Image cold start recommendation ranking method and system

Abstract

This invention relates to the field of cold start technology and discloses an image cold start recommendation and ranking method and system. By acquiring 3D point cloud data and RGB images of the target scene, a spatial topology layout relationship graph, a fine-grained visual feature library, and multi-level feasibility constraint index vectors are constructed. Based on the feature library, multi-granularity representation vectors are generated, which are then decomposed into three independent attribute vectors through de-entanglement learning and graph-level representation vectors are generated. A comprehensive scoring matrix is ​​then constructed, and after three rounds of progressive screening, multi-objective Pareto ranking, and visual confusion disambiguation, the final recommendation sequence is output. This invention effectively solves the problems of visual feature semantic drift and execution constraint mismatch in cold start scenarios, achieving a cognitive leap from "visual similarity" to "functional matching" and then to "executable solutions."

Description

Technical Field

[0001] This invention relates to the field of cold start technology, and more specifically, to an image cold start recommendation and ranking method and system. Background Technology

[0002] Cold start scenarios are a core challenge in the practical application of intelligent recommendation, especially in fields such as smart retail, industrial inspection, and medical imaging. Newly introduced target objects (such as newly listed products, new parts, and rare case images) lack user interaction data in the early stages, making it difficult for traditional recommendation methods that rely on collaborative filtering to establish connections between items. Visual features become the key reasoning basis, and the accuracy of their semantic understanding and the adaptability to execution constraints directly determine the quality and implementation effect of cold start recommendations, becoming the core bottleneck restricting the recommendation performance in such scenarios.

[0003] In the prior art, Chinese Patent No. CN118013135B discloses a graph comparison learning recommendation method based on relational graph convolutional neural networks. This method obtains triplet embedding representations, aggregates and enhances entity-item embeddings of different relationships, and combines cross-view learning to optimize feature representations. Finally, it generates recommendation results based on the similarity score of the enhanced knowledge graph, aiming to alleviate the heterogeneity noise problem of knowledge graphs and improve recommendation accuracy in data-sparse scenarios. Chinese Patent No. CN114004667B discloses a knowledge crowdsourcing cold-start task modeling and recommendation method. This method introduces explicit attributes of cold-start tasks, similar historical tasks, and service providers, uses a self-attention mechanism to capture the coupling relationship between task and service provider features, and learns fusion vector representations to achieve task-service provider matching, focusing on solving the problems of task cold-start and data sparsity in knowledge crowdsourcing scenarios.

[0004] However, existing cold start recommendation technologies, such as CN118013135B, rely on interaction relationship data, making it difficult to build effective associations under zero-interaction conditions during cold starts. Furthermore, they fail to decouple the content attributes (functional essence) and style attributes (appearance) of visual features. When the same visual pattern corresponds to different functional semantics, erroneous recommendations of "visually similar but functionally different" are prone to occur, such as mistakenly recommending high-caffeine energy drinks to users needing sleep aids. Similarly, CN114004667B ignores spatial topology and execution constraints. Even if the recommendation results are visually relevant, they may fail to be implemented due to limited spatial accessibility (e.g., goods are blocked by obstacles), conflicting resource constraints (e.g., zero parts inventory), or excessive operational complexity (e.g., case analysis requires special permissions). When cold start information is incomplete, the semantic ambiguity of visual features and the non-linear changes in execution constraints are coupled together. The system cannot accurately extract functional semantics from visual features, nor can it adapt to dynamic execution requirements. Ultimately, it is difficult to achieve the cognitive leap from "visual similarity" to "functional matching" and then to "executable solutions," severely restricting the accuracy and reliability of recommendations in cold start scenarios. Summary of the Invention

[0005] This invention is applicable to cold-start scenarios such as smart retail (recommendation of newly listed products), industrial inspection (recommendation of new component adaptations), and medical imaging (recommendation of rare case images). It is particularly suitable for scenarios where new target objects lack user interaction data and where there are high requirements for the functional matching and feasibility of the recommended results. For example, in smart retail, complementary products are recommended for newly listed food items, ensuring sufficient inventory and easy access; in industrial inspection, suitable assembly components are recommended for new precision parts, taking into account operational complexity; and in medical imaging, reference cases are matched for rare cases, ensuring image accessibility and compliant analysis. To overcome the above-mentioned deficiencies of existing technologies, this invention provides an image cold-start recommendation ranking method and system. It constructs a spatial topology layout relationship graph, a fine-grained visual feature library, and a multi-level feasibility constraint system through multi-source data modeling. Combined with deentanglement representation learning, it achieves multi-dimensional decoupling of visual semantics, effectively solving the problem of visual feature semantic drift under cold start. Through comprehensive scoring matrix construction, three rounds of progressive screening, multi-objective Pareto ranking, and visual confusion disambiguation, it ensures that the recommendation results take into account both functional matching and feasibility, achieving a cognitive leap from "visual similarity" to "executable solution" without relying on historical interaction data. This invention significantly improves the accuracy, reliability, and feasibility of cold start scenario recommendations, providing an efficient solution for cold start recommendations across multiple fields.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Image cold start recommendation ranking methods include:

[0008] Acquire 3D point cloud data and RGB images of the target recommendation scene; construct a spatial topology layout relationship graph based on the 3D point cloud data; construct a fine-grained visual feature library based on the RGB images; construct a multi-level feasibility constraint index system based on the spatial topology layout relationship graph, and generate a feasibility constraint index vector.

[0009] Based on a fine-grained visual feature library, a global semantic representation vector and a local detail representation vector are generated. Through deentanglement representation learning, the global semantic representation vector and the local detail representation vector are decomposed into three independent attribute vectors. The three attribute vectors are combined with a spatial topology layout relationship graph to generate a graph-level representation vector. The three attribute vectors include content attribute vector, style attribute vector, and state attribute vector.

[0010] Based on global semantic representation vectors, local detail representation vectors, three types of attribute vectors, and graph-level representation vectors, a comprehensive scoring matrix is ​​constructed for each target object node in the spatial topology layout relationship graph. Feasibility constraint index vectors are applied to the comprehensive scoring matrix for three rounds of progressive screening to generate a selected candidate set. Multi-objective Pareto ranking and visual confusion disambiguation are performed on the selected candidate set to output the final recommendation sequence.

[0011] The method for constructing the spatial topology layout relationship diagram includes:

[0012] The RANSAC algorithm is used to segment 3D point cloud data, identify target objects in the target recommendation scene, and segment point cloud clusters for each target object. The 3D bounding box of each target object is recorded, and the target object is marked as a target object node. The 3D coordinates of the centroid of each point cloud cluster are calculated, and the Euclidean distance between the centroids of the point cloud clusters is calculated based on the 3D coordinates of the centroids of the point cloud clusters. When the Euclidean distance between the centroids of two point cloud clusters is less than a preset spatial adjacency threshold, a spatial association edge is established in the spatial topology layout relationship graph to connect the corresponding target object node. The spatial association edge is assigned a direction attribute according to the relative positional relationship between the target object nodes, forming a directed spatial topology layout relationship graph.

[0013] The method for constructing the fine-grained visual feature library includes:

[0014] The 3D bounding box of the target object is projected onto the RGB image of the target recommended scene. The RGB image region of each target object is cropped and defined as the target object RGB image. Each target object RGB image corresponds to a target object node in the spatial topology layout relationship graph. For each target object node in the spatial topology layout relationship graph, a multi-scale feature extraction pipeline is designed to extract hierarchical features from the target object RGB image, from global appearance to local details. The target object RGB image and hierarchical features are stored in a fine-grained visual feature library.

[0015] The hierarchical features are five levels of features, specifically including first-level morphological contour features, second-level texture pattern features, third-level identification information features, fourth-level attribute annotation features, and fifth-level state indicator features.

[0016] The first layer of the multi-level feasibility constraint index system is the physical accessibility constraint layer, the second layer is the resource sufficiency constraint layer, the third layer is the operational complexity constraint layer, and the fourth layer is the risk control constraint layer.

[0017] The method for generating the feasibility constraint index vector includes: calculating a physical accessibility score at the physical accessibility constraint layer, calculating a resource sufficiency score at the resource sufficiency constraint layer, calculating an operational complexity score at the operational complexity constraint layer, calculating a risk control score at the risk control constraint layer, and combining the physical accessibility score, resource sufficiency score, operational complexity score, and risk control score to form the feasibility constraint index vector.

[0018] The methods for generating the global semantic representation vector and the local detail representation vector include:

[0019] The RGB image of the target object in the fine-grained visual feature library is input into the global semantic encoder with VisionTransformer as the backbone network to generate the global semantic representation vector of each target object node.

[0020] Hierarchical features from a fine-grained visual feature library are input into a local detail encoder with a deformable convolutional network at its core to generate local detail representation vectors.

[0021] The method for generating graph-level representation vectors includes: constructing similarity graphs, complementary graphs, causal graphs, and temporal graphs based on spatial topology layout graphs and three types of attribute vectors, and then aggregating them through graph neural networks to generate graph-level representation vectors.

[0022] The method for constructing the comprehensive scoring matrix includes:

[0023] Obtain the user's query requirements and encode them into a query vector;

[0024] Based on the query vector, global semantic representation vector, and local detail representation vector, a semantic relevance score is calculated.

[0025] Attribute matching scores are calculated based on query vectors and three types of attribute vectors.

[0026] For each target object node, a centrality score is calculated based on a similarity graph, a connectivity score is calculated based on a complementarity graph, a causal importance score is calculated based on a causal graph, and a temporal correlation score is calculated based on a temporal graph. The four scores are then fused to generate a graph structure association score.

[0027] The semantic relevance score, attribute matching score, and graph structure association score are organized by column to construct a comprehensive scoring matrix. The rows of the matrix correspond to the target object nodes in the spatial topology layout relationship graph, and the columns correspond to each scoring dimension.

[0028] The first round of the three-round progressive screening is hard constraint screening, the second round is soft constraint weighted screening, and the third round is dynamic threshold screening.

[0029] The method for hard constraint screening is as follows: traverse all rows in the comprehensive scoring matrix M, extract the feasibility constraint index vector of the target object node corresponding to each row, identify hard constraint violation nodes with a physical reachability score, resource sufficiency score, operational complexity score, or risk control score of zero; delete the rows corresponding to the identified hard constraint violation nodes from the comprehensive scoring matrix, and generate the hard constraint screening matrix.

[0030] The soft constraint weighting method is as follows: for each target object node retained in the matrix after hard constraint screening, calculate the geometric mean of the four components of its feasibility constraint index vector as the comprehensive feasibility score, and add the comprehensive feasibility score as a new column to the matrix after hard constraint screening to form a soft constraint weighting matrix.

[0031] The dynamic threshold screening method is as follows: collect time series parameters, spatial location parameters, environmental condition parameters, and concurrent state parameters of the target recommendation scenario to construct a context environment parameter vector; dynamically calculate the screening threshold based on the context environment parameter vector, and select target object nodes whose comprehensive feasibility score exceeds the screening threshold from the soft constraint weighted matrix M2 to form a selected candidate set C.

[0032] An image cold start recommendation and ranking system, used to implement the above-described image cold start recommendation and ranking method, the system comprising:

[0033] Modeling constraint module: used to acquire 3D point cloud data and RGB images of the target recommendation scene, construct a spatial topology layout relationship graph based on the 3D point cloud data; construct a fine-grained visual feature library based on the RGB images; construct a multi-level feasibility constraint index system based on the spatial topology layout relationship graph, and generate a feasibility constraint index vector.

[0034] Representation learning decoupling module: Based on a fine-grained visual feature library, it generates global semantic representation vectors and local detail representation vectors. Through deentanglement representation learning, it decomposes the global semantic representation vectors and local detail representation vectors into three independent attribute vectors. The three attribute vectors are combined with a spatial topology layout relationship graph to generate a graph-level representation vector. The three attribute vectors include content attribute vectors, style attribute vectors, and state attribute vectors.

[0035] The filtering and sorting module constructs a comprehensive scoring matrix for each target object node in the spatial topology layout relationship graph based on global semantic representation vectors, local detail representation vectors, three types of attribute vectors, and graph-level representation vectors. It applies feasibility constraint index vectors to perform three rounds of progressive filtering on the comprehensive scoring matrix to generate a selected candidate set. It then performs multi-objective Pareto ranking and visual confusion disambiguation on the selected candidate set to output the final recommended sequence.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] This invention provides ample alternative data support for cold-start recommendations under zero-interaction data conditions through multi-source data fusion modeling, overcoming the limitations of traditional reliance on user interaction data. By leveraging deentangled representation learning to achieve multi-dimensional decoupling of visual semantics, it accurately distinguishes the functional essence and appearance of target objects, effectively avoiding erroneous recommendations that are "visually similar but functionally different." A multi-level feasibility constraint index system quantifies and integrates spatial, resource, operational, and risk constraints, ensuring that the recommendation results align with the real-world execution environment. A comprehensive scoring and multi-objective ranking mechanism balances recommendation relevance, diversity, and execution efficiency, while visual confusion disambiguation further enhances the distinguishability of results. Ultimately, this significantly improves the accuracy, reliability, and feasibility of recommendations in cold-start scenarios, providing an efficient solution for cold-start recommendations in fields such as smart retail and industrial inspection. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of the image cold start recommendation ranking method provided in an embodiment of the present invention;

[0040] Figure 2 This is a schematic diagram of a hierarchical feature extraction method provided in an embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of the structure of the feasibility constraint index system provided in the embodiments of the present invention;

[0042] Figure 4 A schematic diagram illustrating the principle of a three-round progressive screening of the comprehensive scoring matrix provided for embodiments of the present invention;

[0043] Figure 5 This is a functional block diagram of the image cold start recommendation and ranking system provided in an embodiment of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1

[0046] Please see Figure 1 As shown, this embodiment provides an image cold start recommendation ranking method, including:

[0047] Step S10: Obtain the 3D point cloud data and RGB image of the target recommendation scene; construct a spatial topology layout relationship diagram based on the 3D point cloud data; construct a fine-grained visual feature library based on the RGB image; construct a multi-level feasibility constraint index system based on the spatial topology layout relationship diagram, and generate a feasibility constraint index vector.

[0048] Further, step S10 includes:

[0049] Step S11: Obtain the 3D point cloud data and RGB image of the target recommendation scene. Based on the 3D point cloud data, identify the target objects in the target recommendation scene, record the 3D bounding box of each target object, mark the target objects as target object nodes, and form a directed spatial topology layout relationship graph according to the spatial distance and relative position relationship between the target object nodes.

[0050] In step S11, an RGB-D camera is used to scan the target recommendation scene from multiple angles, simultaneously acquiring RGB images and depth maps. The depth map is converted into 3D point cloud data, or 3D point cloud data can be directly acquired via LiDAR. The two methods can be flexibly selected based on the scene's hardware configuration. 3D point cloud data is a set of 3D spatial coordinates, representing the spatial position and shape of objects in the target recommendation scene. RGB images are 2D images containing color information, used to capture the visual appearance features of objects. The RANSAC algorithm is used to segment the 3D point cloud data, dividing it into point cloud clusters for each independent target object. The 3D coordinates of the centroid of each point cloud cluster are calculated. The Euclidean distance between the centroids of point cloud clusters is calculated based on the 3D coordinates of the centroids. When the Euclidean distance between the centroids of two point cloud clusters is less than a preset spatial adjacency threshold, a spatial association edge is established in the spatial topology layout graph connecting the corresponding target object nodes. The spatial association edge is assigned a direction attribute based on the relative positional relationship (up / down, left / right, front / back) between the target object nodes, forming a directed spatial topology layout graph.

[0051] RANSAC, or Random Sample Consensus Algorithm, constructs a fitting model by randomly sampling a subset of 3D point cloud data. It then filters inliers based on the consistency between data points and the model, thus segmenting point cloud clusters into independent objects. This technique overcomes the shortcomings of traditional point cloud segmentation methods, such as sensitivity to noise and susceptibility to outliers. It accurately separates the spatial data of different target objects, avoiding subsequent association inference biases caused by object confusion, and providing a reliable data foundation for constructing spatial topological relationships. The 3D coordinates of the centroid of the point cloud cluster are obtained by calculating the mean coordinates of all points within the cluster. Euclidean distance is calculated using the common method of taking the square root of the sum of the squares of the differences between the coordinates of two spatial points. The spatial adjacency threshold is determined based on the typical size and functional layout density of target objects in the scene. It is set by statistically analyzing the minimum distance range between objects with actual functional relationships. For example, the spatial adjacency threshold range is 1.5 to 2 times the average size of objects in the scene. When the Euclidean distance between the centroids of two point cloud clusters is less than the spatial adjacency threshold, it indicates that the objects are spatially close. Based on the prior knowledge that "spatially close objects often have complementary or substitutive functional relationships," edge connections are established in the spatial topology layout diagram. The relative positional relationship between target object nodes is determined by the sign and absolute value of the coordinate difference. The vertical relationship is determined by the Z-axis coordinate difference, the horizontal relationship is determined by the X-axis coordinate difference, and the front-back relationship is determined by the Y-axis coordinate difference. After assigning directional attributes to the spatially associated edges, the resulting directed spatial topology layout diagram can more accurately represent the functional orientation relationship between objects. For example, the products on the upper and lower shelves may have a retrieval order relationship, and the equipment in front and the accessories behind may have an assembly dependency relationship. For example, in smart retail scenarios, RGB-D cameras scan shelves to acquire 3D point cloud data and RGB images. The RANSAC algorithm is used to segment the point cloud clusters of each product, and the centroid coordinates of each product are calculated. When the Euclidean distance between the centroids of milk and bread is less than the spatial adjacency threshold, an edge connection is established, and the edge's direction attribute is assigned based on their positional relationship. This allows the spatial topology layout diagram to implicitly suggest the functional association of "buying milk may be paired with bread." In industrial inspection scenarios, multi-sensor systems scan assembly lines, segmenting the point cloud clusters of each component. The spatial topology layout diagram represents the assembly position associations of the components, providing a spatial basis for subsequent component recommendations. The construction of the spatial topology layout diagram transforms physical spatial relationships into structured data. Utilizing the prior knowledge of "spatial proximity implying functional association," it replaces the user interaction data missing in the cold start phase, enabling the system to initially infer potential associations between objects. Without this step, subsequent recommendations would lack spatial reference, making it impossible to distinguish between physically close but functionally unrelated objects, leading to blind association inferences.

[0052] Step S12: Project the 3D bounding box of the target object onto the RGB image of the target recommended scene, and crop out the RGB image region of each target object, defining it as the target object RGB image. Each target object RGB image corresponds to a target object node in the spatial topology layout relationship graph. For each target object node in the spatial topology layout relationship graph, design a multi-scale feature extraction pipeline, see [reference]. Figure 2 Extract hierarchical features from the target object's RGB image, ranging from global appearance to local details, and store the target object's RGB image and hierarchical features in a fine-grained visual feature library.

[0053] The hierarchical features are five levels of features, specifically including first-level morphological contour features, second-level texture pattern features, third-level identification information features, fourth-level attribute annotation features, and fifth-level state indicator features.

[0054] The 3D bounding box of the target object is determined by the coordinate extrema of the point cloud cluster. That is, the maximum and minimum values ​​of the point cloud cluster on the X, Y, and Z coordinate axes are taken as the coordinates of the six vertices of the bounding box. The process of projecting the 3D bounding box onto the RGB image is to convert the three-dimensional coordinates into two-dimensional pixel coordinates through the camera intrinsic parameter matrix, and then crop the RGB image region containing only a single target object based on the two-dimensional pixel coordinates to ensure that the subsequent feature extraction is not affected by the background and other objects. The fine-grained feature extraction pipeline employs a cascaded processing strategy, progressively characterizing the target object's features from global to local: The first-level morphological contour features identify edge pixels in the target object's RGB image using the Canny edge detection algorithm, then map these edge pixels to geometric shape parameters using Hough transform. The first-level features capture the overall structure of the target object, providing a basis for preliminary object category judgment. The second-level texture pattern features describe the grayscale distribution patterns of local regions in the target object's RGB image using local binary patterns. Combined with the filtering responses of the Gabor filter bank at different scales and directions, it captures the surface texture information of the target object. Its function is to distinguish differences in materials, surface treatment processes, etc., such as differentiating between metal and plastic parts. The third-level identification information features extract the target object's identification information using optical character recognition technology. The text labels in the RGB images of the object, such as product names, specifications, and part models, are used to identify visual identifiers such as icons and trademarks through the SIFT feature point matching algorithm. The third-level features are directly related to the functional semantics of the target object and are key to distinguishing objects that are "visually similar but functionally different". The fourth-level attribute label features are implemented based on a pre-trained attribute classifier, which is trained on a large number of image samples labeled with attributes such as color, material, and size. It can automatically identify the physical attributes of the target object and provides clear attribute dimension data for attribute decoupling. The fifth-level status indicator features use image differencing, threshold segmentation and other techniques to identify dynamic elements such as the on / off state of LED indicator lights and the display content of LCD screens, capturing the real-time status information of the target object and laying the foundation for the extraction of status attributes. The five-level feature collaboration comprehensively covers the global structure, surface details, semantic identifiers, physical attributes, and dynamic states of the target object. This overcomes the shortcomings of traditional single-granularity features, which either ignore key details or lack semantic connections. It enables the fine-grained visual feature library to distinguish between objects that look similar but have different functions. For example, it can differentiate between high-caffeine energy drinks and sleep aids through identifier information features and attribute annotation features, providing rich and effective raw feature support for subsequent attribute decoupling and accurate recommendations. After the spatial topology layout relationship graphs in steps S12 and S11 are collaborated, the associated storage of "spatial location - visual features" is achieved, enabling the system to know both "where the object is" and "what characteristics the object has," providing rich raw data for subsequent representation learning.Without this step, subsequent attribute decoupling will lack a feature base, making it impossible to achieve a multi-dimensional understanding of visual semantics, resulting in the continued existence of the erroneous recommendation problem of "visually similar but functionally different".

[0055] Step S13: Collect time-series parameters, spatial location parameters, environmental condition parameters, and concurrent state parameters of the target recommendation scene, and construct a context environment parameter vector;

[0056] Time series parameters include previous time, day of the week, holiday status, and seasonal information. Fourier transform is used to convert these parameters into frequency domain signals, extracting periodic patterns such as demand differences between weekdays and weekends, and demand changes due to seasonal transitions. This uncovers demand patterns over time, enabling recommendations to adapt to periodic changes. Spatial location parameters include the geographic coordinates of the target object node, floor information, functional area labels, and the path distance from the target object node to key locations. Geographic coordinates are obtained from the target object node attributes in the spatial topology diagram. Floor information and functional area labels are assigned values ​​according to the scene's preset spatial division rules. Key locations are determined based on the core functional areas of the scene, such as the cashier in a retail scene or an assembly station in an industrial scene. The method for calculating the path distance from the target object node to the key location is as follows: determine the three-dimensional coordinates of the key location, and filter connected paths based on the spatial topology diagram. When filtering connected paths, only node sequences with spatially related edges in the spatial topology diagram are considered; that is, the target object node must reach the key location through the connectivity of adjacent nodes. For each adjacent node (including the target node, intermediate nodes, and key locations) in a connected path, calculate the Euclidean distance between each pair of nodes based on their centroid's 3D coordinates. Sum the Euclidean distances for each connected path, and the smallest sum is the path distance from the target node to the key location. If there is no connecting path between the target node and the key location that satisfies the condition (i.e., no continuous spatial edge connecting them), the path distance is set to infinity, and the corresponding path reachability sub-index in the physical reachability score is assigned a value of 0. The path distance is used to quantify the spatial reachability cost of the target object, providing data support for physical reachability constraints.

[0057] By deploying temperature, humidity, light, and noise sensors in the target recommendation scenario, environmental parameters such as temperature, humidity, light intensity, and noise level are collected in real time. This data is used to determine whether these parameters exceed normal ranges. The normal range is defined by statistically analyzing the distribution range of environmental parameters under normal scenario conditions. When collected environmental parameters exceed the normal range, they are marked as abnormal environments. These abnormal environments are used to assess the impact of the environment on recommendation needs; for example, in high-temperature environments, the recommendation priority of cooling-related products is increased. The system also statistically analyzes concurrent state parameters such as personnel density, equipment occupancy rate, and network load in the current scenario. These real-time updated concurrent state parameters reflect the resource competition situation in the current scenario, providing dynamic data for resource sufficiency constraints and operational complexity constraints. For example, in high-person-density scenarios, easily accessible and convenient target objects are prioritized for recommendation. When integrating time series parameters, spatial location parameters, environmental condition parameters, and concurrent state parameters into a context environment parameter vector, standardization is used to convert each parameter into a value of the same magnitude, ensuring the computability of the vector. This provides comprehensive environmental context support for subsequent dynamic weight adjustment and constraint index adaptation, enabling the recommendation system to perceive dynamically changing scenario requirements and overcome the requirement mismatch problem caused by static decision-making in traditional recommendation systems.

[0058] The method for integrating time series parameters, spatial location parameters, environmental condition parameters, and concurrent state parameters into a contextual environment parameter vector includes: first, performing "feature decomposition → numerical encoding → 0-1 standardization" on each parameter category sequentially to ensure that all feature magnitudes are uniform under a single parameter category; then, concatenating all standardized features in a fixed order of "time series → spatial location → environmental conditions → concurrent state" to form a one-dimensional computable vector. Taking a scenario at a certain moment in a smart retail supermarket shelf area as an example, the original parameter inputs are as follows: time series parameters are Wednesday 14:30, non-holiday, spring; spatial location parameters are the geographic coordinates of the target product (milk) (X=6m, Y=4m, Z=1.5m), 2nd floor food area, path distance to the checkout counter 8m; environmental condition parameters are temperature 26℃, humidity 55%, light intensity 600 lux, noise 52dB, no anomalies; concurrent state parameters are personnel density 1.8 people / m². 2 The equipment utilization rate is 25%, and the network load is 35%. The numerical and standardization processing of each parameter category is shown in Tables 1-4. The standardized sub-vector of the time series parameter is [0.63, 0.43, 0.00, 0.25], the standardized sub-vector of the spatial location parameter is [0.60, 0.50, 0.50, 0.67, 1.00, 0.00, 0.00, 0.40], the standardized sub-vector of the environmental condition parameter is [0.65, 0.55, 0.60, 0.31, 0.00], and the standardized sub-vector of the concurrent state parameter is [0.36, 0.25, 0.35].

[0059] Table 1. Numericalization and Standardization of Time Series Parameters

[0060]

[0061] Table 2 Numericalization and Standardization of Spatial Location Parameters

[0062]

[0063] Table 3 Numericalization and Standardization of Environmental Condition Parameters

[0064]

[0065] Table 4 Numericalization and Standardization of Concurrent State Parameters

[0066]

[0067] The standardized subvectors of the four types of parameters are concatenated in sequence to form a 20-dimensional context environment parameter vector: [0.63,0.43,0.00,0.25,00.60,0.50,0.50,0.67,1.00,0.00,0.00,0.40,0.65,0.55,0.60,0.31,0.00,0.36,0.25,0.35]. It should be noted that the numerical representation of all parameters must be determined based on the actual situation of the scenario. The temperature (0-40℃), path distance (0-20m), or logical encoding rules mentioned above are merely illustrative examples. The 0-1 standardization process resolves the issue of vector incomputability caused by differences in the magnitude of the original parameters. Furthermore, if new parameters are added subsequently, they can be expanded using the logic of "numerization → standardization → appending to the end of the vector" without altering the core integration rules.

[0068] Step S14: Based on the spatial topology layout relationship diagram and context environment parameter vector, construct a multi-level feasibility constraint index system and generate a feasibility constraint index vector.

[0069] See Figure 3The feasibility constraint index system consists of four layers: the first is the physical accessibility constraint layer, the second is the resource sufficiency constraint layer, the third is the operational complexity constraint layer, and the fourth is the risk control constraint layer. The method for generating the feasibility constraint index vector includes: calculating the physical accessibility score in the physical accessibility constraint layer, the resource sufficiency score in the resource sufficiency constraint layer, the operational complexity score in the operational complexity constraint layer, and the risk control score in the risk control constraint layer; then combining the physical accessibility score, resource sufficiency score, operational complexity score, and risk control score to form the feasibility constraint index vector. The four-layer structure of the multi-level feasibility constraint index system follows a progressive relationship of execution logic. The physical accessibility constraint layer serves as the basic constraint, ensuring that the target object has the prerequisite for acquisition at the spatial level; the resource sufficiency constraint layer serves as the guarantee constraint, ensuring that there is no shortage of the required resources during execution; the operational complexity constraint layer serves as the adaptation constraint, ensuring that the operation behavior matches the scenario capabilities and user capabilities; and the risk control constraint layer serves as the fallback constraint, avoiding potential security or compliance risks during execution. The four layers of constraints are interconnected to form a complete execution guarantee chain, making up for the shortcomings of traditional recommendation systems that only focus on semantic relevance and ignore the feasibility of execution.

[0070] The calculation method for physical accessibility score includes: obtaining the user's current position through the scene positioning module and converting it into user node coordinates in a spatial topology layout diagram; using Dijkstra's algorithm to calculate the shortest path length from the user's current position to each target object node in the spatial topology layout diagram; normalizing the path length to the 0-1 range and marking it as path accessibility; identifying obstacles on the path by comparing the current 3D point cloud data with the initial barrier-free point cloud data, including fixed obstacles such as walls and pillars and dynamic obstacles such as temporary stacked items; calculating path accessibility based on the degree of obstruction of obstacles, with complete blockage scoring 0 points, and partial obstruction scoring linearly according to the difficulty of passage, the degree of obstruction being quantified by the proportion of the cross-sectional area occupied by the obstacle on the path; and then assessing the complexity of retrieving the target object using a preset ergonomic model (based on the average range of limb movement of adults aged 18-60, with a comfortable retrieval height set at 0.8-1.5m, a retrieval angle at 0-45°, and a retrieval weight ≤5kg). The system determines whether the height, angle, and weight of the retrieval are within the range of human comfort. The comfort zone is scored as 1.0 (out of 1). The score is reduced to 0.6-0.8 if uncomfortable actions such as bending over or tiptoeing are required, and to 0.3-0.5 if auxiliary tools such as ladders are needed. Finally, the three sub-indicators of path accessibility, obstacle clearance, and retrieval complexity are weighted and averaged to obtain the physical accessibility score. The weights are set according to the characteristics of the scenario. For example, the path weight is high in the warehouse scenario and the retrieval weight is high in the retail scenario.

[0071] The calculation method for resource sufficiency score includes: In the second layer of resource sufficiency constraint, the inventory management system is queried in real time to obtain the current inventory quantity of each target object node. The inventory quantity is mapped to an inventory sufficiency score, with zero inventory recorded as 0 points (a hard constraint), sufficient inventory as 1 point, and low inventory as a proportional reduction in score. For target objects involving equipment usage, the remaining available time of the equipment is obtained through the equipment monitoring module. The calculation method is the rated service life minus the cumulative running time, and the ratio of the remaining time to the expected usage time is used as the equipment availability score. Simultaneously, network bandwidth availability and computing resource utilization are monitored. For target objects requiring network transmission or computing support, the ratio of the current available bandwidth to the required bandwidth is used as the network resource score, and the ratio of idle computing resources to the required computing volume is used as the computing resource score. The inventory sufficiency score, equipment availability score, network resource score, and computing resource score are calculated using a geometric mean to obtain the overall resource sufficiency score. The geometric mean is used because a severe shortage of any resource will lead to overall infeasibility.

[0072] The calculation method for the operation complexity score includes: In the third layer of operation complexity constraint, firstly, the acquisition or usage process of the target object is decomposed, the number of operation steps required from start to finish is counted, and the complexity mapping function is dynamically adjusted according to the current time parameter in the context environment parameter vector. During late night (0:00-5:00), user fatigue is high and tolerance for complex operations is low, so the step complexity score is calculated as exp(-λ×number of steps), where exp() is an exponential function with the natural logarithm base e, and λ is a decay coefficient. For example, λ=0.3 at night and λ=0.2 during the day, so that the more steps there are, the lower the score. Then, the average time for each step is counted based on historical operation data, and the queuing time is increased according to the personnel density parameter in the context environment parameter vector. For example, when the personnel density exceeds 1.5 people / m², the queuing time is increased. 2When an additional (personnel density × 2) minutes of expected waiting time is added, the total expected time cost is calculated. The time efficiency score is then calculated as (1 - total expected time cost / user-acceptable time limit). The user-acceptable time limit is a preset industry average threshold for the scenario, such as 5 minutes for retail scenarios and 15 minutes for industrial scenarios. Users can customize this threshold through the system interface. When the total expected time cost exceeds the user-acceptable time limit, a score of 0 is recorded, indicating a hard constraint. Next, it is determined whether the operation requires special tools or permissions. The required tool list is identified by querying the target user's operation manual, and the tool availability is determined based on the functional area parameters in the context environment parameter vector. If the target user and the user are not in the same functional area, the tool availability is reduced by 30%. The permission management system is then used to determine whether the user has the necessary permissions. Corresponding permissions are assigned, and the tool permission score is calculated comprehensively. Full compliance is scored as 1 point, while the lack of key tools or permissions is scored as 0 points, indicating a hard constraint. Partial compliance is reduced in score between 0.3 and 0.7 according to importance. Finally, the operation complexity score is calculated by weighted harmonic average of the three sub-indicators: step complexity score, time efficiency score, and tool permission score. The formula is: Operation complexity score = 1 / (w1 / step complexity score + w2 / time efficiency score + w3 / tool ​​permission score), where the weights w1 = 0.3, w2 = 0.4, and w3 = 0.3. The weights are dynamically adjusted according to the environment anomaly markers in the context environment parameter vector. In abnormal environments, the time efficiency weight w2 is increased to 0.5 to promote rapid decision-making. The harmonic average can highlight the weakest link effect, and any sub-indicator that is too low will significantly lower the overall score.

[0073] The risk control score is calculated as follows: In the fourth layer of risk control constraints, the basic operation failure rate of each target object is statistically analyzed based on historical operation logs. A safety score is calculated using an S-shaped function: 1 / (1+exp(5×(basic operation failure rate - 0.3))). This function outputs 0.5 when the basic operation failure rate is 0.3; the higher the basic operation failure rate, the lower the safety score. When assessing the potential losses after an operation failure, the loss of target object value, associated equipment damage losses, and production interruption losses are comprehensively considered. The associated impact is assessed based on the equipment occupancy rate parameter in the context environment parameter vector. In high equipment occupancy rate scenarios, the associated losses from a single point of failure are amplified by (1+equipment occupancy rate). The loss controllability score is calculated as exp(-total loss amount / acceptable loss threshold). The acceptable loss threshold is the annual risk budget allocation value preset for the scenario, categorized by target object type. When the total loss... When the loss amount equals the acceptable loss threshold, the score is approximately 0.37. The score rapidly decays after exceeding the acceptable loss threshold. Each target object is labeled with a safety hazard level according to the scenario safety standards, ranging from no hazard to high risk in 5 levels. The safety level score is calculated as (5 - safety hazard level) / 4, with 1 point for no hazard and 0 points for high risk, indicating a hard constraint. Finally, the three sub-indicators—safety score, loss controllability score, and safety level score—are used to calculate the risk control score using a weighted minimum function. The formula is: Risk Control Score = min(Safety Score, Loss Controllability Score, Safety Level Score) × Dynamic Weighting Coefficient. The dynamic weighting coefficient is adjusted based on the reduced efficiency of accident rescue response at night. It is set to 0.7 at night (20:00-6:00) to improve the risk control standard, and 1.0 during the day. The minimum function is used because risk control follows the barrel principle; any loss of control in any risk dimension is unacceptable.

[0074] After calculating the scores for the four layers of constraints, the system sequentially combines the physical reachability score, resource sufficiency score, operational complexity score, and risk control score into a four-dimensional feasibility constraint index vector. The value range of each component in the vector is normalized to the interval [0,1], where 0 represents a completely infeasible hard constraint violation, such as zero inventory, complete path blocking, lack of necessary permissions, or high-risk prohibited items. 0 to 1 represent different degrees of soft constraint satisfaction, and 1 represents complete satisfaction of the constraint in that dimension. This standardized vector representation makes constraints of different types and dimensions comparable and computable, providing a quantitative basis for constraint perception and screening in the subsequent step S32. By distinguishing between hard and soft constraints, the system can flexibly weigh the importance of different constraints while ensuring basic feasibility, and dynamically adjust the constraint evaluation criteria according to the context environment parameter vector, ensuring that in cold start scenarios, it can both filter out unexecutable target objects and retain a sufficient candidate set for recommendation.

[0075] Step S10, through the collection and structured modeling of multi-source data, constructs a spatial topology layout diagram, a fine-grained visual feature library, a contextual environment parameter vector, and a feasibility constraint index system, forming an alternative data foundation for cold-start scenarios. This solves the information scarcity problem in the cold-start phase. By using multi-dimensional modeling of spatial relationships, visual features, context, and constraints, it replaces the traditional reliance on user interaction data, enabling the system to conduct recommendation reasoning under zero-interaction conditions. The spatial topology layout diagram provides spatial context for visual features, the fine-grained visual feature library provides object attribute basis for constraint evaluation, the contextual environment parameter vector provides a dynamic adjustment basis for feature understanding and constraint calculation, and the feasibility constraint index system ensures the executability of the recommendation results. This multi-dimensional collaboration enables the system to overcome the dual dilemmas of "pure visual feature semantic ambiguity" and "execution constraint mismatch," not only distinguishing objects that are similar in appearance but different in function, but also ensuring that the recommendation results adapt to the dynamics of the scenario and execution conditions, providing comprehensive and reliable foundational support for subsequent representation learning and ranking optimization. Without step S10, subsequent steps will lack data input and structured knowledge support. Feature decoupling will be impossible due to the lack of fine-grained features, and constraint screening will be impossible due to the lack of quantitative indicators. Ultimately, the recommendation will remain at the level of "visual similarity" and will not be able to achieve the cognitive leap to "functional matching" and "executable solutions".

[0076] Step S20: Based on the fine-grained visual feature library, a global semantic representation vector and a local detail representation vector are generated. Through deentanglement representation learning, the global semantic representation vector and the local detail representation vector are decomposed into three independent attribute vectors. The three attribute vectors are combined with the spatial topology layout relationship graph to generate a graph-level representation vector. The three attribute vectors include content attribute vector, style attribute vector, and state attribute vector.

[0077] Further, step S20 includes:

[0078] Step S21: Input the RGB image of the target object in the fine-grained visual feature library into the global semantic encoder to generate a global semantic representation vector for each target object node.

[0079] The global semantic encoder uses VisionTransformer as its backbone network, and its construction process closely relies on the fine-grained visual feature library in step S12. First, the RGB image of the target object is divided into fixed-size image blocks. The size of the image blocks is determined based on the typical size statistics of the target object in the scene, ensuring that each image block contains basic visual information without omitting key feature regions. Each image block is transformed into a fixed-dimensional image block embedding vector through linear projection, achieving a vectorized representation of the image data. A learnable positional encoding is introduced, assigning a positional encoding vector to each image block, randomly initialized according to coordinate rules. The dimension of the positional encoding vector is consistent with the image block embedding vector. The positional encoding vector and the image block embedding vector are added element-wise to obtain an image block embedding vector with positional information. This enables the global semantic encoder to perceive the spatial relationship between different image blocks, compensating for the lack of spatial awareness inherent in the VisionTransformer architecture itself.

[0080] In the self-attention mechanism computation phase, the similarity between each image patch embedding vector with location information and all other image patch embedding vectors with location information is first calculated using dot product operations. This is then normalized using Softmax to obtain an initial attention distribution. Based on this initial attention distribution, all image patch embedding vectors with location information are weighted and fused to generate preliminary image patch feature vectors with global context. Crucially, the system uses a feature matching algorithm to detect whether each image patch contains identifier features or state indicator features from a fine-grained visual feature library. The criterion for "high semantic value features" is that the semantic correlation between the feature and the target object's functional label exceeds a preset correlation threshold. The correlation is determined by calculating the cosine similarity between the feature vector and the predefined functional label embedding vector. The method for constructing the functional label embedding vector is as follows: first, a functional category vocabulary is extracted from the scene domain knowledge base. For example, the retail scene includes over 200 functional labels such as "sleep aid," "energizer," "nutritional supplement," and "cleaning," while the industrial scene includes "assembly," "inspection," and "transmission." Over 150 functional tags, such as "sealed," were identified. A Word2Vec model was then pre-trained on these tags using textual data from product manuals, technical documents, and user reviews within the relevant context. A Skip-gram architecture was used to learn a 300-dimensional embedding vector for each functional tag. During training, a window size of 5 and a negative sampling count of 10 were set, iterating for 100 rounds until the loss function converged. Finally, the generated functional tag embedding vectors were stored as a lookup table. Each image patch feature was mapped to the same 300-dimensional vector through a fully connected layer, and its cosine similarity was calculated with the functional tag embedding vector to determine the association degree between the image patch and each functional category. The association threshold was set based on the statistical distribution of the core semantic features of the functional category; in this example, it was set to 0.62. When such high semantic value features are detected, a preset enhancement coefficient is automatically added to the initial attention value of the image patch. The enhancement coefficient is determined by: firstly, statistically analyzing the percentage increase in accuracy of the identifier information features and state indicator features in historical functional semantic recognition, and then mapping the percentage to the interval [1.1, 1.5] as the initial enhancement coefficient. Subsequently, the model is trained and iteratively adjusted with the recommendation accuracy as the optimization goal, so that the enhanced attention distribution more accurately focuses on high semantic value regions. This allows the weighted fused image patch feature vector to strengthen the function-related semantic information, enabling the encoder to prioritize these regions that can directly reflect the functional essence of the target object when calculating attention.To further enrich the granularity of representation, a multi-scale feature pyramid is constructed. The self-attention outputs of different layers of VisionTransformer are used as feature maps of different scales. The shallow layer output corresponds to local texture details, the middle layer output corresponds to object component relationships, and the deep layer output corresponds to global semantic concepts. Through a cross-layer connection mechanism, the feature maps of different scales are adjusted to the same dimension by upsampling or downsampling and then added element by element. Learnable fusion weights are introduced during the fusion process, enabling the system to dynamically adjust the contribution of features at different scales according to scene requirements. Finally, through this multi-scale fusion process, a global semantic representation vector that takes into account global semantics, key functional features, and spatial structure is generated.

[0081] Step S21 addresses the shortcomings of traditional global feature extraction methods, which either focus solely on overall semantics while ignoring key local information or rely on fixed receptive fields to fail to capture long-range dependencies. It overcomes the spatial limitations of local features by modeling the relationships between arbitrary regions in an image using the self-attention mechanism of VisionTransformer. An attention enhancement mechanism enables the encoder to accurately focus on high-value semantic regions, avoiding semantic dilution caused by indiscriminate attention. Multi-scale fusion preserves feature information at different levels, avoiding the limitations of single-scale representation. This generates a global representation that includes both the overall semantic structure of the target object and highlights key functional features, allowing the system to understand the semantic essence of the target object from a global perspective while not overlooking key local information affecting functional discrimination. For example, in smart retail scenarios, it can simultaneously capture the overall packaging style of a product and the functional labeling text on the packaging. This step, in deep collaboration with the fine-grained visual feature library from step S12, provides clear target region guidance for attention enhancement. This gives the previously general VisionTransformer architecture scene-specific feature focusing capabilities, allowing the encoder to no longer blindly learn global features but instead selectively strengthen region representations related to functional semantics. This significantly improves the discriminative power of the global semantic representation vector, laying a precise foundation for subsequent attribute decoupling. Without this step, subsequent attribute decoupling would lack global semantic support, resulting in the inference of attributes only from local details, failing to form a complete attribute understanding, and significantly reducing the generalization of feature representations.

[0082] Step S22: Input the hierarchical features from the fine-grained visual feature library into the local detail encoder to generate a local detail representation vector;

[0083] A local detail encoder and a global semantic encoder are constructed in parallel, with a deformable convolutional network at its core, to accurately capture fine-grained differences in target objects. The core of deformable convolution is learning an offset for each sampling point of the standard convolutional kernel, enabling the kernel to adaptively change its shape to better match the geometry of the target object in the fine-grained visual feature library. In constructing the deformable convolution, an offset prediction network is first designed. This network consists of convolutional layers and activation functions. The input is hierarchical features from the fine-grained visual feature library, and the output is an offset vector with the same number of sampling points as the standard convolutional kernel. Each element of the offset vector corresponds to the offset value of a sampling point in the x-axis and y-axis directions. The offset sampling point coordinates are mapped onto the feature map using bilinear interpolation. After obtaining the corresponding feature values, a weighted sum is performed with the learnable convolutional weights to obtain the output features of the deformable convolution. The local detail encoder dynamically adjusts the shape of the receptive field by parsing the context environment parameter vector. For example, when the light intensity in the environmental condition parameter is lower than the preset range, the offset prediction network will generate a larger offset to expand the receptive field and capture more detailed information. When the personnel density in the concurrent state parameter is high, the receptive field will be reduced to focus on the core area of ​​the target object and avoid interference from the surrounding environment.

[0084] The local detail encoder uses a region proposal mechanism based on a five-level hierarchical feature library in a fine-grained visual feature set to dynamically determine regions of interest (ROIs). First, a sliding window traverses the RGB image of the target object, generating multiple candidate regions. The information entropy of each candidate region is calculated, based on the probability distribution of pixel grayscale values ​​within the region, reflecting the richness of information contained within. When the information entropy of a candidate region exceeds a region information entropy threshold, it is marked as a region of interest. The process for determining the region information entropy threshold is as follows: Information entropy is collected from over 1000 typical key detail regions in the scene (such as part model identifiers and product function parameter annotations), and 1.2 times the average is taken as the threshold to ensure that only regions with information density higher than the average level are selected. Feature alignment is performed on the features of each ROI, and standardization is used to map the features of different ROIs to the same dimension and distribution space, ensuring feature comparability. A fine-grained attention mechanism is introduced to distinguish between primary features and decorative features. Primary features refer to the core functional attributes of the target object nodes, while decorative features refer to appearance and ornamentation elements. The fine-grained attention mechanism distinguishes between main features and decorative features by training a binary classifier. The training samples for the binary classifier are image feature samples labeled with core functional attributes (such as the assembly interface of a part and the core function of a product) and appearance decoration elements (such as the surface pattern of a part and the packaging pattern of a product). After training, the class judgment result is output for each local feature. The classification confidence is calculated by the output probability of the binary classifier (the closer the output probability is to 1, the higher the confidence). The main features are assigned a higher importance score. The importance score is obtained by mapping the classification confidence to the 0-1 range through the sigmoid function. Finally, all local features are weighted and fused to generate a local detail representation vector.

[0085] Step S22 addresses the shortcomings of traditional convolutional networks, such as fixed receptive fields, inability to adapt to target objects of different shapes and sizes, and difficulty in distinguishing core functional details from non-critical decorative details. It utilizes adaptive sampling through deformable convolution to adapt to complex geometric structures, a region proposal mechanism to accurately locate high-value detail regions, and a fine-grained attention mechanism to strengthen the representation of core functional features. This accurately captures local detail differences in target objects, especially function-related details, effectively distinguishing between similar-looking but functionally different target objects. For example, it differentiates industrial parts by the texture differences between metal and plastic, or beverages by the text labels "sleep aid" and "energy" on packaging. Collaboration with contextual parameter vectors enables dynamic adaptation of detail extraction to changes in the scene environment, avoiding detail loss or interference caused by environmental fluctuations. This collaboration allows the local detail encoder to stably extract key functional details even in dynamically changing scenes. For example, in industrial inspection scenarios, when equipment vibrates, the robustness of detail extraction is improved by narrowing the receptive field to focus on local stable areas. The local detail representation vector provides rich detail support for attribute decoupling, enabling attribute decoupling to be based not only on global semantics but also on local details, improving the accuracy of attribute separation. Without this step, the system will be unable to capture the fine-grained differences of the target objects, resulting in the system only being able to judge similarity from the global appearance. It will be unable to distinguish between objects that are "visually similar but have different detailed functions," causing attribute decoupling to lose its basis at the detailed level, which in turn affects the accuracy of the recommendations.

[0086] Step S23: Input the global semantic representation vector and the local detail representation vector into the attribute decoupling module, and decompose them into mutually independent content attribute vector, style attribute vector and state attribute vector through deentanglement representation learning;

[0087] The attribute decoupling module contains three parallel attribute decoders: a content attribute decoder, a style attribute decoder, and a state attribute decoder. All three decoders employ a multi-layer fully connected network structure. The number of layers is determined based on the complexity of the attributes in the scene, with the number of neurons in each layer decreasing progressively from the input dimension to the output dimension. The ReLU activation function is used to introduce non-linear expressiveness, while the output layer uses a linear activation function to generate a fixed-dimensional attribute vector. The content attribute decoder receives a global semantic representation vector and a local detail representation vector, and extracts features relevant to the functional essence of the target object through weighted bias initialization to generate a content attribute vector. The style attribute decoder receives the same input, extracts visually relevant features, and generates a style attribute vector. The state attribute decoder also receives a global semantic representation vector and a local detail representation vector, identifying dynamically changing features. State attribute vectors are generated. To ensure the independence of the three types of attribute vectors, three adversarial training mechanisms are introduced: content-style discriminator, style-state discriminator, and content-state discriminator. Each discriminator consists of fully connected layers, with the input being one type of attribute vector and the output being the predicted value of the other type of attribute vector. During training, the attribute decoder and the discriminator engage in adversarial competition. The attribute decoder aims to prevent the discriminator from accurately predicting the other type of attribute vector from its own output attribute vector, while the discriminator aims to predict as accurately as possible. Training optimization is achieved by minimizing the discriminator's prediction loss and maximizing the decoder's adversarial loss using a joint loss function: joint loss = discriminator prediction loss + 0.5 × decoder adversarial loss. The joint loss is minimized using the Adam optimizer, where the weight coefficient of 0.5 is used to balance the contributions of the two types of losses, thus decoupling the attribute vectors.

[0088] Step S23 addresses the shortcomings of traditional feature representations, which often involve a mix of attributes and make it difficult to control any single attribute dimension, leading to inaccurate matching of specific needs in recommendation decisions. It utilizes a parallel decoder to achieve targeted extraction of single attributes, adversarial training to ensure the independence of attributes, and information bottleneck control to maintain the simplicity of the representation. The mixed feature representation is decomposed into independent content, style, and state attribute vectors, allowing for flexible selection of the attribute dimension to focus on based on recommendation needs. For example, in function-oriented scenarios, matching can be based solely on content attribute vectors, while in aesthetic-oriented scenarios, style attribute vectors are heavily referenced, improving the interpretability and controllability of recommendations. Global semantic representation vectors provide a global basis for attribute extraction, while local detail representation vectors supplement the detailed support for attribute extraction. This decoupling of attributes is not only comprehensive but also precise. This synergy allows the three types of attribute vectors to simultaneously reflect the global and local attribute features of the target object. For example, the content attribute vector includes both overall functional positioning and local functional details, making functional matching more accurate. The decoupled attribute vectors enable subsequent matching calculations to focus specifically on the attribute dimensions relevant to user needs, avoiding interference from irrelevant attributes and improving matching accuracy. Without this step, the problem of mixed attributes in feature representation will persist, causing the system to be unable to distinguish the essential differences between "visually similar but functionally different" features. Recommendation decisions will still be affected by irrelevant attributes such as appearance style, and the cognitive leap from "visual similarity" to "functional matching" will not be achieved.

[0089] Step S24: Based on the spatial topology layout relationship graph and the three types of attribute vectors, construct similarity relationship graph, complementary relationship graph, causal relationship graph and temporal relationship graph, and aggregate them through graph neural network to generate graph-level representation vectors.

[0090] Step S24 uses a spatial topology layout relationship graph as a structured skeleton and content attribute vectors, style attribute vectors, and state attribute vectors as semantic cores. Through a progressive process of constructing multi-type relationship graphs, graph neural network aggregation, and graph pooling integration, it generates graph-level representation vectors that integrate spatial structure and semantic association. This step primarily addresses the problem of accurately capturing complex relationships between target objects when there is a lack of interactive data in cold start scenarios, providing structured association knowledge support for subsequent recommendations.

[0091] The node connections in the similarity graph are based on content attribute vectors. By calculating the cosine similarity of the content attribute vectors of any two target object nodes, an edge connection is established when the similarity is higher than a preset similarity threshold, and the edge weight is directly assigned to the cosine similarity. The similarity threshold is not a fixed value, but is determined based on the statistical results of the content attribute vector similarity of functionally similar target objects in the scene. By analyzing the attribute vector distribution of known functionally related objects in the scene, a critical value that can distinguish between functionally similar and functionally dissimilar objects is extracted as the similarity threshold to ensure that edge connections are only established between nodes that are truly functionally similar. The edge determination of the complementary relationship graph is based on the spatial co-occurrence pattern in the spatial topology layout graph. Spatial co-occurrence is determined by analyzing the adjacency density of target object nodes in the spatial topology layout graph. When two target object nodes frequently appear as common neighbors in space (i.e., there are multiple common adjacent nodes), or when there are multiple short path connections between them, they are determined to be potential complementary relationships. For scenarios with some existing operational data, the co-occurrence frequency in the short term can be supplemented as an auxiliary criterion. The co-occurrence threshold is set according to the typical spatial layout pattern of functionally complementary objects in the scenario to ensure that only spatially closely related nodes are determined to be complementary relationships, avoiding false complementary associations caused by coincidental spatial layout. The construction of directed edges in the causal relationship graph integrates the relative positional relationships and functional logic of the spatial topology layout relationship graph. First, based on the relative positions of nodes in the spatial topology layout relationship graph (up / down, left / right, front / back), potential functional orientation relationships are initially determined. Then, functional semantic verification is performed using content attribute vectors. For example, a device node in front in space might have a content attribute vector representing the "main device" function, while a node behind it might represent the "supporting accessory" function. Therefore, directed edges are established from the main device node to the accessory node, accurately representing the causal logic of the assembly sequence or usage process. The construction of edges in the temporal relationship graph uses the time series parameters from step S13 as the time dimension, recording the state changes of the target object nodes at different time stages. The edge weights are set as the cosine similarity of the node state attribute vectors at different time stages. Higher similarity indicates a smoother state change, while lower similarity indicates a significant state change. This relationship graph captures the dynamic evolutionary patterns of the target object.

[0092] For each type of relational graph, a suitable graph neural network architecture is selected for feature aggregation to ensure that different types of relational information are accurately captured. The core requirement for similarity graphs is to equally aggregate the functional semantics of all similar nodes. Therefore, a graph convolutional network is used to strengthen the functional semantic associations between similar objects by averaging the content attribute vectors of the neighboring nodes of each target object node, enabling node features to absorb the common semantic information of similar objects. The core requirement for complementary graphs is to distinguish the importance of different complementary nodes. Therefore, a graph attention network is used to learn the attention weights of each neighboring node, giving higher weights to neighboring nodes that have higher functional supplementary value to the current node. This highlights the influence of key complementary objects during aggregation and avoids semantic dilution caused by indiscriminate aggregation. The core requirement for causal graphs is to preserve the directional logic of functional orientation. Therefore, a directed graph network is used to pass information only along the direction of directed edges during feature aggregation, ensuring that the causal logic from the preceding node to the subsequent node is not lost, and that the aggregated features reflect the sequential dependencies of functional execution. The core requirement of temporal relationship graphs is to capture the dynamic changes of states. Therefore, temporal graph networks are adopted, and a time attention mechanism is introduced to assign different time weights to the features of nodes at different time stages. The weight of recent states is higher than that of distant states, so as to accurately capture the dynamic evolution of the features of the target object.

[0093] To further capture higher-order association patterns between target objects, a meta-path-guided random walk strategy is introduced. Meta-path templates are predefined based on the core association patterns of the scene. Exemplary meta-path templates include "similar-complementary-causal" and "temporal-similar-complementary," with each template corresponding to a typical multi-type association combination. Random walks are performed along the meta-path in a multi-relationship graph, generating a sequence of target object nodes containing multiple types of associations. The embedding representation of the node sequence is learned through a Skip-gram model, encoding higher-order association patterns as latent features of the nodes, achieving deep mining from direct to indirect associations. When the system detects an undefined association combination in the node sequence generated by the walk, i.e., a new higher-order association pattern, it automatically analyzes the relational logic and semantic associations of the nodes in the sequence, extracts a new meta-path template, and updates it to the template library, enabling the system to adapt to changes in association patterns in the scene.

[0094] Graph pooling is used to generate graph-level representation vectors. The input consists of attribute vectors of all target object nodes after graph neural network aggregation and meta-path embedding learning, including content, style, and state attribute vectors. A pooling matrix is ​​constructed, initially based on the node feature importance weights output by the graph neural network and the higher-order association contributions obtained from meta-path embedding learning. The dimensional design strictly matches the total number of target object nodes and the dimension of the attribute vectors; the number of rows equals the total number of target object nodes, and the number of columns equals the total dimension of the attribute vectors, ensuring that each node's attributes correspond to unique weight coefficients. The weights of the pooling matrix are learned through model training, with the accuracy of recommendation decisions as the optimization objective. This allows the pooling matrix to automatically learn the contribution of different nodes to recommendation decisions, prioritizing nodes with high contributions. The selected node attribute vectors are then weighted and summed, with weights consistent with the learning results of the pooling matrix. Finally, a fixed-dimensional graph-level representation vector is generated. This vector integrates the semantic attributes, spatial associations, higher-order associations, and dynamic evolution information of the target object, comprehensively reflecting the target object's position and value within the entire association network.

[0095] Step S24 addresses the shortcomings of traditional recommendation systems that rely solely on single relationships, failing to comprehensively capture complex relationships between target objects and lacking interaction data to support association inference in cold-start scenarios. It constructs a multi-relationship graph to comprehensively cover similar, complementary, causal, and temporal relationships, and uses a graph neural network to aggregate structured knowledge. Meta-paths guide the capture of higher-order relationships. Dispersed target object nodes are integrated into structured knowledge through various relationships, generating graph-level representation vectors containing global relationship information. This enables the discovery of implicit relationships between target objects, establishing connections even without direct interaction records through multi-step association inference. For example, a meta-path of "similar products - complementary products" can recommend complementary products to newly listed products. Spatial topology provides a basic framework for relationships, while three types of attribute vectors enrich the semantic connotation of relationships, upgrading them from simple spatial proximity to composite relationships combining semantics and space. This synergy can distinguish between spurious relationships that are spatially adjacent but functionally unrelated, such as adjacent snacks and tools on a shelf. Low similarity filtering of content attribute vectors avoids erroneous complementary relationship inferences. In conjunction with the graph structure association score calculation in step S31, the graph-level representation vector provides a comprehensive basis for score calculation, enabling the score to reflect the importance and relevance of the target object in the entire association network, thus improving the rationality of the recommendation ranking. Without this step, the system will be unable to utilize the complex relationships between target objects, and in cold start scenarios, it will only be able to rely on the features of a single object for recommendations, resulting in a lack of diversity and relevance in the recommendation results, failing to meet the potential needs of users.

[0096] Step S20, through multi-granularity feature extraction (global and local), attribute decoupling and separation, and aggregation of structured associations, forms a hierarchical, semantically clear, and logically defined representation system. This solves the problems of poor feature representation generalization, semantic ambiguity, and missing associations in cold-start scenarios when explicit annotations and interactive data are lacking. Multi-granularity representation takes into account both global semantics and local details, avoiding the limitations of single-granularity representation; attribute decoupling achieves precise semantic separation, resolving the core contradiction of "visually similar but functionally different"; graph-level representation integrates complex associations, providing support for association inference in cold-start scenarios. In deep collaboration with the spatial topology layout relationship graph, fine-grained visual feature library, and contextual environment parameter vectors constructed in step S10, representation learning can fully utilize alternative data in the scene to replace the traditionally relied-upon user interaction data. It also seamlessly connects with the comprehensive scoring matrix construction, constraint filtering, and ranking optimization in step S30, providing accurate feature basis for scoring calculation, interpretable attribute support for constraint filtering, and a structured relational foundation for ranking optimization. This drives the entire recommendation scheme to a cognitive leap from "visual feature reasoning" to "functional relational reasoning" and then to "executable solution reasoning." Without step S20, subsequent steps would lack high-quality feature support, the comprehensive scoring matrix would lack discriminative basis, constraint filtering would fail to accurately match functional requirements, and ranking optimization would struggle to balance relevance and diversity. Ultimately, this would cause the recommendation to remain at the superficial stage of "visual similarity," failing to address the semantic drift and execution constraint mismatch problems in cold-start scenarios.

[0097] Step S30: Based on the global semantic representation vector, local detail representation vector, three types of attribute vectors, and graph-level representation vector, construct a comprehensive scoring matrix for each target object node in the spatial topology layout relationship graph; apply the feasibility constraint index vector to perform three rounds of progressive screening on the comprehensive scoring matrix to generate a selected candidate set; perform multi-objective Pareto ranking and visual confusion disambiguation on the selected candidate set to output the final recommendation sequence.

[0098] Further, step S30 includes:

[0099] Step S31: Based on the global semantic representation vector, local detail representation vector, three types of attribute vectors, and graph-level representation vector, construct a comprehensive scoring matrix for each target object node in the spatial topology layout relationship graph;

[0100] Further, step S31 includes:

[0101] Step S311: Obtain the user's query requirements and encode the query requirements into a query vector;

[0102] Step S312: Calculate the semantic relevance score based on the query vector, global semantic representation vector, and local detail representation vector. Calculate the global relevance score by performing cosine similarity calculation between the query vector and the global semantic representation vector of each target object node in the spatial topology layout relationship diagram. Calculate the detail relevance score by performing similarity calculation between the query vector and the local detail representation vector of each target object node. Calculate the semantic relevance score by weighted fusion of the global relevance score and the detail relevance score.

[0103] Step S313: Calculate the attribute matching score based on the query vector and the three types of attribute vectors: Extract the attribute feature components of the query vector, calculate the matching degree with the content attribute vector, style attribute vector and state attribute vector of each target object node respectively, and obtain the content matching degree, style matching degree and state matching degree. Dynamically determine the content weight, style weight and state weight according to the context environment parameter vector, and generate the attribute matching score by weighted combination of the three types of matching degrees.

[0104] Step S314: For each target object node, calculate the centrality score based on the similarity graph, the connectivity score based on the complementarity graph, the causal importance score based on the causal graph, and the temporal correlation score based on the temporal graph. Then, merge the four scores to generate a graph structure association score.

[0105] Step S315: Organize the semantic relevance score, attribute matching score, and graph structure association score by column to construct a comprehensive scoring matrix M, where the rows of the matrix correspond to the target object nodes in the spatial topology layout relationship graph, and the columns correspond to each scoring dimension.

[0106] The core logic of step S31 is to comprehensively evaluate the fit between the target object and the user's needs from three dimensions: semantic matching, attribute adaptation, and structural association. The query requirement is a text description entered by the user through the system's interactive interface, such as "sleep aid drink," or an image example, such as a packaging image of a sleep aid drink. The two input modalities correspond to different demand expression scenarios. The text modal accurately conveys functional requirements, while the image modal intuitively presents appearance preferences. The encoding process of the query vector needs to take into account the advantages of both modalities: text input is converted into fixed-dimensional text embedding vectors through a pre-trained language model, such as BERT. This model learns semantic associations through a large amount of text corpus and can capture the semantic commonalities of synonyms such as "sleep aid" and "calming." Image input is converted into image embedding vectors through the global semantic encoder in step S21 to ensure that the feature extraction logic of the image modal is consistent with the global semantic representation of the target object, avoiding dimensional differences and distribution shifts in cross-modal features. The fusion of the two embedding vectors employs a weighted summation strategy, with the fusion weights dynamically adjusted based on the query type. When text queries dominate, the text embedding vector has a higher weight than the image embedding vector, and vice versa when image queries dominate. The resulting unified query vector comprehensively covers the user's functional needs and aesthetic preferences, avoiding misunderstandings caused by single-modal input. Step S311 addresses the shortcomings of traditional single-modal query encoding in simultaneously addressing functional and aesthetic requirements by using bimodal fusion to capture all aspects of the needs, generating a query vector that includes both semantic essence and visual preferences. This ensures that subsequent matching calculations accurately align with the user's complex needs.

[0107] The semantic relevance score aims to quantify the degree of matching between query requirements and target object representations. Its core logic is to balance global semantic association with local detail fit. The global relevance score uses a cosine similarity algorithm to calculate the correlation between the query vector and the global semantic representation vector. This algorithm reflects the consistency of semantic direction by measuring the cosine value of the angle between the two vectors, effectively capturing the fit between the overall function of the target object and the query requirements, and avoiding overall judgment bias caused by interference from local features. The detail relevance score also uses a cosine similarity algorithm, focusing on the matching degree between the query vector and the local detail representation vector, accurately capturing the correspondence between key details of the target object and the query requirements, such as the assembly interface dimensions of parts and functional labeling text on product packaging. The weighted fusion of the global relevance score and the detail relevance score is achieved through learnable weight parameters α and β. α and β are determined through model training and optimization. The training process aims at recommendation accuracy, enabling α and β to adapt to scene characteristics. For example, in industrial inspection scenarios, when global structural matching is more important, α takes a higher value; in retail scenarios, when detailed functional labeling matching is key, β takes a higher value. This dual-dimensional weighted fusion method breaks through the limitations of traditional single-dimensional similarity calculation. It enables semantic relevance scores to reflect overall functional fit while not ignoring key detail matching, ensuring effective differentiation of visually similar but functionally different objects. For example, when searching for "sleep aid products", it can exclude beverage categories unrelated to sleep aid function through global relevance scores, and identify key labels such as "sleep aid" and "soothing" on the packaging through detail relevance scores, avoiding misjudging energy drinks with similar appearances as matching objects.

[0108] The core of attribute matching score calculation is to achieve precise alignment between the required attributes and the target object attributes, which requires extracting the attribute feature components of the query vector. These attribute feature components are extracted through an attribute decomposition module, which is built upon the three-dimensional attribute vectors from step S23. This module separates content attribute components, style attribute components, and state attribute components from the query vector, corresponding to the user's functional requirements, appearance preferences, and state requirements for the target object, respectively. Content matching degree, style matching degree, and state matching degree are all calculated using a cosine similarity algorithm, quantifying the degree of fit between the corresponding attribute components of the query vector and the three types of attribute vectors of the target object. The dynamic calculation of content weight (w_content), style weight (w_style), and state weight (w_state) relies on contextual environment parameter vectors: by parsing the functional area labels of the context, such as the "assembly area" in an industrial scenario or the "food area" in a retail scenario, the scenario type is determined. For function-oriented scenarios, such as the assembly area, w_content is adjusted to the highest value. Through user behavior cues in environmental condition parameters, such as users paying more attention to appearance in the "fashion accessories area" of a retail scenario, aesthetic-oriented needs are determined, and w_style is adjusted to the highest value. Through time-series parameters and concurrent state parameters, such as the state monitoring needs during device operation, dynamic state needs are determined, and w_state is adjusted to the highest value. This method addresses the shortcomings of traditional attribute matching, which uses fixed weights and cannot adapt to dynamic changes in scenarios. By dynamically adjusting weights, it achieves scenario adaptation for attribute matching, enabling attribute matching scores to accurately respond to the priority of needs in different scenarios. In function-oriented scenarios, it focuses on function matching; in aesthetic-oriented scenarios, it focuses on style matching, improving the targeting and accuracy of attribute matching. The decoupled attribute vectors provide a foundation for accurately calculating the matching degree of each dimension, enabling attribute matching to get rid of interference between different attributes. Even if the target objects look similar (similar style attribute vectors), objects with inconsistent functions can be excluded by the low matching degree of the content attribute vectors, effectively solving the core problem of "visually similar but functionally different".

[0109] The graph structure association score is calculated based on the four types of relationship graphs constructed in step S24, comprehensively capturing the value and status of the target object in the association network. Centrality score is calculated using the PageRank algorithm, which quantifies the coreness of the target object in a functionally similar network based on the connection strength and propagation path of nodes in the similarity graph. Target objects corresponding to core nodes often possess more typical functional characteristics and are recommended with higher priority. Connectivity score is calculated using a path counting algorithm, counting the number of effective connection paths between the target object node and other nodes in the complementary relationship graph. A higher number of paths indicates broader functional complementarity, better meeting potential user needs. Causal importance score is calculated using a directed graph traversal algorithm, quantifying the indispensability of the target object in the functional execution chain based on the directional edge logic of the causal relationship graph. For example, in the assembly process, the causal importance score of the upstream core equipment is higher than that of the supporting accessories. Temporal relevance score is calculated using a dynamic time warping algorithm, comparing the changes in the state attribute vectors of the target object node at different time stages in the temporal relationship graph with the temporal adaptability of the query requirements. For example, the temporal relevance score of seasonal demand products changes dynamically with seasonal parameters. The four types of scores are weighted and fused to generate a graph-structured association score. The weights are set according to the core needs of the scenario. In the industrial scenario, the causal importance score has a higher weight, while in the retail scenario, the connectivity score has a higher weight. This multi-dimensional structural score fusion method allows the recommendation to not only focus on the matching of individual objects with needs, but also take into account the value of objects in the association network. This provides support for mining potential associations in cold start scenarios. For example, a newly launched sleep aid tea, although it has no historical interaction data, obtains a high centrality score through the connection with other sleep aid products in the similarity relationship graph, and obtains a high connectivity score through the connection with pillows and aromatherapy in the complementary relationship graph, thereby improving the recommendation priority.

[0110] The comprehensive scoring matrix M is constructed with target object nodes as rows and each scoring dimension as columns. Semantic relevance scores, attribute matching scores, and graph structure association scores are organized by column, forming a structured quantitative matrix. The matrix's dimensional design strictly matches the number of target object nodes and the number of scoring dimensions, ensuring that the recommendation value of each target object can be comprehensively represented through multi-dimensional scores. The matrix construction integrates scattered matching indicators into a unified quantitative system, providing a standardized data foundation for subsequent constraint screening and ranking optimization, and avoiding computational conflicts caused by inconsistent indicator dimensions.

[0111] The contextual parameter vector collaboration between steps S31 and S13 enables weight adjustments to adapt to dynamic scenarios, avoiding environmental adaptation biases caused by static scoring. For example, in weekday retail scenarios with high personnel density, adjusting the connectivity score weight in the graph structure association score through contextual parameters prioritizes recommending complementary products to popular items, thus improving recommendation conversion rates. Without this step, subsequent constraint filtering would lack clear quantitative basis, and ranking optimization would fail to accurately weigh the recommendation value across different dimensions. This would cause recommendation decisions to remain at the qualitative judgment level, failing to achieve precise priority ranking and effectively linking multi-source representation features with user needs. Consequently, the representation learning results of step S20 would be difficult to apply, and the recommendation system would still be unable to overcome the superficial matching dilemma of "visual similarity." The core value of step S31 lies in constructing a multi-dimensional, quantifiable recommendation value evaluation system, solving the technical problem of lacking quantitative standards for matching needs and features in cold start scenarios. Multimodal query encoding achieves a unified representation of different types of needs, overcoming the limitations of single-modal queries; dual-dimensional semantic relevance calculation takes into account both global and local factors, effectively distinguishing between "visual similarity" and "functional matching"; dynamic weighted attribute matching scores achieve precise alignment between requirement attributes and object attributes; graph structure association scores uncover potential relationships between objects, compensating for the lack of interaction data in cold start scenarios. The synergy of these methods enables the comprehensive scoring matrix to fully and accurately quantify the recommendation value of each target object, providing reliable data support for subsequent filtering and ranking, driving the upgrade of recommendation logic from "feature matching" to "value matching," while providing flexible quantitative adjustment space for adaptation to different scenarios and different need types, significantly improving the generalization ability and accuracy of the recommendation system.

[0112] Step S32: Apply the feasibility constraint index vector to perform three rounds of progressive screening on the comprehensive scoring matrix to generate a selected candidate set;

[0113] Further, see Figure 4 Step S32 includes:

[0114] Step S321, First round of hard constraint screening: Traverse all rows in the comprehensive scoring matrix M, extract the feasibility constraint index vector of the target object node corresponding to each row, and identify hard constraint violation nodes with physical reachability score, resource sufficiency score, operational complexity score or risk control score of zero; delete the rows corresponding to the identified hard constraint violation nodes from the comprehensive scoring matrix M, and generate the hard constraint screening matrix M1.

[0115] Step S322, Second round of soft constraint weighting: For each target object node retained in matrix M1 after hard constraint screening, calculate the geometric mean of the four components of its feasibility constraint index vector as the comprehensive feasibility score, and add the comprehensive feasibility score as a new column to matrix M1 after hard constraint screening to form soft constraint weighting matrix M2.

[0116] Step S323, third round of dynamic threshold screening: dynamically calculate the screening threshold based on the context environment parameter vector, and select target object nodes whose comprehensive feasibility score exceeds the screening threshold from the soft constraint weighted matrix M2 to form a selected candidate set C.

[0117] The core logic of step S32 is to progressively filter and condense a high-quality candidate set based on the feasibility constraint index vector, thereby ensuring the executability and quality of the recommendation results:

[0118] The first round of hard constraint screening: Hard constraints refer to the core conditions that ensure the recommendation results have basic feasibility. Once violated, the target object will be completely unfeasible for implementation. A hard constraint violation node is a target object node where any component of the feasibility constraint index vector is zero. A physical accessibility score of zero indicates that the target object is completely unavailable in space, such as being completely blocked by non-removable obstacles, being in a closed area without access rights; a resource sufficiency score of zero indicates that the core resources required for execution are completely lacking, such as the target object having zero inventory, key equipment being fully occupied and having no alternative; an operational complexity score of zero indicates that the operation is completely impossible to complete, such as the required special permissions being unavailable, or the operation steps exceeding the limits of the scenario's capabilities; a risk control score of zero indicates that there are serious security or compliance risks, such as the target object being a high-risk prohibited item, or the loss from operational failure exceeding an acceptable threshold. The hard constraint screening process involves traversing each row of the comprehensive scoring matrix M, extracting the feasibility constraint index vector for each target object node, and verifying whether the physical reachability score, resource sufficiency score, operational complexity score, and risk control score are zero. If any zero value exists, it is determined to be a node that violates the hard constraint, and the row corresponding to that node is deleted from matrix M, generating the hard constraint-screened matrix M1. This screening process directly eliminates completely infeasible target objects, avoiding subsequent waste of computational resources, while ensuring the basic feasibility of the recommendation results. This solves the problem of "feasible recommendations but ineffective implementation" caused by traditional recommendation systems neglecting execution constraints.

[0119] The second round of soft constraint weighting: The core of constructing the soft constraint weighting matrix M2 is to quantify the feasibility satisfaction of the target object. The comprehensive feasibility score is calculated using a geometric mean algorithm. This algorithm is obtained by calculating the fourth root of the product of the four components of the feasibility constraint index vector. Its advantage lies in its sensitivity to extreme values, highlighting the impact of weak constraints. If any constraint component has a low score, it will directly lower the overall comprehensive feasibility score, which conforms to the "barrel principle" of execution constraints—execution feasibility is determined by the weakest constraint. For example, if a target object has high scores in physical accessibility, resource sufficiency, and risk control, but a low score in operational complexity, the geometric mean calculation will accurately reflect the object's shortcomings at the operational level, avoiding the problem of shortcomings being masked by the arithmetic mean. The comprehensive feasibility score is added as a new column to matrix M1, so that the recommendation value of each target object simultaneously includes two core dimensions: matching degree and feasibility. This provides a more comprehensive quantitative basis for subsequent screening. This process achieves an initial integration of matching value and execution feasibility, making the recommendation decision not only focus on "whether it matches" but also on "whether it is feasible".

[0120] The third round of dynamic threshold screening: The dynamic screening threshold is determined based on the context environment parameter vector from step S13, and scenario adaptation is achieved through a threshold mapping function. The threshold mapping function is constructed using a multiple linear regression model: personnel density, equipment occupancy rate, and abnormal environmental conditions in the context environment parameter vector are used as independent variables, and the optimal screening threshold in historical scenarios (the threshold that best balances recommendation accuracy and efficiency) is used as the dependent variable. The model is trained using a gradient descent algorithm to obtain regression coefficients, and the function output is the dynamic screening threshold. For example, when the personnel density and equipment occupancy rate in the context environment parameter vector are high, corresponding to a high-load scenario, the screening threshold is increased to reduce the size of the candidate set, prioritize the recommendation of highly feasible target objects, and reduce the system's operating pressure; when the environmental parameters correspond to exploratory scenarios, such as the new product launch period or the new scenario activation period, the screening threshold is decreased to expand the size of the candidate set, increase recommendation diversity, and explore potential effective objects; when the environmental parameters show abnormal conditions, such as high temperature or equipment failure, the screening threshold is increased to focus on core feasible target objects and avoid recommendation risks. The dynamic screening process involves traversing the soft-constraint weighted matrix M2, retaining target object nodes whose comprehensive feasibility score exceeds the dynamic screening threshold, and forming a refined candidate set C. This process achieves scenario-adaptive screening criteria, avoids fluctuations in candidate set quality caused by fixed thresholds, and ensures that a reasonable number of reliable candidate objects can be screened under different scenario conditions.

[0121] The synergistic effect of the comprehensive scoring matrix in steps S32 and S31 ensures that the screening process retains high-matching and high-feasibility target objects while eliminating low-matching or low-feasibility objects. This significantly reduces the search space for subsequent ranking, improving ranking efficiency and quality. Without this step, the recommendation process would be unable to filter infeasible objects, resulting in recommendations containing a large number of unfeasible targets, severely impacting user trust. Furthermore, an excessively large unfiltered candidate set would increase the computational complexity of subsequent ranking, reducing system response speed. In addition, the lack of dynamic threshold adjustment would cause the screening criteria to become detached from the actual scenario; an excessively large candidate set in high-load scenarios would lead to system overload, while an excessively small candidate set in exploratory scenarios would result in monotonous recommendations, failing to meet diverse scenario requirements. The core technical value of step S32 lies in establishing a three-level screening mechanism of "hard constraint fallback - soft constraint weighting - dynamic threshold adaptation," solving the technical problems of recommendation feasibility and scenario adaptability in cold-start scenarios. Hard constraint screening ensures the basic execution baseline of the recommendation results, avoiding completely ineffective recommendations; soft constraint weighting quantifies comprehensive feasibility through a geometric average algorithm, highlighting the impact of weak constraints and making feasibility assessment more accurate; dynamic threshold screening enables scenario-adaptive screening criteria, balancing the quality and diversity of the candidate set. This three-tiered screening mechanism progressively ensures the execution feasibility of candidate objects while also considering the dynamic changes in the scenario, significantly narrowing the search range for subsequent ranking and providing a high-quality candidate foundation for ranking optimization. Hard constraint screening eliminates invalid objects in advance, reducing the waste of subsequent computational resources and improving the overall system efficiency; the synergy between dynamic thresholds and the context environment enables the recommendation system to proactively adapt to scenario changes, maintaining stable recommendation quality in complex and dynamic scenarios; the soft constraint weighted geometric average algorithm makes feasibility assessment more consistent with actual execution logic, avoiding candidate object quality deviations caused by misjudgments based on a single constraint dimension, laying a reliable foundation for subsequent multi-objective ranking.

[0122] Step S33: Define the selected candidate set as the search space of a multi-objective optimization problem. With the optimization objectives of maximizing relevance, maximizing diversity, and maximizing execution efficiency, solve the Pareto front and select the optimal solution from the Pareto front as the initial recommendation sequence based on the context environment parameter vector.

[0123] The multi-objective optimization problem is constructed based on three core objectives. The definition and calculation of each objective rely on the quantitative indicators constructed earlier to ensure consistency with the overall solution. The relevance maximization objective focuses on the fit between the recommendation results and user needs. Its calculation is based on the semantic relevance score and attribute matching score in the comprehensive scoring matrix of step S31. A weighted sum is used to obtain the comprehensive relevance value, and the weights are dynamically adjusted according to the demand type parameter in the contextual environment parameter vector. Under function-oriented needs, the attribute matching score weight is increased; under appearance-oriented needs, the semantic relevance score weight is increased. This objective directly responds to the core user demands, avoiding a disconnect between recommendation results and needs. The diversity maximization objective aims to avoid homogenization of recommendation sequences. It calculates the mean cosine similarity of the content attribute vector, style attribute vector, and state attribute vector of any two target object nodes in the candidate set, and then subtracts this mean from 1 to obtain the diversity index. The lower the mean, the greater the attribute differences, and the higher the diversity index. This objective breaks through the problem of single-source recommendations caused by the "relevance-first" approach in traditional recommendations, satisfying the potential diverse needs of users. The goal of maximizing execution efficiency is based on the feasibility constraint index vector in step S14. It is obtained by calculating the arithmetic mean of the physical reachability score, resource sufficiency score, operational complexity score, and risk control score. This goal ensures that each object in the recommendation sequence has the characteristics of efficient execution and avoids affecting the user experience due to low execution efficiency.

[0124] The optimization of the three objective functions employs the NSGA-II algorithm. This algorithm was chosen based on its ability to balance the distribution and convergence of solutions in multi-objective optimization. The specific implementation process is as follows: In the initialization phase, the population size is dynamically adjusted based on the size of the selected candidate set. When the candidate set is large, the population size increases proportionally; when the candidate set is small, the population size remains at its minimum to ensure optimization efficiency and effectiveness. The population consists of a sequence of nodes randomly selected from the selected candidate set. The length of each sequence is set according to the scenario recommendation requirements, such as recommending 10 objects in a retail scenario and 5 objects in an industrial scenario. In the fitness value calculation phase, the comprehensive relevance value, diversity index, and average execution efficiency corresponding to each sequence are used as the fitness values ​​for the three objectives, directly quantifying the sequence's performance on each objective. In the non-dominated sorting phase, Pareto front solutions are identified by comparing the fitness values ​​of sequences. If sequence A has a fitness value no lower than sequence B on all objectives, and a fitness value higher than sequence B on at least one objective, then sequence B is dominated by sequence A. The Pareto front solution is the set of optimal solutions not dominated by any other sequence. This process ensures the optimality and non-redundancy of the front solution. In the crowding calculation phase, the distribution density of solutions is measured by calculating the distance between each front solution and its neighboring solutions in the objective space. Higher crowding indicates a denser distribution of solutions, and vice versa. This step ensures a uniform distribution of Pareto front solutions, providing a rich set of trade-offs for subsequent selection. In the genetic operation phase, the crossover operation employs a two-point crossover strategy, randomly selecting two crossover points from two parent sequences and exchanging node fragments between these points to generate the offspring sequence. The mutation operation uses a random replacement strategy, randomly selecting some nodes from the offspring sequence and replacing them with other objects not included in the selected candidate set. The crossover and mutation probabilities are determined through model training optimization. The training process aims to ensure the uniformity of the distribution of the frontier solutions, avoiding sequence chaos caused by excessively high crossover probabilities or slow convergence caused by excessively low crossover probabilities. In the iterative optimization phase, after each iteration, the parent and offspring sequences are merged, and non-dominated sorting and crowding calculation are re-executed. The optimal frontier solution is retained for the next iteration until the number of iterations reaches a preset threshold or the change in the frontier solutions between two consecutive iterations is less than a preset convergence threshold. The convergence threshold is determined based on the scenario's requirements for recommendation stability, and is set by statistically analyzing the distribution of changes during convergence in historical optimization processes to ensure the stability and reliability of the optimization results.

[0125] The selection of Pareto optimal solutions requires dynamic adaptation to the contextual environment parameter vector. For example, in high-load scenarios, solutions with superior execution efficiency are prioritized; in exploratory scenarios, solutions with superior diversity are prioritized; and in everyday scenarios, solutions that balance the three objectives are selected. This selection logic enables the recommendation sequence to accurately adapt to the dynamic needs of the scenario, avoiding scenario mismatch caused by fixed selections. This approach addresses the biased recommendation shortcomings caused by traditional single-objective optimization by achieving a balance of multi-dimensional values ​​through multi-objective modeling. It generates multiple optimal recommendation sequences with different trade-off characteristics. The recommendation results can accurately match the user's core needs, provide diverse choices, and ensure high execution efficiency. For example, in a smart retail scenario, the recommendation sequence includes products highly relevant to the user's query "sleep aid drinks," as well as sleep aid drinks from different brands and packaging (diversity), all of which are products with sufficient inventory and easy access (execution efficiency). In conjunction with the selected candidate set in step S32, which filters out infeasible objects, multi-objective optimization no longer needs to consider infeasible objects, significantly improving optimization efficiency. This collaboration allows the optimization process to focus on the trade-offs of high-quality candidates, avoiding interference from invalid objects and simultaneously improving the quality and computational efficiency of the Pareto front solution. In conjunction with the contextual parameter vector in step S13, the dynamic selection strategy enables the recommended sequence to adapt to changes in the scenario. For example, during weekday lunch breaks when personnel density is high, the system automatically selects solutions with superior execution efficiency, prioritizing easily accessible and convenient products to improve user experience. Without this step, the recommended sequence would fail to balance multi-dimensional needs, potentially resulting in high relevance but insufficient diversity, or high diversity but low execution efficiency. This would prevent the selected candidate set in step S32 from fully realizing its value, and subsequent disambiguation processing would lack a high-quality sequence foundation.

[0126] Step S34: Perform visual confusion detection and disambiguation on the preliminary recommended sequence, and output the final recommended sequence.

[0127] Further, step S34 includes:

[0128] Step S341: Define each pair of target object nodes in the preliminary recommendation sequence S0 as a node pair, calculate the similarity of the global semantic representation vector and the similarity of the content attribute vector for each node pair. If the similarity of the global semantic representation vector is higher than the preset semantic threshold and the similarity of the content attribute vector is lower than the preset content threshold, calculate the difference between the similarity of the global semantic representation vector and the similarity of the content attribute vector as the confusion index.

[0129] Step S342: Node pairs with a confusion index greater than a preset confusion threshold are identified as potential confusion pairs. Disambiguation processing is performed on each potential confusion pair to obtain the final recommended sequence.

[0130] The core of visual confusion detection is to accurately identify target object node pairs that are "similar in global semantics but have different content attributes". First, the definition of confusion is "the similarity of the global semantic representation vector is higher than the preset semantic threshold, and the similarity of the content attribute vector is lower than the preset content threshold". This definition directly addresses the essence of visual confusion and avoids false detection caused by single-dimensional judgment. Global semantic similarity is calculated by the cosine similarity of the global semantic representation vectors of two target object nodes, reflecting the degree of similarity at the appearance level. Content attribute similarity is calculated by the cosine similarity of the content attribute vectors of two target object nodes, reflecting the degree of functional compatibility. The confusion index is calculated by the difference between "global semantic similarity and content attribute similarity". This calculation is only performed when the preconditions are met, namely, the global semantic similarity is higher than the semantic threshold and the content attribute similarity is lower than the content threshold, to ensure that a high confusion score is obtained only when "global semantic similarity is high and content attribute similarity is low". For example, when the global semantic similarity is 0.9 (higher than the threshold of 0.7) and the content attribute similarity is 0.2 (lower than the threshold of 0.4), the confusion score is 0.7. If the preconditions are not met, such as the global semantic similarity is 0.5 (lower than the threshold) or the content attribute similarity is 0.6 (higher than the threshold), the confusion score is not calculated and the pair is directly determined to be non-confused. The semantic threshold and content threshold are determined as follows: the semantic threshold is set by the global semantic similarity distribution of visually similar objects in the statistical scene to ensure that it can cover the vast majority of object pairs with similar appearances; the content threshold is set by the content attribute similarity distribution of objects with different functions in the statistical scene to ensure that it can distinguish object pairs with very different functions. For example, the semantic threshold is set to 0.8 times the average similarity of visually similar objects in the scene, and the content threshold is set to 1.2 times the average similarity of objects with different functions. The detection process traverses all node pairs in the preliminary recommendation sequence and calculates the global semantic similarity, content attribute similarity, and confusion index. When the confusion index is greater than the preset confusion threshold, the node pair is included in the potential confusion pair set. The confusion threshold is set based on the statistical tolerance of users to visual confusion in the scene. For example, in the retail scene, users have a low tolerance for confusion, so the threshold is set to a small value, such as [0.2, 0.4]. In the industrial scene, the requirement for functional accuracy is extremely high, so the threshold is set to an even smaller value, such as [0.1, 0.3].

[0131] The disambiguation strategy employs a three-tiered progressive approach to address different levels of confusion, forming a complete disambiguation chain of "detail differentiation → feature description → position adjustment." The first tier is a detail difference enhancement strategy, which strengthens the detail differentiation of confused pairs by calculating the local detail difference degree. The local detail difference degree is obtained by subtracting the cosine similarity of the local detail representation vectors of the two target object nodes from 1. The local detail representation vector contains fine-grained features such as texture, logo, and local structure of the target object. These features are key to distinguishing objects that are similar in appearance but different in function. For example, there are differences in the packaging logos and texture details between high-caffeine energy drinks and sleep aids. The local detail difference degree can accurately capture these differences, thereby improving the distinguishability of confused pairs.

[0132] The second level is a discriminative feature extraction strategy. For confusing pairs with insufficient detail differences, semantic descriptions compensate for the limitations of visual differentiation. This strategy relies on the fine-grained visual feature library in step S12 to extract the identifying information features (such as product names, functional labels, and part models) and attribute label features (such as material, color, and size) of potential confusing pairs, generating discriminative descriptions in natural language, such as "the former contains high caffeine, the latter contains melatonin" or "the former is made of metal, the latter is made of plastic." This description directly presents the essential functional differences, enabling users to quickly identify the functional differences of target objects through textual descriptions even when visual differentiation is difficult, thus assisting users in rapid identification and avoiding misselection.

[0133] The third level is the sequence position optimization strategy, which serves as a fallback when the first two levels fail to adequately disambiguate. It reduces the risk of confusion through spatial isolation. This strategy first calculates the positional distance between confused pairs in the initial recommendation sequence. When adjacent confused pairs are detected, nodes with low relevance between the content attribute vector and the query vector are moved at least two positions backward, ensuring sufficient spacing between confused pairs in the sequence. This processing is based on users' linear browsing habits in recommendation sequences; adjacent objects are easily compared and selected directly. Spacing breaks this direct comparison, reducing the probability of misselection due to visual similarity. The three-level strategy works synergistically: enhanced detail differences provide a quantitative basis for judgment; discriminative feature extraction strengthens differentiation through semantic supplementation; and sequence position optimization avoids remaining risks through layout adjustments. This progressive processing ensures that visual confusion of varying degrees can be effectively disambiguated, ultimately generating a recommendation sequence that maintains both high relevance and high discriminability.

[0134] Step S34 addresses the lack of proactive obfuscation handling mechanisms in traditional recommendation systems by employing a layered strategy of "detection-detail enhancement-feature description-position optimization" to comprehensively manage obfuscation risk. This accurately identifies and weakens the interference of visually confusing pairs, significantly improving the distinguishability of the recommended sequence and reducing the risk of users making incorrect choices due to visual obfuscation. Step S22's local detail representation vector provides precise feature support for calculating detail differences, enabling the capture of subtle functional differences. Even objects with highly similar appearances can be effectively distinguished through local detail differences, avoiding disambiguation failure caused by extremely high appearance similarity. Discriminative feature extraction directly reuses the identifier information features and attribute annotation features from the library, ensuring the accuracy and relevance of discriminative descriptions and avoiding system redundancy caused by additional feature extraction. In cold start scenarios, where the lack of user feedback data prevents iterative optimization to correct obfuscation issues, this proactive disambiguation mechanism ensures the reliability of recommendation results even with zero interaction data, overcoming the technical bottleneck of uncontrollable obfuscation risk in cold start scenarios. Without this step, visually confusing objects in the initial recommendation sequence may lead to user misselection, especially in cold start scenarios where there is no user feedback for correction. This will severely impact user trust and also compromise the multi-objective optimization results of step S33 due to the risk of confusion, preventing the final transition from "high-quality candidates" to "reliable recommendations." Step S34 ultimately outputs a recommendation sequence that satisfies core user needs, balances diversity and execution efficiency, and eliminates the risk of visual confusion. This comprehensively addresses the multi-dimensional coupling problem of visual feature semantic drift and execution constraint mismatch in cold start scenarios.

[0135] Example 2

[0136] This embodiment, based on embodiment 1, provides an image cold start recommendation and ranking system, such as... Figure 5 As shown, it includes:

[0137] Modeling constraint module: used to acquire 3D point cloud data and RGB images of the target recommendation scene, construct a spatial topology layout relationship graph based on the 3D point cloud data; construct a fine-grained visual feature library based on the RGB images; construct a multi-level feasibility constraint index system based on the spatial topology layout relationship graph, and generate a feasibility constraint index vector.

[0138] Representation learning decoupling module: Based on a fine-grained visual feature library, it generates global semantic representation vectors and local detail representation vectors. Through deentanglement representation learning, it decomposes the global semantic representation vectors and local detail representation vectors into three independent attribute vectors. The three attribute vectors are combined with a spatial topology layout relationship graph to generate a graph-level representation vector. The three attribute vectors include content attribute vectors, style attribute vectors, and state attribute vectors.

[0139] The filtering and sorting module constructs a comprehensive scoring matrix for each target object node in the spatial topology layout relationship graph based on global semantic representation vectors, local detail representation vectors, three types of attribute vectors, and graph-level representation vectors. It applies feasibility constraint index vectors to perform three rounds of progressive filtering on the comprehensive scoring matrix to generate a selected candidate set. It then performs multi-objective Pareto ranking and visual confusion disambiguation on the selected candidate set to output the final recommended sequence.

[0140] Furthermore, in the screening and ranking module, the method of using feasibility constraint index vectors to perform three rounds of progressive screening on the comprehensive scoring matrix includes:

[0141] Step S321, First round of hard constraint screening: Traverse all rows in the comprehensive scoring matrix M, extract the feasibility constraint index vector of the target object node corresponding to each row, and identify hard constraint violation nodes with physical reachability score, resource sufficiency score, operational complexity score or risk control score of zero; delete the rows corresponding to the identified hard constraint violation nodes from the comprehensive scoring matrix M, and generate the hard constraint screening matrix M1.

[0142] Step S322, Second round of soft constraint weighting: For each target object node retained in matrix M1 after hard constraint screening, calculate the geometric mean of the four components of its feasibility constraint index vector as the comprehensive feasibility score, and add the comprehensive feasibility score as a new column to matrix M1 after hard constraint screening to form soft constraint weighting matrix M2.

[0143] Step S323, third round of dynamic threshold screening: dynamically calculate the screening threshold based on the context environment parameter vector, and select target object nodes whose comprehensive feasibility score exceeds the screening threshold from the soft constraint weighted matrix M2 to form a selected candidate set C.

[0144] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0145] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0146] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An image cold start recommendation ranking method, characterized in that, The method comprises: acquiring three-dimensional point cloud data of a target recommendation scene and an RGB image of the target recommendation scene, constructing a spatial topology layout relationship graph based on the three-dimensional point cloud data; based on the RGB image, constructing a fine-grained visual feature library; based on the spatial topology layout relationship graph, constructing a multi-level feasibility constraint index system to generate a feasibility constraint index vector; based on the fine-grained visual feature library, generating a global semantic representation vector and a local detail representation vector, decomposing the global semantic representation vector and the local detail representation vector into three types of attribute vectors that are independent of each other through entanglement disentangling representation learning, and combining the three types of attribute vectors with the spatial topology layout relationship graph to generate a graph-level representation vector; the three types of attribute vectors include a content attribute vector, a style attribute vector and a state attribute vector; based on the global semantic representation vector, the local detail representation vector, the three types of attribute vectors and the graph-level representation vector, constructing a comprehensive score matrix for each target object node in the spatial topology layout relationship graph; applying the feasibility constraint index vector to perform three rounds of progressive screening on the comprehensive score matrix to generate a selected candidate set; performing multi-object Pareto sorting and visual confusion disambiguation on the selected candidate set to output a final recommendation sequence.

2. The image cold start recommendation ranking method according to claim 1, characterized in that, The construction method of the spatial topology layout relationship graph comprises: segmenting the three-dimensional point cloud data using a RANSAC algorithm, identifying target objects in the target recommendation scene, segmenting point cloud clusters of the target objects, recording the 3D bounding boxes of the target objects, and marking the target objects as target object nodes; calculating the three-dimensional coordinates of the centroids of the point cloud clusters, calculating the Euclidean distances between the centroids of the point cloud clusters according to the three-dimensional coordinates of the centroids of the point cloud clusters, and establishing a spatial correlation edge in the spatial topology layout relationship graph to connect the corresponding target object nodes when the Euclidean distance between two centroids of the point cloud clusters is less than a preset spatial adjacency threshold; and according to the relative positional relationship between the target object nodes, giving the spatial correlation edge a directional attribute to form a directed spatial topology layout relationship graph.

3. The image cold start recommendation ranking method according to claim 2, characterized in that, The construction method of the fine-grained visual feature library comprises: projecting the 3D bounding boxes of the target objects onto the RGB image of the target recommendation scene, cropping the RGB image regions of the target objects to define target object RGB images, each target object RGB image corresponding to a target object node in the spatial topology layout relationship graph; for each target object node in the spatial topology layout relationship graph, designing a multi-scale feature extraction pipeline to extract hierarchical features from the target object RGB image, from global appearance to local details, and storing the target object RGB image and the hierarchical features into the fine-grained visual feature library.

4. The image cold start recommendation ranking method according to claim 3, characterized in that, The hierarchical features are five-level features, specifically including first-level morphological contour features, second-level texture pattern features, third-level identification information features, fourth-level attribute annotation features and fifth-level state indicator features.

5. The image cold start recommendation ranking method according to claim 4, characterized in that, The first layer of the multi-level feasibility constraint index system is a physical accessibility constraint layer, the second layer is a resource sufficiency constraint layer, the third layer is an operation complexity constraint layer, and the fourth layer is a risk control constraint layer.

6. The image cold start recommendation ranking method according to claim 5, characterized in that, The method for generating the feasibility constraint indicator vector comprises: calculating a physical accessibility score in a physical accessibility constraint layer, calculating a resource sufficiency score in a resource sufficiency constraint layer, calculating an operation complexity score in an operation complexity constraint layer, calculating a risk control score in a risk control constraint layer, and combining the physical accessibility score, the resource sufficiency score, the operation complexity score and the risk control score to form the feasibility constraint indicator vector.

7. The image cold start recommendation ranking method according to claim 6, characterized in that, The method for generating the global semantic representation vector and the local detail representation vector comprises: inputting a target object RGB image in a fine-grained visual feature library into a global semantic encoder with a VisionTransformer as a backbone network to generate a global semantic representation vector of each target object node; inputting a hierarchical feature in the fine-grained visual feature library into a local detail encoder with a deformable convolution network as a core to generate a local detail representation vector.

8. The image cold start recommendation ranking method according to claim 7, characterized in that, The method for generating the graph-level representation vector comprises: based on the spatial topology layout relationship graph and the three types of attribute vectors, constructing a similarity relationship graph, a complementary relationship graph, a causal relationship graph and a time sequence relationship graph, and generating a graph-level representation vector through a graph neural network aggregation.

9. The image cold start recommendation ranking method according to claim 8, characterized in that, The method for constructing the comprehensive score matrix comprises: obtaining a query requirement of a user and encoding the query requirement into a query vector; based on the query vector, the global semantic representation vector and the local detail representation vector, calculating a semantic correlation score; based on the query vector and the three types of attribute vectors, calculating an attribute matching score; for each target object node, based on the similarity relationship graph, calculating a centrality score, based on the complementary relationship graph, calculating a connectivity score, based on the causal relationship graph, calculating a causal importance score, and based on the time sequence relationship graph, calculating a time sequence correlation score, and fusing the four types of scores to generate a graph structure association score; organizing the semantic correlation score, the attribute matching score and the graph structure association score by column to construct a comprehensive score matrix, wherein the rows of the matrix correspond to the target object nodes in the spatial topology layout relationship graph, and the columns correspond to each score dimension.

10. The image cold start recommendation ranking method according to claim 9, characterized in that, The first round of the three-round progressive screening is hard constraint screening, the second round is soft constraint weighting, and the third round is dynamic threshold screening. The method for the hard constraint screening comprises: traversing all rows in the comprehensive score matrix M, extracting a feasibility constraint indicator vector of a target object node corresponding to each row, identifying a hard constraint violation node with a physical accessibility score, a resource sufficiency score, an operation complexity score or a risk control score of zero, deleting a row corresponding to the identified hard constraint violation node from the comprehensive score matrix, and generating a hard constraint screened matrix.

11. The image cold start recommendation ranking method according to claim 10, characterized in that, The method for the soft constraint weighting comprises: for each target object node retained in the hard constraint screened matrix, calculating a geometric mean value of four components of a feasibility constraint indicator vector thereof as a feasibility comprehensive score, and adding the feasibility comprehensive score as a new column to the hard constraint screened matrix to form a soft constraint weighted matrix; The method of the dynamic threshold screening is: collecting time sequence parameters, spatial position parameters, environmental condition parameters and concurrent state parameters of a target recommendation scene, constructing a context environment parameter vector; dynamically calculating a screening threshold based on the context environment parameter vector, screening target object nodes with a feasibility comprehensive score exceeding the screening threshold from a soft constraint weighted matrix M2, and composing a selected candidate set C.

12. An image cold start recommendation ranking system for implementing the image cold start recommendation ranking method of any one of claims 1-11, characterized in that, The system comprises: A modeling constraint module is configured to acquire three-dimensional point cloud data of a target recommendation scene and an RGB image of the target recommendation scene, construct a spatial topology layout relationship graph based on the three-dimensional point cloud data, construct a fine-grained visual feature library based on the RGB image, construct a multi-level feasibility constraint index system based on the spatial topology layout relationship graph, and generate a feasibility constraint index vector; A representation learning decoupling module is configured to generate a global semantic representation vector and a local detail representation vector based on the fine-grained visual feature library, decompose the global semantic representation vector and the local detail representation vector into three types of independent attribute vectors through entanglement-free representation learning, combine the three types of attribute vectors with the spatial topology layout relationship graph to generate a graph-level representation vector, and the three types of attribute vectors include a content attribute vector, a style attribute vector and a state attribute vector; A screening and sorting module is configured to construct a comprehensive score matrix for each target object node in the spatial topology layout relationship graph based on the global semantic representation vector, the local detail representation vector, the three types of attribute vectors and the graph-level representation vector, perform three rounds of progressive screening on the comprehensive score matrix by using the feasibility constraint index vector to generate a selected candidate set, perform multi-object Pareto sorting and visual confusion disambiguation on the selected candidate set, and output a final recommendation sequence.

Images