Search

CN-121998749-A - Enterprise cross-line fund prediction method based on multidimensional feature fusion and computer equipment

CN121998749ACN 121998749 ACN121998749 ACN 121998749ACN-121998749-A

Abstract

The invention relates to the technical field of financial science and technology and artificial intelligence, and discloses a multi-dimensional feature fusion-based enterprise cross-line fund prediction method and computer equipment, wherein the method firstly integrates enterprise data and screens high-value target guest groups based on transaction characteristics; the method comprises the steps of extracting a multi-dimensional feature set covering time sequence trend, transaction network topology and enterprise background, constructing a Panel mode joint prediction model based on TimeMixers architecture, realizing cross-enterprise feature transfer learning by sharing bottom LSTM-CNN joint encoder parameters, training the model by utilizing a staged strategy, and fusing classification and regression model output by a dynamic weighting algorithm to generate a final predicted value. The method effectively solves the problems of insufficient generalization capability and single characteristic caused by the limitation of data sparseness in the traditional single enterprise modeling, and improves the prediction precision and the model maintenance efficiency of the enterprise cross-line fund flow by mining the universal mode and the multi-source heterogeneous characteristic of the industry.

Inventors

  • WANG LI

Assignees

  • 中国建设银行股份有限公司江苏省分行

Dates

Publication Date
20260508
Application Date
20251223

Claims (10)

  1. 1. The enterprise cross-line fund prediction method based on multidimensional feature fusion is characterized by comprising the following steps of: S1, data acquisition and preprocessing, namely integrating account transaction data of enterprise clients and enterprise basic information data to establish a unified data view, and performing exploration, cleaning and format standardization processing on original data to generate an initial modeling data set; S2, carrying out client group hierarchical screening, namely receiving the initial modeling data set, dividing clients into different groups according to transaction characteristic indexes of enterprises, screening out high-stability value client groups and potential value client groups as target modeling objects, and eliminating non-value clients and low-value client groups; S3, carrying out multi-dimensional feature engineering construction, namely extracting a multi-dimensional feature set covering time sequence features, transaction behavior features and enterprise background features aiming at the target modeling object, wherein the multi-dimensional feature set covers time sequence rules, transaction modes and enterprise attribute three-dimensional information; S4, constructing a Panel mode joint prediction model, namely constructing a three-dimensional hybrid neural network based on TimeMixers architecture, fusing the time sequence characteristics, the transaction behavior characteristics and the enterprise background characteristics by the three-dimensional hybrid neural network, performing joint modeling on multi-client data by adopting a Panel mode, and sharing the neural network joint encoder parameters of the bottom layer by sample data of all target modeling objects in the Panel mode to realize feature transfer learning; And S5, carrying out model training and dynamic fusion prediction, namely optimizing the Panel mode joint prediction model by utilizing a staged training strategy, running a fund grade classification model and a fund amount regression model in parallel when the prediction is carried out, dynamically calculating and adjusting weights of output results of the two models according to the confidence coefficient of the fund grade classification model and the verification set error index of the fund amount regression model by a dynamic weighted fusion algorithm, and obtaining a cross-line fund change predicted value of a future time period of an enterprise by weighted calculation.
  2. 2. The method for predicting business cross-line funds based on multi-dimensional feature fusion according to claim 1, wherein in the step S1, the collected data fields comprise client numbers, transaction dates, transaction times, lender marks, transaction amounts, fund usage fields, basic features of the business, scale of business management assets and information of financial products held by the business; the method comprises the steps of cleaning original data, namely completing non-key fields with missing values in a default value filling or average value filling mode, eliminating records of key identification fields with missing values, and filtering logical error data with negative transaction amount.
  3. 3. The multi-dimensional feature fusion-based enterprise cross-line fund prediction method of claim 1, wherein in step S2, the trade characteristic index comprises a trade viscosity index, a trade activity index, and a trade stability index; The transaction viscosity index is determined based on the number of months in which transaction records exist in a statistical period; the transaction activity index is determined based on daily average transaction frequency of clients; The transaction stability indicator is determined based on a statistical variance of the historical transaction amount of the customer.
  4. 4. The multi-dimensional feature fusion-based enterprise cross-line fund prediction method of claim 3, wherein the step S2 further comprises: Defining the clients with the transaction month number less than a preset first threshold and unstable transaction state as the non-valued clients; Defining clients with unstable transaction states, the number of the transaction months of which is between the first threshold value and a preset second threshold value, as the low-value guest group; Defining customers with stable transaction states, the number of which is between a preset third threshold value and a preset fourth threshold value, as the high-stability value customer groups; Defining a customer whose number of transaction months is between the second threshold and the fourth threshold and whose transaction records last exist for a last preset month as the potential value guest group; And reserving the data of the high-stability value guest group and the potential value guest group as input data of the target modeling object.
  5. 5. The multi-dimensional feature fusion-based enterprise cross-line fund prediction method of claim 1, wherein in step S3, the extracting of the timing features comprises: And processing the historical fund flowing sequence by adopting a long-short-term memory network sliding window technology, setting the sliding window to move gradually on a time axis to intercept continuous time slice data, analyzing the fund change sequence in the window by a gating mechanism of a long-short-term memory network unit, and calculating and extracting dynamic indexes of the fund inflow slope and the fund outflow slope.
  6. 6. The multi-dimensional feature fusion-based enterprise cross-line funds prediction method according to claim 5, wherein in the step S3, the extraction of the transaction behavior features comprises: constructing a transaction time distribution matrix, wherein the transaction time distribution matrix is a two-dimensional matrix defined by time periods and transaction frequencies; constructing a fund flow network diagram, wherein nodes in the fund flow network diagram represent transaction accounts, and edges represent fund flow directions; Performing feature extraction on the transaction time distribution matrix and the fund flow network graph by using a convolutional neural network, and capturing local mode features and node centrality topological structure features by performing convolutional operation through a convolutional kernel; The extraction of the enterprise background features comprises the steps of obtaining discrete type classification data of the position of an industrial chain and the industry attribute of an enterprise, inputting the discrete type classification data into an embedded layer, and converting the classification features into low-dimensional dense real vectors.
  7. 7. The multi-dimensional feature fusion-based enterprise cross-line fund prediction method of claim 1, wherein in step S4, the TimeMixers architecture comprises a long-short-term memory network and convolutional neural network joint encoder; the joint encoder comprises parallel processing branches, wherein the processing branches are respectively used for processing fund flow time sequence data, transaction time distribution matrix, fund flow network diagram data and embedded vectors of enterprise background characteristics, which are generated by a sliding window; The parameter sharing mechanism in the Panel mode is configured such that the weight parameters and bias parameters in the joint encoder are shared between all the input client samples, and only parameters adapting to the personalized differences of different clients are reserved in the full connection layer.
  8. 8. The multi-dimensional feature fusion-based enterprise cross-line fund prediction method of claim 1, wherein in step S5, the staged training strategy comprises: the pre-training stage is to input a full-scale enterprise client data set into a model and update the shared layer parameters in the Panel mode joint prediction model in a full-scale manner; And a fine tuning stage, namely freezing the shared layer parameters, updating the full-connection layer parameters of the Panel mode joint prediction model only, and setting the learning rate of the fine tuning stage to be lower than that of the pre-training stage.
  9. 9. The multi-dimensional feature fusion-based enterprise cross-line fund prediction method of claim 1, wherein in step S5, the dynamic weighted fusion algorithm comprises: obtaining the maximum probability value output by the fund level classification model after Softmax function processing as the classification confidence coefficient; acquiring the average absolute error of the fund amount regression model on the verification set as a regression accuracy index; Calculating a dynamic weight coefficient based on the index function value of the classification confidence and the index function value related to the regression accuracy index, so that the dynamic weight coefficient increases when the classification confidence increases and decreases when the average absolute error decreases; And carrying out weighted summation on the prediction probability or the mapping grade value output by the fund grade classification model and the prediction specific monetary value output by the fund monetary regression model according to the dynamic weight coefficient, and generating a final cross-line fund change prediction result.
  10. 10. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the enterprise cross-line funds prediction method based on multi-dimensional feature fusion of any one of claims 1 to 9 when the computer program is executed by the processor.

Description

Enterprise cross-line fund prediction method based on multidimensional feature fusion and computer equipment Technical Field The invention relates to the technical fields of financial science and technology and artificial intelligence, in particular to an enterprise cross-line fund prediction method and computer equipment based on multidimensional feature fusion. Background At present, with the deep development of financial science and technology and the advancement of enterprise-level funds circulation strategy, the demands of commercial banks on fine management of public customer funds flow are becoming urgent. Particularly, aiming at entity enterprises such as large and medium-sized manufacturing industry, the fund turnover chain is long and high in frequency, and complex cross-bank interaction is involved, the cross-bank fund transfer-in and transfer-out conditions in future time periods of the enterprises are accurately pre-judged, and the method becomes a core link for banks to improve the in-vivo fund bearing capacity, strengthen the liability quality management and formulate accurate and stable storage-increasing strategies. For the above application scenario of fund liquidity prediction, the prior art generally adopts a statistical-based time series analysis or a basic machine learning algorithm for processing. In the implementation, the system mainly uses a single enterprise as an independent research object, utilizes algorithms such as an autoregressive moving average model (ARIMA) or a long and short term memory network (LSTM) and the like to perform fitting and extrapolation according to the linear change trend of the historical transaction amount of the enterprise, or adopts an integrated learning method such as XGBoost and the like to combine the static financial index and the historical flow statistic value of the enterprise to construct a univariate model so as to realize regression analysis or simple fluctuation classification of future fund states. However, the above-described prior art has limitations in facing a huge number of business guest groups, and the transaction behavior is complicated. On one hand, the mode of independent modeling of single client is adopted to cause the number of models to linearly increase along with the scale of users, not only is the redundancy of calculation power and storage resources caused, but also the model generalization capability is weak because parameter sharing and characteristic migration across a main body cannot be realized, so that general fund flow rules (such as supply chain cash back period) of industries are difficult to capture when facing a single enterprise with sparse transaction data. On the other hand, the existing method is often limited to single-dimension time sequence feature mining, and key behavior modes such as topological structures, node centrality and the like of enterprises in a fund interaction network are ignored, so that potential risk features which are at the edge of a transaction network although running water is difficult to effectively identify. In addition, a dynamic fusion mechanism aiming at qualitative fund grades and quantitative specific amounts is lacking, a single output mode is difficult to consider the accuracy of numerical value prediction while the risk early warning accuracy is ensured, and the actual requirements of fine fund management cannot be met. Therefore, the invention provides an enterprise cross-line fund prediction method and computer equipment based on multidimensional feature fusion, which solve the defects in the prior art. Disclosure of Invention Aiming at the defects of the prior art, the invention provides an enterprise cross-line fund prediction method and computer equipment based on multidimensional feature fusion, which solve the problems of weak model generalization capability, transaction network structure information loss and low prediction precision in a data sparse scene caused by single-body independent modeling and single feature dimension in the traditional enterprise fund prediction. In order to achieve the above purpose, the invention is realized by the following technical scheme: In a first aspect, the invention provides an enterprise cross-line fund prediction method based on multidimensional feature fusion, which comprises the following steps: S1, data acquisition and preprocessing, namely integrating account transaction data of enterprise clients and enterprise basic information data to establish a unified data view, and performing exploration, cleaning and format standardization processing on original data to generate an initial modeling data set; S2, carrying out client group hierarchical screening, namely receiving the initial modeling data set, dividing clients into different groups according to transaction characteristic indexes of enterprises, screening out high-stability value client groups and potential value client groups as target modeling objects, and elimin