Search

CN-121998752-A - Debt portrait construction method and system based on repayment willingness mining

CN121998752ACN 121998752 ACN121998752 ACN 121998752ACN-121998752-A

Abstract

The application provides a debt person portrait construction method and system based on repayment willingness mining, wherein the method comprises the steps of collecting multiple kinds of data of debt persons to form a fusion data set; the method comprises the steps of carrying out feature mining on the basis of a fusion data set to obtain a multidimensional repayment willing feature set, carrying out cross-domain alignment and causal screening on the multidimensional repayment willing feature set to form a causal feature library, carrying out weight calibration on features in the causal feature library on the basis of interference effect analysis, pertinently adjusting parameters of feature mining according to a weight calibration result, verifying the effectiveness of the causal feature library through anti-facts reasoning to obtain a core feature library, and constructing a debtor portrait model on the basis of the core feature library. The application can enable the portrait to draw the repayment related attributes of the debtor from static will, dynamic trend and correlation influence multidimensional, realize the real-time generation of the portrait and the accurate matching of the repayment strategy, and effectively improve the overall efficiency and repayment rate of the credit repayment service.

Inventors

  • SUN QING
  • SUN CHENGCHENG

Assignees

  • 数社(深圳)科技有限公司

Dates

Publication Date
20260508
Application Date
20260116

Claims (10)

  1. 1. A debt portrait construction method based on repayment willingness mining is characterized by comprising the following steps: Collecting multiple kinds of data of debtors, and mapping heterogeneous data formed by preprocessing the multiple kinds of data to a unified feature space to form a fusion data set; performing cross-domain alignment and causal screening on the multidimensional repayment willing feature set, and eliminating features which are not directly causally related with repayment behaviors to form a causal feature library; The method comprises the steps of taking feedback data of a collection service as increment input, carrying out weight calibration on features in a causal feature library based on interference effect analysis, pertinently adjusting parameters of feature mining according to weight calibration results, verifying the effectiveness of the causal feature library through inverse facts reasoning, and obtaining a core feature library; and constructing a debt portrait model based on the core feature library, carrying out light weight processing on the model, packaging the model into a standardized interface, and embedding the model into the existing collection promoting system to realize the real-time generation of the debt portrait and the accurate matching of collection promoting strategies.
  2. 2. The method of claim 1, wherein the feature mining includes semantic driven hidden state decoding of repayment intent, time-sequential driven trend modeling of repayment intent, associated person driven mining of repayment intent conduction effects.
  3. 3. The method of claim 2, wherein the multiple types of data include interactive semantic data characterizing a repayment intent semantic signal, time sequence associated data reflecting intent dynamic changes, associated person data reflecting associated influences, and business label data supporting portrait base judgment; The preprocessing comprises noise elimination and keyword extraction of interaction semantic data, time stamp ordering and three-dimensional sequence integration of time sequence associated data, label standardization of associated person data, and compliance desensitization and redundant data filtering of all data.
  4. 4. The method for constructing the human figure of debt according to claim 2, wherein the semantically driven hidden state decoding of the repayment intent comprises the steps of calculating an initial hidden state of the repayment intent and a state transition probability based on service tag data, extracting semantic tags and polarity scores from interactive semantic data, and adjusting the initial state transition probability by taking the semantic tags and the polarity scores as correction factors to obtain static repayment intent characteristics; Aligning the static repayment willingness hidden state with an execution time sequence of an induced charge strategy and an external event time sequence to form a three-dimensional data stream, segmenting and counting willingness state transition probability according to a preset key time interval, generating a repayment willingness trend curve, marking a high willingness window period, and obtaining dynamic repayment willingness characteristics; Calculating a conduction effect value based on a relation type, a performance state and a response behavior in the data of the associated person, and adjusting a static repayment willingness hidden state of the debtor according to the conduction effect value to supplement hidden association influence characteristics; And integrating the static repayment willingness features, the dynamic repayment willingness features and the hidden association influence features to form a multidimensional repayment willingness feature set containing static willingness, dynamic trend and association influence.
  5. 5. The method of claim 1, wherein a cross-domain aligned feature range is determined, the feature range including business-related features related to the construction of the liability profile in addition to the multi-dimensional repayment intent feature set; Mapping the multidimensional repayment willingness feature set and the business association features to the same hidden space, constructing a causal association graph between the features in the hidden space and repayment behaviors of debtors, and identifying direct causal relations between the features in the causal association graph and the repayment behaviors; and removing the pseudo-relevant features which are not directly causally related to the repayment behaviors in the causal relationship graph, reserving core features with the direct causal relationship, and integrating to form a causal feature library.
  6. 6. The method of claim 1, wherein determining the type of the feedback data of the collection service includes a payment achievement rate, a performance age, a second overdue condition, and a communication objection content after the collection policy is executed; Setting an intervention variable, wherein the intervention variable is a variable related to the execution of an induction strategy and comprises at least one of induction time selection, induction strategy type and related human intervention mode; adopting a sample matching algorithm to match samples of a control group with similar characteristics and without intervention for the sample of the debtor applying the intervention, and eliminating sample selection deviation to obtain sample data; And giving weight to the corresponding feature according to the magnitude of the intervention effect value, wherein the weight corresponding to the feature is larger as the intervention effect value is higher, so as to finish the weight calibration of the feature in the causal feature library.
  7. 7. The method for constructing a representation of a debt person according to claim 2, wherein the process of obtaining the core feature library specifically comprises: Extracting target features with weight fluctuation amplitude exceeding a preset threshold value from a weight calibration result, and carrying out parameter adjustment on feature mining links corresponding to the target features; Establishing a causal relation network between the features in the causal relation library and the repayment behaviors of the debtors, and determining the association direction of the features and the repayment behaviors; And comparing the deviation value with a preset validity threshold, reserving the valid features of which the deviation value exceeds the threshold, removing the invalid features of which the deviation value does not reach the threshold, and integrating the valid features to form a core feature library.
  8. 8. The debt portrait construction method according to claim 2, wherein parameter adjustment is performed for a feature mining link corresponding to the target feature, specifically including: If the target features are derived from time sequence driven repayment intent trend modeling, repartitioning a key time interval and adjusting the calculation weight of repayment intent state transition probability in the interval; if the target characteristic is derived from the repayment willingness conduction effect mining driven by the associated person, recalibrating the calculation coefficient of the conduction effect of the associated person; If the target features are derived from semantic-driven repayment willingness hidden state decoding, optimizing a matching rule of the semantic tag and adjusting the correction amplitude of the polarity score to the initial state transition probability.
  9. 9. The liability person portrait construction method according to claim 1 is characterized in that a model architecture adapting to the collection service scene is selected, a portrait model is constructed based on the model architecture, and the model architecture comprises a gradient lifting tree model, a random forest model, a single-layer perceptron and a shallow neural network.
  10. 10. A debt portrait construction system based on repayment willingness mining, comprising: The data acquisition unit is used for acquiring multiple types of data of debtors, and mapping heterogeneous data formed by preprocessing the multiple types of data to a unified feature space to form a fusion data set; The causal feature unit is used for carrying out feature mining based on the fusion data set to obtain a multidimensional repayment willing feature set, carrying out cross-domain alignment and causal screening on the multidimensional repayment willing feature set, and eliminating features which are not directly causally related with repayment behaviors to form a causal feature library; The weight calibration unit is used for taking feedback data of the collecting service as increment input, and carrying out weight calibration on the features in the causal feature library based on interference effect analysis; and the portrait construction unit is used for constructing a debt portrait model based on the core feature library, carrying out light weight processing on the model, packaging the model into a standardized interface, and embedding the model into the existing collection promoting system to realize the real-time generation of the debt portrait and the accurate matching of collection promoting strategies.

Description

Debt portrait construction method and system based on repayment willingness mining Technical Field The application relates to the field of debt portrait construction, in particular to a debt portrait construction method and system based on repayment willing mining. Background In the risk management and collection treatment links of financial credit business, debt portrait construction is a core technical support for realizing accurate collection and improving the withdrawal efficiency. The existing debt portrait construction technology takes basic business data (such as credit line, overdue duration, historical performance record and the like) of debts as a core, and portrait labels are generated through conventional feature extraction and model training to assist in the business decision-making. However, existing debt portrait construction techniques have significant shortcomings. For example, the mining of the repayment willingness of the liability person flows on the surface, the hidden characteristics related to the repayment willingness are not fully captured, so that the portraits cannot truly reflect the repayment subjective tendency of the liability person, and the portraits model is mainly constructed statically, dynamic iterative optimization is not carried out by combining the feedback data of the repayment urging business, and the dynamic characteristics of the repayment willingness of the liability person, which change along with time and external events, are difficult to adapt. Therefore, there is a need for a method for creating a representation of a debt person that can deeply mine repayment will to solve the problems in the prior art and improve the efficiency and effect of credit-induced payment services. Disclosure of Invention Based on the problems existing in the prior art, the application provides a debt portrait construction method and a system based on repayment willing mining. The specific scheme is as follows: A debt portrait construction method based on repayment willingness mining is characterized by comprising the following steps: Collecting multiple kinds of data of debtors, and mapping heterogeneous data formed by preprocessing the multiple kinds of data to a unified feature space to form a fusion data set; performing cross-domain alignment and causal screening on the multidimensional repayment willing feature set, and eliminating features which are not directly causally related with repayment behaviors to form a causal feature library; The method comprises the steps of taking feedback data of a collection service as increment input, carrying out weight calibration on features in a causal feature library based on interference effect analysis, pertinently adjusting parameters of feature mining according to weight calibration results, verifying the effectiveness of the causal feature library through inverse facts reasoning, and obtaining a core feature library; and constructing a debt portrait model based on the core feature library, carrying out light weight processing on the model, packaging the model into a standardized interface, and embedding the model into the existing collection promoting system to realize the real-time generation of the debt portrait and the accurate matching of collection promoting strategies. In some embodiments, the feature mining includes semantically driven payoff intent hidden state decoding, time-series driven payoff intent trend modeling, associated person driven payoff intent conduction effect mining. In some specific embodiments, the multiple types of data include interactive semantic data representing semantic signals of repayment willing, time sequence associated data reflecting dynamic changes of willing, associated person data reflecting associated influence, and business label data supporting basic judgment of portraits; The preprocessing comprises noise elimination and keyword extraction of interaction semantic data, time stamp ordering and three-dimensional sequence integration of time sequence associated data, label standardization of associated person data, and compliance desensitization and redundant data filtering of all data. In some specific embodiments, the semantic-driven hidden state decoding of the repayment intent comprises the steps of calculating an initial hidden state and state transition probability of the repayment intent based on service tag data, extracting semantic tags and polarity scores from interactive semantic data, and adjusting the initial state transition probability by taking the semantic tags and the polarity scores as correction factors to obtain static repayment intent characteristics; Aligning the static repayment willingness hidden state with an execution time sequence of an induced charge strategy and an external event time sequence to form a three-dimensional data stream, segmenting and counting willingness state transition probability according to a preset key time interval, generating a repayment wil