate, is the weight matrix of the candidate memory, σ is the Sigmoid function, whose output range is [0, 1], , , and They are the forget gate, input gate, output gate, and bias term of candidate memory; Calculate the weight of each time point, the calculation formula is:

[0038] in, is the attention weight, Represents the hidden state vector at time point t, the hidden state of the current step, which is used to capture all historical information before this time point. Represents the hidden state vector at time point j, which is the hidden state of each historical time point j, used to compare and calculate with the state at the current time point t. Represents the risk rating vector at time point j, which contains risk-related features and enters the attention mechanism to help the model better understand the risk status at each time point. For conversion , making it compatible with other vectors, For conversion , making it compatible with other vectors, Used to transform the risk rating vector , allowing it to be efficiently incorporated into the model's computation, is the bias vector, is the weight vector of the attention mechanism; The hidden states are weighted summed using the attention weights to obtain the context vector c:

[0039] The context vector c is input into the fully connected layer. The calculation formula of the fully connected layer is:

[0040] in, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer, and z is the linear transformation result without the activation function; The linear transformation result z is input into the softmax activation function to obtain the probability of the category to which the user payment risk belongs. The user payment risk is divided into 5 levels.

[0041] in, represents the probability that a given input x belongs to category k, represents the linear transformation result of category k, l represents the category index traversed by the current loop, and A, B, C, D and E represent different risk categories.

[0042] The working principle and effect of the above technical solution are as follows: Feature vector construction, first, the features extracted from the training set are time-tagged and converted into feature vectors, which are organized along the time series so as to capture time correlation and patterns and provide input data for the subsequent deep learning model; LSTM layer processing, the feature vector of each time point is input into the LSTM layer. LSTM processes data by using gated units (input gate, forget gate, output gate and candidate memory unit) to capture long-term dependencies in time series data. The hidden state vector generated at each time point encodes all historical information up to that time point. The attention mechanism used here can make the model focus on the historical events most relevant to the current prediction task. By comparing the hidden states at different time points, the model can learn which past information points are most useful for the current prediction, and calculate the attention weight accordingly. The context vector and the fully connected layer use the attention weight to weight the sum of all hidden states to form the context vector c, which represents the comprehensive representation of the most important historical information for the current prediction. The context vector is then fed into a fully connected layer, which converts the context vector into an output vector z as the basis for predicting the user's payment risk category, Softmax activation and risk category prediction. Finally, the output of the fully connected layer is transformed by the softmax function to obtain the probability distribution of different risk levels. Softmax ensures that the sum of the output probabilities is 1, and the probability of each category reflects the possibility that the user belongs to that risk level. The combination of LSTM and attention mechanism can effectively capture the long-term dependencies and importance of time series data, thereby improving the accuracy of prediction results; the attention mechanism enables the model to dynamically focus on the historical information most relevant to the current prediction task, so as to adjust its prediction in real time according to the changes in historical data; by dividing the user's payment risk into multiple levels (A, B, C, D, E), the model can provide more detailed risk assessment and help take corresponding preventive or collection measures; optimize decision support, the deep insights provided by this technical solution help to formulate more effective user management strategies, such as formulating differentiated collection strategies for different risk levels, and this technical solution provides a powerful tool to improve collection efficiency and optimize user risk management by utilizing advanced machine learning models and attention mechanisms.

[0043] An embodiment of the present invention is a property collection monitoring and early warning method based on artificial intelligence, which uses a cross entropy loss function to measure the difference between the predicted probability of a time-weighted attention mechanism and the actual risk category, including: Determine the weight of each category, wherein the weight of each category can be set based on the degree of sample imbalance, including: if category A, i.e., high-reputation customers, rarely have serious consequences of prediction errors, a lower weight is given, while category E, i.e., high-risk customers, are given a higher weight due to serious consequences; In each batch of model training, there are multiple samples. For each sample, the model outputs the predicted probability that it belongs to each category. These predicted probabilities and the actual category labels are used to calculate the loss function value through the weighted cross entropy loss function. Specifically, the cross entropy loss function is:

[0044] Among them, L represents the loss function, m represents the category index, from 1 to 5, corresponding to risk rating categories A to E, is the weight associated with category m, which is used to adjust the penalty for misclassification of different categories, including: misclassification of high-risk customers may bring greater costs, so this category can be given a higher weight, N is the total number of samples, is the binary representation of the true label. If sample i belongs to category m, then is 1, otherwise it is 0. is the probability that the model predicts that the sample belongs to category m; The calculated loss value is used to update the model weights through the back-propagation algorithm to minimize the loss function; Repeat the above process until the performance of the model on the validation set no longer improves.

[0045] The working principle and effect of the above technical solution are as follows: This solution is a deep learning model training process based on classification weights and sample rebalancing. It mainly optimizes the unbalanced data categories through weighted loss functions. The weights are set according to the actual business impact and sample distribution of each category. For example, for the high-risk category E (category with serious consequences), a higher weight is given so that the model will pay more attention to reducing the prediction error of this category; model training, in each training batch, the model predicts the probability of each sample belonging to each category; the weighted cross entropy loss function is used to calculate the error between the true category label and the model prediction. This loss function takes into account the category weight, thereby giving different categories different importance in the loss calculation; loss function optimization, the model weights are updated according to the loss function value through the back-propagation algorithm to optimize the model performance; the training process is repeated until the performance of the model on the validation set is no longer significantly improved to avoid overfitting. Improve the risk sensitivity of the model. By increasing the weight of high-risk categories, the model can more sensitively identify high-risk cases, which is particularly critical for risk prevention and mitigation of potential losses. Improve the category imbalance problem. In many practical applications, the categories of data are often unbalanced. Through this weighted strategy, the model can pay more attention to a few categories and improve the classification performance of the model on all categories. Optimize the model prediction accuracy. The weighted loss function helps the model focus on those incorrect predictions that have a great impact on the business, thereby optimizing the overall prediction accuracy, which is especially valuable in the field of key business decisions. Implement customized risk management strategies. Different risk levels can be personalized according to actual business needs and the severity of the consequences. The weight adjustment makes the model output more in line with the actual risk management needs of the enterprise. This technical solution improves the processing capabilities of unbalanced data by introducing category weights, and makes the model pay more attention to those categories that have a greater impact on the business, thereby providing an effective risk assessment tool that helps enterprises formulate more accurate and targeted risk management strategies.

[0046] One embodiment of the present invention is a property collection monitoring and early warning method based on artificial intelligence, which determines a collection strategy based on predicted risk probability, including: Input the user's current time series feature data into the model to obtain the predicted probability; For users with a predicted probability of A, obtain the user's historical payment method and payment status on the payment date. If the user has not paid on the payment date, make a reminder on the payment date based on the user's historical payment method. If the user's historical payment method is bank transfer, remind the user via SMS. If the user's historical payment method is online payment, remind the user via APP notification. If the user's historical payment method is cash, remind the user by calling and provide value-added services. For users with a predicted probability of B, obtain the user's historical payment method and payment status three days before the payment date. If the user has not paid three days before the payment date, make reminders three days before the payment date and on the payment date based on the user's historical payment method. For users with a predicted probability of C, obtain the user's historical payment method and payment status five days before the payment date. If the user has not paid five days before the payment date, make reminders five days, three days, and one day before the payment date, respectively, based on the user's historical payment method; For users with a predicted probability of D, obtain the user's historical payment method and payment status seven days before the payment date. If the user has not paid seven days before the payment date, make reminders seven days, five days, and three days before the payment date based on the user's historical payment method. If the user has not paid on the payment date, post a reminder at the user's door. For users with a predicted probability of E, the user's historical payment method and payment status ten days before the payment date are obtained. If the user has not paid ten days before the payment date, reminders will be made ten days, seven days, and five days before the payment date respectively based on the user's historical payment method. If the user has not paid three days before the payment date, a reminder will be posted at the user's door. If the user has not paid on the payment date, a message will be sent to the property staff that door-to-door collection is required.

[0047] One embodiment of the present invention is a property collection monitoring and early warning system based on artificial intelligence, such as Figure 2 As shown, the system comprises: A data preprocessing module is used to obtain user historical time series data, preprocess the user historical time data, and extract features from the preprocessed data; Establish a preliminary model module for building a temporal weighted attention mechanism model based on the features extracted from the partitioned training set; A loss function-based model adjustment module is used to measure the difference between the predicted probability of the time-weighted attention mechanism and the actual risk category using a cross-entropy loss function, and adjust the time-weighted attention mechanism model; A collection suggestion module is provided to obtain user time series feature data, input the user time series feature data into the adjusted time-weighted attention mechanism model, and adopt a collection strategy according to the model prediction results.

[0048] In one embodiment of the present invention, a property collection monitoring and early warning system based on artificial intelligence, the data preprocessing module includes: A module for obtaining user historical time series data is used to obtain user historical time series data, wherein the historical time series data includes payment records, property service requests and property fee adjustment records. The payment records include the date and amount of each payment, the payment method, whether there is overdue payment, the amount of overdue payment and the number of overdue days. The property service request includes the date and type of each service request and the date and result of service completion. The property fee adjustment record includes the date of adjustment, the amount of adjustment and the reason. A preprocessing module is used to preprocess the user's historical data, including: for missing payment records, using forward filling to process missing values; converting the payment date into a unified time format; standardizing the payment amount and the overdue amount to have the same scale; aligning the time series data of each user to ensure that all sequences have the same length; The data set partitioning module is used to divide the preprocessed data into a training set and a validation set; A payment feature extraction module is used to extract payment record features from the preprocessed payment records, wherein the payment record features include payment frequency, average payment amount, unique hot code of payment method, number of overdue payments, average overdue days, maximum days and overdue payment amount ratio; A property service request feature extraction module is used to extract property service request features from pre-processed property service request records, wherein the property service request features include the total number of service requests, a unique hot code of the service request type, a response time, and a service satisfaction score; Extracting property expense adjustment record features is used to extract property expense adjustment record features from preprocessed property expense adjustment records, wherein the property expense adjustment record features include the number of expense adjustments, the average adjustment amount, the expense increase ratio, and the unique hot coding of the adjustment reason.

[0049] In one embodiment of the present invention, a property debt collection monitoring and early warning system based on artificial intelligence, the module for establishing a preliminary model includes: The vector construction module is used to construct the feature vector x based on time from the features extracted from the training set t , the feature vectors constitute the time series feature data {x1,x2,x3,…,x T}, input the time series feature data into the LSTM layer; Generate hidden state vector module, the LSTM layer processes the time series feature data through the following formula to generate a hidden state vector for each time point t , the hidden state vector {h1,h2,h3,…,h T};

[0050] in, is the input feature vector at time t, It's the Gate of Forgetfulness. is the input gate, is the output gate, is a candidate memory cell, is the weight matrix of the forget gate, is the weight matrix of the input gate, is the weight matrix of the output gate, is the weight matrix of the candidate memory, σ is the Sigmoid function, whose output range is [0, 1], , , and They are the forget gate, input gate, output gate, and bias term of candidate memory; The time point weight calculation module is used to calculate the weight of each time point. The calculation formula is:

[0051] in, is the attention weight, Represents the hidden state vector at time point t, the hidden state of the current step, which is used to capture all historical information before this time point. Represents the hidden state vector at time point j, which is the hidden state of each historical time point j, used to compare and calculate with the state at the current time point t. Represents the risk rating vector at time point j, which contains risk-related features and enters the attention mechanism to help the model better understand the risk status at each time point. For conversion , making it compatible with other vectors, For conversion , making it compatible with other vectors, Used to transform the risk rating vector , allowing it to be efficiently incorporated into the model's computation, is the bias vector, is the weight vector of the attention mechanism; The hidden state weighting module is used to weight the hidden state using the attention weights to obtain the context vector c:

[0052] The input fully connected layer module is used to input the context vector c into the fully connected layer. The calculation formula of the fully connected layer is:

[0053] in, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer, and z is the linear transformation result without the activation function; The payment risk probability module is used to input the linear transformation result z into the softmax activation function to obtain the probability of the category to which the user payment risk belongs. The user payment risk is divided into 5 levels.

[0054] in, represents the probability that a given input x belongs to category k, represents the linear transformation result of category k, l represents the category index traversed by the current loop, and A, B, C, D and E represent different risk categories.

[0055] In one embodiment of the present invention, a property debt collection monitoring and early warning system based on artificial intelligence is provided, wherein the loss function-based model adjustment module comprises: A weight setting module is used to determine the weight of each category, and the weight of each category can be set based on the degree of sample imbalance, including: if category A, i.e., high-reputation customers, rarely have serious consequences of prediction errors, a lower weight is given, while category E, i.e., high-risk customers, are given a higher weight due to serious consequences; The loss calculation module is used for each batch of model training. There are multiple samples. For each sample, the model outputs the predicted probability that it belongs to each category. These predicted probabilities and the actual category labels are used to calculate the loss function value through the weighted cross entropy loss function. Specifically, the cross entropy loss function is:

[0056] Among them, L represents the loss function, m represents the category index, from 1 to 5, corresponding to risk rating categories A to E, is the weight associated with category m, which is used to adjust the penalty for misclassification of different categories, including: misclassification of high-risk customers may bring greater costs, so this category can be given a higher weight, N is the total number of samples, is the binary representation of the true label. If sample i belongs to category m, then is 1, otherwise it is 0. is the probability that the model predicts that the sample belongs to category m; Back propagation module, the calculated loss value is used to update the model weights through the back propagation algorithm to minimize the loss function; The iterative optimization module is used to repeat the above process until the performance of the model on the validation set no longer improves.

[0057] In one embodiment of the present invention, a property debt collection monitoring and early warning system based on artificial intelligence, the module for providing debt collection suggestions includes: The prediction probability acquisition module is used to input the user's current time series feature data into the model to obtain the prediction probability; The module with predicted probability A is used to obtain the user's historical payment method and payment status on the payment date for the user with predicted probability A. If the user has not paid on the payment date, a reminder will be made on the payment date based on the user's historical payment method. If the user's historical payment method is bank transfer, the user will be reminded via SMS. If the user's historical payment method is online payment, the user will be reminded via APP notification. If the user's historical payment method is cash, the user will be reminded by phone and value-added services will be provided. The module with predicted probability B is used to obtain the user's historical payment method and payment status three days before the payment date for the user with predicted probability B. If the user has not paid three days before the payment date, a reminder will be made three days before the payment date and on the payment date according to the user's historical payment method; The module with a predicted probability of C is used to obtain the user's historical payment method and payment status five days before the payment date for the user with a predicted probability of C. If the user has not paid five days before the payment date, a reminder will be given five days, three days and one day before the payment date respectively according to the user's historical payment method; The module with a predicted probability of D is used to obtain the user's historical payment method and payment status seven days before the payment date for the user with a predicted probability of D. If the user has not paid seven days before the payment date, reminders will be made seven days, five days and three days before the payment date respectively according to the user's historical payment method. If the user has not paid on the payment date, a warm reminder will be posted at the user's door; The module with predicted probability E is used to obtain the user's historical payment method and payment status ten days before the payment date for the user with predicted probability E. If the user has not paid ten days before the payment date, reminders will be made ten days, seven days and five days before the payment date respectively according to the user's historical payment method. If the user has not paid three days before the payment date, a reminder will be posted at the user's door. If the user has not paid on the payment date, a message will be sent to the property staff that door-to-door collection is required.

[0058] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A property collection monitoring and early warning method based on artificial intelligence, characterized in that: The method comprises: Acquire user historical time series data, preprocess the user historical time data, and extract features from the preprocessed data; A temporal weighted attention mechanism model is established based on the features extracted from the partitioned training set; Using a cross entropy loss function to measure the difference between the predicted probability of the time-weighted attention mechanism and the actual risk category, and adjusting the time-weighted attention mechanism model; Obtain user time series feature data, input the user time series feature data into the adjusted time-weighted attention mechanism model, and adopt a collection strategy based on the model prediction results.

2. According to the artificial intelligence-based property collection monitoring and early warning method of claim 1, it is characterized in that: Acquire user historical time series data, preprocess the user historical time series data, and extract features from the preprocessed data, including: Acquire historical time series data of the user, the historical time series data including: payment records, property service requests and property fee adjustment records, the payment records including: the date and amount of each payment, the payment method, whether there is overdue payment, the overdue amount and the number of overdue days, the property service request including: the date and type of each service request and the date and result of service completion; the property fee adjustment record including: the date of adjustment, the amount of adjustment and the reason; Preprocessing the user historical data includes: for missing payment records, using forward filling to process missing values; converting the payment date into a unified time format; standardizing the payment amount and the overdue amount to have the same scale; aligning the time series data of each user to ensure that all sequences have the same length; Divide the preprocessed data into training set and validation set; Extract payment record features from the pre-processed payment records, wherein the payment record features include payment frequency, average payment amount, unique hot code of payment method, number of overdue payments, average overdue days, maximum days, and overdue payment amount ratio; Extracting property service request features from the pre-processed property service request records, wherein the property service request features include the total number of service requests, a unique hot encoding of the service request type, a response time, and a service satisfaction score; The property fee adjustment record features are extracted from the pre-processed property fee adjustment records, wherein the property fee adjustment record features include the number of fee adjustments, the average adjustment amount, the fee increase ratio and the adjustment reason unique hot coding.

3. According to the artificial intelligence-based property collection monitoring and early warning method of claim 1, it is characterized in that: A temporal weighted attention mechanism model is established based on the features extracted from the divided training set, including: The features extracted from the training set are used to construct the feature vector x according to time. t , the feature vectors constitute the time series feature data {x1,x2,x3,…,x T }, input the time series feature data into the LSTM layer; The LSTM layer processes the time series feature data through the following formula to generate a hidden state vector for each time point t , the hidden state vector {h1,h2,h3,…,h T }; in, is the input feature vector at time t, It's the Gate of Forgetfulness. is the input gate, is the output gate, is a candidate memory cell, is the weight matrix of the forget gate, is the weight matrix of the input gate, is the weight matrix of the output gate, is the weight matrix of the candidate memory, σ is the Sigmoid function, whose output range is [0, 1], , , and They are the forget gate, input gate, output gate, and bias term of candidate memory; Calculate the weight of each time point, the calculation formula is: in, is the attention weight, Represents the hidden state vector at time point t, the hidden state of the current step, which is used to capture all historical information before this time point. Represents the hidden state vector at time point j, which is the hidden state of each historical time point j, used to compare and calculate with the state at the current time point t. Represents the risk rating vector at time point j, which contains risk-related features and enters the attention mechanism to help the model better understand the risk status at each time point. For conversion , making it compatible with other vectors, For conversion , making it compatible with other vectors, Used to transform the risk rating vector , allowing it to be efficiently incorporated into the model's computation, is the bias vector, is the weight vector of the attention mechanism; The hidden states are weighted summed using the attention weights to obtain the context vector c: The context vector c is input into the fully connected layer. The calculation formula of the fully connected layer is: in, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer, and z is the linear transformation result without the activation function; The linear transformation result z is input into the softmax activation function to obtain the probability of the category to which the user payment risk belongs. The user payment risk is divided into 5 levels. in, represents the probability that a given input x belongs to category k, represents the linear transformation result of category k, l represents the category index traversed by the current loop, and A, B, C, D and E represent different risk categories.

4. According to the artificial intelligence-based property collection monitoring and early warning method of claim 1, it is characterized in that: The cross entropy loss function is used to measure the difference between the predicted probability of the time-weighted attention mechanism and the actual risk category, including: Determine the weight of each category, wherein the weight of each category can be set based on the degree of sample imbalance, including: if category A, i.e., high-reputation customers, rarely have serious consequences of prediction errors, a lower weight is given, while category E, i.e., high-risk customers, are given a higher weight due to serious consequences; In each batch of model training, there are multiple samples. For each sample, the model outputs the predicted probability that it belongs to each category. These predicted probabilities and the actual category labels are used to calculate the loss function value through the weighted cross entropy loss function. Specifically, the cross entropy loss function is: Among them, L represents the loss function, m represents the category index, from 1 to 5, corresponding to risk rating categories A to E, is the weight associated with category m, which is used to adjust the penalty for misclassification of different categories, including: misclassification of high-risk customers may bring greater costs, so this category can be given a higher weight, N is the total number of samples, is the binary representation of the true label. If sample i belongs to category m, then is 1, otherwise it is 0. is the probability that the model predicts that the sample belongs to category m; The calculated loss value is used to update the model weights through the back-propagation algorithm to minimize the loss function; Repeat the above process until the performance of the model on the validation set no longer improves.

5. According to the artificial intelligence-based property collection monitoring and early warning method of claim 1, it is characterized in that: Determine the collection strategy based on the predicted risk probability, including: Input the user's current time series feature data into the model to obtain the predicted probability; For users with a predicted probability of A, obtain the user's historical payment method and payment status on the payment date. If the user has not paid on the payment date, make a reminder on the payment date based on the user's historical payment method. If the user's historical payment method is bank transfer, remind the user via SMS. If the user's historical payment method is online payment, remind the user via APP notification. If the user's historical payment method is cash, remind the user by calling and provide value-added services. For users with a predicted probability of B, obtain the user's historical payment method and payment status three days before the payment date. If the user has not paid three days before the payment date, make reminders three days before the payment date and on the payment date based on the user's historical payment method. For users with a predicted probability of C, obtain the user's historical payment method and payment status five days before the payment date. If the user has not paid five days before the payment date, make reminders five days, three days, and one day before the payment date, respectively, based on the user's historical payment method; For users with a predicted probability of D, obtain the user's historical payment method and payment status seven days before the payment date. If the user has not paid seven days before the payment date, make reminders seven days, five days, and three days before the payment date based on the user's historical payment method. If the user has not paid on the payment date, post a reminder at the user's door. For users with a predicted probability of E, the user's historical payment method and payment status ten days before the payment date are obtained. If the user has not paid ten days before the payment date, reminders will be made ten days, seven days, and five days before the payment date respectively based on the user's historical payment method. If the user has not paid three days before the payment date, a reminder will be posted at the user's door. If the user has not paid on the payment date, a message will be sent to the property staff that door-to-door collection is required.

6. A property collection monitoring and early warning system based on artificial intelligence, characterized in that: The system comprises: A data preprocessing module is used to obtain user historical time series data, preprocess the user historical time data, and extract features from the preprocessed data; Establish a preliminary model module for building a temporal weighted attention mechanism model based on the features extracted from the partitioned training set; A loss function-based model adjustment module is used to measure the difference between the predicted probability of the time-weighted attention mechanism and the actual risk category using a cross-entropy loss function, and adjust the time-weighted attention mechanism model; A collection suggestion module is provided to obtain user time series feature data, input the user time series feature data into the adjusted time-weighted attention mechanism model, and adopt a collection strategy according to the model prediction results.

7. According to claim 6, a property collection monitoring and early warning system based on artificial intelligence is characterized in that: The data preprocessing module comprises: A module for obtaining user historical time series data is used to obtain user historical time series data, wherein the historical time series data includes payment records, property service requests and property fee adjustment records. The payment records include the date and amount of each payment, the payment method, whether there is overdue payment, the amount of overdue payment and the number of overdue days. The property service request includes the date and type of each service request and the date and result of service completion. The property fee adjustment record includes the date of adjustment, the amount of adjustment and the reason. A preprocessing module is used to preprocess the user's historical data, including: for missing payment records, using forward filling to process missing values; converting the payment date into a unified time format; standardizing the payment amount and the overdue amount to have the same scale; aligning the time series data of each user to ensure that all sequences have the same length; The data set partitioning module is used to divide the preprocessed data into a training set and a validation set; A payment feature extraction module is used to extract payment record features from the preprocessed payment records, wherein the payment record features include payment frequency, average payment amount, unique hot code of payment method, number of overdue payments, average overdue days, maximum days and overdue payment amount ratio; A property service request feature extraction module is used to extract property service request features from pre-processed property service request records, wherein the property service request features include the total number of service requests, a unique hot code of the service request type, a response time, and a service satisfaction score; Extracting property expense adjustment record features is used to extract property expense adjustment record features from preprocessed property expense adjustment records, wherein the property expense adjustment record features include the number of expense adjustments, the average adjustment amount, the expense increase ratio, and the unique hot coding of the adjustment reason.

8. According to claim 6, a property collection monitoring and early warning system based on artificial intelligence is characterized in that: The preliminary model building module comprises: The vector construction module is used to construct the feature vector x based on time from the features extracted from the training set t , the feature vectors constitute the time series feature data {x1,x2,x3,…,x T }, input the time series feature data into the LSTM layer; Generate hidden state vector module, the LSTM layer processes the time series feature data through the following formula to generate a hidden state vector for each time point t , the hidden state vector {h1,h2,h3,…,h T }; in, is the input feature vector at time t, It's the Gate of Forgetfulness. is the input gate, is the output gate, is a candidate memory cell, is the weight matrix of the forget gate, is the weight matrix of the input gate, is the weight matrix of the output gate, is the weight matrix of the candidate memory, σ is the Sigmoid function, whose output range is [0, 1], , , and They are the forget gate, input gate, output gate, and bias term of candidate memory; The time point weight calculation module is used to calculate the weight of each time point. The calculation formula is: in, is the attention weight, Represents the hidden state vector at time point t, the hidden state of the current step, which is used to capture all historical information before this time point. Represents the hidden state vector at time point j, which is the hidden state of each historical time point j, used to compare and calculate with the state at the current time point t. Represents the risk rating vector at time point j, which contains risk-related features and enters the attention mechanism to help the model better understand the risk status at each time point. For conversion , making it compatible with other vectors, For conversion , making it compatible with other vectors, Used to transform the risk rating vector , allowing it to be efficiently incorporated into the model's computation, is the bias vector, is the weight vector of the attention mechanism; The hidden state weighting module is used to weight the hidden state using the attention weights to obtain the context vector c: The input fully connected layer module is used to input the context vector c into the fully connected layer. The calculation formula of the fully connected layer is: in, is the weight matrix of the fully connected layer, is the bias vector of the fully connected layer, and z is the linear transformation result without the activation function; The payment risk probability module is used to input the linear transformation result z into the softmax activation function to obtain the probability of the category to which the user payment risk belongs. The user payment risk is divided into 5 levels. in, represents the probability that a given input x belongs to category k, represents the linear transformation result of category k, l represents the category index traversed by the current loop, and A, B, C, D and E represent different risk categories.

9. According to claim 6, a property collection monitoring and early warning system based on artificial intelligence is characterized in that: The loss function-based model adjustment module includes: A weight setting module is used to determine the weight of each category, and the weight of each category can be set based on the degree of sample imbalance, including: if category A, i.e., high-reputation customers, rarely have serious consequences of prediction errors, a lower weight is given, while category E, i.e., high-risk customers, are given a higher weight due to serious consequences; The loss calculation module is used for each batch of model training. There are multiple samples. For each sample, the model outputs the predicted probability that it belongs to each category. These predicted probabilities and the actual category labels are used to calculate the loss function value through the weighted cross entropy loss function. Specifically, the cross entropy loss function is: Among them, L represents the loss function, m represents the category index, from 1 to 5, corresponding to risk rating categories A to E, is the weight associated with category m, which is used to adjust the penalty for misclassification of different categories, including: misclassification of high-risk customers may bring greater costs, so this category can be given a higher weight, N is the total number of samples, is the binary representation of the true label. If sample i belongs to category m, then is 1, otherwise it is 0. is the probability that the model predicts that the sample belongs to category m; Back propagation module, the calculated loss value is used to update the model weights through the back propagation algorithm to minimize the loss function; The iterative optimization module is used to repeat the above process until the performance of the model on the validation set no longer improves.

10. According to claim 6, a property collection monitoring and early warning system based on artificial intelligence is characterized in that: The module for providing collection suggestions includes: The prediction probability acquisition module is used to input the user's current time series feature data into the model to obtain the prediction probability; The module with predicted probability A is used to obtain the user's historical payment method and payment status on the payment date for the user with predicted probability A. If the user has not paid on the payment date, a reminder will be made on the payment date based on the user's historical payment method. If the user's historical payment method is bank transfer, the user will be reminded via SMS. If the user's historical payment method is online payment, the user will be reminded via APP notification. If the user's historical payment method is cash, the user will be reminded by phone and value-added services will be provided. The module with predicted probability B is used to obtain the user's historical payment method and payment status three days before the payment date for the user with predicted probability B. If the user has not paid three days before the payment date, a reminder will be made three days before the payment date and on the payment date according to the user's historical payment method; The module with a predicted probability of C is used to obtain the user's historical payment method and payment status five days before the payment date for the user with a predicted probability of C. If the user has not paid five days before the payment date, a reminder will be given five days, three days and one day before the payment date respectively according to the user's historical payment method; The module with a predicted probability of D is used to obtain the user's historical payment method and payment status seven days before the payment date for the user with a predicted probability of D. If the user has not paid seven days before the payment date, reminders will be made seven days, five days and three days before the payment date respectively according to the user's historical payment method. If the user has not paid on the payment date, a warm reminder will be posted at the user's door; The module with predicted probability E is used to obtain the user's historical payment method and payment status ten days before the payment date for the user with predicted probability E. If the user has not paid ten days before the payment date, reminders will be made ten days, seven days and five days before the payment date respectively according to the user's historical payment method. If the user has not paid three days before the payment date, a reminder will be posted at the user's door. If the user has not paid on the payment date, a message will be sent to the property staff that door-to-door collection is required.

Images