A global optimization method and apparatus for multimodal data quality

Abstract

This invention discloses a global optimization method and apparatus for multimodal data quality, belonging to the field of data quality management. The method includes: performing data abstraction processing on acquired raw multi-source heterogeneous data to obtain a set of atomic data units containing weights and unit relationships; sequentially instantiating rules and binding weights to the atomic data unit set according to preset cleaning rules to obtain a set of constraint instances representing the cleaning rules and violation costs for each type of data; performing global cost optimization modeling and approximate solution processing on the atomic data unit set based on the constraint instance set to obtain an optimal cleaning strategy that minimizes the sum of the total modification cost and the total violation cost of the data unit set; modifying or deleting corresponding units in the atomic data unit set according to the optimal cleaning strategy to obtain a clean atomic data unit set that meets preset quality requirements and a cleaning process report. This invention can improve the cleaning efficiency and quality of multimodal data.

Description

Technical Field

[0001] This invention relates to the field of data quality management technology, and in particular to a global optimization method and apparatus for multimodal data quality. Background Technology

[0002] Data fusion cleaning is a systematic data optimization process that addresses various quality issues arising after the fusion of multi-source heterogeneous modal data, such as structured tables, unstructured documents, and knowledge graphs. It is a crucial step before multi-source data fusion applications. Its core objective is to resolve issues such as inconsistencies, non-standard formats, logical conflicts, and data redundancy in the fused data. Through a series of methods including verification, repair, modification, and deletion, it achieves cross-modal data consistency assurance and collaborative optimization, ultimately improving the accuracy, completeness, consistency, and usability of the fused data, laying a high-quality data foundation for subsequent data analysis, mining, and business applications.

[0003] Existing technologies have many key shortcomings in the data cleaning process. Traditional methods often result in low-quality cleaned data when dealing with multi-modal data because it is difficult to quantify the credibility of different data sources or the strictness of different cleaning rules. At the same time, since the absolutely optimal cleaning solution is a computationally difficult problem, traditional cleaning solutions lack efficient algorithms for finding the optimal solution.

[0004] Therefore, there is an urgent need for a global optimization method and apparatus for multimodal data quality to solve the above-mentioned technical problems. Summary of the Invention

[0005] This invention provides a global optimization method and apparatus for multimodal data quality, which can simultaneously improve the cleaning efficiency and quality of multimodal data. The technical solution is as follows: On the one hand, a global optimization method for multimodal data quality is provided, the method comprising: The acquired raw multi-source heterogeneous data is subjected to data abstraction processing to obtain a set of atomic data units containing weights and unit relationships; According to the preset cleaning rules, the atomic data unit set is sequentially instantiated and weighted to obtain a set of constraint instances that characterize the cleaning rules and violation costs of each type of data. Based on the set of constraint instances, global cost optimization modeling and approximate solution processing are performed on the set of atomic data units to obtain the optimal cleaning strategy that minimizes the sum of the total modification cost and the total violation cost of the data unit set. According to the optimal cleaning strategy, the corresponding units in the atomic data unit set are modified or deleted to obtain a clean atomic data unit set and a cleaning process report that meet the preset quality requirements.

[0006] On the other hand, a global optimization device for multimodal data quality is provided, the device comprising: The processing module is used to perform data abstraction processing on the acquired raw multi-source heterogeneous data to obtain a set of atomic data units containing weights and unit relationships; The binding module is used to sequentially instantiate rules and bind weights to the set of atomic data units according to preset cleaning rules, so as to obtain a set of constraint instances that characterize the cleaning rules and violation costs of each type of data. The optimization module is used to perform global cost optimization modeling and approximate solution processing on the atomic data unit set based on the constraint instance set, so as to obtain the optimal cleaning strategy that minimizes the sum of the total modification cost and the total violation cost of the data unit set. The cleaning module is used to modify or delete corresponding units in the atomic data unit set according to the optimal cleaning strategy, so as to obtain a clean atomic data unit set and a cleaning process report that meet the preset quality requirements.

[0007] On the other hand, a computer device is provided, the computer device including a memory and a processor, the memory for storing a computer program, and the processor for executing the computer program stored in the memory to implement the steps of the global optimization method for multimodal data quality described above.

[0008] On the other hand, a computer-readable storage medium is provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements the steps of the global optimization method for multimodal data quality described above.

[0009] On the other hand, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the global optimization method for multimodal data quality described above.

[0010] The technical solution provided by this invention offers at least the following beneficial effects: First, by abstracting and processing multi-source heterogeneous data, it solves the fundamental problem of the difficulty in collaborative processing of cross-modal data. Second, by formalizing business rules into configurable violation constraints and quantifying business knowledge and integrating it into the cleaning process, it greatly enhances the system's flexibility and business adaptability. Simultaneously, by constructing the cleaning problem as a global programming model and employing an approximate algorithm based on linear programming relaxation for solution, it can find a theoretically sound near-optimal solution within the target time, achieving a breakthrough balance between global optimization of cleaning quality and processing efficiency. Finally, by executing interpretable cleaning decisions and generating detailed cleaning reports, it achieves visualization and traceability of the cleaning process. This method effectively overcomes the shortcomings of traditional technologies, such as local optimization, inability to balance multiple objectives, and reliance on human experience, providing a globally optimized, automated, efficient, and interpretable intelligent multi-modal data cleaning system and method. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a global optimization method for multimodal data quality provided in an embodiment of the present invention; Figure 2 This is a structural diagram of a global optimization device for multimodal data quality provided in an embodiment of the present invention; Figure 3 This is a hardware architecture diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0014] As mentioned earlier, existing cleaning methods, when processing multi-source data, either sacrifice theoretical quality assurance for processing speed or sacrifice efficiency for cleaning effect, making it difficult to simultaneously meet the requirements of quality and efficiency.

[0015] Based on this, the concept of the present invention is to achieve data unification by abstracting and processing multi-source data, and to generate a high-quality and efficient cleaning strategy for multi-source data by constructing a global planning model.

[0016] The specific implementation of the above concept is described below.

[0017] Please refer to Figure 1 This invention provides a global optimization method for multimodal data quality, the method comprising: Step 100: Perform data abstraction processing on the acquired raw multi-source heterogeneous data to obtain a set of atomic data units containing weights and unit relationships; Step 102: Instantiate rules and bind weights to the set of atomic data units in sequence according to the preset cleaning rules to obtain a set of constraint instances that characterize the cleaning rules and violation costs of each type of data. Step 104: Perform global cost optimization modeling and approximate solution processing on the atomic data unit set based on the constraint instance set to obtain the optimal cleaning strategy that minimizes the sum of the total modification cost and the total violation cost of the data unit set. Step 106: Modify or delete the corresponding units in the atomic data unit set according to the optimal cleaning strategy to obtain a clean atomic data unit set and a cleaning process report that meet the preset quality requirements.

[0018] In this embodiment of the invention, the fundamental problem of difficult collaborative processing of cross-modal data is solved by abstracting and processing multi-source heterogeneous data. Secondly, by formalizing business rules into configurable violation constraints and quantifying business knowledge and integrating it into the cleaning process, the system's flexibility and business adaptability are greatly enhanced. Simultaneously, the cleaning problem is constructed as a global programming model and solved using an approximate algorithm based on linear programming relaxation, thereby finding a theoretically sound near-optimal solution within the target time, achieving a breakthrough balance between global optimization of cleaning quality and processing efficiency. Finally, by executing interpretable cleaning decisions and generating detailed cleaning reports, the cleaning process is visualized and traceable. This method effectively overcomes the shortcomings of traditional technologies, such as local optimization, inability to balance multiple objectives, and reliance on human experience, providing a globally optimized, automated, efficient, and interpretable intelligent multi-modal data cleaning system and method.

[0019] The following description Figure 1 The execution method of each step is shown.

[0020] First, for step 100, the acquired original multi-source heterogeneous data is subjected to data abstraction processing to obtain a set of atomic data units containing weights and unit relationships.

[0021] When the information from tabular data, document descriptions, and graph relationships is inconsistent, there is a lack of a unified mathematical model and algorithm to determine which modality's information should be used, or how to generate a compromise, globally optimal, and consistent version.

[0022] Therefore, this embodiment abstracts heterogeneous data such as tables, documents, and graphs into a unified "weighted data unit" representation, and formalizes various cleaning rules (within a single modality and across modalities) into "weighted constraints," thereby characterizing all data quality issues with the same framework.

[0023] In this embodiment of the invention, the data abstraction process includes: First, the original multi-source heterogeneous data is parsed to obtain data objects that can be recognized by computers for each type of data.

[0024] Specifically, the system accesses and parses data from different sources and formats. In this embodiment, the multi-source heterogeneous data includes structured tables (CSV, SQL), semi-structured documents (JSON, XML), unstructured text (resumes, reports), and knowledge graphs. This provides structured input for subsequent unified transformation, converting raw byte streams or messy text into data structures that the program can recognize and process.

[0025] Furthermore, all the data objects are abstracted to obtain data units composed of entity identifiers, attributes, and values, and unit weights are added to the corresponding data units according to the credibility of the data object's source and the importance of the data.

[0026] Specifically, the system breaks down and transforms all data into unified "atomic data units." Each data unit is a triple consisting of <entity identifier, attribute, value>, along with metadata such as source and timestamp. The transformation rules include: For structured records (such as a row in a database), each row corresponds to an entity, each column is an attribute, and each cell is a data unit, thus generating a triplet unit <record primary key, field name, field value> for each field; For unstructured text, entities, attributes, and values ​​are extracted using natural language processing techniques (such as named entity recognition and relation extraction) to generate units <entity ID, attribute type, mention value>. For a knowledge graph, generate a cell <node ID, attribute name, attribute value> for each node attribute, and generate a cell <source node ID, relation name, target node ID> for each edge to represent a relation.

[0027] For example, the original content of a resume text is "Zhang San, graduated from the Department of Computer Science at Tsinghua University in 2015, with 8 years of software development experience." After parsing, the following atomic data units are obtained: Zhang San Name Zhang San CV Zhang San Graduation year 2015 CV Zhang San Graduation Institution Tsinghua University CV Zhang San major Computer Science Department CV Zhang San Years of service 8 CV After parsing the atomic data units, they are assigned unit weights. These weights combine the credibility of the data source (e.g., higher weights for authoritative data sources) and the importance of the data itself (e.g., higher weights for core business fields). In this embodiment, these weights are normalized to a fixed range (e.g., [0,1]). Specifically, the weights are assigned according to the following principles: Source credibility: Authoritative databases (such as publicly available government data) have a high weight (e.g., 0.9~1.0); ordinary business systems have a medium weight (0.5~0.8); web crawling or user-generated content has a low weight (0.1~0.4). Credibility can be pre-assessed or dynamically adjusted based on historical accuracy. Data importance: Core business fields (such as ID number and amount) have high weight; auxiliary information (such as remarks and descriptions) has low weight. Configurable business priority mapping is available. Weight normalization: It is recommended that the weights of all data units be normalized to the interval [0,1] or [0,100] to ensure consistent measurement with constraint violation weights. For example, a weight of 0.9 is used for high-confidence sources + core fields, and 0.2 is used for low-confidence sources + secondary fields.

[0028] Furthermore, the data units are identified based on entity identifiers, and all data units referring to the same entity are aggregated into the same group of data units based on the identification results; different attribute units in the same group of data units are associated based on attribute names, and dependency relationships are established for attributes with inference relationships; reference associations are established between entities mentioned in unstructured text and corresponding nodes in the knowledge graph based on the data units; a preliminary scan of all associated units is performed to identify conflicting associations in the association relationships; and the set of atomic data units is generated based on the data units of all data objects and all association relationships.

[0029] Specifically, establishing relationships between elements, such as same-point relationships ("Zhang San" mentioned in different places refers to the same person) and derivative relationships ("age" can be derived from "date of birth"), lays the foundation for subsequent constraint modeling.

[0030] Relationship building includes: Entity alignment: Identifying units that refer to the same entity in different data sources. Aggregating units of the same entity into the same group based on entity identifiers (such as IDs) or similarity matching (such as fuzzy name matching, attribute value combination matching).

[0031] Attribute association: For the same entity, associate different attribute units by attribute name; for inferable attributes (such as "age" and "year of birth"), establish functional dependencies.

[0032] Cross-modal reference links: For entities mentioned in unstructured text, entity linking technology is used to associate them with nodes in the knowledge graph, establishing reference relationships between text units and graph units.

[0033] Conflict pre-detection: A preliminary scan of all associated units is performed to identify candidate conflict pairs that clearly violate the rules (e.g., inconsistent values ​​of the same attribute from different sources within the same entity). These conflict pairs serve as input variables for the subsequent optimization model. The pre-association results are stored as a unit relationship graph, where nodes represent data units and edges indicate association types (e.g., common association, derivation, conflict).

[0034] Then, for step 102, the atomic data unit set is sequentially instantiated and weighted according to the preset cleaning rules to obtain a set of constraint instances that characterize the cleaning rules and violation costs of each type of data.

[0035] Existing technologies for cleaning multi-modal data either strictly adhere to rules, leading to excessive data modification or discarding, or tolerate violations, resulting in low data quality. They cannot perform refined, weighted selection based on the importance of data items and the credibility of the rules.

[0036] Therefore, this embodiment defines a total cost function that includes "data modification / deletion cost" and "constraint violation cost". By balancing the weights of data units and constraints, it seeks a cleaning scheme that minimizes the total cost, thus achieving global optimization.

[0037] In this embodiment of the invention, the definition process includes: The user-input natural language constraints are converted according to a preset declarative rule language, and the conversion results are encapsulated in a template to obtain data cleaning rules that can be recognized by computers; wherein, the declarative rule language includes rule scope, logical conditions, and violation weights.

[0038] Specifically, the implementation of cleaning rule definition and template involves formally describing the business knowledge, data quality requirements, and compliance rules provided by the user through a declarative rule language, and then using pre-set templates for parameterized encapsulation, ultimately resulting in a set of formal cleaning rule definitions that can be automatically parsed and executed by the system.

[0039] First, users input informal business knowledge, data quality requirements, and compliance rules. These are usually in the form of natural language documents, expert experience, or lists of business requirements, such as "employees' ages should be positive integers," "the contact number for the same customer should be unique in the system," and "the monetary values ​​in financial statements cannot be negative."

[0040] Next, a dedicated, near-natural language declarative rule language was designed and provided, allowing users or data administrators to precisely describe constraints. Each rule consists of three core parts: Rule scope: Specifies which data the rule applies to. It filters specific sets of data units using conditions similar to the WHERE clause in SQL (e.g., "attribute == 'ID number'" or "entity == 'order' AND attribute IN ('amount', 'tax')").

[0041] Violation logic condition: Describes the specific logical condition that constitutes a "violation" or "data quality problem". It consists of a series of predicates (such as =, !=, >, <, IN, regular expression matching, etc.) and logical connectors (AND, OR, NOT) (e.g., "value ISNULL" or "value NOT MATCH 'regular expression'").

[0042] Violation weight: This assigns a weight value representing the cost of violating the rule. This weight will be directly used in the subsequent calculation of the objective function of the optimization model.

[0043] In addition, to lower the barrier to entry for users, improve efficiency, and ensure the quality of rule definitions, the system provides a series of pre-built common constraint templates. Users do not need to write rules from scratch; they only need to select a suitable template and fill in the key parameters. The template types mentioned in the document include: Single-model rule templates: used for constraints within a single data type, such as field format, numeric range, uniqueness, and functional dependencies.

[0044] Cross-modal rule templates: used to specify consistency between different data modalities, such as "the entity attributes mentioned in the text should be consistent with the corresponding attribute values ​​in the knowledge graph".

[0045] Logical rule templates: used to describe more complex business logic, such as the implication of "if A then B", "mutually exclusive" relationships, etc.

[0046] The process described above ultimately produces a set of formalized, structured cleansing rule definitions. These definitions are stored in a system-parseable format (such as specific configuration files, internal DSL code, or database records). They specify "on which data," "what conditions are checked," and "what the penalty is for violations," providing a directly executable blueprint for subsequent rule instantiation.

[0047] Further, the atomic data unit set is traversed and filtered according to the rule scope to obtain candidate unit groups consisting of all data units applicable to the same rule scope; the logical conditions corresponding to the rule scope are attached to the corresponding candidate unit groups, and the corresponding violation weights are assigned to all candidate unit groups according to the violation acceptance degree of each rule, to obtain hard constraint instances and soft constraint instances containing rule conditions and violation costs; wherein, the hard constraint instance represents that violating the rule is prohibited, and the soft constraint instance represents that a cost must be paid when violating the rule.

[0048] Specifically, in rule traversal and constraint instantiation, the system applies the formal rule definition to specific data. The process of generating a set of constraint instances from the filtered unit groups is an "instantiation" or "binding" process from abstract rules to specific data constraints.

[0049] First, the data cleaning rules obtained from the above process are traversed one by one to determine the scope of each rule. This scope is essentially a query or filtering condition used to select the specific units governed by the rule from the full set of data units.

[0050] Next, a match is made among all data units. All data units that simultaneously meet the scope condition are grouped into the same candidate unit group. For example: For the rule "Scope: attribute == 'Employee ID'", the system will find all data cells whose attribute field value is "Employee ID". Assuming 100 cells are found, these 100 cells constitute a candidate cell group corresponding to this rule.

[0051] For the rule "Scope: Functional Dependency Department Code → Department Name", the system first identifies all units with the attribute "Department Code" and their values, and then groups them according to the "Department Code" value. Units with the same "Department Code" value are grouped together, and units belonging to the same entity and with the attribute "Department Name" are also associated with that group. Thus, each different "Department Code" value (e.g., "DEPT01", "DEPT02") generates a corresponding candidate unit group, containing all entities with that code and their "Department Name" units.

[0052] Subsequently, for each candidate unit group obtained, the system creates an independent constraint instance. This instance is the "manifestation" of the rule on specific data. Each constraint instance contains the following key information: List of involved units: Clearly record which specific data units (from the candidate unit group) are constrained by this instance.

[0053] The logical conditions for violations are defined by the user and describe what kind of relationship between units within a group constitutes a violation (for example, for uniqueness constraints, the condition is "the values ​​of all units within the group must be the same pairwise"; for functional dependencies, the condition is "the values ​​of all department name units within the group must be the same").

[0054] Violation weight: The source of the weight can be explicitly configured by the user, that is, the user directly specifies the weight value (such as 5, 10, ∞) when defining the rule; based on the default value of the rule type, that is, the system predefines the default weight range for different rule types (such as compliance rules with high weight, format rules with medium weight); dynamic learning, that is, the weight is automatically adjusted based on historical expert feedback.

[0055] Constraint type label: The constraint type is determined by the user adjusting the corresponding weight values ​​according to actual requirements, marking whether the instance is a "hard constraint" or a "soft constraint". The weight of a hard constraint is set to infinity (e.g., weight +∞ or represented by a very large number such as 1e9), indicating that it must be satisfied and cannot be violated.

[0056] The weights of soft constraints should be set to finite positive numbers, with a recommended range of [0.1, 100], consistent with the range of data cell weights (if cell weights are normalized to [0,1], then constraint weights can be amplified and mapped to [0.1, 100] to highlight differences in importance). Specific settings are based on: Rule importance levels: divided into three levels: high, medium, and low, corresponding to violation weight ranges of [50, 100], [10, 50], and [0.1, 10].

[0057] Rule strictness: Users can fine-tune the rules according to their business tolerance. For example, the weight of compliance rules can be set to 80, and the weight of format rules can be set to 20.

[0058] Coordinating with data unit weights: Constraint weights should not be so small that the cost of modifying the data unit is negligible, nor should they be so large that any modification cannot compensate for the violation. Typically, constraint weights are of the same magnitude as or slightly larger than the unit weights, to encourage prioritizing the modification of lower-weight units rather than retaining the violation in the event of a conflict.

[0059] For example, assuming a data unit weight is 0 to 1, the violation weight of a critical compliance constraint can be set to 100, meaning that it would be better to delete 10 high-weight units (each with a weight of 0.9) than to tolerate the violation; while a format constraint can be set to 5, meaning that it is possible to choose to modify a unit with a weight of 0.6 to eliminate the violation.

[0060] Finally, all the constraint instances are stored in a constraint library used for managing constraint rules, resulting in a constraint instance set.

[0061] Specifically, the generated constraint instances are stored in a constraint library. Each constraint instance contains a unique ID, the set of cells involved, the violation condition, the weight, and a flag indicating whether it is a hard constraint. The constraint library supports adding, deleting, modifying, and querying, facilitating subsequent model optimization.

[0062] For step 104, global cost optimization modeling and approximate solution processing are performed on the atomic data unit set based on the constraint instance set to obtain the optimal cleaning strategy that minimizes the sum of the total modification cost and the total violation cost of the data unit set.

[0063] Existing cleaning strategies often lack global optimization capabilities. For example, modifying a key piece of data to fix an inconsistency may trigger more inconsistencies, lacking the ability to minimize the overall "cleaning cost" from a global perspective.

[0064] Therefore, this embodiment innovatively proposes an efficient approximation algorithm based on linear programming relaxation, which provides a cleaning scheme close to the optimal solution in polynomial time and provides an explicit theoretical approximation ratio guarantee.

[0065] In this invention, the optimal cleaning strategy is obtained as follows: For each element in the constraint instance set and the atomic data unit set, a binary decision variable is constructed to characterize the final cleaning result, so as to quantify all cleaning actions and violation handling status into computable mathematical variables.

[0066] Specifically, define a variable x for each data unit to represent its final state (e.g., retain the original value, modify to a new value, or delete). Define a variable y for each possible instance of violation of each constraint to indicate whether the violation is eliminated.

[0067] Next, the logical relationship between data units and constraint violations is encoded as a linear inequality. For example, to eliminate a constraint violation involving two conflicting data units, at least one of the units must be modified (or deleted). Let I be the set of all data units, and V... h V is the set of hard-constrained instances. s This is a set of instances of soft constraints. Define the decision variables: For each unit t ∈ I, define a binary decision variable x. t :x t = 1 indicates that the cell is deleted / modified (i.e., the cell is cleaned up), x t = 0 means retain the original value.

[0068] For each soft constraint instance V ∈ V s Define binary decision variable y V y V= 1 indicates that the violation still exists (i.e., it has not been eliminated), y V = 0 indicates that the violation has been eliminated.

[0069] All of the above variables take the value 0 or 1.

[0070] Furthermore, based on the logical relationship between constraint instances and corresponding decision variables, an optimization decision model is established to characterize the minimization of the total cost of global data cleaning under multiple constraints.

[0071] For any instance of a hard constraint, the logic is that it "must be eliminated." This means that at least one data unit in the set involved in that instance must be modified or deleted. Mathematically, this can be expressed as: This inequality guarantees that as long as there is an x ​​in the element associated with the hard constraint. t If it is set to 1 (i.e. modified), the inequality is true, satisfying the requirement of "at least one is modified".

[0072] For a soft constraint instance, the logical relationship is more complex: a violation exists (y V =1) if and only if all associated units have not been modified (all x t =0); conversely, as long as one associated unit is modified (any x = 0); t =1), the violation is eliminated (y V =0). Therefore, it can be expressed by the following formula: The correctness of this formula can be verified as follows: if all x t = 0 (no unit modified), then The inequality becomes y V ≥ 1. Because y V It is a binary variable, and it can only be equal to 1. This means that violations are preserved (y). V =1). Similarly, if there is at least one x t = 1 (at least one unit was modified), then The inequality becomes y V ≥ Some non-positive number. Driven by minimizing the objective function, y V It will automatically take the minimum possible value of 0, indicating that the violation has been eliminated.

[0073] After determining the constraints, it is necessary to construct an objective function, which quantifies the total cost of the "cleaning scheme". The optimization objective is to find the decision combination that minimizes the total cost.

[0074] Since the total cost = total cost of data modification + total cost of constraint violation, the total cost of data modification is the sum of the weights of all modified / deleted units, i.e.: The total cost of constraint violation is the sum of the weights of all retained soft constraint violations, i.e.: In the above formula, t represents a data unit; T represents the set of data units. Let be the cell weight of cell t; The violation weight of the soft constraint V.

[0075] Therefore, the objective function is shown in the following formula: In summary, a complete optimization decision-making model has been constructed.

[0076] Finally, the optimized decision model is subjected to linear programming relaxation and threshold rounding to calculate the optimal cleaning strategy that meets the rule requirements.

[0077] Specifically, since both decision variables are binary variables, the optimization problem is a combinatorial optimization problem, and it is computationally infeasible to directly find the exact optimal solution in large-scale data scenarios.

[0078] Therefore, this embodiment adopts a two-stage approximation algorithm framework of "relaxation-rounding". First, the problem is made easier to solve by relaxing the integer restrictions on the variables; then, a rounding rule is designed to transform the obtained "fractional solution" into a high-quality, feasible integer solution.

[0079] First, restrict the values ​​of the original decision variable from x t y V The relaxation condition for x ∈ {0, 1} is 0 ≤ x t ≤ 1 and 0 ≤ y V ≤ 1. That is, the variable is allowed to take any real value (fraction) in the interval [0, 1], thus transforming the original problem into a linear programming problem.

[0080] The system invokes an efficient linear programming solver to quickly compute and obtain the optimal solution to this relaxation problem. Because the variable constraints are relaxed, the variable values ​​in this solution are likely to be fractions; therefore, it is called the fractional optimal solution, denoted as . and This solution is a lower bound of the optimal solution to the original integer programming problem, and its objective function value is less than or equal to the optimal value of the original problem.

[0081] Furthermore, a deterministic threshold rounding method based on theoretical analysis is designed to "round" fractional solutions to integer solutions of 0 or 1, while ensuring that the solutions satisfy all constraints and the total cost is controllable.

[0082] Iterate through all constraint instances (including hard and soft constraints), find the maximum number of data units involved in a single constraint instance, denoted as λ. Set a uniform rounding threshold θ = 1 / (λ+1), where λ is the maximum number of units involved in all constraint instances (i.e., ...). ).

[0083] Round the decision variables of the data unit: like Then let (Reserve the unit); otherwise, let (Modify / delete unit).

[0084] After getting all After the integer value is obtained, the integer value of the soft constraint variable is no longer independently rounded, but is based on the rounded integer value. Recalculate: This formula directly reflects the logic of soft constraints: only when all elements involved in the constraint are preserved (all...) Only when this condition is met does the violation exist. ); as long as one unit is modified (at least one) 1) Violations will be eliminated. ).

[0085] This rounding method guarantees that the rounded solution is a feasible solution to the original problem, i.e., it satisfies all hard and soft constraints; at the same time, the total cost satisfies the approximation ratio: final integer solution ( and The total cost of the solution shall not exceed (λ+1) times the total cost of the theoretically optimal solution to the original problem.

[0086] It's worth noting that in common scenarios, many constraints involve only two data units (such as conflicts between two inconsistent values), i.e., λ=2. In this case, the algorithm can provide a stronger 2-approximation guarantee, meaning the quality of the solution is at least within 50% of the optimal solution.

[0087] Finally, output a feasible cleaning decision scheme with 0-1 integers. and This scheme explicitly specifies whether each data unit should be retained, modified, or deleted, and whether each soft constraint violation should be eliminated.

[0088] For step 106, the corresponding units in the atomic data unit set are modified or deleted according to the optimal cleaning strategy to obtain a clean atomic data unit set and a cleaning process report that meet the preset quality requirements.

[0089] In this embodiment of the invention, based on the cleaning strategy obtained through the above process, the system automatically performs corresponding data operations, such as retaining the original value, modifying it to a specific new value, or deleting data units. Furthermore, it can support generating cleaning results with different strategy preferences (such as the "least modification" or "least violations" schemes) for users to choose from. The system automatically generates a detailed cleaning report, clarifying which conflicts were found, which rules were applied, which specific cleaning operations were performed and their reasons (e.g., to reduce overall cost), and provides a comparison of data quality indicators before and after cleaning. This makes the entire decision-making process transparent and auditable.

[0090] After obtaining a set of clean atomic data units that meet the preset quality requirements and a cleaning process report, the process also includes: conducting expert review and adjustment feedback based on the cleaning process report to obtain optimized violation weight values ​​and new cleaning rules.

[0091] Specifically, the system allows data experts to review and adjust the cleaning solutions recommended by the system (such as rejecting a modification or adjusting weights), and feed the adjusted decisions back to the system as new constraints. Simultaneously, the system records the experts' feedback and uses machine learning models to dynamically optimize the weights of rules in the rule base or suggest new rule templates, making the system increasingly intelligent with use.

[0092] Please refer to Figure 2 This invention provides a global optimization device for multimodal data quality, the device comprising: The processing module 200 is used to perform data abstraction processing on the acquired raw multi-source heterogeneous data to obtain a set of atomic data units containing weights and unit relationships; The binding module 202 is used to sequentially instantiate rules and bind weights to the set of atomic data units according to preset cleaning rules, so as to obtain a set of constraint instances that characterize the cleaning rules and violation costs of each type of data. Optimization module 204 is used to perform global cost optimization modeling and approximate solution processing on the atomic data unit set according to the constraint instance set, so as to obtain the optimal cleaning strategy that minimizes the sum of the total modification cost and the total violation cost of the data unit set. The cleaning module 206 is used to modify or delete corresponding units in the atomic data unit set according to the optimal cleaning strategy, so as to obtain a clean atomic data unit set and a cleaning process report that meet the preset quality requirements.

[0093] In this embodiment of the invention, the step of performing data abstraction processing on the acquired original multi-source heterogeneous data to obtain a set of atomic data units containing weights and unit relationships includes: The original multi-source heterogeneous data is parsed to obtain data objects that can be recognized by computers for each type of data; wherein, the multi-source heterogeneous data includes structured tables, semi-structured documents, unstructured text, and knowledge graphs; All the data objects are abstracted to obtain data units consisting of entity identifiers, attributes, and values, and unit weights are added to the corresponding data units according to the credibility of the data object's source and the importance of the data. The data units are identified based on entity identifiers, and all data units referring to the same entity are aggregated into the same group of data units based on the identification results. Associate different attribute units in the same group of data units based on attribute names, and establish dependency relationships for attributes that have inference relationships. Based on the data units, establish reference associations between entities mentioned in unstructured text and corresponding nodes in the knowledge graph; A preliminary scan of all associated units is performed to identify conflicting relationships. The atomic data unit set is generated based on the data units of all data objects and all their relationships.

[0094] In this embodiment of the invention, the step of sequentially instantiating rules and binding weights to the set of atomic data units according to preset cleaning rules to obtain a set of constraint instances representing the cleaning rules and violation costs for each type of data includes: The user-input natural language constraints are converted according to a preset declarative rule language, and the conversion results are encapsulated in a template to obtain data cleaning rules that can be recognized by computers; wherein, the declarative rule language includes rule scope, logical conditions, and violation weights; The atomic data unit set is traversed and filtered according to the rule scope to obtain a candidate unit group consisting of all data units applicable to the same rule scope; The logical conditions corresponding to the scope of the rule are attached to the corresponding candidate unit group, and the corresponding violation weights are assigned to all candidate unit groups according to the violation acceptance level of each rule, so as to obtain hard constraint instances and soft constraint instances containing rule conditions and violation costs; wherein, the hard constraint instance represents that the rule is prohibited from being violated, and the soft constraint instance represents that a cost must be paid when the rule is violated; All the constraint instances are stored in a constraint library used for managing constraint rules, resulting in a constraint instance set.

[0095] In this embodiment of the invention, global cost optimization modeling and approximate solution processing are performed on the atomic data unit set based on the constraint instance set to obtain the optimal cleaning strategy that minimizes the sum of the total modification cost and the total violation cost of the data unit set, including: For each element in the constraint instance set and the atomic data unit set, a binary decision variable is constructed to characterize the final cleaning result, so as to quantify all cleaning actions and violation handling status into computable mathematical variables. Based on the logical relationship between constraint instances and corresponding decision variables, an optimization decision model is established to characterize the minimization of the total cost of global data cleaning under multiple constraints. The optimal cleaning strategy that meets the rule requirements is calculated by performing linear programming relaxation and threshold rounding on the optimization decision model.

[0096] In this embodiment of the invention, establishing an optimization decision model based on the logical relationship between constraint instances and corresponding decision variables to characterize minimizing the total cost of global data cleaning under multiple constraints includes: Based on the logical relationship between hard constraint instances and decision variables, the first constraint condition is established as shown in the following inequality: In the formula, V represents the decision variable for the set of atomic data units; a value of 1 indicates deletion or modification of the unit, while a value of 0 indicates retention of the original value. V represents the constraint instance. This is an instance of a hard constraint; Based on the logical relationship between soft constraint instances and decision variables, the second constraint condition is established as shown in the following inequality: In the formula, For soft constraint instances, the decision variable is set to 1 to indicate that the violation has not been eliminated, and 0 to indicate that the violation has been eliminated. This is an instance of a soft constraint; To minimize the global cost of data cleaning, the objective function is established as shown in the following formula: In the formula, t is a data unit; T is the set of data units; Let be the cell weight of cell t; The violation weight of the soft constraint V.

[0097] In this embodiment of the invention, after obtaining a set of clean atomic data units and a cleaning process report that meet preset quality requirements, the method further includes: Based on the cleaning process report, expert review and adjustment feedback were conducted to obtain optimized violation weight values ​​and new cleaning rules.

[0098] It should be noted that the global optimization device for multimodal data quality provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the global optimization device for multimodal data quality provided in the above embodiments and the global optimization method embodiments for multimodal data quality belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0099] Embodiments of this application also provide a computer device, please refer to... Figure 3 The computer device includes a processor and a memory, the memory storing at least one instruction, at least one program, code set, or instruction set, the at least one instruction, at least one program, code set, or instruction set being loaded and executed by the processor to implement the global optimization method for multimodal data quality provided in the above-described method embodiments.

[0100] Embodiments of this application also provide a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, at least one program, code set, or instruction set is loaded and executed by a processor to implement the global optimization method for multimodal data quality provided in the above-described method embodiments.

[0101] Embodiments of this application also provide a computer program product, which includes a computer program. A processor of a computer device reads the computer program from a computer-readable storage medium and executes the computer program, causing the computer device to perform any of the global optimization methods for multimodal data quality described in the above embodiments.

[0102] For ease of description, the above systems or devices are described separately as various modules or units based on their functions. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware components.

[0103] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0104] Finally, it should be noted that in this document, relational terms such as first, second, third, and fourth are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0105] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A global optimization method for multimodal data quality, characterized in that, The method includes: The acquired raw multi-source heterogeneous data is subjected to data abstraction processing to obtain a set of atomic data units containing weights and unit relationships; According to the preset cleaning rules, the atomic data unit set is sequentially instantiated and weighted to obtain a set of constraint instances that characterize the cleaning rules and violation costs of each type of data. Based on the set of constraint instances, global cost optimization modeling and approximate solution processing are performed on the set of atomic data units to obtain the optimal cleaning strategy that minimizes the sum of the total modification cost and the total violation cost of the data unit set. According to the optimal cleaning strategy, the corresponding units in the atomic data unit set are modified or deleted to obtain a clean atomic data unit set and a cleaning process report that meet the preset quality requirements.

2. The method as described in claim 1, characterized in that, The process of abstracting the acquired raw multi-source heterogeneous data to obtain a set of atomic data units containing weights and unit relationships includes: The original multi-source heterogeneous data is parsed to obtain data objects that can be recognized by computers for each type of data; wherein, the multi-source heterogeneous data includes structured tables, semi-structured documents, unstructured text, and knowledge graphs; All the data objects are abstracted to obtain data units consisting of entity identifiers, attributes, and values, and unit weights are added to the corresponding data units according to the credibility of the data object's source and the importance of the data. The data units are identified based on entity identifiers, and all data units referring to the same entity are aggregated into the same group of data units based on the identification results. Associate different attribute units in the same group of data units based on attribute names, and establish dependency relationships for attributes that have inference relationships. Based on the data units, establish reference associations between entities mentioned in unstructured text and corresponding nodes in the knowledge graph; A preliminary scan of all associated units is performed to identify conflicting relationships. The atomic data unit set is generated based on the data units of all data objects and all their relationships.

3. The method as described in claim 1, characterized in that, The atomic data unit set is sequentially instantiated and weighted according to preset cleaning rules to obtain a set of constraint instances representing the cleaning rules and violation costs for each type of data, including: The user-input natural language constraints are converted according to a preset declarative rule language, and the conversion results are encapsulated in a template to obtain data cleaning rules that can be recognized by computers; wherein, the declarative rule language includes rule scope, logical conditions, and violation weights; The atomic data unit set is traversed and filtered according to the rule scope to obtain a candidate unit group consisting of all data units applicable to the same rule scope; The logical conditions corresponding to the scope of the rule are attached to the corresponding candidate unit group, and the corresponding violation weights are assigned to all candidate unit groups according to the violation acceptance level of each rule, so as to obtain hard constraint instances and soft constraint instances containing rule conditions and violation costs; wherein, the hard constraint instance represents that the rule is prohibited from being violated, and the soft constraint instance represents that a cost must be paid when the rule is violated; All the constraint instances are stored in a constraint library used for managing constraint rules, resulting in a constraint instance set.

4. The method as described in claim 3, characterized in that, Based on the set of constraint instances, global cost optimization modeling and approximate solution processing are performed on the set of atomic data units to obtain the optimal cleaning strategy that minimizes the sum of the total modification cost and the total violation cost of the data unit set, including: For each element in the constraint instance set and the atomic data unit set, a binary decision variable is constructed to characterize the final cleaning result, so as to quantify all cleaning actions and violation handling status into computable mathematical variables. Based on the logical relationship between constraint instances and corresponding decision variables, an optimization decision model is established to characterize the minimization of the total cost of global data cleaning under multiple constraints. The optimal cleaning strategy that meets the rule requirements is calculated by performing linear programming relaxation and threshold rounding on the optimization decision model.

5. The method as described in claim 4, characterized in that, The step of establishing an optimization decision model based on the logical relationship between constraint instances and corresponding decision variables to characterize minimizing the total cost of global data cleaning under multiple constraints includes: Based on the logical relationship between hard constraint instances and decision variables, the first constraint condition is established as shown in the following inequality: In the formula, V represents the decision variable for the set of atomic data units; a value of 1 indicates deletion or modification of the unit, while a value of 0 indicates retention of the original value. V represents the constraint instance. This is an instance of a hard constraint; Based on the logical relationship between soft constraint instances and decision variables, the second constraint condition is established as shown in the following inequality: In the formula, For soft constraint instances, the decision variable is set to 1 to indicate that the violation has not been eliminated, and 0 to indicate that the violation has been eliminated. This is an instance of a soft constraint; To minimize the global cost of data cleaning, the objective function is established as shown in the following formula: In the formula, t is a data unit; T is the set of data units; Let be the cell weight of cell t; The violation weight of the soft constraint V.

6. The method as described in claim 1, characterized in that, After obtaining a clean set of atomic data cells that meet the preset quality requirements and a cleaning process report, the following is also included: Based on the cleaning process report, expert review and adjustment feedback were conducted to obtain optimized violation weight values ​​and new cleaning rules.

7. A global optimization device for multimodal data quality, characterized in that, The device includes: The processing module is used to perform data abstraction processing on the acquired raw multi-source heterogeneous data to obtain a set of atomic data units containing weights and unit relationships; The binding module is used to sequentially instantiate rules and bind weights to the set of atomic data units according to preset cleaning rules, so as to obtain a set of constraint instances that characterize the cleaning rules and violation costs of each type of data. The optimization module is used to perform global cost optimization modeling and approximate solution processing on the atomic data unit set based on the constraint instance set, so as to obtain the optimal cleaning strategy that minimizes the sum of the total modification cost and the total violation cost of the data unit set. The cleaning module is used to modify or delete corresponding units in the atomic data unit set according to the optimal cleaning strategy, so as to obtain a clean atomic data unit set and a cleaning process report that meet the preset quality requirements.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the steps of the method according to any one of claims 1-6.

Images