An industrial algorithm mall-oriented cross-modal retrieval and accurate recommendation method

Abstract

This invention discloses a cross-modal retrieval and precise recommendation method for industrial algorithm marketplaces, comprising: collecting candidate algorithm data and user behavior data; generating algorithm feature vectors and user demand feature vectors, mapping them to generate standardized input feature vectors; constructing algorithm call adjacency sequences and algorithm industry association sequences; constructing an improved liquid state machine model and setting connection paths for the industry association sequences; constructing and optimizing an adaptive differential evolution algorithm model; inputting the standardized feature vectors into the optimized liquid state machine model to generate dynamic semantic state representations; performing similarity matching between the query dynamic state representations and the dynamic semantic state representations of candidate algorithms to generate initial retrieval results; reweighting and outputting personalized recommendation results. This invention achieves efficient algorithm model optimization for cross-modal retrieval and precise recommendation, improves recommendation accuracy and personalized services, and is applicable to intelligent recommendation scenarios in multiple fields.

Description

Technical Field

[0001] This invention relates to the field of industrial algorithm retrieval and intelligent recommendation technology, and in particular to a cross-modal retrieval and accurate recommendation method for industrial algorithm marketplaces. Background Technology

[0002] Industrial algorithm marketplaces aggregate a large number of algorithm resources. Common organizational methods rely on algorithm description texts, tag classifications, and simple call relationships for index management. The retrieval process often uses keyword matching or vector similarity calculations for initial screening, and the recommendation process is mainly based on user historical behavior or static rules for ranking. Existing systems, in terms of multimodal data processing, typically extract features from text, images, and structured information separately and then directly concatenate them, lacking deep relational modeling within a unified semantic space. Regarding algorithm relationship modeling, most systems only utilize call counts or tag co-occurrence to construct simple associations, failing to form a structured expression that reflects complex dependencies. In terms of model computation, conventional neural networks or vector retrieval models struggle to simultaneously characterize the topological relationships and multi-layered semantic information between algorithms, resulting in limited semantic representation capabilities.

[0003] Existing solutions often employ fixed parameters or single optimization strategies during model optimization, lacking a hierarchical optimization mechanism to address differences in model structure, making it difficult to coordinate adjustments among structural, connectivity, and state parameters. In the retrieval and ranking stage, ranking is typically based on single similarity or simple weighting, without considering algorithm call relationships, industry relevance, and user preferences for joint modeling, resulting in insufficient response to complex needs. In the re-ranking stage, there is a lack of coupled analysis and dynamic correction mechanisms for multi-source information, making it difficult to achieve unified optimization of personalization and structural relevance.

[0004] Therefore, how to provide a cross-modal retrieval and accurate recommendation method for industrial algorithm marketplaces 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 cross-modal retrieval and accurate recommendation method for industrial algorithm marketplaces. This invention constructs an improved liquid state machine model and combines it with an adaptive differential evolution optimization strategy to achieve multimodal feature fusion and algorithm relationship modeling, and completes dynamic semantic matching and reweighted ranking of retrieval results. It has the advantages of strong semantic expression ability, full utilization of structural associations and high recommendation accuracy.

[0006] A cross-modal retrieval and accurate recommendation method for an industrial algorithm marketplace according to an embodiment of the present invention includes: Collect candidate algorithm data and user behavior data from the industrial algorithm marketplace; Generate algorithm feature vectors from candidate algorithm data, generate user demand feature vectors from user behavior data, and map the algorithm feature vectors and user demand feature vectors to a unified dimensional space to generate standardized input feature vectors. Construct an algorithm call adjacency sequence and an algorithm industry association sequence based on candidate algorithm data; An improved liquid state machine model is constructed, including a ring-cluster topology reservoir structure, a shell input region, a core input region, and a heterogeneous neuron group structure. Based on the algorithm, the cross-cluster jump connection path in the ring-cluster topology reservoir structure is set, and based on the algorithm, the penetration connection path from the shell input region to the core input region is set. An adaptive differential evolution algorithm is used to optimize the improved liquid state machine model. During the differential evolution iteration process, the crossover probability and mutation scaling ratio are dynamically adjusted. A multi-level information fusion mechanism is applied to optimize different levels of structure independently, and the results are fused to dynamically adjust the optimization weights of each level. The standardized input feature vector is fed into the optimized liquid state machine model to generate the dynamic semantic state representation of the candidate algorithm, and the query dynamic state representation is generated based on the user's retrieval request. The query dynamic state representation is matched with the candidate algorithm's dynamic semantic state representation to generate initial search results; The initial search results are reweighted to generate personalized recommendation results, which are then output to the industrial algorithm marketplace.

[0007] Optionally, the candidate algorithm data includes algorithm description text, demonstration images, interface parameters, application cases, industry tags, algorithm call relationship data, and unique algorithm identifiers; the user behavior data includes user search requests, uploaded images, filtering conditions, browsing history, click history, favorites history, and call history.

[0008] Optionally, generating the standardized input feature vector includes: Multimodal feature extraction is performed on the candidate algorithm data to generate text feature vectors, image feature vectors and structured label vectors. The fusion result of text feature vectors and image feature vectors is defined as outer semantic information, and the structured label vectors are defined as core semantic information. Perform behavioral feature modeling on user behavior data to generate user demand feature vectors; The outer semantic information and the core semantic information are concatenated in a preset order to generate an algorithm feature vector; The algorithm feature vectors are mapped to user-required feature vectors in a unified dimension to generate standardized input feature vectors.

[0009] Optionally, constructing the algorithm call adjacency sequence and algorithm industry association sequence based on candidate algorithm data includes the following steps: Based on the algorithm call relationships in the candidate algorithm data, we construct the call relationships between algorithms and generate an algorithm call adjacency sequence, where the call relationships are weighted by call strength and time decay; Based on the industry tag information in the candidate algorithm data, we construct industry association relationships between algorithms and generate an algorithm industry association sequence. The algorithm call adjacency sequence and the algorithm industry association sequence are indexed and aligned according to the algorithm identifier.

[0010] Optionally, constructing the improved liquid state machine model includes the following steps: The basic structure of the improved liquid state machine model is initialized, a ring cluster topology reservoir is constructed, neurons are divided into multiple ring cluster units, unidirectional ring connection paths are established in a fixed order within the unit, and a unique identifier is assigned to each neuron. Establish a mapping rule from algorithm nodes to neurons: map each algorithm identifier in the algorithm call adjacency sequence to a unique neuron node in the reserve pool, so that the algorithm call relationship is transformed into the connection relationship between neurons; In the reserve pool, a cross-cluster connection candidate set is constructed. All neuron pairs between different ring cluster units are traversed. Based on the adjacency algorithm and call relationship weight in the algorithm call adjacency sequence, algorithm pairs with weights greater than a preset threshold are selected and mapped as cross-cluster jump connection paths. The corresponding cross-cluster connection edges are filtered and retained from the cross-cluster connection candidate set. After linear normalization according to the call relationship weight, the connection strength value is assigned and written into the weight parameter. Construct a shell input region and a kernel input region. The shell input region receives the outer semantic information from the standardized input feature vector, and the kernel input region receives the core semantic information from the standardized input feature vector. Assign an identifier to each input node. Based on the industry association algorithms and weights in the industry association sequence, algorithm pairs with weights greater than the median are selected and mapped to penetrating connection paths from the shell input region to the kernel input region. Penetrating connection edges are established, and after normalization according to the industry association weight interval, connection strength values ​​are assigned and written into the weight parameters. A heterogeneous neuron group structure is constructed in the ring-cluster topology reservoir. The standardized input feature vectors are assigned to the corresponding input nodes, and the input connection relationship between the input nodes and the reservoir neurons is established to form a complete liquid state machine model structure.

[0011] Optionally, the model optimization using the adaptive differential evolution algorithm includes the following steps: The traditional parameter optimization is improved into structure-aware optimization. The parameters to be optimized are divided into structural layer parameters, connection layer parameters, and neuron state layer parameters. The parameters are arranged in order of their position in the model structure to generate a parameter sequence, and the parameter sequence is encoded with real numbers to form an initial population. The parameter sequences in the initial population are input into the liquid state machine model, and dynamic semantic state representations of candidate algorithms are generated based on the standardized input feature vectors. These representations are then matched with the corresponding reference semantic state representations to obtain fitness values. An improved structural divide-and-conquer search and coupling optimization is adopted. During the differential evolution iteration process, hierarchical differential mutation operation and hierarchical crossover operation are performed on the structural layer parameters, connection layer parameters and neuron state layer parameters respectively. A collaborative update relationship is established between different parameter layers so that each parameter layer is updated synchronously in the same iteration round. The fitness values ​​of the candidate sequences generated by the hierarchical generation are calculated and compared with the fitness values ​​of the original parameter sequences. When the fitness value of the candidate sequence is better than that of the original parameter sequence, it is replaced to form the next generation of population. The fitness change trend of each parameter layer is recorded during the iteration process. Based on the fitness change trend of each parameter layer, an adaptive strategy is executed to adjust and a multi-level information fusion mechanism is introduced. The fitness contribution values ​​of structural layer parameters, connection layer parameters and neuron state layer parameters are calculated respectively. The overall fitness is weighted and fused based on the fitness contribution values ​​of each parameter layer to obtain the overall optimization evaluation value. When the overall optimization evaluation value reaches the preset termination condition, the parameter sequence with the best fitness in the current population is output, and the parameter sequence is written into the liquid state machine model to obtain the optimized liquid state machine model.

[0012] Optionally, the dynamic semantic state representation of the candidate generation algorithm includes the following steps: The standardized input feature vectors are grouped according to the algorithm and input into the liquid state machine model. The outer semantic information is input into the shell input region, the core semantic information is input into the kernel input region, and then passed to the ring cluster topology reservoir. State updates and cross-cluster propagation are performed in the ring-cluster topology reserve pool to generate dynamic semantic state representations of candidate algorithms; Feature extraction is performed on the user's search request and input into the liquid state machine model to generate a dynamic semantic state representation of the query; The dynamic semantic state representations of candidate algorithms and query dynamic semantic state representations are output along a unified dimension.

[0013] Optionally, generating the initial search results includes the following steps: The dynamic semantic state representations of the candidate algorithms and the dynamic semantic state representations of the queries are dimensionally aligned, and the Euclidean distance and cosine similarity are calculated and fused to obtain a comprehensive similarity value. Candidate algorithms are ranked based on comprehensive similarity values ​​and filtered through a preset threshold to generate a candidate selection set; The density value of the candidate algorithm is calculated based on the similarity neighborhood in the candidate screening set, and then combined with the comprehensive similarity value for fusion and ranking. Output the sorted list of candidate algorithms and generate the initial search results.

[0014] Optionally, the reweighting of the initial search results includes the following steps: Based on the initial search results, obtain the comprehensive similarity value, ranking score, industry relevance weight, call impact value, and user preference matching value of the candidate algorithms; The above indicators are integrated, and the integration result is corrected based on the coupling relationship between industry-related weights and call impact values ​​to generate a reweighted score. Candidate algorithms are reordered based on reweighted scores, and stability adjustments are performed based on the differences in sorting positions. Output the adjusted candidate algorithm ranking results to generate the final recommendation results.

[0015] The beneficial effects of this invention are: This invention extracts multimodal features from candidate algorithm data and constructs algorithm call adjacency sequences and algorithm industry association sequences. It achieves the fusion expression of algorithm features and user demand features within a unified semantic space. Furthermore, it introduces an improved liquid state machine model to perform layered input and structured propagation of outer-layer semantic information and core semantic information. This enables unified modeling of call relationships and industry association relationships between algorithms in dynamic semantic state representation, effectively solving the problem of weak semantic expression capabilities caused by fragmented multimodal features and insufficient utilization of relationships in traditional methods.

[0016] This invention divides the parameters to be optimized in a liquid state machine model into structural layer parameters, connection layer parameters, and neuron state layer parameters. During differential evolution, hierarchical mutation and crossover operations are performed on different parameter layers respectively, and inter-layer collaborative update relationships are established. At the same time, the crossover probability and mutation scaling ratio are adaptively adjusted based on the fitness change trend, and a multi-level information fusion mechanism is combined to uniformly evaluate the fitness contribution of each parameter layer. This achieves collaborative optimization of model structure and parameters, avoiding the problem of limited optimization effect caused by the separation of structure and weight in traditional single-parameter optimization methods.

[0017] This invention performs comprehensive similarity matching between the dynamic semantic state representation of candidate algorithms and the dynamic semantic state representation of queries. Based on the initial retrieval results, it integrates industry association weights, call influence values, and user preference matching values ​​for reweighting. At the same time, it introduces coupling relationship correction and ranking stability adjustment mechanisms to form the final recommendation results. This makes the recommendation process take into account semantic matching, structural association, and user preference information, effectively improving the accuracy and personalization of retrieval results and solving the problems of single ranking and insufficient relevance of results in traditional recommendation methods. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They explain the invention together with the embodiments of the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a cross-modal retrieval and accurate recommendation method for an industrial algorithm marketplace proposed in this invention; Figure 2 This is a schematic diagram of the structure of an improved liquid state machine model for a cross-modal retrieval and precise recommendation method for an industrial algorithm marketplace proposed in this invention. Figure 3 This is a schematic diagram illustrating the adaptive differential evolution algorithm optimization process of a cross-modal retrieval and accurate recommendation method for an industrial algorithm marketplace proposed in this invention. Detailed Implementation

[0019] 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.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A cross-modal retrieval and accurate recommendation method for industrial algorithm marketplaces includes: Collect candidate algorithm data and user behavior data from the industrial algorithm marketplace; Generate algorithm feature vectors from candidate algorithm data, generate user demand feature vectors from user behavior data, and map the algorithm feature vectors and user demand feature vectors to a unified dimensional space to generate standardized input feature vectors. Construct an algorithm call adjacency sequence and an algorithm industry association sequence based on candidate algorithm data; An improved liquid state machine model is constructed, including a ring-cluster topology reservoir structure, a shell input region, a core input region, and a heterogeneous neuron group structure. Based on the algorithm, the cross-cluster jump connection path in the ring-cluster topology reservoir structure is set, and based on the algorithm, the penetration connection path from the shell input region to the core input region is set. An adaptive differential evolution algorithm is used to optimize the improved liquid state machine model. During the differential evolution iteration process, the crossover probability and mutation scaling ratio are dynamically adjusted. A multi-level information fusion mechanism is applied to optimize different levels of structure independently, and the results are fused to dynamically adjust the optimization weights of each level. The standardized input feature vector is fed into the optimized liquid state machine model to generate the dynamic semantic state representation of the candidate algorithm, and the query dynamic state representation is generated based on the user's retrieval request. The query dynamic state representation is matched with the candidate algorithm's dynamic semantic state representation to generate initial search results; The initial search results are reweighted to generate personalized recommendation results, which are then output to the industrial algorithm marketplace.

[0021] In this embodiment, the candidate algorithm data includes algorithm description text, demonstration images, interface parameters, application cases, industry tags, algorithm call relationship data, and unique algorithm identifiers; the user behavior data includes user search requests, uploaded images, filtering conditions, browsing history, click history, favorites history, and call history.

[0022] In this embodiment, generating the standardized input feature vector includes: Multimodal feature extraction is performed on the candidate algorithm data to generate text feature vectors, image feature vectors, and structured label vectors. The fusion result of the text feature vector and image feature vector is defined as the outer semantic information, and the structured label vector is defined as the core semantic information. Specifically, the multimodal feature extraction on the candidate algorithm data involves: The algorithm description text is processed by word segmentation and semantic encoding. The demonstration images in the candidate algorithm data are processed by size unification and pixel normalization. The interface parameters and application cases are processed by field parsing and structured encoding. The semantic representation of the text is extracted based on the bidirectional encoding representation model BERT to obtain the text feature vector. The semantic representation of the image is extracted based on the input visual transformer model to obtain the image feature vector. The parameter name, parameter type and parameter value range are converted into structured label data and vectorized encoding is performed to obtain the structured label vector. The text feature vector and image feature vector are fused. During the fusion process, they are concatenated or weighted according to the preset feature dimension order to form a unified expression of text semantics and visual semantics. The fused result is defined as the outer semantic information. The structured label vector is retained separately and used as the expression of the algorithm's internal attributes and constraints, which is defined as the core semantic information. Perform behavioral feature modeling on user behavior data to generate user demand feature vectors. Specifically, performing behavioral feature modeling on user behavior data involves: The historical browsing records, click records, call records and filtering conditions in the user behavior data are sorted by time series, and each behavior type is assigned a corresponding behavior weight. A time decay coefficient is introduced, and the behavior weights are decayed to obtain a weighted behavior sequence. The weighted behavior sequence is input into the bidirectional encoding representation model BERT for semantic encoding processing to obtain a behavior semantic representation and normalize it to generate a user demand feature vector. The outer semantic information and the core semantic information are concatenated in a preset order to generate an algorithm feature vector; The algorithm feature vectors are mapped to user-required feature vectors in a unified dimension to generate standardized input feature vectors.

[0023] In this embodiment, the construction of the algorithm call adjacency sequence and the algorithm industry association sequence based on candidate algorithm data includes: Based on the algorithm call relationships in the candidate algorithm data, call relationships between algorithms are constructed and an algorithm call adjacency sequence is generated. The call relationships are weighted by call strength and time decay, specifically as follows: Call records are extracted from candidate algorithm data. Directed connections are established according to the correspondence between the calling algorithm and the called algorithm. The number of calls to the same algorithm pair within the statistical period is accumulated to obtain the call intensity. The call intensity is used to characterize the frequency of dependency and the degree of use between algorithms. By uniformly normalizing the number of calls to different algorithm pairs, the call relationships can be compared and calculated within the same numerical range. Based on the call strength, a time decay process is introduced to reflect the timeliness of the call relationship. The time interval between the call occurrence time and the current time is calculated for each call record, and the call strength is adjusted according to a preset exponential decay function so that call records with shorter time intervals retain higher weights, while call records with longer time intervals gradually have reduced influence. The result after time decay is combined with the normalized call strength to obtain the final call relationship weight, and an algorithm call adjacency sequence is generated accordingly. The adjacency sequence reflects both call frequency and time variation characteristics. Based on the industry tag information in the candidate algorithm data, we construct industry association relationships between algorithms and generate an algorithm industry association sequence. The algorithm call adjacency sequence and the algorithm industry association sequence are indexed and aligned according to the algorithm identifier.

[0024] In this embodiment, constructing the improved liquid state machine model includes: The basic structure of the improved liquid state machine model is initialized, and a ring cluster topology reservoir is constructed. Neurons are divided into multiple ring cluster units, and unidirectional ring connection paths are established within each unit in a fixed order. Each neuron is assigned a unique identifier. Specifically, the construction of the ring cluster topology reservoir involves: When constructing the ring cluster topology reservoir, the total number of neurons in the reservoir is determined to be 1,024. All neurons are divided into 16 ring cluster units according to a fixed partitioning rule. Each ring cluster unit contains 64 neurons. All neurons are assigned a unique identifier in a sequential numbering manner, with the numbering range from 1 to 1,024. The neurons in each ring cluster unit are arranged in numerical order to form a locally ordered structure and establish a unified index basis. Within each ring cluster unit, unidirectional circular connection paths are established according to the order of neuron numbering, that is, connections are formed in the direction of increasing numbering. The last neuron establishes a closed loop connection with the first neuron to form a complete unidirectional ring structure. The initial connection strength of the connection edge within the ring is set to 0.5, and the connections within the ring are normalized. The sum of the input connection strength of a single neuron is kept to one, and information is stably cyclically propagated within the ring cluster along a fixed path. During the construction of the ring cluster topology, cross-cluster connection positions are reserved between different ring cluster units. Each ring cluster unit has eight preset cross-cluster connection interface positions, and the corresponding numbers are selected in the neurons within the ring in an equally spaced manner for the establishment of subsequent cross-cluster jump connection paths. Non-local connection relationships are formed between different ring clusters, and the reserve pool has both local loop propagation capability and global topology expansion capability. The ring cluster topology reserve pool forms a composite structure with multiple local closed loops and reserved cross-ring connections, which ensures that information maintains a stable path when propagating within the ring and has controllable jump capability when propagating across the ring. It provides a structural foundation for the connection mapping of adjacency sequences and industry-related sequences based on algorithm calls and enhances the model's ability to express multi-layer semantic information. Establish a mapping rule from algorithm nodes to neurons: Map each algorithm identifier in the algorithm call adjacency sequence to a unique neuron node in the reserve pool, transforming the algorithm call relationship into a connection relationship between neurons. Specifically, the mapping rule from algorithm nodes to neurons is as follows: When establishing the mapping rules from algorithm nodes to neurons, all algorithm identifiers are extracted based on the algorithm call adjacency sequence and sorted according to the numerical order or lexicographical order of the algorithm identifiers to form an ordered list of algorithm identifiers. A one-to-one mapping relationship is then constructed based on the unique identifier range of neurons in the reserve pool, specifically: The first sorted algorithm identifier is mapped to the neuron node numbered one, and so on until all algorithm identifiers are mapped to neuron nodes, so that each algorithm identifier obtains a unique position identifier in the reserve pool and establishes a stable index relationship. Complete the basic mapping, transform the call relationship in the algorithm call adjacency sequence into the connection relationship between neurons. Specifically, for any call record, the initiating algorithm and the called algorithm are established in the reserve pool according to their corresponding neuron identifiers, and the call relationship weight is mapped to the connection strength value of the corresponding connection edge. All connection strength values ​​are normalized in intervals so that the sum of the input connection strength of each neuron node is kept to one. The call relationship between algorithms is structurally embedded into the neuron connection network. During the mapping process, algorithm identifiers with frequent call relationships are preferentially assigned to highly connected positions in the ring cluster unit, while algorithm identifiers with sparse call relationships are assigned to edge positions in the ring cluster. This position allocation strategy enables high-frequency call algorithms to have stronger information propagation capabilities in the reserve pool and keeps low-frequency call algorithms from being disturbed, thus realizing the expression of differences in algorithm call relationships at the structural level. A cross-cluster connection candidate set is constructed in the reserve pool. All neuron pairs between different ring cluster units are traversed. Based on the adjacency algorithm and call relationship weights in the algorithm call adjacency sequence, algorithm pairs with weights greater than a preset threshold are selected and mapped as cross-cluster jump connection paths. Corresponding cross-cluster connection edges are filtered and retained from the cross-cluster connection candidate set, linearly normalized according to the call relationship weights, and assigned connection strength values ​​as weight parameters. Specifically, traversing all neuron pairs between different ring cluster units involves: Obtain the set of numbers of all ring cluster units, and generate different ring cluster unit pairs by pairwise combination according to the ring cluster unit numbers. For any ring cluster unit pair, extract the unique identifier set of all neurons in the two ring cluster units respectively. Generate a set of neuron pairs by performing a full combination operation on the two sets, that is, pair each neuron in the first ring cluster unit with each neuron in the second ring cluster unit to form a cross-cluster neuron pair set. Merge the neuron pair sets corresponding to all ring cluster unit pairs to obtain a complete cross-cluster connection candidate set. The size of the cross-cluster connection candidate set is equal to the number of pairwise combinations after the number of ring cluster units is reduced by one, multiplied by the square of the number of neurons in each ring cluster unit. Algorithm pairs with weights greater than a preset threshold are selected and mapped as cross-cluster skip paths, specifically: After generating a set of cross-cluster neuron pairs, the algorithm identifier corresponding to each neuron pair is back-mapped. The corresponding algorithm identifier is obtained through the neuron identifier. The call relationship between the algorithm pairs and the corresponding call relationship weight are found in the algorithm call adjacency sequence. All call relationship weights are statistically calculated to obtain the weight mean and standard deviation. The preset threshold is set as the sum of the mean and standard deviation. The call relationship weight of each algorithm pair is compared with the preset threshold. Algorithm pairs with weights greater than the preset threshold are selected, and the corresponding neuron pairs are marked as valid cross-cluster connection candidates. After linear normalization based on the call relationship weights, a connection strength value is assigned and written into the weight parameters, specifically: The selected set of effective cross-cluster connection candidates is processed to build connections. Directed cross-cluster connections are established between the corresponding neuron pairs. The call weights of all connections are linearly normalized so that the normalized connection strength values ​​are distributed in the range of zero to one. During the normalization process, the minimum weight of 0.7 and the maximum weight of 1 in the candidate set are used as the interval boundaries. The connection strength values ​​of each connection are calculated by interval scaling and written into the connection weight parameters in the reserve pool. This forms a cross-cluster jump connection path structure driven by the algorithm call relationship, and semantically related connections are established between different ring cluster units. A shell input region and a kernel input region are constructed. The shell input region receives the outer semantic information from the standardized input feature vector, and the kernel input region receives the core semantic information from the standardized input feature vector. An identifier is assigned to each input node. Specifically, the construction of the shell and kernel input regions is as follows: Based on the dimensional structure of the standardized input feature vector, the outer semantic information and the core semantic information are divided into different input regions. The outer semantic information is set to 512 dimensions, and the core semantic information is set to 128 dimensions. A corresponding number of input node sets are established on the input side of the reserve pool. The shell input region contains 512 input nodes, and the kernel input region contains 128 input nodes. Each input node is assigned a unique identifier according to the feature dimension order. The identifier range is consecutively numbered for the shell input region and the kernel input region, so that the input node and the feature dimension form a one-to-one correspondence. The input nodes of the shell input region are connected to the neurons in the ring cluster topology reservoir. Each shell input node is connected to eight neuron nodes in the corresponding ring cluster unit. The connection method adopts a uniform distribution strategy. The outer semantic information can cover all ring cluster units and form a wide propagation path in the reservoir. The input nodes of the core input region are connected to the core neuron nodes in the corresponding ring cluster unit. Each core input node is connected to four neuron nodes. This allows the core semantic information to be concentrated into the reservoir with a high intensity, forming a structural feature of wide-area diffusion input of outer semantic information and local reinforcement input of core semantic information. The input connection weights of the shell input region and the kernel input region are initialized respectively. The connection weight of the shell input region is set to 0.3 and the connection weight of the kernel input region is set to 0.7. The input connection weight of each neuron is normalized so that the sum of all input connection weights is one, which ensures the stable transmission of the input signal in the reservoir and makes the outer semantic information and core semantic information form a differentiated influence effect during the propagation process. Based on the industry-related algorithms and weights in the industry-related sequence, algorithm pairs with weights greater than the median are selected and mapped as penetrating connection paths from the shell input region to the kernel input region. Penetrating connection edges are established, and after normalization according to the industry-related weight range, connection strength values ​​are assigned and written into the weight parameters. Specifically, the mapping as a penetrating connection path from the shell input region to the kernel input region is as follows: Based on the algorithm industry association sequence, all algorithm pairs and their corresponding industry association weights are extracted. All weights are sorted and the median is calculated. The median is the average of the weights at positions N / 2 and N / 2+1 after sorting. The filtering threshold is set to this median value. The industry association weight of each algorithm pair is compared with the threshold. Algorithm pairs with weights greater than the threshold are filtered out, and the corresponding algorithm identifier set is obtained. The corresponding shell input node and kernel input node number are determined through the mapping relationship between the aforementioned algorithm identifier and neuron identifier, thereby forming a candidate penetration connection path set. Establish a through-connection edge, specifically as follows: In the candidate penetration connection path set, a directional connection relationship is established between the corresponding shell input node and the core input node of each algorithm pair. The connection direction is from the shell input area to the core input area, so that the outer semantic information can enter the core semantic processing area along the penetration path. During the connection establishment process, the connection strength allocation basis is determined according to the industry association weight of the algorithm pair. The filtered industry association weights are normalized in interval. The lower limit of the normalization interval is the median, and the upper limit is the maximum value among all industry association weights. The connection strength value in the range of 0 to 1 is calculated by linear scaling, and the connection strength value is written into the weight parameter of the corresponding connection edge, thereby forming a penetration connection structure driven by industry association. A heterogeneous neuron population structure is constructed within a ring-cluster topological reservoir. Standardized input feature vectors are assigned to corresponding input nodes, establishing input connections between input nodes and reservoir neurons, thus forming a complete liquid state machine model structure. Specifically, the construction of the heterogeneous neuron population structure involves: All neurons are classified into types according to a preset ratio. The proportion of linear response neurons is set to 0.4, the proportion of nonlinear response neurons is 0.4, and the proportion of gated response neurons is 0.2. The types are assigned in sequence according to the neuron identification order so that different types of neurons are evenly distributed in each ring cluster unit. Each ring cluster unit is divided into three types of neurons in the same proportion. In the linear response neuron, the state update method is set as weighted summation update. The input signal and the historical state are linearly combined according to the weight coefficients, which are initialized to 0.6 and 0.4 and dynamically adjusted in subsequent optimization. In the nonlinear response neuron, the state update method is set as activation function transformation update. The hyperbolic tangent function is selected as the activation function to perform nonlinear compression processing on the input signal. In the gated response neuron, the state update method is set as gated modulation update. Two control parameters, the input gate and the forget gate, are introduced. The initial value of the input gate is set to 0.5, and the initial value of the forget gate is set to 0.5. The retention ratio of the input signal and the historical state is adjusted through the gating mechanism. By combining different types of neurons according to the topological connection relationship of ring clusters, a neuron group structure with multi-level dynamic response capability is formed, which enables the reservoir to handle linear and nonlinear relationships simultaneously during state propagation and to selectively control the information flow.

[0025] In this embodiment, the model optimization using the adaptive differential evolution algorithm includes: Traditional parameter optimization is improved to structure-aware optimization. The parameters to be optimized are structurally partitioned and encoded, into structural layer parameters, connection layer parameters, and neuron state layer parameters. These parameters are arranged in order of their position within the model structure to generate a parameter sequence, which is then encoded with real numbers to form an initial population. Specifically, the partitioning of the parameters to be optimized into structural layer parameters, connection layer parameters, and neuron state layer parameters is as follows: The parameters to be optimized include structural layer parameters, connection layer parameters, and neuron state layer parameters; The structural layer parameters include the number of ring clusters, the number of neurons in each ring cluster unit, the connection order index of the unidirectional ring connection path, and the existence identifier and connection position index of the cross-cluster skip connection path. The connectivity parameters include the weight of the internal connectivity edges of the ring cluster, the connectivity strength value of the cross-cluster connectivity edges, and the penetration connectivity strength value from the shell input region to the core input region. The neuron state layer parameters include the weighting coefficients of linear response neurons, the activation function parameters of nonlinear response neurons, and the input gate parameters and forget gate parameters of gated response neurons; The parameter sequences in the initial population are input into the liquid state machine model, and dynamic semantic state representations of candidate algorithms are generated based on the standardized input feature vectors. These representations are then matched with the corresponding reference semantic state representations to obtain fitness values. An improved structural divide-and-conquer search and coupled optimization approach is adopted. During the differential evolution iteration process, hierarchical differential mutation and hierarchical crossover operations are performed on the structural layer parameters, connection layer parameters, and neuron state layer parameters, respectively. A collaborative update relationship is established between different parameter layers, ensuring synchronous updates of each parameter layer within the same iteration round. Specifically, the hierarchical differential mutation and hierarchical crossover operations performed on the structural layer parameters, connection layer parameters, and neuron state layer parameters are as follows: The parameters to be optimized are divided into three subspaces: structural parameters, connection parameters, and neuron state parameters. Differential mutation and crossover operations are performed on the three subspaces in the same differential evolution iteration round. The structural subspace uses discrete position indices as basic units for neighborhood perturbation search, the connection subspace uses continuous weight vectors as basic units for scaling differential mutation, and the neuron state subspace uses state parameter vectors as basic units for random perturbation and combination update. Each subspace adopts its own adapted encoding method and mutation strategy in the independent search process, forming a divide-and-conquer search path in different parameter dimensions. The three types of parameters, namely structural topology, connection strength, and state update, can be fully explored in their respective spaces. Establish collaborative update relationships between different parameter layers, specifically as follows: Based on divide-and-conquer search, a coupled optimization mechanism is introduced to simultaneously evaluate candidate solutions in the three subspaces within the same iteration round. The candidate parameter sequences generated by each subspace are combined into a complete parameter sequence and input into the liquid state machine model to calculate the fitness value. The update direction of each subspace is adjusted based on the overall fitness result. When the parameter update of a subspace improves the overall fitness, the corresponding update result is retained. When the parameter update of a subspace causes the overall fitness to decrease, the parameter state of the previous round is reverted. The overall fitness constraint realizes the cooperative constraint relationship between the subspaces, so that different parameter layers maintain a consistent convergence direction during the optimization process. In the collaborative update process, a unified iteration step size of 1 round is set. In each iteration, the three subspaces simultaneously perform differential mutation and crossover operations, and complete fitness evaluation and update decision within the same round. This ensures that the structural layer parameters, connection layer parameters, and neuron state layer parameters always maintain a synchronous evolution state. Through the combination of divide-and-conquer search and coupling constraints, the model has both local fine search capability and global consistency adjustment capability during the optimization process. The fitness values ​​of the candidate sequences generated hierarchically are calculated and compared with the fitness values ​​of the original parameter sequences. When a candidate sequence has a better fitness value than the original parameter sequence, it is replaced to form the next generation population. During the iteration process, the fitness change trend of each parameter layer is recorded. Specifically, the calculation of the fitness values ​​of the candidate sequences generated hierarchically is as follows: Candidate subsequences corresponding to structural layer parameters, connection layer parameters, and neuron state layer parameters are concatenated into a complete parameter sequence in a predetermined order. The complete parameter sequence is then written into a liquid state machine model to complete a parameter instantiation process. The standardized input feature vector is input into the model to perform state propagation calculation, thereby obtaining the dynamic semantic state representation of the candidate algorithm. The user demand feature vector is input into the same model to generate a reference semantic state representation. Euclidean distance and cosine similarity are used to jointly calculate the matching degree index, and the matching degree is normalized to an fitness value in the range of 0 to 1 to characterize the overall performance of the current candidate parameter sequence in the semantic matching task. A hierarchical contribution evaluation mechanism is introduced in the fitness calculation process. The complete fitness value is decomposed in reverse. Based on the contribution ratio of changes in structural layer parameters, connection layer parameters, and neuron state layer parameters to the improvement of matching degree, the local fitness contribution values ​​of the three parameter layers are calculated respectively. The contribution values ​​are obtained by normalizing the difference between the current candidate sequence and the previous sequence in each parameter layer and the corresponding fitness increment. The fitness values ​​are compared with those of the original parameter sequence, specifically: After completing the fitness calculation, the overall fitness value of the candidate parameter sequence is compared with the fitness value of the original parameter sequence. When the fitness value of the candidate sequence is higher than that of the original sequence, a replacement operation is performed to generate the next generation population. During the iteration process, the fitness change trend of each parameter layer is recorded, specifically as follows: The fitness change sequences of structural layer parameters, connection layer parameters, and neuron state layer parameters are recorded synchronously during continuous iteration. The change sequences are calculated with the most recent 5 iterations as a window to measure the change amplitude and fluctuation degree, which are used to characterize the convergence trend and stable state of each parameter layer. Based on the fitness change trends of each parameter layer, an adaptive strategy adjustment is executed, and a multi-level information fusion mechanism is introduced. Fitness contribution values ​​are calculated for structural layer parameters, connection layer parameters, and neuron state layer parameters respectively. The overall fitness is then weighted and fused based on the fitness contribution values ​​of each parameter layer to obtain the overall optimization evaluation value. Specifically, the adaptive strategy adjustment based on the fitness change trends of each parameter layer is as follows: The fitness change sequences of structural layer parameters, connection layer parameters, and neuron state layer parameters in five consecutive iterations were statistically analyzed. The change amplitude and fluctuation index of each parameter layer were calculated. The change amplitude was obtained by averaging the absolute values ​​of the differences between the current fitness value and the previous fitness value. The fluctuation index was calculated by the standard deviation of the fitness values ​​in five iterations. The change amplitude threshold was set to 0.01, and the fluctuation threshold was set to 0.02. When the change amplitude of the structural layer parameters was lower than 0.01, the mutation scaling ratio was reduced from the initial value of 0.5 to 0.3. When the change amplitude of the connection layer parameters was higher than 0.01, the crossover probability was increased from the initial value of 0.6 to 0.8. When the fluctuation index of the neuron state layer parameters was higher than 0.02, the crossover probability was adjusted to 0.7 and the mutation scaling ratio was adjusted to 0.4. By setting trigger conditions and adjustment strategies for different parameter layers, the differential evolution process can automatically adjust the search intensity and exploration range at different stages. A multi-level information fusion mechanism is introduced, specifically as follows: A three-dimensional contribution vector is constructed based on the fitness contribution values ​​of structural layer parameters, connection layer parameters, and neuron state layer parameters. The three types of contribution values ​​are normalized so that their sum is 1. A linear scaling method is used in the normalization process. The corresponding weight is calculated according to the contribution ratio of each parameter layer to the overall fitness improvement in the current iteration. When the contribution value of structural layer parameters is greater than 0.4, its weight is set to 0.5, the weight of connection layer parameters is set to 0.3, and the weight of neuron state layer parameters is set to 0.2. When the contribution value of connection layer parameters is the largest, its weight is set to 0.5. The weights of the other two types of parameter layers are allocated at 0.25. When the contribution values ​​of the three types of parameter layers are close to balanced, the weights are uniformly set to 0.33. The fitness contribution value of each parameter layer is multiplied with its corresponding weight and summed to obtain the overall optimization evaluation value. This evaluation value is used to uniformly measure the comprehensive performance of the current parameter sequence at the three levels of structure, connectivity and state. When the overall optimization evaluation value reaches the preset termination condition, the parameter sequence with the best fitness in the current population is output, and the parameter sequence is written into the liquid state machine model to obtain the optimized liquid state machine model. Specifically, the parameter sequence with the best fitness in the current population is output as follows: Statistical analysis is performed on the overall optimization evaluation value sequence in 10 consecutive iterations to calculate the evaluation value change range and convergence degree. The change range is obtained by averaging the absolute values ​​of the difference between the current round evaluation value and the previous round evaluation value. The convergence judgment threshold is set to 0.001. When the evaluation value change range is less than 0.001 in 10 consecutive iterations, it is judged as convergence. At the same time, the maximum number of iterations is set to 200. When the number of iterations reaches 200, a termination judgment is triggered even if the convergence condition is not met, thus forming a dual termination mechanism based on convergence state and iteration upper limit. The parameter sequence is written into the liquid state machine model, specifically as follows: After the termination condition is met, the parameter sequence with the highest fitness value is extracted from the current population as the optimal parameter sequence. This parameter sequence contains a complete combination of structural layer parameters, connection layer parameters and neuron state layer parameters. The parameters are parsed according to their position order in the model. The structural layer parameters are written into the ring cluster topology reserve pool structure configuration to determine the ring cluster partitioning and cross-cluster jump connection path. The connection layer parameters are written into the weight parameters of various connection edges to determine the information transmission intensity. The neuron state layer parameters are written into the state update function of the corresponding type of neuron to determine the dynamic response mode, thereby completing the parameter reconstruction of the liquid state machine model. The optimized liquid state machine model is obtained as follows: A complete forward state propagation computation is performed on the model. The standardized input feature vector is input to generate a dynamic semantic state representation of the candidate algorithm. At the same time, a reference semantic state representation is generated for the user requirement feature vector. The matching results of the two semantic state representations are verified. When the matching degree remains stable and is better than the initial model, the optimization result is confirmed to be effective, and the final optimized liquid state machine model is obtained.

[0026] In this embodiment, the dynamic semantic state representation of the candidate generation algorithm includes: The standardized input feature vectors are grouped according to the algorithm and input into the liquid state machine model. The outer semantic information is input into the shell input region, the core semantic information is input into the kernel input region, and then passed to the ring cluster topology reservoir. State updates and cross-cluster propagation are performed in the ring-cluster topology reserve pool to generate dynamic semantic state representations of candidate algorithms; Feature extraction is performed on the user's search request and input into the liquid state machine model to generate a dynamic semantic state representation of the query; The dynamic semantic state representations of candidate algorithms and query dynamic semantic state representations are output along a unified dimension.

[0027] In this embodiment, generating the initial search results includes: The dynamic semantic state representations of the candidate algorithms and the query dynamic semantic state representations are dimensionally aligned, and Euclidean distance and cosine similarity are calculated and fused to obtain a comprehensive similarity value. Specifically, the calculation of Euclidean distance and cosine similarity is as follows: The dynamic semantic state representations of the candidate algorithms and the dynamic semantic state representations of the queries are matched dimension by dimension in a unified order. Interval normalization is performed on the feature values ​​of each dimension so that all feature values ​​fall within the range of 0 to 1. Then, the feature differences in the corresponding dimensions are calculated, and the differences are squared and summed. The square root of the sum is then taken to obtain the Euclidean distance value. The smaller the Euclidean distance, the smaller the difference between the two semantic state representations. Inverse normalization is performed on the Euclidean distance to map the distance value to a distance similarity value in the range of 0 to 1, where the minimum distance is mapped to 1 and the maximum distance is mapped to 0. The dynamic semantic state representations of the candidate algorithms and the dynamic semantic state representations of the queries are vector dot products to obtain the cumulative result of the product of the feature values ​​of the corresponding dimensions. At the same time, the magnitude of the two vectors is calculated separately. The magnitude is obtained by taking the square root of the sum of the squares of the feature values ​​of each dimension. The dot product result is divided by the product of the two magnitudes to obtain the cosine similarity value. The cosine similarity value ranges from -1 to 1. For a uniform scale, the cosine similarity is linearly transformed to map to the interval of 0 to 1, where 1 represents complete similarity and 0 represents complete dissimilarity, thus obtaining a similarity index with the same dimension as the distance similarity. The fusion yields a comprehensive similarity value, specifically: The two similarity indices are fused. The weight of Euclidean distance similarity is set to 0.4 and the weight of cosine similarity is set to 0.6. The two are then weighted and summed to obtain the comprehensive similarity value. The fusion result reflects both the degree of feature difference and the consistency of direction. Candidate algorithms are ranked based on comprehensive similarity values ​​and then filtered using a preset threshold to generate a candidate selection set. Specifically, the preset threshold filtering process is as follows: The comprehensive similarity values ​​of all candidate algorithms are statistically analyzed, and the mean and standard deviation of the comprehensive similarity values ​​are calculated. A preset threshold is set as the sum of the mean and standard deviation. The comprehensive similarity value of each candidate algorithm is compared with this threshold. When the comprehensive similarity value is greater than or equal to the threshold, the corresponding candidate algorithm is retained. When the comprehensive similarity value is less than the threshold, the corresponding candidate algorithm is removed, thus obtaining the preliminary candidate set. After the initial screening is completed, the size of the screening set is controlled. When the number of candidates after screening is less than 10, the threshold is lowered to the mean level and the screening is re-executed. When the number of candidates after screening is greater than 100, the threshold is raised to the mean plus 1.5 times the standard deviation and the screening is re-executed, so that the size of the candidate set is stabilized in the range of 10 to 100. In the candidate selection set, the density value of the candidate algorithm is calculated based on the similarity neighborhood, and then combined with the comprehensive similarity value for fusion and ranking. Specifically, the calculation of the density value of the candidate algorithm based on the similarity neighborhood in the candidate selection set is as follows: The similarity values ​​of all candidate algorithms in the candidate selection set are calculated pairwise to form a similarity matrix. For each candidate algorithm, the similarity vector between it and other candidate algorithms is extracted. The similarity neighborhood threshold is set to 0.8. Candidate algorithms with a comprehensive similarity value greater than or equal to 0.8 are regarded as neighborhood members of the candidate algorithm. The number of neighborhood members is counted to obtain the neighborhood size. The similarity values ​​within the neighborhood are summed to obtain the neighborhood strength value. The neighborhood size and neighborhood strength values ​​are weighted and combined to obtain the density value. The neighborhood size weight is set to 0.5 and the neighborhood strength weight is set to 0.5, thus forming a quantitative representation of the local clustering degree of the candidate algorithms. The ranking is based on a combination of comprehensive similarity scores, specifically as follows: The density values ​​and comprehensive similarity values ​​of candidate algorithms are processed in a unified manner so that they can be weighted and fused within the same dimension. The density values ​​and comprehensive similarity values ​​are normalized to a range of 0 to 1. The density values ​​and comprehensive similarity values ​​are then weighted and fused according to a preset weight ratio of 0.7 for density values ​​and 0.3 for comprehensive similarity values ​​to obtain the fused ranking score. All candidate algorithms are then ranked from high to low according to the fused ranking score. Output the sorted list of candidate algorithms and generate the initial search results.

[0028] In this embodiment, the reweighting of the initial search results includes: Based on the initial search results, obtain the comprehensive similarity value, ranking score, industry relevance weight, call impact value, and user preference matching value of the candidate algorithms; The above indicators are integrated, and the integration result is corrected based on the coupling relationship between industry-related weights and call impact values ​​to generate a reweighted score. Specifically, the integration of the above indicators involves: The normalized comprehensive similarity value, ranking score, industry association weight, call impact value, and user preference matching value are weighted and fused according to preset weights. The weights of comprehensive similarity value and ranking score are set to 0.4 and 0.3, respectively; the weights of industry association weight and call impact value are set to 0.15 and 0.1, respectively; and the weight of user preference matching value is set to 0.05. All weights are summed to 1. After weighted fusion, a preliminary fusion score is obtained, which represents the comprehensive performance of the candidate algorithm in various dimensions. The reweighted score is generated as follows: The algorithm is modified by analyzing the coupling relationship between industry association weight and call impact value. If the industry association weight is high and the call impact value is large, the applicability of the algorithm matches the user's needs well, and the fusion score is appropriately increased. When the coupling between industry association weight and call impact value is low, the modified score will decrease. The modification process adopts a scaling method, applying the coupling modification coefficient to the initial fusion score value, so that it is dynamically adjusted within the range of 0 to 1, and finally generating a reweighted score value. Candidate algorithms are reordered based on reweighted scores, and stability adjustments are performed based on the differences in sorting positions. Output the adjusted candidate algorithm ranking results to generate the final recommendation results.

[0029] Example 1: To verify the feasibility of this invention in practical use, it was applied in an industrial algorithm marketplace. A technical team used the method of this invention to improve recommendation efficiency and accuracy when needing to select the most suitable recommendation algorithm from a massive pool of algorithms for a project. The team's requirement was to select a suitable deep learning algorithm for a project aimed at optimizing automated production. The project required the algorithm to perform well in image processing, data analysis, and real-time feedback. This project required screening a large number of candidate algorithms and, based on the user's historical behavior and needs, personalized recommendations of the algorithm that best met their requirements.

[0030] The team collected data on 3,625 candidate algorithms from an industrial algorithm marketplace, as well as behavioral data from 1,027 users, including historical search records, browsing history, click records, and favorites records. For this candidate algorithm data, the system used a multimodal feature extraction module to extract features from the text description, demonstration images, and interface parameters of each algorithm, generating text feature vectors, image feature vectors, and structured label vectors, respectively. After unified dimensional mapping and standardization, the algorithm feature vectors and user behavior feature vectors formed the system's standardized input feature vectors, which were then input into the optimized liquid state machine model.

[0031] Algorithm call adjacency sequences and algorithm industry association sequences, constructed based on candidate algorithm data, were used to initialize the basic structure of the liquid state machine model. The model's ring-cluster topology reservoir structure was built, and cross-cluster and penetrating connection paths were established using the algorithm call adjacency sequences and industry association sequences. Each algorithm node was mapped to a neuron node in the model, and connections formed the relationships between neurons in the network, further optimizing the model's structure.

[0032] An adaptive differential evolution algorithm was used to optimize the model. Through the iterative process of differential evolution, the crossover probability and mutation scaling ratio were dynamically adjusted, enabling the structural layer parameters, connection layer parameters, and neuron state layer parameters to be updated collaboratively, thus improving the model's convergence speed. A multi-level information fusion mechanism weighted and fused the fitness contributions of each layer in each iteration, yielding the optimal overall optimization evaluation value.

[0033] After a certain number of iterations, and when the overall optimization evaluation value reaches the preset termination condition, the system outputs the parameter sequence with optimal fitness. This parameter sequence is then applied to the liquid state machine model to obtain the final optimized recommendation model. With the support of this model, the team can generate dynamic semantic state representations based on user search requests and execute precise algorithmic recommendations accordingly. The system ultimately outputs a list of algorithms that best meet user needs, receiving high praise from users.

[0034] Table 1. Performance Comparison Data Between the Method of the Invention and Traditional Methods

[0035] Compared to traditional methods, the method of this invention significantly improves the performance of industrial algorithm recommendation. In terms of algorithm recommendation accuracy, the method of this invention improves by 16.7%, meaning that the recommended algorithms better meet user needs and improve the accuracy of the algorithm. Furthermore, the time required for users to select an algorithm is significantly reduced; the traditional method requires 88 minutes, while the method of this invention only requires 30 minutes, a reduction of 65.9%, indicating a significant improvement in recommendation efficiency. In terms of matching degree of the recommendation algorithm and success rate of personalized recommendations, the method of this invention also improves by 15.4% and 14.6%, respectively. This means that the recommendation algorithm is better able to match the actual needs of users and provide more personalized recommendation services. In terms of the ranking accuracy of initial search results and the relevance of recommendation results, the method of this invention also shows significant improvements, with accuracy and relevance increasing by 15.8% and 14.7%, respectively. These improvements directly impact the user experience, making the recommended results more in line with actual needs.

[0036] In terms of the time required to compute recommendation results, traditional methods take 5 hours, while the method of this invention only takes 1.5 hours, reducing the time by 70%. This demonstrates that the present invention can generate recommendation results more efficiently when processing large-scale data. The improved final ranking accuracy and increased user satisfaction also indicate that the present invention has significant advantages in providing accurate, personalized, and timely recommendation results. The diversity of recommendation algorithms has been enhanced, increasing from 61.3% in traditional methods to 82.0% in the method of this invention. This means that the present invention can provide more diverse algorithmic recommendations, better meeting the needs of different users.

[0037] Overall, the method of this invention demonstrates significant advantages over traditional methods in cross-modal retrieval and accurate recommendation in industrial algorithm marketplaces, in terms of recommendation accuracy, efficiency, personalization, and user satisfaction, proving its feasibility and effectiveness in practical applications.

[0038] 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 cross-modal retrieval and precise recommendation method for industrial algorithm marketplaces, characterized in that, Includes the following steps: Collect candidate algorithm data and user behavior data from the industrial algorithm marketplace; Generate algorithm feature vectors from candidate algorithm data, generate user demand feature vectors from user behavior data, and map the algorithm feature vectors and user demand feature vectors to a unified dimensional space to generate standardized input feature vectors. Construct an algorithm call adjacency sequence and an algorithm industry association sequence based on candidate algorithm data; An improved liquid state machine model is constructed, including a ring-cluster topology reservoir structure, a shell input region, a core input region, and a heterogeneous neuron group structure. Based on the algorithm, the cross-cluster jump connection path in the ring-cluster topology reservoir structure is set, and based on the algorithm, the penetration connection path from the shell input region to the core input region is set. An adaptive differential evolution algorithm is used to optimize the improved liquid state machine model. During the differential evolution iteration process, the crossover probability and mutation scaling ratio are dynamically adjusted. A multi-level information fusion mechanism is applied to optimize different levels of structure independently, and the results are fused to dynamically adjust the optimization weights of each level. The standardized input feature vector is fed into the optimized liquid state machine model to generate the dynamic semantic state representation of the candidate algorithm, and the query dynamic state representation is generated based on the user's retrieval request. The query dynamic state representation is matched with the candidate algorithm's dynamic semantic state representation to generate initial search results; The initial search results are reweighted to generate personalized recommendation results, which are then output to the industrial algorithm marketplace.

2. A cross-modal retrieval and precise recommendation method for an industrial algorithm marketplace as described in claim 1, characterized in that, The candidate algorithm data includes algorithm description text, demonstration images, interface parameters, application cases, industry tags, algorithm call relationship data, and unique algorithm identifiers; the user behavior data includes user search requests, uploaded images, filtering conditions, browsing history, click history, favorites history, and call history.

3. A cross-modal retrieval and precise recommendation method for an industrial algorithm marketplace as described in claim 1, characterized in that, The generation of the standardized input feature vector includes: Multimodal feature extraction is performed on the candidate algorithm data to generate text feature vectors, image feature vectors and structured label vectors. The fusion result of text feature vectors and image feature vectors is defined as outer semantic information, and the structured label vectors are defined as core semantic information. Perform behavioral feature modeling on user behavior data to generate user demand feature vectors; The outer semantic information and the core semantic information are concatenated in a preset order to generate an algorithm feature vector; The algorithm feature vectors are mapped to user-required feature vectors in a unified dimension to generate standardized input feature vectors.

4. A cross-modal retrieval and precise recommendation method for an industrial algorithm marketplace as described in claim 1, characterized in that, The process of constructing the algorithm call adjacency sequence and algorithm industry association sequence based on candidate algorithm data includes the following steps: Based on the algorithm call relationships in the candidate algorithm data, we construct the call relationships between algorithms and generate an algorithm call adjacency sequence, where the call relationships are weighted by call strength and time decay; Based on the industry tag information in the candidate algorithm data, we construct industry association relationships between algorithms and generate an algorithm industry association sequence. The algorithm call adjacency sequence and the algorithm industry association sequence are indexed and aligned according to the algorithm identifier.

5. A cross-modal retrieval and precise recommendation method for an industrial algorithm marketplace as described in claim 1, characterized in that, The construction of the improved liquid state machine model includes the following steps: The basic structure of the improved liquid state machine model is initialized, a ring cluster topology reservoir is constructed, neurons are divided into multiple ring cluster units, unidirectional ring connection paths are established in a fixed order within the unit, and a unique identifier is assigned to each neuron. Establish a mapping rule from algorithm nodes to neurons: map each algorithm identifier in the algorithm call adjacency sequence to a unique neuron node in the reserve pool, so that the algorithm call relationship is transformed into the connection relationship between neurons; In the reserve pool, a cross-cluster connection candidate set is constructed. All neuron pairs between different ring cluster units are traversed. Based on the adjacency algorithm and call relationship weight in the algorithm call adjacency sequence, algorithm pairs with weights greater than a preset threshold are selected and mapped as cross-cluster jump connection paths. The corresponding cross-cluster connection edges are filtered and retained from the cross-cluster connection candidate set. After linear normalization according to the call relationship weight, the connection strength value is assigned and written into the weight parameter. Construct a shell input region and a kernel input region. The shell input region receives the outer semantic information from the standardized input feature vector, and the kernel input region receives the core semantic information from the standardized input feature vector. Assign an identifier to each input node. Based on the industry association algorithms and weights in the industry association sequence, algorithm pairs with weights greater than the median are selected and mapped to penetrating connection paths from the shell input region to the kernel input region. Penetrating connection edges are established, and after normalization according to the industry association weight interval, connection strength values ​​are assigned and written into the weight parameters. A heterogeneous neuron group structure is constructed in the ring-cluster topology reservoir. The standardized input feature vectors are assigned to the corresponding input nodes, and the input connection relationship between the input nodes and the reservoir neurons is established to form a complete liquid state machine model structure.

6. A cross-modal retrieval and precise recommendation method for an industrial algorithm marketplace as described in claim 1, characterized in that, The model optimization using the adaptive differential evolution algorithm includes the following steps: The traditional parameter optimization is improved into structure-aware optimization. The parameters to be optimized are divided into structural layer parameters, connection layer parameters, and neuron state layer parameters. The parameters are arranged in order of their position in the model structure to generate a parameter sequence, and the parameter sequence is encoded with real numbers to form an initial population. The parameter sequences in the initial population are input into the liquid state machine model, and dynamic semantic state representations of candidate algorithms are generated based on the standardized input feature vectors. These representations are then matched with the corresponding reference semantic state representations to obtain fitness values. An improved structural divide-and-conquer search and coupling optimization is adopted. During the differential evolution iteration process, hierarchical differential mutation operation and hierarchical crossover operation are performed on the structural layer parameters, connection layer parameters and neuron state layer parameters respectively. A collaborative update relationship is established between different parameter layers so that each parameter layer is updated synchronously in the same iteration round. The fitness values ​​of the candidate sequences generated by the hierarchical generation are calculated and compared with the fitness values ​​of the original parameter sequences. When the fitness value of the candidate sequence is better than that of the original parameter sequence, it is replaced to form the next generation of population. The fitness change trend of each parameter layer is recorded during the iteration process. Based on the fitness change trend of each parameter layer, an adaptive strategy is executed to adjust and a multi-level information fusion mechanism is introduced. The fitness contribution values ​​of structural layer parameters, connection layer parameters and neuron state layer parameters are calculated respectively. The overall fitness is weighted and fused based on the fitness contribution values ​​of each parameter layer to obtain the overall optimization evaluation value. When the overall optimization evaluation value reaches the preset termination condition, the parameter sequence with the best fitness in the current population is output, and the parameter sequence is written into the liquid state machine model to obtain the optimized liquid state machine model.

7. A cross-modal retrieval and precise recommendation method for an industrial algorithm marketplace as described in claim 1, characterized in that, The dynamic semantic state representation of the candidate generation algorithm includes the following steps: The standardized input feature vectors are grouped according to the algorithm and input into the liquid state machine model. The outer semantic information is input into the shell input region, the core semantic information is input into the kernel input region, and then passed to the ring cluster topology reservoir. State updates and cross-cluster propagation are performed in the ring-cluster topology reserve pool to generate dynamic semantic state representations of candidate algorithms; Feature extraction is performed on the user's search request and input into the liquid state machine model to generate a dynamic semantic state representation of the query; The dynamic semantic state representations of candidate algorithms and query dynamic semantic state representations are output along a unified dimension.

8. A cross-modal retrieval and precise recommendation method for an industrial algorithm marketplace as described in claim 1, characterized in that, The process of generating initial search results includes the following steps: The dynamic semantic state representations of the candidate algorithms and the dynamic semantic state representations of the queries are dimensionally aligned, and the Euclidean distance and cosine similarity are calculated and fused to obtain a comprehensive similarity value. Candidate algorithms are ranked based on comprehensive similarity values ​​and filtered through a preset threshold to generate a candidate selection set; The density value of the candidate algorithm is calculated based on the similarity neighborhood in the candidate screening set, and then combined with the comprehensive similarity value for fusion and ranking. Output the sorted list of candidate algorithms and generate the initial search results.

9. A cross-modal retrieval and precise recommendation method for an industrial algorithm marketplace as described in claim 1, characterized in that, The reweighting of the initial search results includes the following steps: Based on the initial search results, obtain the comprehensive similarity value, ranking score, industry relevance weight, call impact value, and user preference matching value of the candidate algorithms; The above indicators are integrated, and the integration result is corrected based on the coupling relationship between industry-related weights and call impact values ​​to generate a reweighted score. Candidate algorithms are reordered based on reweighted scores, and stability adjustments are performed based on the differences in sorting positions. Output the adjusted candidate algorithm ranking results to generate the final recommendation results.

Images