Machine learning decision modeling method for on-premise multi-source heterogeneous data

Abstract

The present application belongs to the technical field of machine learning and multi-source heterogeneous data processing, and specifically relates to an internally deployed multi-source heterogeneous data machine learning decision modeling method. The method acquires multi-source heterogeneous data and field semantics, dimensions, time granularity and associated objects, generates source domain feature descriptions, constructs a semantic anchor point set according to the stable corresponding relationship of cross-source features, and decomposes each source domain feature into a public semantic component and a source domain residual component. Conflict codes are generated for data with inconsistent values, states or time sequences under the same semantic anchor point, and semantic anchor point coverage information, source domain residual components, conflict codes and data source missing identifiers are input into a gated machine learning decision model to obtain a decision result. The present application retains cross-source commonality, source-specific information and inter-source contradictions, reduces decision deviation caused by field error merging and conflict data averaging, and enables the model to maintain stable determination and decision tracing under conditions of field changes, source missing and inconsistent records.

Description

Technical Field

[0001] This invention belongs to the field of machine learning and multi-source heterogeneous data processing technology, specifically a machine learning decision modeling method for internally deployed multi-source heterogeneous data. Background Technology

[0002] The existing internal deployment of multi-source heterogeneous data machine learning decision modeling typically begins with a data access program reading fields from relational tables, log records, file data, and existing business databases. Then, it merges the fields based on their names, data types, and manually maintained mapping tables. For numerical data, the program performs unit conversion, normalization, missing value imputation, and time window aggregation. For categorical data, it converts the data into discrete vectors using an encoding dictionary. For data with different time granularities, it typically performs upsampling or downsampling according to a preset statistical period. After preprocessing, the fields from each data source are concatenated into a unified feature vector, which is then input into classification, regression, or ranking models for training. Field mapping relationships mainly rely on name similarity, fixed encoding rules, and manual configuration. When multiple sources have the same or different names, different or the same names, or changes in statistical definitions, the program generally maintains data format consistency by modifying the mapping table, adding conversion rules, or deleting abnormal fields. The training set error is then used as the basis for updating model parameters.

[0003] Another existing approach uses a unified representation method to process multi-source data. This involves setting up separate feature encoding processes for each data source, then projecting the encoding results into a shared vector space. Distance constraints, correlation constraints, or joint loss are used to ensure similar samples receive similar representations. Encoding results from different sources are summed, concatenated, attention-weighted, or gated in the shared space, and the fused result is used to train a machine learning decision model. For cases of inconsistent values, mutually exclusive states, or misaligned recording times between sources, existing methods typically classify records with significant differences as outliers and reduce the impact of conflicting data on the training process through mean substitution, trusted source coverage, weight decay, or sample deletion. If a data source is temporarily missing, zero-value imputation, historical value filling, or retraining the model under the missing source conditions are used to maintain the output. While this approach can reduce data format differences, the shared representation process often focuses on compressing differences between sources, lacking independent preservation of source-specific information and cross-source conflict relationships.

[0004] When multiple source fields are mapped to a unified feature space, semantically consistent information, source-specific information, and contradictory information between sources are not differentiated, causing machine learning decision models to be unable to consistently identify the causes of input differences. Similar field names, units, or time granularities do not equate to consistent technical meaning; fixed mappings may incorrectly merge fields with different statistical definitions; shared vector compression may eliminate change patterns that only exist in specific sources; averaging, overwriting, or deleting conflicting records can also lose the difference structures that characterize state abrupt changes, record delays, and source anomalies. When data sources are missing, field definitions change, or contradictory records appear from multiple sources, changes in the unified feature vector are directly transmitted to the model's decision boundary. The model struggles to determine whether the change originates from the true state, source characteristics, or data conflicts, leading to problems such as decision output drift, inconsistent judgments of similar samples, and uncontrollable model behavior under conditions of missing sources. Summary of the Invention

[0005] The purpose of this invention is to provide an on-premises multi-source heterogeneous data machine learning decision modeling method that can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The internally deployed multi-source heterogeneous data machine learning decision modeling method includes: acquiring heterogeneous data from multiple data sources and their corresponding data types, field semantics, dimensions, time granularity, and related object descriptions, and generating source domain feature descriptions for each data source. Construct a set of semantic anchors based on the stable correspondence between cross-source features; Each source domain feature is decomposed into a common semantic component mapped to the set of semantic anchors and a source domain residual component that retains the source features. Generate conflict codes for data mapped to the same semantic anchor but with inconsistent values, states, or time sequences; The semantic anchor coverage information, source domain residual components, conflict codes, and data source missing identifiers are input into the gated machine learning decision model to obtain the decision modeling results.

[0007] Preferably, the construction of the semantic anchor set includes: forming a candidate anchor group based on the semantic similarity relationship of fields, and calculating the cross-source correspondence matrix in the candidate anchor group based on the overlapping relationship of value intervals, the unit conversion relationship, the time synchronization relationship and the consistency relationship of associated objects; Extract bidirectional correspondence edges that remain stable within multiple historical time windows from the cross-source correspondence matrix, and merge source domain features that share the same stable correspondence subgraph into semantic anchors; For source domain features that cannot form stable corresponding edges, retain independent anchor point identifiers, and record the data source set, feature source path and effective time range corresponding to each semantic anchor point.

[0008] Preferably, the generation of the common semantic component and the source domain residual component includes: setting shared projection parameters and source domain-specific projection parameters for each data source, and projecting the source domain features to the common semantic subspace and the source domain residual subspace respectively; Semantic conservation constraints are established based on the contribution direction of cross-source features to the target label under the same semantic anchor point, and residual separation constraints are established based on the correlation matrix between the common semantic components and the source domain residual components. The shared projection parameters and the source domain-specific projection parameters are jointly adjusted, and cross-source consistency information is written into the common semantic subspace. Source-specific changes that are not interpreted by semantic anchors are written into the corresponding source domain residual subspace.

[0009] Preferably, the construction of the conflict coding and the gated machine learning decision model includes: calculating the value difference, state repulsion relationship, time offset relationship and source reliability difference of cross-source data under the same semantic anchor point, and generating a conflict type vector, a conflict intensity vector and a conflict source combination identifier accordingly; Set up a common semantic decision branch, a source domain residual decision branch, and a conflict-sensitive decision branch; Branch weights are generated based on a combination of semantic anchor coverage, conflict intensity, residual energy, and missing data sources. The outputs of each decision branch are then weighted and fused to form the decision modeling result.

[0010] Preferably, the generation of the stable correspondence subgraph includes: constructing a cross-source semantic relationship graph using the source domain features as nodes and the candidate correspondence relationships as directed edges; Calculate the semantic loop closure consistency along the forward field mapping path and the reverse field restoration path respectively, and mark the candidate corresponding edges that do not meet the loop closure consistency condition as edges to be verified; The edge to be verified is updated by utilizing the changes in conditional mutual information, the co-occurrence stability of associated objects, and the reversibility of dimensional transformation in different historical time windows. Only corresponding edges that simultaneously satisfy the conditions of bidirectional closure consistency and cross-window stability are included in the stable correspondence subgraph, and the field mapping path, verification window, and correspondence source are saved for each corresponding edge.

[0011] Preferably, the establishment of the semantic conservation constraint and the residual separation constraint includes: constructing paired sample groups between cross-source samples corresponding to the same semantic anchor point, and calculating the gradient direction of the common semantic component with respect to the target label and the discrimination direction of the source domain residual component with respect to the source identifier, respectively; The common semantic components with opposite gradient directions are reprojected to the corresponding source domain residual subspace, and the anchor point center consistency constraint is applied to the cross-source common semantic components. Identify residual components that can characterize the source identifiers of other data sources, perform cross-source leakage suppression on the identified residual components, and limit the write path of each source domain residual component according to the data source attribution relationship.

[0012] Preferably, the generation of the conflict type vector includes: establishing conflict primitives according to numerical deviation, category mutual exclusion, time sequence reversal, and asynchronous source state; Match cross-source data corresponding to the same semantic anchor point with each conflict primitive to form anchor-level conflict nodes; Based on the dependencies between the semantic anchors, connect the anchor-level conflict nodes to obtain a conflict propagation graph; The conflict propagation graph is used to distinguish between secondary conflicts generated by upstream semantic anchors and original conflicts occurring at the current semantic anchor, and the conflict origin, propagation path, data source involved, and duration are encoded into the conflict type vector.

[0013] Preferably, the generation of the branch weights includes: constructing a source availability mask for different combinations of missing data sources, and determining, based on the source availability mask, effective anchors that still have cross-source support and degenerate anchors that only have single-source support from the semantic anchor set; Calculate the effective anchor coverage, the proportion of degraded anchor residuals, and the achievable conflict intensity in the conflict propagation diagram, respectively. The effective anchor coverage, the proportion of degraded anchor residuals, the achievable conflict intensity, and the source reliability sequence are input into the weight generation network, which outputs the normalized weights of the common semantic decision branch, the source domain residual decision branch, and the conflict-sensitive decision branch.

[0014] Preferably, the update of the semantic anchor set includes: when a new source domain feature is added, calculating the bidirectional semantic loop consistency, cross-window correspondence stability, and dimensional transformation invertibility between the new source domain feature and the existing semantic anchors respectively; New source domain features that simultaneously satisfy all correspondences are incorporated into the corresponding semantic anchors, and new source domain features that only form locally stable correspondence edges are established as candidate sub-anchors. When a stable corresponding subgraph within the same semantic anchor point splits into multiple disconnected subgraphs, the semantic anchor point is split according to each disconnected subgraph, and the inheritance mapping relationship between the anchor point identifier before splitting and each candidate sub-anchor point is saved.

[0015] Preferably, the training of the gated machine learning decision model includes: generating counterfactual training samples based on the original conflict nodes, secondary conflict nodes and missing data sources in the conflict propagation graph. The counterfactual training samples are formed by shielding the conflict origin data, truncating the conflict propagation path or removing the specified data source. The branch weights and decision outputs of the original samples and counterfactual training samples are obtained respectively. Joint training constraints are constructed based on the output consistency relationship of the unmasked semantic anchors, the response difference of the conflict-sensitive decision branches to the original conflict, and the output preservation relationship of the source domain residual decision branches corresponding to the degenerate anchors. The model parameters of the weight generation network and each decision branch are then adjusted.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a set of semantic anchor points jointly defined by field semantics, value range, dimension, time granularity, associated objects, and stable cross-window correspondences, cross-source features with the same technical meaning can be mapped to a common semantic space, while changes that cannot be explained by the common semantics are preserved in the corresponding source domain residual space. Semantic conservation constraints ensure that cross-source features under the same semantic anchor point maintain a consistent contribution direction to the target judgment, residual separation constraints restrict the re-mixing of source-specific components into the common semantic components, and bidirectional loop closure consistency and cross-window stability verification reduce erroneous merging caused by homonymous or heteronymous fields. Therefore, the input received by the machine learning decision model can simultaneously retain cross-source commonalities and source differences. When field definitions are adjusted, local source fluctuations occur, or a single source is missing, the common semantic components will not be directly replaced by source-specific changes, the non-target perturbations on the decision boundary are limited, and similar data can maintain a relatively consistent judgment structure under different source combinations.

[0017] 2. For data mapped to the same semantic anchor but with inconsistent values, states, or temporal order, conflict primitives, conflict propagation graphs, and mechanisms for distinguishing between primary and secondary conflicts are used to transform previously averaged, covered, or deleted discrepancies into conflict codes that can be used for modeling. The gated machine learning decision model then allocates weights to common semantic decision branches, source domain residual decision branches, and conflict-sensitive decision branches based on a combination of semantic anchor coverage, source domain residual proportion, reachable conflict intensity, source reliability, and data source missingness. Combined with counterfactual training samples formed by source masking, conflict origin truncation, and propagation path truncation, the model can maintain the output relationships of unaffected semantic anchors when the conflict source changes or some data sources are missing, and adjust the response of conflict-sensitive branches for primary conflicts. This reduces the likelihood of strong conflicts being normalized to ordinary feature differences, enabling decision results to distinguish between real-state changes, source-specific changes, and cross-source contradictions, enhancing the repeatability and traceability of model training, updating, and inference processes in internal deployment environments. Attached Figure Description

[0018] Figure 1 Overall flowchart for machine learning decision modeling of multi-source heterogeneous data deployed internally; Figure 2 This is a trend chart of cross-source semantic correspondence scores. Figure 3 This is a distribution map of semantic anchor point conflict intensity. Figure 4 This is a graph showing the changes in gating weights for the three types of decision branches; Figure 5 Flowchart for constructing semantic anchors and decomposing common semantic-source domain residuals; Figure 6 Flowchart for cross-source conflict propagation, gating decision-making, and counterfactual training; Figure 7 New source domain features are incorporated into the scoring trend chart. Detailed Implementation

[0019] In one embodiment, please refer to Figure 1 The internally deployed multi-source heterogeneous data machine learning decision modeling method includes: acquiring heterogeneous data from multiple data sources and their corresponding data types, field semantics, dimensions, time granularity, and related object descriptions; generating source domain feature descriptions for each data source; constructing a set of semantic anchors based on stable correspondences between cross-source features; decomposing each source domain feature into a common semantic component mapped to the set of semantic anchors and a source domain residual component retaining the source features; generating conflict codes for data mapped to the same semantic anchor but with inconsistent values, states, or time orders; inputting semantic anchor coverage information, source domain residual components, conflict codes, and data source missing identifiers into a gated machine learning decision model to obtain decision modeling results. In this embodiment, multi-source heterogeneous data processing is confined to the internally deployed data boundary. Original records, field descriptions, mapping relationships, training samples, model parameters, and decision results all flow within the data boundary. The external environment only receives authorized decision outputs and does not participate in semantic anchor generation, source domain residual learning, or conflict code calculation.

[0020] Specifically, multiple data sources can be data sets formed by internal databases, file storage, message records, or existing data services. The data access process does not directly concatenate field names, but instead establishes a description record with a version identifier for each source domain feature. The description record includes at least the field identifier, field definition, data type, allowed value range, dimensional expression, sampling or summarizing period, associated object identifier, source path, missing value expression, and effective time. The field definition is jointly parsed by internal metadata, data dictionary, and historical processing rules. The value range is characterized by the quantile position within the sliding time window, discrete category set, and anomaly truncation boundary. The associated object identifier is used to distinguish the data entities corresponding to fields with the same name. The source path is used to track the cleaning, conversion, and derivation processes that the field has undergone, so that subsequent machine learning modeling can jointly determine the cross-source correspondence based on semantics, numerical values, time series, and object relationships.

[0021] In this embodiment, the source domain feature description is encoded as a description vector with a fixed structure. The textual semantic components in the description vector are generated by an internally maintained terminology lexicon and field relationship graph. The numerical components are generated by the value distribution, missing value distribution, and time change statistics. The dimensional components are generated by the basic dimensional units and conversion paths. The temporal components are generated by the sampling period, aggregation period, and record delay distribution. The object components are generated by the associated object types and upstream and downstream association chains. During encoding, an independent position is reserved for each type of component to avoid directly replacing dimensional consistency or object consistency with textual similarity. If the field definitions are the same but the associated objects are different, the object components keep the corresponding relationships in a candidate state instead of merging them directly. If the field names are different but the dimensional, object, and historical change relationships are stable and consistent, the description vectors are allowed to enter the verification process of the same semantic anchor point.

[0022] The semantic correspondence score between cross-source feature i and cross-source feature j is calculated according to the following formula: ; in, Indicates semantic correspondence score, Indicates the similarity of field definitions. Indicates the degree of overlap in the value distribution. Indicates the consistency of dimensional conversion. Indicates time synchronization degree. Indicates the consistency of related objects. Represents the r-th coefficient. to The combination coefficients are determined on the internal training set and remain unchanged during a single anchor point construction cycle, for example, taking , , , , and take , , , , At that time, The score is used to verify stable correspondences rather than directly as a merging conclusion, thus avoiding erroneous merging caused by similar single field names.

[0023] Please refer to Figure 2 , Figure 2 The horizontal axis represents the sampling window number, ranging from 1 to 500, and the vertical axis represents the semantic correspondence score between cross-source features. The solid line in the legend to the right of the curve corresponds to the... The resulting scoring sequence uses dashed lines and dotted lines to represent the first and second warning thresholds, respectively. This scoring sequence incorporates the similarity of the defined fields. Value distribution overlap Consistency of dimensional conversion Time synchronization and consistency of associated objects Therefore, an increase in the curve indicates that the candidate features are simultaneously converging in five aspects: semantics, numerical value, dimensionality, temporal sequence, and object relationship.

[0024] from Figure 2 It can be seen that most sampling windows Above the first warning threshold, the candidate cross-source correspondence has the foundation to enter the stable verification region. Around the 20th, 190th, and 380th windows, the curve approaches or exceeds the second warning threshold, indicating that the field meaning, time synchronization, and object constraints are all relatively strong within these intervals, making them suitable as priority candidates for stable correspondence edges. The curve declines around the 80th, 330th, and 470th windows, reflecting potential changes in dimensional conversion paths, differences in object identification, or a decrease in the overlap of value distributions within local windows.

[0025] The semantic anchor set consists of cross-source feature nodes and stable corresponding edges. The processing program repeatedly calculates candidate cross-source relationships within multiple historical time windows and checks whether the candidate relationships maintain directional consistency under conditions of field version changes, sample quantity changes, and time segmentation changes. Cross-source features that meet the stability condition are included in the same anchor candidate group. Within the candidate group, bidirectional mapping verification and object consistency verification are performed. After the verification passes, a semantic anchor identifier is generated. The semantic anchor stores the corresponding data source set, source domain feature set, common definition, dimensional benchmark, time benchmark, object constraint, and effective time range. Source domain features that cannot be stably mapped retain independent anchor identifiers and are not merged into the common space through forced normalization.

[0026] When each source domain feature is decomposed into a common semantic component mapped to the set of semantic anchors and a source domain residual component that retains the source features, each data source is configured with a common projection path and a dedicated projection path. The common projection path shares the parameter structure used to express the same semantic anchor point, while the dedicated projection path only receives the input from the corresponding data source and outputs the source domain residual. The original source domain features are entered into the two projection paths after missing identifier splicing, dimensional benchmark conversion, and time benchmark resampling. The common semantic component is used to express cross-source stable commonalities, while the source domain residual component is used to preserve local changes that the common semantic component cannot reconstruct. The reconstruction process jointly restores the two types of components to the source domain feature space and compares them with the input data, thereby limiting the simultaneous loss of input information by the common semantic component and the source domain residual component.

[0027] The joint training objective of the common semantic components and the source domain residual components is constructed according to the following formula: ; in, This represents the common semantic components of the nth sample. The center vector representing the semantic anchor point to which the sample belongs. This represents a set of cross-source sample pairs mapped to the same semantic anchor. and Let P represent the gradient direction of the common semantic component with respect to the target label, and let P and Z represent the intra-batch common semantic matrix and the source domain residual matrix, respectively. This represents the nth original input source domain feature. This represents the source domain features reconstructed from two types of components. Indicates the joint training target value. to The constraint coefficient represents a constraint that maintains a fixed meaning. This indicates transpose; for example, when the gradient cosine of two sources with respect to the same label is negative, the second term increases and drives the readjustment of the common semantic components. As the matrix norm increases, the third term restricts the common semantics and the source domain residuals, resulting in information overlap.

[0028] When generating conflict codes for data mapped to the same semantic anchor but with inconsistent values, states, or time sequences, the processing program does not delete conflicting records, nor does it cover other sources with a single source. Instead, it forms alignment groups for records under the same anchor based on object identifiers and comparable time intervals. Within each alignment group, it calculates the numerical deviation after dimensional conversion, the category-state incompatibility, the event sequence, the record delay difference, and the source reliability difference. The numerical deviation is normalized using the anchor's historical scale, the category incompatibility is determined based on the internal state relationship table, the time sequence conflict is determined based on the dual time axis of event time and processing time, and the source reliability is formed by historical missing, abnormal, delayed, and feedback consistent records. All conflicting components retain the source combination and corresponding record index.

[0029] semantic anchor The intensity of the conflict is calculated according to the following formula: ; in, Indicates the intensity of the conflict. This indicates deviation of the normalized value. Indicates the degree of repulsion between states. This indicates a deviation in the chronological order. Indicates differences in source reliability. This represents the s-th coefficient. to For combination coefficients of conflict types, such as when the values ​​at the same anchor point are small but the states are mutually exclusive and the time sequence is reversed, a value can be taken as... , , , The intensity of the conflict is jointly dominated by state repulsion and time deviation, and the processing program retains the conflict type instead of classifying the record as ordinary numerical noise.

[0030] Please refer to Figure 3 The horizontal axis represents the sampling window number, ranging from 1 to 500, and the vertical axis represents the conflict intensity of the semantic anchor points. The solid line in the right-hand legend represents the... The calculated anchor-level conflict intensity sequence, with the long dashed line and dotted line corresponding to two levels of warning thresholds. This indicator is derived from the deviation of normalized values. Degree of state repulsion Time sequence deviation Differences in source reliability The curve is formed collectively, so it is not a display of a single outlier, but a comprehensive expression of the degree of inconsistency between records from multiple sources within the same semantic anchor point.

[0031] Figure 3 middle, The data remained mostly within the medium range, indicating that even after dimensional conversion and object alignment, there were still periodic differences in state and time offsets in the cross-source data, but overall, a sustained high-intensity conflict had not yet formed. Within several local windows, the curve approached the first warning threshold, indicating that state incompatibility needed to be checked at this stage. Deviation from chronological order Whether they increase together. If the curve only shows small oscillations, it can be understood as a deviation from ordinary values. Or differences in source reliability The short-term changes; if the continuous window crosses a higher threshold, it indicates that the anchor point should retain the original conflict type and enter the conflict propagation analysis.

[0032] When semantic anchor coverage information, source domain residual components, conflict codes, and data source missing identifiers are input into a gated machine learning decision model, the semantic anchor coverage information describes the number of common anchors, the number of cross-source supports, and the anchor completeness in the current sample. The source domain residual components are indexed by source and anchor. The conflict codes are organized by the original conflict location and propagation location. The data source missing identifier is expressed by the source availability mask. The gated machine learning decision model includes a common semantic decision branch, a source domain residual decision branch, and a conflict-sensitive decision branch. The three branches use the same target label definition but receive different data representations. The common semantic decision branch receives common semantic components, the source domain residual decision branch receives residual components summarized by source, and the conflict-sensitive decision branch receives conflict codes and common semantic components of related anchors.

[0033] The gating weights for each decision branch are obtained using the following formula: ; Where k represents the common semantic decision branch, the source domain residual decision branch, and the conflict-sensitive decision branch, respectively, and h represents the gated input formed by concatenating semantic anchor coverage, residual energy, conflict intensity, source reliability sequence, and source availability mask. This represents the original confidence score for the k-th branch. Indicates the first Branch original confidence score, and This represents the gating parameters corresponding to the k branches. This indicates that the normalized weights sum to 1, for example, when the common anchor point coverage is complete, the residual energy is low, and the conflict intensity is close to 0, the common semantic decision branch corresponds to... If a larger value is obtained, and the original conflict persists, then the conflict-sensitive decision branch corresponds to... As the conflict input changes, the final decision modeling result is output from three branches according to... Weighted synthesis.

[0034] Please refer to Figure 4 The horizontal axis represents the sampling window number, ranging from 1 to 500, and the vertical axis represents the normalized gating weights of the three types of decision branches. Figure 4 The solid line represents the common semantic decision branch, the dashed line with a circular marker represents the source domain residual decision branch, and the dotted line with a square marker represents the conflict-sensitive decision branch; the three curves correspond to the following within any sampling window: The sum remains 1. The gating weights are determined by... The normalized output is obtained, and It also varies with the semantic anchor coverage, residual energy, conflict intensity, source reliability sequence, and source availability mask in the gated input h.

[0035] from Figure 4 It can be seen that within windows with sufficient anchor coverage and weak conflict, the weight of the common semantic decision branch remains relatively high, indicating that the model mainly relies on cross-source stable commonalities to generate decision results. When a local window experiences source loss, increased residual energy, or increased conflict intensity, the weights of the source domain residual branch and conflict-sensitive branch rise accordingly, while the weight of the common semantic branch decreases. This change reflects... Immediate response to input states: Instead of setting a fixed ratio for the three branches, dynamic allocation is performed based on the available sources of the current samples, degradation anchors, and conflict propagation states.

[0036] In a preferred embodiment, please refer to Figure 5The construction of the semantic anchor set includes: forming candidate anchor groups based on field semantic similarity relationships; calculating cross-source correspondence matrices within the candidate anchor groups based on overlapping value intervals, dimensional conversion relationships, time synchronization relationships, and consistency relationships of associated objects; extracting bidirectional correspondence edges that remain stable within multiple historical time windows from the cross-source correspondence matrix, and merging source domain features that share the same stable correspondence subgraph into semantic anchors; retaining independent anchor identifiers for source domain features that cannot form stable correspondence edges, and simultaneously recording the data source set, feature source path, and effective time range corresponding to each semantic anchor; after the candidate anchor group is formed, the existing anchor structure is not immediately changed, but candidate relationships are saved in an isolated anchor verification area; only candidate relationships that have completed cross-window verification and bidirectional mapping verification are written into the currently effective anchor version.

[0037] In this embodiment, the rows and columns of the cross-source correspondence matrix correspond to source domain features in different data sources. The matrix elements store semantic correspondence scores, dimensional conversion paths, time alignment offsets, object consistency identifiers, and verification status. If a source domain feature has high semantic correspondence scores with multiple candidate features, a one-to-one relationship or a one-to-many derivation relationship is retained based on object constraints and bidirectional mapping results. The derivation relationship requires that the merged field can be reconstructed from the source field according to the registered aggregation rules. Multiple candidate features that cannot be reconstructed do not enter the same semantic anchor point. The stable correspondence subgraph adopts connectivity and path consistency constraints to avoid indirectly merging fields with excessively large semantic distances through transitive relationships.

[0038] Preferably, the generation of the stable correspondence subgraph includes: constructing a cross-source semantic relationship graph with source domain features as nodes and candidate correspondences as directed edges; calculating semantic loop closure consistency along the forward field mapping path and the reverse field restoration path respectively, and marking candidate correspondence edges that do not meet the loop closure consistency condition as edges to be verified; updating the edges to be verified using the changes in conditional mutual information, the co-occurrence stability of associated objects, and the reversibility of dimensional transformation in different historical time windows; including only the correspondence edges that simultaneously meet the bidirectional loop closure consistency and cross-window stability conditions into the stable correspondence subgraph, and saving the field mapping path, verification window, and correspondence source for each correspondence edge. The forward field mapping path records the transformation sequence of the source domain value after conversion, aggregation, or encoding to reach the target field, and the reverse field restoration path records the transformation sequence of the target field restored to the source domain expression within the allowed information granularity.

[0039] The semantic loop closure consistency between cross-source feature i and cross-source feature j is calculated according to the following formula: ; in, Indicates semantic loop consistency. This represents the positive field mapping from feature i to feature j. This indicates that feature j returns the inverse field reconstruction of feature i. This represents the robustness of feature i within the validation window. This represents the original input feature vector. This indicates a fixed small quantity that prevents the denominator from being zero. This indicates the result of a joint verification of the object and its dimensions, with values ​​between 0 and 1. For example, it represents a value obtained by forward conversion and then reverse conversion that is close to the original value and has the same object and dimensions. If the names are similar but the objects are different, then... Set the value to 0 and prevent the corresponding edge from entering the stable corresponding subgraph.

[0040] In a preferred embodiment, multiple historical time windows are composed of mutually staggered fixed-length intervals and boundary intervals containing field version change points. Within each window, the processor calculates the conditional mutual information, object co-occurrence frequency, and loop closure consistency of candidate corresponding edges, and calculates the directional consistency of results between windows. If a candidate edge appears only in a single window and disappears in adjacent windows, it is retained as an edge to be verified. If a candidate edge maintains the same mapping direction before and after a field version change but the dimension conversion path changes, a new valid time period is established for the same semantic anchor point and the old time period is retained. If the mapping direction and object relationship change simultaneously, the old corresponding edge is terminated and a new candidate anchor point relationship is generated.

[0041] In a preferred embodiment, the generation of the common semantic component and the source domain residual component includes: setting shared projection parameters and source domain-specific projection parameters for each data source, projecting source domain features to the common semantic subspace and the source domain residual subspace respectively; establishing semantic conservation constraints based on the contribution direction of cross-source features to the target label under the same semantic anchor point, and establishing residual separation constraints based on the correlation matrix between the common semantic component and the source domain residual component; jointly adjusting the shared projection parameters and the source domain-specific projection parameters, writing cross-source consistency information into the common semantic subspace, and writing source-specific changes not explained by the semantic anchor point into the corresponding source domain residual subspace. The shared projection parameters are organized according to the semantic anchor point rather than the data source, and the source domain-specific projection parameters are organized according to the cross-index of the data source and the anchor point, allowing different anchor points from the same source to use different residual expressions.

[0042] Specifically, the public projection path adopts a shared target dimension and anchor center for inputs from different sources under the same semantic anchor point, but retains source adaptation transformation to handle differences in expression after data type, dimensional conversion, and temporal resampling. The source adaptation transformation only completes reversible scaling transformation and missing mask combination, and does not participate in the target decision output. The dedicated projection path receives the reconstruction error, original source domain features, and source identifiers of the public projection path, and generates source domain residual components using a restricted coding process. The restricted coding process prohibits reading source identifiers from other sources to avoid dedicated residuals learning public semantic content through cross-source identifiers.

[0043] Preferably, the establishment of the semantic conservation constraint and residual separation constraint includes: constructing paired sample groups between cross-source samples corresponding to the same semantic anchor point, calculating the gradient direction of the common semantic component with respect to the target label and the discrimination direction of the source domain residual component with respect to the source identifier respectively; reprojecting the common semantic components with opposite gradient directions to the corresponding source domain residual subspace, and applying the anchor point center consistency constraint to the cross-source common semantic components; identifying residual components that can characterize the source identifiers of other data sources, performing cross-source leakage suppression on the identified residual components, and limiting the writing path of each source domain residual component according to the data source attribution relationship; the paired sample groups are matched together according to the associated object, comparable time interval and target label state; when there is no real label, the historical decision state that has passed internal verification is used to form a weak pairing identifier, and the weak pairing identifier does not participate in the calculation of the source residual discrimination direction.

[0044] Furthermore, the determination of opposite gradient directions is not based on the results of a single batch, but rather on the cumulative vector of the cross-source gradient direction of the same anchor point in consecutive training batches. When the cumulative vectors continuously form a negative angle, the processing program extracts the feature dimension that causes the reverse contribution and redistributes it to the source domain residual subspace. After redistribution, a local reconstruction check is performed. Only when the common semantic reconstruction error does not exceed the historical scale of the current anchor point and the source residual can recover the assigned dimension is it written into the model version. Cross-source leakage suppression checks whether the residual of a source can identify other source identifiers through the source discriminator. If an identifiable component appears, the reverse gradient is applied to the corresponding residual dimension and its writing path is restricted.

[0045] In a preferred embodiment, please refer to Figure 6The construction of the conflict encoding and gated machine learning decision model includes: calculating the value difference, state repulsion relationship, time offset relationship and source reliability difference of cross-source data under the same semantic anchor point, and generating conflict type vector, conflict intensity vector and conflict source combination identifier accordingly; setting up common semantic decision branch, source domain residual decision branch and conflict sensitive decision branch; generating branch weights based on semantic anchor coverage, conflict intensity, residual energy and data source missing combination, and weighting and fusing the outputs of each decision branch to form the decision modeling result. The conflict source combination identifier is jointly encoded by ordered source pairs and anchor identifiers to distinguish the formation path of the same conflict type in different source combinations.

[0046] Specifically, source reliability does not set a permanent fixed score for the data source. Instead, it establishes locally reliable records for the data source, semantic anchors, and effective time periods. The locally reliable records are updated jointly by field integrity, time delay stability, historical anomaly density, records consistent with the confirmed target state, and mapping path reversibility. During the update, a limited increment is used for single anomalies to avoid short-term fluctuations that cause sudden changes in source reliability. Conflict coding retains the common semantic components, source domain residual components, and source reliable records before and after the conflict, enabling conflict-sensitive decision branches to distinguish source-specific offsets, record delays, and state incompatibilities.

[0047] Preferably, the generation of the conflict type vector includes: establishing conflict primitives according to numerical deviation, category mutual exclusion, time sequence reversal, and source state asynchrony; matching cross-source data corresponding to the same semantic anchor with each conflict primitive to form anchor-level conflict nodes; connecting the anchor-level conflict nodes according to the dependency relationship between semantic anchors to obtain a conflict propagation graph; distinguishing between secondary conflicts formed by upstream semantic anchors and original conflicts occurring at the current semantic anchor along the conflict propagation graph, and encoding the conflict starting point, propagation path, involved data sources, and duration into the conflict type vector. The dependency relationship between semantic anchors is jointly determined by the stable temporal sequence, derived field relationship, and target contribution relationship in historical data. Anchor relationships that are simply related but lack a stable direction are not written into the conflict propagation graph.

[0048] In this embodiment, the determination of primary conflicts is based on whether the source domain record corresponding to the current anchor point has already become incompatible when no conflict has occurred at the upstream anchor point. The determination of secondary conflicts is based on whether the conflict occurred later than the upstream conflict and was propagated along the registered dependency edge. If the same anchor point receives multiple upstream conflicts at the same time, an independent path identifier and path cumulative strength are saved for each propagation path. The conflict-sensitive decision branch takes the primary conflict node as the main input and the secondary conflict path as the context input to avoid multiple downstream conflicts being repeatedly interpreted as multiple independent sources of anomalies.

[0049] In a preferred embodiment, the generation of the branch weights includes: constructing a source availability mask for different combinations of missing data sources, and determining, based on the source availability mask, effective anchors that still have cross-source support and degenerate anchors that only have single-source support from the semantic anchor set; calculating the effective anchor coverage, the proportion of degenerate anchor residuals, and the reachable conflict intensity in the conflict propagation graph; inputting the effective anchor coverage, the proportion of degenerate anchor residuals, the reachable conflict intensity, and the source reliability sequence into the weight generation network, and outputting the normalized weights of the common semantic decision branch, the source domain residual decision branch, and the conflict-sensitive decision branch. The source availability mask is generated according to the sample time point rather than according to the entire dataset, so that the availability status of the same data source in different time intervals can participate in the gating calculation separately.

[0050] The residual proportion of degenerate anchors is determined according to the ratio between the source domain residual energy corresponding to the degenerate anchor and its common semantic reconstruction energy. If a degenerate anchor supported by a single source can still be reconstructed by historically stable common semantic components, the common semantic decision branch retains its input. If the residual energy continues to occupy the majority, the source domain residual decision branch receives the corresponding input. During training, the weight generation network receives complete source combinations, random missing combinations, and historical true missing combinations. Complete source combinations are used to learn the baseline allocation under common semantic conditions, random missing combinations are used to form source substitution relationships, and historical true missing combinations are used to correct the distribution difference between random masking and actual missing data.

[0051] In a preferred embodiment, the update of the semantic anchor set includes: when a new source domain feature is added, calculating the bidirectional semantic loop consistency, cross-window correspondence stability, and dimensional transformation invertibility between the new source domain feature and the existing semantic anchors respectively; incorporating the new source domain feature that simultaneously satisfies all correspondence relationships into the corresponding semantic anchor, and establishing the new source domain feature that only forms a locally stable correspondence edge as a candidate sub-anchor; when the stable correspondence subgraph within the same semantic anchor is split into multiple disconnected subgraphs, splitting the semantic anchor according to each disconnected subgraph, and saving the inheritance mapping relationship between the anchor identifier before splitting and each candidate sub-anchor, the update process is completed in the new anchor version, and the old model continues to read the old anchor version until the new anchor version completes training and consistency verification.

[0052] New source domain features Incorporate into existing semantic anchors The updated score is calculated using the following formula: ; in, This represents the semantic correspondence score between the newly added source domain features and the semantic anchor points. Indicates the consistency of two-way semantic loopbacks. This indicates the stability across windows. Indicates the invertibility of dimensional transformation. This indicates an update to the rating value. to This indicates an update to the rating combination coefficients, for example, when a new field has a similar definition to the anchor point but only forms a stable correspondence in certain windows. and It is possible to obtain higher values. When a lower value is obtained, the processing program creates candidate sub-anchor points instead of directly merging them into existing anchor points, and recalculates them after subsequent window updates. .

[0053] Please refer to Figure 7 The horizontal axis represents the sampling window number, ranging from 1 to 500, and the vertical axis represents the updated score where newly added source domain features are incorporated into existing semantic anchors. The solid line in the right-hand legend represents the change from... The calculated merging score is represented by the long dashed line and the dotted line, which indicate the candidate observation threshold and the stable merging threshold, respectively. This score also considers the semantic correspondence between the newly added features and the anchor points. Bidirectional semantic loop consistency Cross-window corresponding stability And the degree of invertibility of dimensional transformation.

[0054] Figure 7 show, The fact that the curve remains near the candidate observation threshold for most windows and gradually approaches the stable threshold indicates a continuously strengthening correspondence between the newly added source domain features and existing semantic anchors; when the curve briefly falls back, it usually signifies cross-window stability. Not yet fully accumulated, or the degree of reversibility of dimensional transformation. The local window is affected by changes in field version, aggregation period, or conversion path. New features are only suitable for writing to the new anchor version when the curve is consistently above a higher threshold within a continuous window; if a high value is only reached in a local window, it should be reserved as a candidate sub-anchor and await subsequent data supplementation.

[0055] In a preferred embodiment, anchor point splitting is based on the connected components of the stable corresponding subgraph, and the distance between the common semantic centers, the target contribution direction, and the effective time range of each connected component are checked. If the split is only caused by a short-term missing edge, the original anchor point is retained and the missing edge is marked as paused. If multiple connected components remain separated within a continuous window and have different target contribution directions, multiple candidate sub-anchor points are generated. The mapping relationship is inherited to record the projection relationship from the old anchor point center vector to the center vector of each sub-anchor point, the historical sample affiliation, and the model parameter migration path. During historical decision playback, the old anchor point or sub-anchor point is selected according to the effective time period of the sample.

[0056] In a preferred embodiment, the training of the gated machine learning decision model includes: generating counterfactual training samples based on the original conflict nodes, secondary conflict nodes, and missing data source combinations in the conflict propagation graph. The counterfactual training samples are formed by masking conflict origin data, truncating conflict propagation paths, or removing specified data sources. The branch weights and decision outputs of the original samples and the counterfactual training samples are obtained respectively. Joint training constraints are constructed based on the output consistency relationship of unmasked semantic anchors, the response differences of conflict-sensitive decision branches to original conflicts, and the output preservation relationship of source domain residual decision branches corresponding to degenerate anchors. The model parameters of the weight generation network and each decision branch are adjusted. The counterfactual training samples and the original samples share associated objects and target labels, and only the conflict origin, propagation path, or source availability status is changed to prevent the sample construction from introducing semantic changes unrelated to the original decision objective.

[0057] The joint training constraints are constructed according to the following formula: ; Where y and These represent the decision outputs of the original sample and the counterfactual training sample, respectively. This represents the joint training constraint, and M represents the output mask corresponding to the unmasked semantic anchor. This represents element-wise multiplication. and This represents the branch weights for the two classes of samples. and This indicates the response of conflict-sensitive decision branches to the primary conflict. Indicates the preset response interval. and This represents the output of the source domain residual decision branch corresponding to the degenerate anchor point. to Training coefficients with fixed meanings, such as the outputs of unaffected anchor points after masking the original conflict initiation point, should remain close, while the responses of conflict-sensitive decision branches should produce outputs no less than [a certain value]. The changes in these constraints cause different constraints to act on the stable region and the conflict region respectively.

[0058] Specifically, when masking conflict origin data, the missing identifier of the corresponding anchor point is retained and the conflict component from the origin source is removed. When truncating the conflict propagation path, the original conflict node is retained while the secondary conflict input after the specified dependency edge is removed. When removing the specified data source, the source availability mask, the set of valid anchor points, and the set of degenerate anchor points are updated simultaneously. The training batch consists of original samples, single counterfactual samples, and multi-path counterfactual samples. The weight generation network independently calculates the gating input for each sample in the same batch. The common semantic decision branch performs output consistency constraints on unmasked anchor points. The conflict-sensitive decision branch performs response difference constraints before and after the original conflict is removed. The source domain residual decision branch performs output preservation constraints when the data source is missing but the degenerate anchor point still exists.

[0059] In this embodiment, after the model training is completed, a model version is generated that is bound to the semantic anchor version, field description version, source reliability record version, and conflict propagation graph version. When the internal deployment environment receives the sample to be decided, it selects the corresponding set of semantic anchors according to the sample time and field version, completes the calculation of common semantic components and source domain residual components, generates conflict codes and source availability masks, and then outputs the decision modeling results by the gated machine learning decision model. At the same time, it saves the anchor identifiers, source combinations, conflict paths, branch weights, and model versions involved in the decision. If there are subsequent changes in fields, missing sources, or changes in conflict status, the corresponding decision process can be replayed according to the saved input relationship, and the replay process uses the anchor versions and model parameters within the original valid time range.

[0060] Furthermore, the internally deployed data access process generates batch fingerprints for each batch of heterogeneous data. The batch fingerprint is calculated jointly by the source identifier, field version, data time range, number of records, cleaning rule version, and derivation relationship version. Before generating the source domain feature description, the processing program checks the correspondence between the batch fingerprint and the registered version. If the field structure remains unchanged but the data distribution has drifted, the field version is used and the value distribution components are updated. If the field definition, unit, or associated object changes, a new field version is established and the validity period of the old version is terminated. Historical samples are not overwritten by the new version. When machine learning training reads samples, it also reads the field version corresponding to the sample time, so that the same field identifier forms distinguishable source domain features at different definition stages. The inheritance relationship between field versions is only established after the bidirectional semantic loopback verification is completed.

[0061] In a preferred embodiment, data with inconsistent time granularity are not aligned using unconditional upsampling or downsampling. Instead, a comparable interval is established centered on the time reference registered by the semantic anchor point. The continuous numerical field stores interval statistics, direction of change, and most recent observation time within the comparable interval. The discrete state field stores the state duration interval, number of state transitions, and most recent transition position. The event field stores the event time and processing time. If multiple sources have different numbers of records within the same comparable interval, the common semantic projection path reads the interval representation that can be jointly interpreted among the sources. The source domain-specific projection path reads the record density, local fluctuations, and delay distribution, thereby preventing high-frequency sources from directly dominating the common semantic component due to the large number of records.

[0062] Furthermore, for cases where derived fields and original fields coexist, the semantic relationship graph registers the calculation source, aggregation range, and conversion rules of the derived fields as directed derived edges. The candidate anchor verification process checks whether the derived field can be reconstructed from its source field within the same time range. If one source provides a total field and another source provides multiple component fields, a one-to-many correspondence is established only when the component set is complete and the aggregation rules are stable. When the one-to-many relationship enters the common semantic subspace, the aggregated anchor is used for expression. The deviation of each component from the total is written into the corresponding source domain residual subspace. If the component set is missing, the total field and the remaining component fields are not forced into the same common representation. The missing state is written into the source availability mask and degenerate anchor record.

[0063] In this embodiment, machine learning training samples are constructed according to associated objects and target observation intervals. The processing program saves a semantic anchor snapshot for each sample. The semantic anchor snapshot records the cross-source support relationships, degenerate anchors, independent anchors, conflict nodes, and field versions that the sample actually possesses. Training batches are extracted hierarchically according to the anchor coverage structure, so that complete source combinations, partially missing combinations, and conflict combinations all participate in parameter updates. The training samples of the common semantic decision branch are required to have at least one semantic anchor that has been verified by stable correspondence. The source domain residual decision branch allows independent anchors and degenerate anchors to participate in training. The conflict-sensitive decision branch requires the retention of the original record index that forms the conflict to prevent the use of only the aggregated conflict intensity to replace the conflict formation process.

[0064] Preferably, the dependent edges of the conflict propagation graph are stored according to the effective time range. The processing program uses the stable temporal relationship and derivation relationship between anchors to determine the edge direction, and uses the historical target contribution direction to exclude pseudo propagation relationships formed only by common upstream factors. When the upstream anchor conflict ends but the downstream anchor continues to be in conflict, the propagation path is truncated at the conflict end position and a new original conflict candidate is generated. The new original conflict candidate must remain incompatible after the upstream conflict input is removed in order to be confirmed as an original conflict. If the downstream conflict disappears with the removal of the upstream conflict, it is still marked as a secondary conflict. The conflict type vector thus stores the verifiable start point, path and end position.

[0065] Furthermore, the gated machine learning decision model generates branch weights for each sample during the inference phase, instead of using the average weights from the training batches. The output of the common semantic decision branch and the output of the source domain residual decision branch are in the same decision space. The conflict-sensitive decision branch outputs both candidate decisions and conflict response vectors under conflict conditions. The processing procedure first fuses the candidate decisions according to normalized weights, and then checks the consistency between the conflict response vector and the conflict type vector of the corresponding anchor point. If the conflict response points to a non-existent source combination or propagation path, the current sample is marked as an input to be verified and the original branch output is saved. The decision modeling result is still synthesized from the branch outputs supported by the existing inputs, and conflict responses that cannot be traced back to the input data are not used.

[0066] In a preferred embodiment, historical sample replay is performed in the internal validation area before the model version is released. The replay process fixes the field description version, semantic anchor version, conflict propagation graph version, and source reliable record version. Complete source samples, samples restored according to historical missing states, and counterfactual samples formed by original conflict truncation are input respectively. The processing program compares whether the ownership of common semantic components, source domain residual writing path, conflict starting position, and changes in gating weights are consistent with the training constraints. If the anchor update causes historical samples to be assigned to different common semantic centers, the historical anchor identifier is converted according to the inheritance mapping relationship. Samples that cannot be restored through inheritance mapping are not included in the release set of the new model version.

[0067] Furthermore, the training process employs an anchor version locking mechanism. After a single training task is started, the same set of semantic anchors and the same set of field descriptions are read. New field versions, new candidate corresponding edges, or new conflict propagation edges that appear during training are written into the next verification version without changing the data attribution of the current batch. Parameter updates are executed alternately according to the dependency order of common projection parameters, source domain-specific projection parameters, three decision branch parameters, and weight generation network parameters. When updating common projection parameters, the source domain-specific projection parameters are frozen and semantic conservation constraints are checked. When updating source domain-specific projection parameters, the common projection parameters are frozen and residual reconstruction and cross-source leakage are checked. When updating decision branches, the generation paths of the two types of representations are fixed. When updating the weight generation network, the outputs of each decision branch are fixed. After completing one alternation cycle, the training objective is calculated jointly. If the anchor center attribution, residual writing source, and conflict starting point are inconsistent in consecutive verification batches, the previous parameter check point is restored and the update range of the corresponding parameter group is narrowed.

[0068] In a preferred embodiment, reliable source records are maintained separately according to semantic anchors. Field integrity is generated by the correspondence between expected and actual records. Time delay stability is generated by the offset sequence between event time and processing time. Historical anomaly density is generated by anomaly records identified through field constraints, dimensional constraints, and object constraints. Records consistent with the target state are only written after the target label has been version confirmed. Mapping path reversibility is generated by bidirectional semantic loopback results. The processing program assembles the above records into a reliability sequence according to the effective time range and inputs it into the gating network. The reliability sequence is updated using a restricted cumulative method. A single batch only changes the records corresponding to the current data source, current anchor, and current time period, without modifying reliable records from other anchors or other sources. When switching field versions, the new version inherits the basic records of the old version but recalculates the mapping path reversibility.

[0069] Specifically, missing data processing does not generate alternative public semantic values. If an anchor point is still supported by other sources across sources, the position of the missing source is written into the source availability mask. Public semantic components are generated by available sources according to the anchor point center constraint. The original source domain residual position of the missing source is set to an unwritable state. If the anchor point degenerates into a single source support, the public projection result and exclusive residual result of the existing source are retained, and the supporting source, missing source, and degeneration start time are written into the degenerate anchor point record. If all sources are missing, the corresponding anchor point is removed from the set of valid anchor points of the current sample, and only the missing identifier is retained. The gating input is thus reduced to reduce the reference to the representation of the non-existent anchor point.

[0070] Furthermore, the decision objective can be a classification objective, a continuous value prediction objective, or an ordered decision objective. The three decision branches use the same objective space and calculate the training error corresponding to the objective type respectively. The classification objective outputs the unnormalized decision quantity of each candidate category, the continuous value prediction objective outputs the center estimate and the internal discrete interval representation, and the ordered decision objective outputs the boundary quantity between adjacent levels. The branch weighting is completed in the unnormalized decision space. Then, normalization, interval restoration, or level boundary resolution is performed according to the objective type. When the objective label is revised, the old label version and revision source are retained. The training samples are selected according to the label effective time to avoid the new label definition directly changing the semantic anchor point of the historical field.

Claims

1. An internally deployed multi-source heterogeneous data machine learning decision modeling method, characterized in that, include: Obtain heterogeneous data from multiple data sources, along with their corresponding data types, field semantics, dimensions, time granularity, and descriptions of associated objects, and generate source domain feature descriptions for each data source. Construct a set of semantic anchors based on the stable correspondence between cross-source features; Each source domain feature is decomposed into a common semantic component mapped to the set of semantic anchors and a source domain residual component that retains the source features. Generate conflict codes for data mapped to the same semantic anchor but with inconsistent values, states, or time sequences; The semantic anchor coverage information, source domain residual components, conflict codes, and data source missing identifiers are input into the gated machine learning decision model to obtain the decision modeling results.

2. The on-premises multi-source heterogeneous data machine learning decision modeling method according to claim 1, characterized in that, The construction of the semantic anchor set includes: forming candidate anchor groups based on field semantic similarity relationships, and calculating the cross-source correspondence matrix within the candidate anchor groups based on value interval overlap relationships, unit conversion relationships, time synchronization relationships, and associated object consistency relationships; Extract bidirectional correspondence edges that remain stable within multiple historical time windows from the cross-source correspondence matrix, and merge source domain features that share the same stable correspondence subgraph into semantic anchors; For source domain features that cannot form stable corresponding edges, retain independent anchor point identifiers, and record the data source set, feature source path and effective time range corresponding to each semantic anchor point.

3. The on-premises multi-source heterogeneous data machine learning decision modeling method according to claim 1, characterized in that, The generation of the common semantic component and the source domain residual component includes: setting shared projection parameters and source domain-specific projection parameters for each data source, and projecting the source domain features to the common semantic subspace and the source domain residual subspace respectively; Semantic conservation constraints are established based on the contribution direction of cross-source features to the target label under the same semantic anchor point, and residual separation constraints are established based on the correlation matrix between the common semantic components and the source domain residual components. The shared projection parameters and the source domain-specific projection parameters are jointly adjusted, and cross-source consistency information is written into the common semantic subspace. Source-specific changes that are not interpreted by semantic anchors are written into the corresponding source domain residual subspace.

4. The on-premises multi-source heterogeneous data machine learning decision modeling method according to claim 1, characterized in that, The construction of the conflict coding and the gated machine learning decision model includes: calculating the value difference, state repulsion relationship, time offset relationship and source reliability difference of cross-source data under the same semantic anchor point, and generating conflict type vector, conflict intensity vector and conflict source combination identifier accordingly; Set up a common semantic decision branch, a source domain residual decision branch, and a conflict-sensitive decision branch; Branch weights are generated based on a combination of semantic anchor coverage, conflict intensity, residual energy, and missing data sources. The outputs of each decision branch are then weighted and fused to form the decision modeling result.

5. The on-premises multi-source heterogeneous data machine learning decision modeling method according to claim 2, characterized in that, The generation of the stable correspondence subgraph includes: constructing a cross-source semantic relationship graph using the source domain features as nodes and the candidate correspondence relationships as directed edges; Calculate the semantic loop closure consistency along the forward field mapping path and the reverse field restoration path respectively, and mark the candidate corresponding edges that do not meet the loop closure consistency condition as edges to be verified; The edge to be verified is updated by utilizing the changes in conditional mutual information, the co-occurrence stability of associated objects, and the reversibility of dimensional transformation in different historical time windows. Only corresponding edges that simultaneously satisfy the conditions of bidirectional closure consistency and cross-window stability are included in the stable correspondence subgraph, and the field mapping path, verification window, and correspondence source are saved for each corresponding edge.

6. The on-premises multi-source heterogeneous data machine learning decision modeling method according to claim 3, characterized in that, The establishment of the semantic conservation constraint and the residual separation constraint includes: constructing paired sample groups between cross-source samples corresponding to the same semantic anchor point, and calculating the gradient direction of the common semantic component with respect to the target label and the discrimination direction of the source domain residual component with respect to the source identifier, respectively. The common semantic components with opposite gradient directions are reprojected to the corresponding source domain residual subspace, and the anchor point center consistency constraint is applied to the cross-source common semantic components. Identify residual components that can characterize the source identifiers of other data sources, perform cross-source leakage suppression on the identified residual components, and limit the write path of each source domain residual component according to the data source attribution relationship.

7. The on-premises multi-source heterogeneous data machine learning decision modeling method according to claim 4, characterized in that, The generation of the conflict type vector includes: establishing conflict primitives according to numerical deviation, category mutual exclusion, time sequence reversal, and asynchronous source state; Match cross-source data corresponding to the same semantic anchor point with each conflict primitive to form anchor-level conflict nodes; Based on the dependencies between the semantic anchors, connect the anchor-level conflict nodes to obtain a conflict propagation graph; The conflict propagation graph is used to distinguish between secondary conflicts generated by upstream semantic anchors and original conflicts occurring at the current semantic anchor, and the conflict origin, propagation path, data source involved, and duration are encoded into the conflict type vector.

8. The on-premises multi-source heterogeneous data machine learning decision modeling method according to claim 4, characterized in that, The generation of the branch weights includes: constructing a source availability mask for different combinations of missing data sources, and determining, based on the source availability mask, effective anchors that still have cross-source support and degenerate anchors that only have single-source support from the set of semantic anchors. Calculate the effective anchor coverage, the proportion of degraded anchor residuals, and the achievable conflict intensity in the conflict propagation diagram, respectively. The effective anchor coverage, the proportion of degraded anchor residuals, the achievable conflict intensity, and the source reliability sequence are input into the weight generation network, which outputs the normalized weights of the common semantic decision branch, the source domain residual decision branch, and the conflict-sensitive decision branch.

9. The on-premises multi-source heterogeneous data machine learning decision modeling method according to claim 1, characterized in that, The update of the semantic anchor set includes: when a new source domain feature is added, calculating the bidirectional semantic loop consistency, cross-window correspondence stability, and dimensional transformation invertibility between the new source domain feature and the existing semantic anchors respectively; New source domain features that simultaneously satisfy all correspondences are incorporated into the corresponding semantic anchors, and new source domain features that only form locally stable correspondence edges are established as candidate sub-anchors. When a stable corresponding subgraph within the same semantic anchor point splits into multiple disconnected subgraphs, the semantic anchor point is split according to each disconnected subgraph, and the inheritance mapping relationship between the anchor point identifier before splitting and each candidate sub-anchor point is saved.

10. The on-premises multi-source heterogeneous data machine learning decision modeling method according to claim 8, characterized in that, The training of the gated machine learning decision model includes generating counterfactual training samples based on the original conflict nodes, secondary conflict nodes and missing data sources in the conflict propagation graph. The counterfactual training samples are formed by shielding the conflict origin data, cutting off the conflict propagation path or removing the specified data source. The branch weights and decision outputs of the original samples and counterfactual training samples are obtained respectively. Joint training constraints are constructed based on the output consistency relationship of the unmasked semantic anchors, the response difference of the conflict-sensitive decision branches to the original conflict, and the output preservation relationship of the source domain residual decision branches corresponding to the degenerate anchors. The model parameters of the weight generation network and each decision branch are then adjusted.

Images