A Deep Learning-Based Multi-Dimensional Intelligent Optimization Method for Evaluation

Abstract

This invention discloses a multi-dimensional intelligent optimization method for evaluation based on deep learning, comprising the following steps: acquiring and preprocessing historical and current evaluation data to generate a multi-dimensional input feature set; constructing a hierarchical Pareto frontier learning network, with the bottom-level target sub-network outputting a set of target sub-representations for each evaluation dimension; executing multi-task gradient balancing and hypervolume maximization strategies in the middle-level frontier aggregator; performing frontier point screening and interpolation fusion on the Pareto frontier candidate set in the top-level decision layer to generate a comprehensive scoring result and contribution mapping; and outputting evaluation ranking and selection recommendations based on the comprehensive scoring result to generate an evaluation decision result set. This invention uses a hierarchical Pareto frontier learning network to achieve multi-dimensional intelligent optimization for evaluation, possessing advantages such as adaptive weights, interpretable results, and stable review decisions.

Description

Technical Field

[0001] This invention relates to the field of intelligent bidding evaluation, and in particular to a multi-dimensional intelligent optimization method for bidding evaluation based on deep learning. Background Technology

[0002] Existing evaluation techniques mostly employ weighted averages, hierarchical analysis, and rule base constraints. In recent years, deep learning-based text extraction and knowledge graph matching have also been introduced to assist in technical bid identification and scoring. However, these methods generally focus on a single objective or static weights, lacking collaborative modeling of multi-dimensional indicators such as price, technology, service, performance credit, and schedule within the network. They also struggle to adaptively handle objective conflicts during training and have not developed a non-dominated solution set generation and quality control mechanism oriented towards multiple objectives.

[0003] Within the aforementioned framework, common issues include the sensitivity of weights to changes in scenario and sample distribution, instability in learning due to mutual interference between gradients of different dimensions, insufficient coverage and uneven distribution of non-dominated solutions, and a lack of traceable explanation for the comprehensive score. Existing technologies have not yet established an end-to-end link from the bottom-level target sub-network—the mid-level front aggregator—to the top-level decision layer. They lack Pareto front approximation based on multi-task gradient balancing and hypervolume maximization, and they do not output comprehensive score results and contribution mappings at the top level through front point screening and interpolation fusion, resulting in insufficient objectivity and stability in ranking and selection recommendations.

[0004] Therefore, how to provide a multi-dimensional intelligent optimization method for evaluation based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a multi-dimensional intelligent optimization method for evaluation based on deep learning. This invention uses a hierarchical Pareto frontier learning network to achieve multi-dimensional intelligent optimization of evaluation, which has the advantages of adaptive weights, interpretable results, and stable evaluation decisions.

[0006] According to an embodiment of the present invention, a multi-dimensional intelligent optimization method for evaluation based on deep learning includes the following steps:

[0007] Acquire historical and current bidding data and unify their format, time, and semantics to generate a multi-dimensional input feature set;

[0008] A hierarchical Pareto frontier learning network is constructed, with the bottom target sub-network outputting a set of target sub-representations for each evaluation dimension;

[0009] A multi-task gradient balancing and hypervolume maximization strategy is executed in the mid-level front aggregator to generate an iteratively updated Pareto front candidate set.

[0010] At the top-level decision-making level, the Pareto frontier candidate set is screened and interpolated to generate a comprehensive score result and contribution mapping.

[0011] Based on the comprehensive scoring results, output the bid evaluation ranking and selection recommendation, and generate a bid evaluation decision result set.

[0012] Optionally, the generation of the multidimensional input feature set specifically includes:

[0013] Collect historical and current bid evaluation data, including bid documents, technical parameters, price records, service indicators, performance credit, schedule and expert scoring records, establish an original bid evaluation data set, and record the source identifier and time identifier;

[0014] The original bid evaluation data set was subjected to structured parsing and field unification. Numerical fields, category fields and text fields were standardized in format and aligned with time base to form a standardized field table.

[0015] The continuous numerical fields in the normalized field table are normalized, and the original values ​​of each field are linearly mapped to the closed interval between zero and one according to the minimum and maximum values, so as to obtain the numerical feature sequence.

[0016] The text fields in the normalized field table are segmented, terminology normalized, and semantically embedded. Multiple text embeddings of the same sample are aggregated into a unified representation to generate a semantic embedding feature sequence.

[0017] Anomaly detection and missing detection are performed on the numerical feature sequence and semantic embedded feature sequence. Abnormal samples are screened out, and linear interpolation is performed on the missing detection positions to generate a multidimensional feature sequence after consistency correction.

[0018] Based on the evaluation dimensions of price, technology, service, performance credit, and schedule cycle, the consistency-corrected multidimensional feature sequences are dimensionally mapped and indexed to construct a multidimensional input candidate set consisting of a subset of numerical features, a subset of categorical features, and a subset of semantic embeddings.

[0019] Optionally, the generation of the target sub-representation set specifically includes:

[0020] Using a multidimensional input feature set as network input, an input adapter for a hierarchical Pareto frontier learning network is established. Channel binding and time index alignment are performed on the numerical feature subset, the category feature subset, and the semantic embedding subset. Data type labeling, default placeholders, and order rearrangement are completed according to the predetermined field mapping relationship to generate adapted network input units.

[0021] In the hierarchical Pareto frontier learning network, a shared encoder of the underlying target subnetwork is configured to perform deep feature extraction on the adapted network input units to obtain a shared representation. The shared representation is uniformly generated by the shared encoder under multi-channel conditions.

[0022] The shared encoder consists of a multimodal input adaptation layer, a multi-layer feature extraction block, and a cross-modal fusion and compression layer. It extracts time series and semantic features through convolutional units and bidirectional gated recurrent units, achieves cross-modal fusion and feature weighting under a multi-head attention mechanism, maintains gradient stability through residual connections and layer normalization, and generates a shared representation.

[0023] In the underlying target sub-network, a dimension mapping module is configured for the evaluation dimensions of price, technology, service, performance credit and schedule cycle. The shared representation and the prior embedding of the corresponding dimension are gated and reweighted. The feature subspace representation of the dimension is generated by conditional affine transformation and intra-dimensional attention aggregation. After residual alignment and normalization, the dimension feature representation is output.

[0024] In the underlying target sub-network, target heads are configured for each evaluation dimension, nonlinear transformation and scale alignment are performed on the corresponding dimension feature representations, and dimension homogenization and time homogenization are performed to form a target sub-representation set.

[0025] Batch forward execution and stability constraint verification are performed on the underlying target subnetwork. The target head output distribution range, batch statistics and convergence flags of each evaluation dimension are recorded to form the runtime metadata of the underlying target subnetwork. Together with the target sub-representation set, it is written into the state register of the hierarchical Pareto frontier learning network, and a version number and time window identifier are established.

[0026] Optionally, the generation of the iteratively updated Pareto front candidate set specifically includes:

[0027] The target sub-representation set is input into the mid-level frontier aggregator of the hierarchical Pareto frontier learning network. Under a unified time index, the target sub-representations of each evaluation dimension are concatenated at the dimension level to form a multi-dimensional task feature matrix.

[0028] Multi-task gradient sampling is performed in the mid-level front aggregator. The gradient vector of the target sub-representation of each evaluation dimension is obtained based on the multi-dimensional task feature matrix, and the gradient angle and direction correlation between each task are calculated to form a task gradient distribution set.

[0029] Based on the task gradient distribution set, normalization and direction projection operations are performed on the gradient vectors of each evaluation dimension. Conflicting gradient directions are projected and corrected to obtain a set of gradient vectors after gradient balance.

[0030] By analyzing gradient direction, conflict degree and convergence trend, the weight factors of each evaluation dimension in the multi-task optimization process are dynamically calculated. The gradient vector set after gradient balance is multiplied with the weight factors and weighted summation is performed to generate a comprehensive gradient representation.

[0031] In the mid-level front aggregator, the comprehensive gradient representation is used as a constraint to calculate the objective function value of the evaluation dimension and map it to the multi-dimensional objective coordinate system. The hypervolume value of the multi-dimensional objective space is obtained by accumulating the increment of the objective function of each dimension.

[0032] Pareto front approximation is performed on the multidimensional task space based on the hypervolume maximization criterion. After calculating the hypervolume contribution of all candidate samples, the samples are sorted according to the hypervolume value. Non-dominated samples that are not dominated by other samples in all evaluation dimensions are selected as the Pareto front of the current round, forming a Pareto front candidate set.

[0033] For each candidate point in the Pareto front candidate set, calculate its sparsity, crowding and balance indices in the multidimensional target space, and filter out highly crowded samples according to the set threshold to obtain the front balance candidate set;

[0034] Dynamic sampling and updating are performed on the Pareto front equilibrium candidate set. The newly added candidate points are merged with the previous Pareto front point set and deduplicated to form an iteratively updated Pareto front point candidate set. At the same time, its hierarchical position and index identifier in the multidimensional target space are recorded.

[0035] Optionally, the generation of the mapping between the comprehensive scoring result and the contribution level specifically includes:

[0036] Receive the iteratively updated Pareto frontier candidate set and target sub-representation set, establish a bidirectional mapping relationship between frontier points and evaluation dimension indexes at the top-level decision layer, and form a frontier point index table to be decided;

[0037] Based on the index table of frontier points to be decided, the coordinates of each frontier point in terms of price, technology, service, performance credit and schedule cycle are extracted to form a multi-dimensional target coordinate set of frontier points, and the coordinates are registered with scale consistency.

[0038] Based on the multidimensional target coordinate set of the leading edge points, the interpolation coefficients are calculated according to the relative positions of the leading edge points in the multidimensional target coordinate system. The interpolation coefficients are normalized for the set of leading edge points corresponding to the same evaluation object to obtain the interpolation coefficient vector.

[0039] The weight factors obtained by the mid-level front aggregator during the multi-task optimization process are called to generate comprehensive score candidates by weighted combination of the multi-dimensional target coordinate set of the front points. The weighted combination first uses the interpolation coefficient of each front point as the outer weight, and then uses the weight factor of each evaluation dimension as the inner weight. The objective function values ​​of each front point in each evaluation dimension are weighted and summed to obtain comprehensive score candidates.

[0040] The stability check is performed on the comprehensive score candidates, including the boundary constraint check of the interpolation coefficient vector and the abnormal drift check of the weight factor. After the check is passed, it is recorded as the comprehensive score result, and the comprehensive score result and the corresponding interpolation coefficient vector and weight factor snapshot are registered in the top decision layer.

[0041] Based on the comprehensive scoring results and the multidimensional target coordinate set of the frontier points, the attribution ratio of each evaluation dimension to the comprehensive scoring results is calculated, and a contribution mapping is generated.

[0042] Optionally, the generation of the evaluation decision result set specifically includes:

[0043] The top-level decision-making level receives the comprehensive scoring results and contribution mapping, establishes the correspondence between the evaluation object identifier and the comprehensive scoring results and contribution mapping, and forms a list of objects to be ranked.

[0044] Construct a bid evaluation constraint rule base, which includes eligibility constraints, compliance boundary constraints, risk threshold constraints, and budget boundary constraints. Perform constraint verification on the list of objects to be ranked and output a constraint verification matrix and compliance pass flags.

[0045] The comprehensive score result is penalized and corrected based on the constraint verification matrix. Risk penalty factors are defined and associated with compliance through tags. For items that do not meet the constraints, the corresponding scores are deducted from the comprehensive score result according to the graded deduction rules to obtain the corrected comprehensive score result.

[0046] A dimensional preference weight vector is generated based on the contribution mapping, and the corrected comprehensive score is coupled with the dimensional preference weight vector to obtain a weighted score.

[0047] The weighted scores are subjected to stability checks, which include checking the fluctuation range of the interpolation coefficient vector boundary and the snapshot of the weight factor of the middle front aggregator, generating stability markers, and fine-tuning the consistency of the weighted scores to obtain the scores after stability checks.

[0048] The scores after stability verification are sorted in descending order to form the evaluation ranking. For evaluation objects with the same score, the set of dimensional contribution items in the contribution mapping is used to disambiguate the scores and output a unique evaluation ranking.

[0049] Based on the unique evaluation ranking and budget boundary constraints, quota boundary constraints and minimum compliance requirements, a winning recommendation is generated, including a shortlist, a reserve list and a rejection list. Each recommendation is then traceably linked to its corresponding comprehensive score result, the revised comprehensive score result, the weighted score and the stability marker.

[0050] The evaluation ranking and the winning recommendation are merged into an evaluation decision result set, which, together with the comprehensive scoring results, contribution mapping, constraint verification matrix, risk penalty factor, dimension preference weight vector, weight factor snapshot and stability label, are written into the decision archive of the top-level decision-making layer.

[0051] The beneficial effects of this invention are:

[0052] This invention constructs a hierarchical Pareto frontier learning network, establishing end-to-end information flow between the bottom-level target sub-network, the mid-level frontier aggregator, and the top-level decision layer, achieving adaptive coordination and non-dominated solution optimization for multi-dimensional evaluation objectives. Through multi-task gradient balancing and hypervolume maximization strategies, this invention can dynamically adjust the optimization direction and contribution ratio of each evaluation dimension, enabling efficient collaboration among multiple objectives such as price, technology, service, performance credit, and schedule cycle within the same network, overcoming the bias inherent in traditional static weight models when indicators conflict. Simultaneously, through frontier point selection and interpolation fusion mechanisms, this invention not only generates comprehensive scoring results in the multi-dimensional objective space but also outputs interpretable contribution mappings, clarifying the sources of influence of different dimensions on the final score, achieving transparency and traceability in the evaluation process. Overall, this method significantly outperforms existing technologies in terms of the objectivity, stability, and interpretability of evaluation results, providing an efficient and scalable deep learning solution for intelligent, multi-dimensional evaluation optimization. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is a flowchart of a multi-dimensional intelligent optimization method for evaluation based on deep learning proposed in this invention;

[0055] Figure 2 This diagram illustrates the iterative updating of the Pareto front candidate set generation for a deep learning-based multi-dimensional evaluation intelligent optimization method proposed in this invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figures 1-2 A multi-dimensional intelligent optimization method for evaluation based on deep learning includes the following steps:

[0058] Acquire historical and current bidding data and unify their format, time, and semantics to generate a multi-dimensional input feature set;

[0059] A hierarchical Pareto frontier learning network is constructed, with the bottom target sub-network outputting a set of target sub-representations for each evaluation dimension;

[0060] A multi-task gradient balancing and hypervolume maximization strategy is executed in the mid-level front aggregator to generate an iteratively updated Pareto front candidate set.

[0061] At the top-level decision-making level, the Pareto frontier candidate set is screened and interpolated to generate a comprehensive score result and contribution mapping.

[0062] Based on the comprehensive scoring results, output the bid evaluation ranking and selection recommendation, and generate a bid evaluation decision result set.

[0063] In this embodiment, the generation of the multidimensional input feature set specifically includes:

[0064] Collect historical and current bid evaluation data, including bid documents, technical parameters, price records, service indicators, performance credit, schedule and expert scoring records, establish an original bid evaluation data set, and record the source identifier and time identifier;

[0065] The original bid evaluation data set was subjected to structured parsing and field unification. Numerical fields, category fields and text fields were standardized in format and aligned with time base to form a standardized field table.

[0066] The continuous numerical fields in the normalized field table are normalized, and the original values ​​of each field are linearly mapped to the closed interval between zero and one according to the minimum and maximum values, so as to obtain the numerical feature sequence.

[0067] The text fields in the normalized field table are segmented, terminology normalized, and semantically embedded. Multiple text embeddings of the same sample are aggregated into a unified representation to generate a semantic embedding feature sequence.

[0068] Anomaly detection and missing detection are performed on the numerical feature sequence and semantic embedded feature sequence. Abnormal samples are screened out, and linear interpolation is performed on the missing detection positions to generate a multidimensional feature sequence after consistency correction.

[0069] Based on the evaluation dimensions of price, technology, service, performance credit, and schedule cycle, the consistency-corrected multidimensional feature sequences are dimensionally mapped and indexed to construct a multidimensional input candidate set consisting of a subset of numerical features, a subset of categorical features, and a subset of semantic embeddings.

[0070] In this embodiment, the generation of the target sub-representation set specifically includes:

[0071] Using a multidimensional input feature set as network input, an input adapter for a hierarchical Pareto frontier learning network is established. Channel binding and time index alignment are performed on the numerical feature subset, the category feature subset, and the semantic embedding subset. Data type labeling, default placeholders, and order rearrangement are completed according to the predetermined field mapping relationship to generate adapted network input units.

[0072] In the hierarchical Pareto frontier learning network, a shared encoder of the underlying target subnetwork is configured to perform deep feature extraction on the adapted network input units to obtain a shared representation. The shared representation is uniformly generated by the shared encoder under multi-channel conditions.

[0073] The shared encoder consists of a multimodal input adaptation layer, a multi-layer feature extraction block, and a cross-modal fusion and compression layer. It extracts time series and semantic features through convolutional units and bidirectional gated recurrent units, achieves cross-modal fusion and feature weighting under a multi-head attention mechanism, maintains gradient stability through residual connections and layer normalization, and generates a shared representation.

[0074] In the underlying target sub-network, a dimension mapping module is configured for the evaluation dimensions of price, technology, service, performance credit and schedule cycle. The shared representation and the prior embedding of the corresponding dimension are gated and reweighted. The feature subspace representation of the dimension is generated by conditional affine transformation and intra-dimensional attention aggregation. After residual alignment and normalization, the dimension feature representation is output.

[0075] In the underlying target sub-network, target heads are configured for each evaluation dimension, nonlinear transformation and scale alignment are performed on the corresponding dimension feature representations, and dimension homogenization and time homogenization are performed to form a target sub-representation set.

[0076] Batch forward execution and stability constraint verification are performed on the underlying target subnetwork. The target head output distribution range, batch statistics and convergence flags of each evaluation dimension are recorded to form the runtime metadata of the underlying target subnetwork. Together with the target sub-representation set, it is written into the state register of the hierarchical Pareto frontier learning network, and a version number and time window identifier are established.

[0077] In this embodiment, the generation of the iteratively updated Pareto front candidate set specifically includes:

[0078] The target sub-representation set is input into the mid-level frontier aggregator of the hierarchical Pareto frontier learning network. Under a unified time index, the target sub-representations of each evaluation dimension are concatenated at the dimension level to form a multi-dimensional task feature matrix. The multi-dimensional task feature matrix retains the task independence of evaluation dimensions such as price, technology, service, performance credit and schedule cycle, and establishes feature correspondence on the dimension index axis.

[0079] Multi-task gradient sampling is performed in the mid-level front aggregator. The gradient vector of the target sub-representation of each evaluation dimension is obtained based on the multi-dimensional task feature matrix, and the gradient angle and direction correlation between each task are calculated to form a task gradient distribution set.

[0080] Based on the task gradient distribution set, normalization and direction projection operations are performed on the gradient vectors of each evaluation dimension. Conflicting gradient directions are projected and corrected to obtain a set of gradient vectors after gradient balance.

[0081] By analyzing gradient direction, conflict degree and convergence trend, the weight factors of each evaluation dimension in the multi-task optimization process are dynamically calculated. The gradient vector set after gradient balance is multiplied with the weight factors and weighted summation is performed to generate a comprehensive gradient representation.

[0082] In the mid-level front aggregator, the comprehensive gradient representation is used as a constraint to calculate the objective function value of the evaluation dimension and map it to the multi-dimensional objective coordinate system. The hypervolume value of the multi-dimensional objective space is obtained by accumulating the increment of the objective function of each dimension.

[0083] Pareto front approximation is performed on the multidimensional task space based on the hypervolume maximization criterion. After calculating the hypervolume contribution of all candidate samples, the samples are sorted according to the hypervolume value. Non-dominated samples that are not dominated by other samples in all evaluation dimensions are selected as the Pareto front of the current round, forming a Pareto front candidate set.

[0084] For each candidate point in the Pareto front candidate set, calculate its sparsity, crowding and balance indices in the multidimensional target space, and filter out highly crowded samples according to the set threshold to obtain the front balance candidate set;

[0085] The sparsity is obtained by calculating the average distance between the front edge point and its neighboring points in each evaluation dimension; the crowding is defined by the ratio of the number of samples in the neighborhood of the front edge point to the neighborhood volume; and the balance is obtained by comparing the ratio of the variance to the covariance of the objective function in each evaluation dimension.

[0086] Dynamic sampling and updating are performed on the Pareto front equilibrium candidate set. The newly added candidate points are merged with the previous Pareto front point set and deduplicated to form an iteratively updated Pareto front point candidate set. At the same time, its hierarchical position and index identifier in the multidimensional target space are recorded.

[0087] In this embodiment, the generation of the comprehensive scoring result and the contribution mapping specifically includes:

[0088] Receive the iteratively updated Pareto frontier candidate set and target sub-representation set, establish a bidirectional mapping relationship between frontier points and evaluation dimension indexes at the top-level decision layer, and form a frontier point index table to be decided;

[0089] Based on the index table of frontier points to be decided, the coordinates of each frontier point in terms of price, technology, service, performance credit and schedule cycle are extracted to form a multi-dimensional target coordinate set of frontier points, and the coordinates are registered with scale consistency.

[0090] Based on the multidimensional target coordinate set of the leading edge points, the interpolation coefficients are calculated according to the relative positions of the leading edge points in the multidimensional target coordinate system. The interpolation coefficients are normalized for the set of leading edge points corresponding to the same evaluation object to obtain the interpolation coefficient vector.

[0091] The weight factors obtained by the mid-level front aggregator during the multi-task optimization process are called to generate comprehensive score candidates by weighted combination of the multi-dimensional target coordinate set of the front points. The weighted combination first uses the interpolation coefficient of each front point as the outer weight, and then uses the weight factor of each evaluation dimension as the inner weight. The objective function values ​​of each front point in each evaluation dimension are weighted and summed to obtain comprehensive score candidates.

[0092] The stability check is performed on the comprehensive score candidates, including the boundary constraint check of the interpolation coefficient vector and the abnormal drift check of the weight factor. After the check is passed, it is recorded as the comprehensive score result, and the comprehensive score result and the corresponding interpolation coefficient vector and weight factor snapshot are registered in the top decision layer.

[0093] Based on the comprehensive scoring results and the multidimensional target coordinate set of the frontier points, the attribution ratio of each evaluation dimension to the comprehensive scoring results is calculated, and a contribution mapping is generated.

[0094] The attribution percentage is calculated by using the sum of the objective function values ​​of the evaluation dimension weighted by the weighting factors of one evaluation dimension and the interpolation coefficients of all frontier points as the numerator, and the sum of the objective function values ​​of each evaluation dimension weighted by the weighting factors of all evaluation dimensions and the interpolation coefficients of all frontier points as the denominator, thus obtaining the attribution percentage of that evaluation dimension.

[0095] In this embodiment, the generation of the evaluation decision result set specifically includes:

[0096] The top-level decision-making level receives the comprehensive scoring results and contribution mapping, establishes the correspondence between the evaluation object identifier and the comprehensive scoring results and contribution mapping, and forms a list of objects to be ranked.

[0097] Construct a bid evaluation constraint rule base, which includes eligibility constraints, compliance boundary constraints, risk threshold constraints, and budget boundary constraints. Perform constraint verification on the list of objects to be ranked and output a constraint verification matrix and compliance pass flags.

[0098] The comprehensive score result is penalized and corrected based on the constraint verification matrix. Risk penalty factors are defined and associated with compliance through tags. For items that do not meet the constraints, the corresponding scores are deducted from the comprehensive score result according to the graded deduction rules to obtain the corrected comprehensive score result.

[0099] A dimensional preference weight vector is generated based on the contribution mapping, and the corrected comprehensive score is coupled with the dimensional preference weight vector to obtain a weighted score.

[0100] The weighted scores are subjected to stability checks, which include checking the fluctuation range of the interpolation coefficient vector boundary and the snapshot of the weight factor of the middle front aggregator, generating stability markers, and fine-tuning the consistency of the weighted scores to obtain the scores after stability checks.

[0101] The scores after stability verification are sorted in descending order to form the evaluation ranking. For evaluation objects with the same score, the set of dimensional contribution items in the contribution mapping is used to disambiguate the scores and output a unique evaluation ranking.

[0102] Based on the unique evaluation ranking and budget boundary constraints, quota boundary constraints and minimum compliance requirements, a winning recommendation is generated, including a shortlist, a reserve list and a rejection list. Each recommendation is then traceably linked to its corresponding comprehensive score result, the revised comprehensive score result, the weighted score and the stability marker.

[0103] The evaluation ranking and the winning recommendation are merged into an evaluation decision result set, which, together with the comprehensive scoring results, contribution mapping, constraint verification matrix, risk penalty factor, dimension preference weight vector, weight factor snapshot and stability label, are written into the decision archive of the top-level decision-making layer.

[0104] Example 1:

[0105] To verify the feasibility of this invention in practice, it was applied to a comprehensive bidding evaluation center in the construction sector of a coastal province. This center handles a large number of bidding evaluations annually, covering equipment procurement, construction contracting, and operation and maintenance services. The evaluation dimensions are diverse, including not only price factors but also technical solutions, service response, performance credibility, and schedule. In the past, the weights of each indicator were manually set during the review process, leading to significant differences in the scoring preferences of different experts. This resulted in significant fluctuations in the overall results, with some review results exhibiting strong subjectivity, unbalanced weights, and interference between different dimensions, making it difficult to form a fair, stable, and interpretable basis for decision-making.

[0106] After deploying the proposed deep learning-based multi-dimensional intelligent optimization method for bid evaluation on the center's review platform, the system first collects review data from previous reviews via a data interface, including tender documents, price records, technical parameters, performance records, and expert scores. Format standardization and semantic correction are then performed to generate a multi-dimensional input feature set. The platform subsequently constructs a hierarchical Pareto frontier learning network. The bottom-level target sub-network extracts deep features for each evaluation dimension, forming dimensional feature representations. The middle-level frontier aggregator balances gradient conflicts among multiple objectives, dynamically generating a Pareto frontier candidate set using the hypervolume maximization criterion. The top-level decision layer then obtains the comprehensive scoring result and dimensional contribution mapping through frontier point selection and interpolation fusion, thereby achieving automation, transparency, and traceability of the bid evaluation process.

[0107] During a three-month review process, this method was applied to the review of multiple batches of construction projects and equipment procurement, covering evaluation tasks of varying scale and complexity. Comparison with the manual weighted average method and traditional deep scoring models showed that the stability of the comprehensive scoring results was significantly improved, and the differences between repeated reviews were significantly reduced. Simultaneously, evaluation experts could clearly understand the specific impact of price, technology, and service dimensions on the comprehensive results through contribution mapping, making the review conclusions more convincing. During continuous operation, the system continuously updates its leading edge position through an incremental learning mechanism, enabling the model to adapt to new data characteristics and achieve rapid response to market changes and review standards. Comprehensive evaluation indicates that this invention effectively solves the problems of traditional evaluation methods, such as difficulty in adapting weights, strong subjectivity of results, and insufficient interpretability, achieving intelligent and objective evaluation processes and providing reliable technical support for multi-dimensional comprehensive decision-making.

[0108] Table 1. Performance comparison between deep learning-based multi-dimensional evaluation intelligent optimization method and traditional methods.

[0109] Comparison Methods Traditional weighted average method Based on deep scoring model This invention Overall score stability (volatility%) 14.8 9.7 4.3 Weighted adaptive accuracy (%) 62.3 78.9 92.7 Dimensional correlation interference (lower is better) 0.46 0.33 0.18 Explainability score (out of 10) 4.2 6.1 9.3 Review consistency (expert review consistency rate %) 71.5 83.2 94.5 Average review time (minutes) 38 29 24

[0110] As shown in Table 1, in this comparative experiment, the traditional weighted average method, the general deep scoring model, and the hierarchical Pareto frontier learning network proposed in this invention were selected for testing on multi-dimensional evaluation tasks in the same batch. The results in the table show that this invention achieves significant advantages in all key indicators, especially in terms of overall scoring stability and weight adaptive accuracy. Traditional methods rely on fixed weights and are easily affected by differences in review samples and expert subjectivity, resulting in a scoring volatility as high as 14.8%. In contrast, this invention, by introducing a mid-level frontier aggregator for multi-task gradient balancing, reduces the volatility to 4.3%, indicating that the model possesses good convergence stability and evaluation consistency.

[0111] At the weight learning level, traditional methods require manual determination of dimensional importance and cannot dynamically adjust based on data changes. While depth-based scoring models can learn some patterns, they fail to establish non-dominated optimization relationships among multiple objectives. This invention utilizes Pareto front approximation and hypervolume maximization strategies to create a dynamic game among dimensions during training, achieving adaptive weight allocation and improving weight learning accuracy to 92.7%. Furthermore, this invention reduces dimensionality-related interference by approximately 45% compared to traditional models, indicating its effective elimination of gradient conflicts and redundant effects among indicators in a multi-objective space.

[0112] Regarding interpretability and review consistency, thanks to the frontier point interpolation fusion and contribution mapping mechanism at the top-level decision-making layer, experts can intuitively understand the impact of different dimensions on the final comprehensive score, thereby enhancing scoring transparency and review consistency. The interpretability score improved to 9.3 points, and the expert consistency rate reached 94.5%, demonstrating the high trustworthiness of this method in practical reviews. Meanwhile, because the model implements parallel feature extraction and gradient optimization, the average review time is shortened to 24 minutes, saving approximately one-third of the time compared to traditional processes. In summary, this invention achieves significant performance improvements in stability, intelligence, and interpretability, providing a reliable intelligent optimization approach for complex multi-dimensional evaluation.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-dimensional intelligent optimization method for evaluation based on deep learning, characterized in that, Includes the following steps: Acquire historical and current bidding data and unify their format, time, and semantics to generate a multi-dimensional input feature set; Collect historical and current bid evaluation data, including bid documents, technical parameters, price records, service indicators, performance credit, schedule and expert scoring records, establish an original bid evaluation data set, and record the source identifier and time identifier; A hierarchical Pareto frontier learning network is constructed, with the bottom target sub-network outputting a set of target sub-representations for each evaluation dimension; A multi-task gradient balancing and hypervolume maximization strategy is executed in the mid-level front aggregator to generate an iteratively updated Pareto front candidate set. At the top-level decision-making level, the Pareto frontier candidate set is screened and interpolated to generate a comprehensive score result and contribution mapping. Based on the comprehensive scoring results, output the bid evaluation ranking and selection recommendation, and generate a bid evaluation decision result set; The generation of the iteratively updated Pareto front candidate set specifically includes: The target sub-representation set is input into the mid-level frontier aggregator of the hierarchical Pareto frontier learning network. Under a unified time index, the target sub-representations of each evaluation dimension are concatenated at the dimension level to form a multi-dimensional task feature matrix. Multi-task gradient sampling is performed in the mid-level front aggregator. The gradient vector of the target sub-representation of each evaluation dimension is obtained based on the multi-dimensional task feature matrix, and the gradient angle and direction correlation between each task are calculated to form a task gradient distribution set. Based on the task gradient distribution set, normalization and direction projection operations are performed on the gradient vectors of each evaluation dimension. Conflicting gradient directions are projected and corrected to obtain a set of gradient vectors after gradient balance. By analyzing gradient direction, conflict degree and convergence trend, the weight factors of each evaluation dimension in the multi-task optimization process are dynamically calculated. The gradient vector set after gradient balance is multiplied with the weight factors and weighted summation is performed to generate a comprehensive gradient representation. In the mid-level front aggregator, the comprehensive gradient representation is used as a constraint to calculate the objective function value of the evaluation dimension and map it to the multi-dimensional objective coordinate system. The hypervolume value of the multi-dimensional objective space is obtained by accumulating the increment of the objective function of each dimension. Pareto front approximation is performed on the multidimensional task space based on the hypervolume maximization criterion. After calculating the hypervolume contribution of all candidate samples, the samples are sorted according to the hypervolume value. Non-dominated samples that are not dominated by other samples in all evaluation dimensions are selected as the Pareto front of the current round, forming a Pareto front candidate set. For each candidate point in the Pareto front candidate set, calculate its sparsity, crowding and balance indices in the multidimensional target space, and filter out highly crowded samples according to the set threshold to obtain the front balance candidate set; Dynamic sampling and updating are performed on the Pareto front equilibrium candidate set. The newly added candidate points are merged with the previous Pareto front point set and deduplicated to form an iteratively updated Pareto front point candidate set. At the same time, its hierarchical position and index identifier in the multidimensional target space are recorded.

2. The multi-dimensional intelligent optimization method for evaluation based on deep learning according to claim 1, characterized in that, The generation of the multidimensional input feature set specifically includes: The original bid evaluation data set was subjected to structured parsing and field unification. Numerical fields, category fields and text fields were standardized in format and aligned with time base to form a standardized field table. The continuous numerical fields in the normalized field table are normalized, and the original values ​​of each field are linearly mapped to the closed interval between zero and one according to the minimum and maximum values, so as to obtain the numerical feature sequence. The text fields in the normalized field table are segmented, terminology normalized, and semantically embedded. Multiple text embeddings of the same sample are aggregated into a unified representation to generate a semantic embedding feature sequence. Anomaly detection and missing detection are performed on the numerical feature sequence and semantic embedded feature sequence. Abnormal samples are screened out, and linear interpolation is performed on the missing detection positions to generate a multidimensional feature sequence after consistency correction. Based on the evaluation dimensions of price, technology, service, performance credit, and schedule cycle, the consistency-corrected multidimensional feature sequences are dimensionally mapped and indexed to construct a multidimensional input candidate set consisting of a subset of numerical features, a subset of categorical features, and a subset of semantic embeddings.

3. The multi-dimensional intelligent optimization method for evaluation based on deep learning according to claim 1, characterized in that, The generation of the target sub-representation set specifically includes: Using a multidimensional input feature set as network input, an input adapter for a hierarchical Pareto frontier learning network is established. Channel binding and time index alignment are performed on the numerical feature subset, the category feature subset, and the semantic embedding subset. Data type labeling, default placeholders, and order rearrangement are completed according to the predetermined field mapping relationship to generate adapted network input units. In the hierarchical Pareto frontier learning network, a shared encoder of the underlying target subnetwork is configured to perform deep feature extraction on the adapted network input units to obtain a shared representation. The shared representation is uniformly generated by the shared encoder under multi-channel conditions. The shared encoder consists of a multimodal input adaptation layer, a multi-layer feature extraction block, and a cross-modal fusion and compression layer. It extracts time series and semantic features through convolutional units and bidirectional gated recurrent units, achieves cross-modal fusion and feature weighting under a multi-head attention mechanism, maintains gradient stability through residual connections and layer normalization, and generates a shared representation. In the underlying target sub-network, a dimension mapping module is configured for the evaluation dimensions of price, technology, service, performance credit and schedule cycle. The shared representation and the prior embedding of the corresponding dimension are gated and reweighted. The feature subspace representation of the dimension is generated by conditional affine transformation and intra-dimensional attention aggregation. After residual alignment and normalization, the dimension feature representation is output. In the underlying target sub-network, target heads are configured for each evaluation dimension, nonlinear transformation and scale alignment are performed on the corresponding dimension feature representations, and dimension homogenization and time homogenization are performed to form a target sub-representation set. Batch forward execution and stability constraint verification are performed on the underlying target subnetwork. The target head output distribution range, batch statistics and convergence flags of each evaluation dimension are recorded to form the runtime metadata of the underlying target subnetwork. Together with the target sub-representation set, it is written into the state register of the hierarchical Pareto frontier learning network, and a version number and time window identifier are established.

4. The multi-dimensional intelligent optimization method for evaluation based on deep learning according to claim 1, characterized in that, The generation of the comprehensive scoring result and the contribution mapping specifically includes: Receive the iteratively updated Pareto frontier candidate set and target sub-representation set, establish a bidirectional mapping relationship between frontier points and evaluation dimension indexes at the top-level decision layer, and form a frontier point index table to be decided; Based on the index table of frontier points to be decided, the coordinates of each frontier point in terms of price, technology, service, performance credit and schedule cycle are extracted to form a multi-dimensional target coordinate set of frontier points, and the coordinates are registered with scale consistency. Based on the multidimensional target coordinate set of the leading edge points, the interpolation coefficients are calculated according to the relative positions of the leading edge points in the multidimensional target coordinate system. The interpolation coefficients are normalized for the set of leading edge points corresponding to the same evaluation object to obtain the interpolation coefficient vector. The weight factors obtained by the mid-level front aggregator during the multi-task optimization process are called to generate comprehensive score candidates by weighted combination of the multi-dimensional target coordinate set of the front points. The weighted combination first uses the interpolation coefficient of each front point as the outer weight, and then uses the weight factor of each evaluation dimension as the inner weight. The objective function values ​​of each front point in each evaluation dimension are weighted and summed to obtain comprehensive score candidates. The stability check is performed on the comprehensive score candidates, including the boundary constraint check of the interpolation coefficient vector and the abnormal drift check of the weight factor. After the check is passed, it is recorded as the comprehensive score result, and the comprehensive score result and the corresponding interpolation coefficient vector and weight factor snapshot are registered in the top decision layer. Based on the comprehensive scoring results and the multidimensional target coordinate set of the frontier points, the attribution ratio of each evaluation dimension to the comprehensive scoring results is calculated, and a contribution mapping is generated.

5. The multi-dimensional intelligent optimization method for evaluation based on deep learning according to claim 1, characterized in that, The generation of the evaluation decision result set specifically includes: The top-level decision-making level receives the comprehensive scoring results and contribution mapping, establishes the correspondence between the evaluation object identifier and the comprehensive scoring results and contribution mapping, and forms a list of objects to be ranked. Construct a bid evaluation constraint rule base, which includes eligibility constraints, compliance boundary constraints, risk threshold constraints, and budget boundary constraints. Perform constraint verification on the list of objects to be ranked and output a constraint verification matrix and compliance pass flags. The comprehensive score result is penalized and corrected based on the constraint verification matrix. Risk penalty factors are defined and associated with compliance through tags. For items that do not meet the constraints, the corresponding scores are deducted from the comprehensive score result according to the graded deduction rules to obtain the corrected comprehensive score result. A dimensional preference weight vector is generated based on the contribution mapping, and the corrected comprehensive score is coupled with the dimensional preference weight vector to obtain a weighted score. The weighted scores are subjected to stability checks, which include checking the fluctuation range of the interpolation coefficient vector boundary and the snapshot of the weight factor of the middle front aggregator, generating stability markers, and fine-tuning the consistency of the weighted scores to obtain the scores after stability checks. The scores after stability verification are sorted in descending order to form the evaluation ranking. For evaluation objects with the same score, the set of dimensional contribution items in the contribution mapping is used to disambiguate the scores and output a unique evaluation ranking. Based on the unique evaluation ranking and budget boundary constraints, quota boundary constraints and minimum compliance requirements, a winning recommendation is generated, including a shortlist, a reserve list and a rejection list. Each recommendation is then traceably linked to its corresponding comprehensive score result, the revised comprehensive score result, the weighted score and the stability marker. The evaluation ranking and the winning recommendation are merged into an evaluation decision result set, which, together with the comprehensive scoring results, contribution mapping, constraint verification matrix, risk penalty factor, dimension preference weight vector, weight factor snapshot and stability label, are written into the decision archive of the top-level decision-making layer.

Images