A precise audience matching system for content delivery on short video platforms

By generating high-dimensional scene state tensors and dynamically constrained inverted index trees, the problem of delayed response to user needs on short video platforms in complex environments is solved, achieving accurate matching of content delivery and efficient utilization of system resources.

CN123115461APending Publication Date: 2026-11-17GUIZHOU NORMAL UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611516343.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing short video platform content delivery systems are unable to respond in real time to shifts in user needs when faced with complex and ever-changing spatiotemporal environments and device status changes, resulting in biased search results, wasted resources, and low recommendation accuracy.

Method used

A high-dimensional scene state tensor is generated by a multi-dimensional perception collaboration module, and nonlinear spatial mapping is performed by a dynamic retrieval decoupling module. The traversal depth of the inverted index tree and the attribute filtering logic are dynamically constrained. The system's adaptive optimization is achieved by using a closed-loop evolution collaboration module to perform correlation regression updates based on human-computer interaction feedback.

Benefits of technology

It improved the accuracy of content delivery and the efficiency of system resource coordination, alleviated retrieval delays and resource waste, and ensured the real-time response and accuracy of recommendation strategies.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115461A_ABST
    Figure CN123115461A_ABST
Patent Text Reader

Abstract

This invention relates to a precise audience matching system for content delivery on short video platforms, encompassing the field of semantic retrieval. It couples the user's spatiotemporal viewing window vector with the risk state vector of the device's security topology to generate a high-dimensional scene state tensor. Based on this tensor, a nonlinear mapping is performed to generate a dynamic path constraint vector including depth offset parameters, attribute filtering thresholds, and bias weights. The traversal depth of the inverted index tree is controlled according to the depth offset parameters, and metadata filtering is performed within a specific content attribute subspace to output a set of target candidate videos. Human-computer interaction feedback data is collected and correlation regression is performed to recursively update the high-dimensional scene state tensor and nonlinearly correct the bias weights. This invention achieves dynamic adaptive optimization of the retrieval path through multi-dimensional perception and closed-loop evolution mechanisms, effectively improving the accuracy and resource utilization efficiency of short video content matching in complex scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of semantic retrieval, and more specifically to a precise audience matching system for content delivery on short video platforms. Background Technology

[0002] With the rapid development of the mobile internet and short video industry, the real-time and accurate delivery of massive amounts of audiovisual content has become crucial for improving user experience and platform operational efficiency. This typically relies on static modeling of user interests and retrieval techniques based on large-scale inverted indexes. The introduction of semantic retrieval technology can, to some extent, uncover the deeper connections between content and user preferences.

[0003] Most existing retrieval solutions assume that user demand is constant and predictable in a short period of time. Database retrieval engines typically traverse the index tree along a fixed path to match video metadata. This static retrieval mechanism struggles to capture real-time shifts in user demand under complex and ever-changing spatiotemporal environments or abnormal device access conditions. Once the user's external environment is disturbed, the query path of the static index will fail to respond promptly to these potential state changes, leading to a significant deviation between the candidate set derived from semantic retrieval and the user's true points of interest.

[0004] As the complexity of recommendation scenarios increases, this lag becomes increasingly prominent. When the system cannot perceive and process user environmental noise and scene fluctuations in real time from multiple dimensions, the retrieval engine is forced to perform ineffective searches with high latency and low hit rate within a broad matching range. In the long run, this redundancy in index paths caused by the lack of dynamic association not only results in a significant waste of retrieval resources but also leads to a feedback trap in the recommendation model because interactive feedback is difficult to effectively correct system weights. Even with the introduction of some semantic retrieval optimizations, the accuracy of retrieval results still cannot achieve a true closed-loop improvement due to the lack of adaptive adjustment strategies for scene evolution. Therefore, how to overcome the lag effect of existing recommendation systems when facing complex scene fluctuations and achieve a leap from static passive retrieval to dynamic environmental perception and real-time optimized matching has become a core challenge that urgently needs to be overcome to improve the efficiency of content delivery. Summary of the Invention

[0005] The purpose of this application is to provide a precise audience matching system for content delivery on short video platforms. This system has the advantages of enhancing the dynamic response capability of retrieval logic to fluctuations in users' real-time needs in complex spatiotemporal scenarios, improving the positioning accuracy and computational efficiency of the retrieval space under multi-dimensional attribute mapping, thereby improving the accuracy of content delivery and the closed-loop effect of system resource collaboration.

[0006] The objective of this application can be achieved through the following technical solution: a precise audience matching system for content delivery on short video platforms, comprising the following modules:

[0007] The multi-dimensional perception collaboration module is used to obtain a viewing window vector containing the user's spatiotemporal state distribution characteristics and a risk state vector containing the device's security topology association, and to perform multi-dimensional coupling processing on the viewing window vector and the risk state vector to generate a high-dimensional scene state tensor representing the scene state.

[0008] The dynamic retrieval decoupling module is used to store a preset index path bias mapping matrix. It projects the high-dimensional scene state tensor onto the index path bias mapping matrix through a preset matrix mapping function to perform nonlinear spatial mapping operations, generating a dynamic path constraint vector including depth offset parameters, attribute filtering thresholds and bias weights.

[0009] The dynamic index retrieval module is used to obtain video content metadata stored in the database storage engine within the short video platform, and based on the depth offset parameter extracted by the dynamic path constraint vector, control the traversal depth of the database storage engine on the inverted index tree, and restrict the attribute filtering logic of the video content metadata to a specific content attribute subspace defined by the dynamic path constraint vector, and output a target candidate video set.

[0010] The closed-loop evolutionary coordination module is used to collect human-computer interaction feedback data corresponding to the target candidate video set, perform correlation regression operation on the human-computer interaction feedback data, the high-dimensional scene state tensor and the dynamic path constraint vector to obtain the spatial evolution gradient; recursively update the high-dimensional scene state tensor through the spatial evolution gradient, and nonlinearly correct the bias weights in the index path bias mapping matrix.

[0011] Compared with the prior art, the beneficial effects of this application are:

[0012] 1. By coupling the multi-dimensional perception collaboration module and the dynamic retrieval decoupling module, a high-dimensional tensor representing the scene state is constructed using Fourier transform and graph attention network, and a dynamic path constraint vector is generated accordingly. This achieves precise pruning of the database inverted index tree traversal depth and attribute filtering logic, reducing the redundancy of the retrieval path. In complex and ever-changing spatiotemporal scenarios, it alleviates the retrieval lag and computational resource waste caused by the inability of traditional static retrieval schemes to respond to environmental fluctuations, and improves the response speed and resource allocation efficiency of short video content retrieval.

[0013] 2. By using the closed-loop evolutionary collaboration module to correlate regression and recursive weighted correction mechanism for human-computer interaction feedback, the abnormal system energy efficiency ratio is linked to the feature evolution gradient. This enables nonlinear adaptive closed-loop updates of multidimensional perception parameters and matrix mapping weights, ensuring that the content matching strategy can evolve in real time with user interaction behavior, thereby improving the targeting and recommendation accuracy of content delivery. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the module of the content delivery audience matching system for short video platforms in this application. Detailed Implementation

[0015] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0016] Example 1,

[0017] Application Overview:

[0018] In the current field of short video content recommendation, traditional recommendation systems often suffer from irrational deviations and lags in the retrieval path within the content attribute space when faced with complex spatiotemporal environments that change in real time, interference from secure network topologies of heterogeneous devices, or dynamic fluctuations in terminal computing resources. This is due to a lack of deep perception and logical modeling of the user's immediate context. Existing technologies typically rely solely on static user profile features or basic semantic retrieval models, failing to effectively couple the user's multidimensional spatiotemporal state with the real-time risk state of the device. Furthermore, they lack dynamic constraints and closed-loop evolution compensation mechanisms for the database retrieval path. Consequently, the system cannot correct interest prediction biases caused by environmental noise in real time during the retrieval process, leading to fluctuations in matching results due to scene variations. This reduces the accuracy of content delivery and the reliability of system decisions.

[0019] For example, when delivering short video content to a user in a mobile commuting scenario where the terminal network experiences security access fluctuations, the system obtains the user profile and video metadata using conventional algorithms. Because traditional methods rely solely on semantic similarity for inverted index matching, when analyzing a sequence where the user's immediate interests shift due to a sudden change in environment and temporary network security topology, the system mistakenly matches this immediate retrieval demand caused by a sudden environmental change with the user's long-accumulated static profile. Since the traversal depth of the retrieval path fails to undergo adaptive nonlinear correction based on the actual scene state tensor, at specific moments of high system concurrency and fluctuating network cache hit rates, the candidate videos sampled using preset rules actually deviate from the user's content attribute requirements in that immediate scenario. This leads to periodic abnormal fluctuations in click-through and completion rates of the recommended content, severely masking the user's true preferences and their evolutionary trends in specific environments, thus causing deviations in the recommendation system's operational strategy.

[0020] Without addressing the aforementioned issues, precise content matching based on deep semantic retrieval will be difficult to reliably apply in the high-volume, high-frequency interaction of short video projects. Due to the lag in matching strategies, the system may frequently push out outdated content that is severely inconsistent with the user's current context, leading to a significant decrease in user interaction with the platform, resulting in ineffective waste of traffic resources or insufficient exposure to potentially high-quality content, thereby causing user churn or a decline in the platform's ecosystem quality. In the long run, the uncertainty and low adaptability of delivery results will hinder the large-scale release of the effectiveness of precise content distribution technology, forcing platform operations to remain highly dependent on extensive traffic scheduling based on human experience, failing to fully leverage the core technological advantages of intelligent systems in dynamic environment perception, real-time path optimization, and closed-loop evolution.

[0021] like Figure 1 As shown, the content delivery audience matching system for short video platforms includes the following modules:

[0022] The multi-dimensional perception and collaboration module acquires a viewing window vector containing the user's spatiotemporal state distribution characteristics and a risk state vector containing device security topology associations. It then performs multi-dimensional coupling processing on the viewing window vector and the risk state vector to generate a high-dimensional scene state tensor representing the scene state. The viewing window vector is a numerical representation reflecting the spatiotemporal environment changes of the user over a viewing period. Specifically, it can be implemented by frequency domain transformation of the geographic coordinates and timestamp sequences corresponding to the user's click operations. Its function is to transform the user's location migration and dwell patterns over time into structured data that can participate in computation. The risk state vector is a numerical representation reflecting the network security environment of the user's currently accessed device. Specifically, it can be implemented by topology relationship encoding of the gateway logs associated with the device identifier. Its function is to transform potential abnormal connections, unconventional gateway jumps, and other security states during device access into structured data that can participate in computation. The high-dimensional scene state tensor is a comprehensive scene description data obtained by fusing the above two types of vectors in terms of feature dimensions. Specifically, it can be implemented by concatenating and mapping the two types of vectors using a deep neural network. Its function is to provide a unified scene state input for subsequent retrieval path adjustments.

[0023] The dynamic retrieval decoupling module stores a preset index path bias mapping matrix. It projects the high-dimensional scene state tensor onto this matrix using a preset matrix mapping function, performing nonlinear spatial mapping operations to generate a dynamic path constraint vector including depth offset parameters, attribute filtering thresholds, and bias weights. The index path bias mapping matrix is ​​a pre-configured parameter matrix that can be continuously adjusted during operation. Its rows and columns correspond one-to-one with the node numbers of each branch in the inverted index tree. Each element in the matrix represents the current bias weight value of the corresponding branch, carrying the mapping relationship between the scene state and the adjustment range of the retrieval path. Specifically, it can be implemented using a weight matrix obtained through random initialization and pre-training with historical retrieval logs. Its function is to provide a learnable intermediate carrier for the conversion from scene state to retrieval path parameters. The matrix mapping function refers to the nonlinear operation rule that converts the high-dimensional scene state tensor into a dynamic path constraint vector. Specifically, it can be implemented using a combination of linear transformations with activation functions. Its function is to ensure that the adjustment range of the retrieval path parameters exhibits a nonlinear response to changes in the scene state. In the dynamic path constraint vector, the depth offset parameter refers to the value used to limit the upper limit of the index tree traversal depth; the attribute filtering threshold refers to the criterion used to determine whether the attributes of candidate nodes meet the filtering conditions; and the bias weight refers to the set of weighted coefficients extracted from the index path bias mapping matrix by the matrix mapping function in this round of retrieval, used to adjust the probability of different retrieval branches being selected. That is, the value of the corresponding branch element in the matrix at the current moment. The three together constitute the quantitative constraint on the retrieval path. Therefore, the bias weight has a consistent meaning in this application: its resident value is stored in the corresponding element of the index path bias mapping matrix; at the beginning of each retrieval cycle, it is extracted by the matrix mapping function and participates in the dynamic path constraint vector as part of this round of retrieval; after the retrieval is completed, the calculation result of the closed-loop evolutionary coordination module is written back to the corresponding element in the matrix, forming a complete chain of storage, retrieval, use, and update.

[0024] A dynamic index retrieval module is used to obtain video content metadata stored in the database storage engine of the short video platform, and control the traversal depth of the database storage engine on the inverted index tree based on the depth offset parameter extracted by the dynamic path constraint vector. It also restricts the attribute filtering logic of the video content metadata to a specific content attribute subspace defined by the dynamic path constraint vector, outputting a set of target candidate videos. The inverted index tree in the database storage engine refers to a tree-shaped index structure built according to video content attributes (such as theme tags, duration range, release time, etc.). It can be implemented using existing inverted index storage and traversal methods. This application dynamically constrains its traversal depth and attribute filtering range to avoid redundant searches on branches unrelated to the current scene state. The set of target candidate videos refers to the set of video metadata obtained after filtering by both depth and attribute constraints, used as input for subsequent content placement and ranking.

[0025] The closed-loop evolutionary coordination module is used to collect human-computer interaction feedback data corresponding to the target candidate video set, perform correlation regression calculations on the human-computer interaction feedback data, the high-dimensional scene state tensor, and the dynamic path constraint vector to obtain the spatial evolutionary gradient, recursively update the high-dimensional scene state tensor through the spatial evolutionary gradient, and nonlinearly correct the bias weights in the index path bias mapping matrix. The human-computer interaction feedback data refers to user behavior records such as viewing time, likes, and swipes in the target candidate video set, which can be specifically collected using platform tracking logs. Its function is to provide real-world feedback for adjusting retrieval path parameters. The spatial evolutionary gradient is a numerical value representing the direction and magnitude of the adjustment of scene state and path parameters. It can be specifically implemented using the partial derivative of the loss function with respect to parameters in regression calculations, and its function is to transform the feedback signal into a quantitative basis that can be used for parameter updates.

[0026] The core innovation of this application lies in coupling the spatiotemporal state and the device safety state into a unified high-dimensional scene state tensor through a multi-dimensional perception collaboration module. This scene state tensor is then transformed into a quantitative constraint on the traversal depth of the inverted index tree and the attribute subspace through a nonlinear mapping of the dynamic retrieval decoupling module. This allows the retrieval path to be adjusted according to changes in the scene state. At the same time, the scene state tensor and the index path bias mapping matrix are continuously corrected by human-computer interaction feedback collected by the closed-loop evolution collaboration module. An iterative adjustment mechanism is built between scene perception, path constraints and effect feedback, which alleviates the problem of delayed response to complex scene fluctuations in the fixed path indexing method.

[0027] This application further proposes the following specific process for generating high-dimensional scene state tensors:

[0028] The latitude and longitude coordinates and timestamps of N user clicks are extracted using a sliding time window, and a viewing window vector is generated through Discrete Fourier Transform. A graph attention network is used to encode the gateway security logs associated with the device MAC address using a topological adjacency matrix, generating a risk state vector. A deep fully connected neural network is then used to connect and fuse the viewing window vector and the risk state vector along the feature dimension, and layer normalization is applied to output a high-dimensional scene state tensor. The sliding time window refers to a fixed-duration segment of click records that slides along the time axis, specifically a set of click records generated within the most recent few minutes. Its purpose is to ensure that the spatiotemporal features reflected by the viewing window vector are within a recent valid range.

[0029] ;

[0030] in, ; This represents the total number of clicks within the sliding time window. For the first The latitude and longitude coordinates corresponding to each click are represented in complex form (with longitude as the real part and latitude as the imaginary part). For the first The timestamp corresponding to each click (relative time calculated from the start time of the window). The total duration of the sliding time window is the timestamp sequence. In this transformation, the timestamp sequence does not participate as an independent signal, but is instead substituted into the exponential term as the non-uniform sampling time of each click within the window, which is used to reflect the actual uneven distribution of click behavior on the time axis. These are the complex coefficients of the k-th frequency component obtained through the above non-uniform discrete Fourier transform. The unit is imaginary. View window vector. The amplitude of each frequency component With phase The system is composed of multiple components: the amplitude component characterizes the energy intensity of the user's click behavior at that frequency, i.e., the strength of the periodicity of the behavior pattern; the phase component characterizes the time offset of that frequency component relative to the window's start time. Because... The amplitude and phase of the data have been used as sampling timestamps in the calculation. They contain both the spatial coordinate distribution characteristics and the temporal density distribution characteristics of the click. Together, they transform the original discrete and unstructured coordinate timestamp sequence into a fixed-length numerical vector that can participate in subsequent fusion calculations. Its function is to preserve the temporal characteristics such as movement rhythm and dwell patterns implied in the user's click behavior in the frequency domain.

[0031] ;

[0032] ;

[0033] ;

[0034] In the gateway security log topology adjacency matrix, each node corresponds to a gateway record during a single device access process. The initial feature vector of node i is obtained by encoding log fields such as access time, hop gateway number, and connection duration. For a trainable linear transformation weight matrix, This is the weight vector corresponding to the attention mechanism. This represents vector concatenation. Let i be the set of neighboring nodes that have a jump or communication relationship with node i. The attention coefficient assigned to node i for its neighbor node j is used to characterize the degree of influence of that neighbor node on the current node's risk state. It is a non-linear activation function. This is the feature vector output by node i after aggregating neighbor information. Risk state vector. Features of all nodes The function of pooling or concatenation is to transform the topological anomaly relationships implied in multiple gateway logs that are related to each other during the device access process into a fixed-length numerical representation, so that the correlation between isolated log records can be preserved in the same vector.

[0035] ;

[0036] ;

[0037] in, , This represents the weight matrix of each layer in a deep fully connected neural network. , For the corresponding bias term, For activation function, for The mean across the feature dimensions, for Variance in the feature dimension To prevent small constants with a denominator of zero, , For the learnable scaling and offset parameters corresponding to layer normalization, This is the final output high-dimensional scene state tensor. The purpose of layer normalization is to eliminate the difference in numerical dimensions between the viewing window vector and the risk state vector due to their different sources, so that their value ranges are consistent after fusion, and to avoid one type of vector having a disproportionate influence weight in subsequent matrix mapping operations due to its larger numerical magnitude.

[0038] Specifically, the sliding time window duration T is set to 10 minutes, and a total of N=50 user clicks are recorded within the window. The 50 sets of latitude and longitude coordinates are then assigned their corresponding timestamps. Substituting the sampling time into the non-uniform discrete Fourier transform formula yields 50 frequency components. The amplitude and phase of the 16 components with the highest energy are concatenated to obtain a 32-dimensional viewing window vector. The device MAC address was associated with 8 gateway security log entries during this access process. An 8-node topological adjacency matrix was constructed and input into a graph attention network. After two layers of aggregation, the output was pooled into a 32-dimensional risk state vector. The two 32-dimensional vectors were concatenated into a 64-dimensional input, which was then processed through two layers of fully connected networks (128-dimensional hidden layer and 256-dimensional output layer) and layer normalization to obtain a 256-dimensional high-dimensional scene state tensor, which is used by the dynamic retrieval decoupling module for subsequent mapping operations.

[0039] Through the above technical solution, the temporal coordinate information of user click behavior is transformed into a viewing window vector that can reflect the movement rhythm through frequency domain transformation, and the gateway logs that are related to each other during device access are transformed into a risk state vector that can reflect the relationship between nodes through topological encoding. This reduces the problem of difficulty in uniform representation due to differences in data structure when processing coordinate sequences and security logs separately. After being combined with the content of the aforementioned multi-dimensional perception collaboration module, the high-dimensional scene state tensor obtained by full-connection fusion and layer normalization processing is consistent in numerical dimensions, providing a unified input basis for the stable performance of matrix mapping operations in the subsequent dynamic retrieval decoupling module, and also providing a clear adjustable object for the recursive update of the scene state tensor by the closed-loop evolution collaboration module.

[0040] This application further proposes a process for controlling the traversal depth of the inverted index tree by the database storage engine, including:

[0041] Traversing to the index tree When layering, calculate the current node's attributes and the attribute filtering threshold. The semantic cosine similarity; if the semantic cosine similarity is less than And the current depth Greater than the depth offset parameter If the similarity is greater than or equal to a certain value, then perform asymmetric pruning, discarding all child branches of the current node; only perform pruning on nodes with similarity greater than or equal to a certain value. And depth Less than or equal to The branches are recursively traversed to extract video content metadata from the candidate space. Attribute filtering threshold. The storage and transmission process involves two levels of value selection: one is a semantic reference vector used for similarity comparison with node attributes. The first component is obtained by mapping the numerical components of a threshold from a pre-defined attribute semantic embedding library; the second component is the numerical component used to crop and determine the similarity calculation results. Together, these two constitute a complete attribute filtering threshold. .

[0042] ;

[0043] in, This is the attribute semantic vector corresponding to the d-th layer node currently being traversed (obtained from the metadata such as video theme and tags associated with this node through a preset semantic embedding method). Filter threshold for attributes The corresponding semantic reference vector, The cosine similarity between the two is given by the formula, and its value range is given by the formula. The larger the value, the higher the semantic consistency between the current node's attributes and the target attribute subspace.

[0044] Pruning Discard sub-branches, if and ; Retain and recursively, and ;

[0045] in, Filter threshold for attributes The numerical components, The depth offset parameter is derived from the dynamic path constraint vector.

[0046] In this determination rule, the depth condition Similarity conditions The discarding of all child branches of the current node will only be triggered if both conditions are met simultaneously. That is, asymmetric pruning is only performed on branches that have exceeded the preset depth limit and whose node attributes deviate significantly from the target semantic subspace, thus avoiding the accidental pruning of shallow branches that are still within the depth limit and may still have relevance.

[0047] Specifically, let the depth offset parameter be set. Numerical components of attribute filtering threshold The search engine traverses to a node at level 6 of the index tree and calculates the semantic vector of that node's attributes. The cosine similarity is 0.62. Since 6 is greater than 5 and 0.62 is less than 0.75, the pruning condition is met, so this node and all its sub-branches are discarded. However, when traversing to another node in the 4th layer, even though its similarity is less than 0.60... However, since 4 is less than 5 and does not exceed the depth offset parameter, it is still recursively traversed. Through the above determination method, the traversal process of the inverted index tree is restricted to the candidate space defined by the depth offset parameter and the attribute filtering threshold.

[0048] The above technical solution transforms the comparison between node attributes and attribute filtering thresholds into quantifiable cosine similarity calculations. Combined with the depth offset parameter, it forms a dual judgment condition, reducing the time and computational resources required for the retrieval engine to continue traversing branches that deviate significantly from the target attribute subspace and exceed the depth limit. After being combined with the aforementioned high-dimensional scene state tensor and dynamic path constraint vector generation process, both the depth offset parameter D and the attribute filtering threshold T are obtained from the scene state tensor through matrix mapping operations. This allows the traversal depth and attribute filtering range to be adjusted according to changes in the scene state, providing a specific execution basis for the dynamic index retrieval module to output a set of target candidate videos that match the current scene.

[0049] This application further proposes that the dynamic retrieval decoupling module also monitors the cache hit rate and query response latency of the database storage engine in real time; the cache hit rate and query response latency are fed back to the multi-dimensional perception collaboration module as a dynamic correction benchmark for adjusting the feature fusion weighting value in the high-dimensional scene state tensor, so as to achieve adaptive linkage between retrieval resource allocation and scene state. The cache hit rate refers to the proportion of video metadata that the database storage engine directly retrieves from the cache during the retrieval process, and the query response latency refers to the time elapsed from initiating a retrieval request to returning the target candidate video set. Both can be obtained through the operation log statistics of the database storage engine, and are used to reflect the retrieval efficiency under the current retrieval path constraints.

[0050] ;

[0051] in, In the specific process of generating the high-dimensional scene state tensor, the deep fully connected neural network is used to fuse the weight matrix of the fully connected layers of the viewing window vector and the risk state vector. This represents the current measured cache hit rate. The preset target cache hit rate, This represents the current measured query response latency. The preset target response delay, , These are the hit rate correction coefficient and the delay correction coefficient, respectively, used to control the degree of influence of the two types of feedback signals on the adjustment of the weight matrix. This is the weight matrix, adjusted for retrieval efficiency feedback, used in the next round of fusion computation. When the measured cache hit rate is lower than the target value or the measured response latency is higher than the target value, the overall scaling factor within the square brackets is less than 1. Compared to Overall scaling reduces the weighting of feature components that are detrimental to current retrieval efficiency in the next round of fusion computation of the high-dimensional scene state tensor; conversely, when the retrieval efficiency is better than the target value, the scaling factor is greater than 1, and the weighting of each feature component in the fusion process is correspondingly increased.

[0052] Specifically, the preset target cache hit rate Target response delay Milliseconds, hit rate correction factor Delay correction factor Measured cache hit rate over a certain period of time Actual query response latency Substituting milliseconds into the above formula, we get the scaling factor within the square brackets as 0.936, which is... The weight matrix of the fully connected layer is reduced by about 6.4% in the next round of scene state tensor generation.

[0053] By incorporating the cache hit rate and query response latency during the database storage engine's operation into the scene state tensor generation process, the resource allocation deviation caused by the independence and lack of correlation between retrieval efficiency and scene perception processes is reduced. Combined with the aforementioned specific process of generating high-dimensional scene state tensors, the feature fusion weighting value is dynamically adjusted according to retrieval efficiency feedback, establishing a quantifiable linkage between the multi-dimensional perception collaboration module and the actual operating state of the database storage engine. This provides the scene state input, which is adjusted according to the retrieval resource status, for the subsequent matrix mapping operation of the dynamic retrieval decoupling module.

[0054] This application further proposes that when the dynamic index retrieval module performs a biased traversal of the retrieval path, it performs asymmetric pruning on different branches of the inverted index tree based on the depth offset parameter in the dynamic path constraint vector. If the branch relevance weight is lower than a preset dynamic entropy threshold, the index path weight offset of that branch is passed to the closed-loop evolution coordination module to adjust the spatial evolution gradient. The branch relevance weight refers to the numerical value that measures the relevance between a candidate node under a certain branch and the current scene state. Specifically, it can be obtained from the average attribute similarity value of the candidate nodes under that branch. The dynamic entropy threshold is a judgment criterion that changes with the distribution of relevance weights of each branch under the current node, and is used to replace a fixed threshold to prune and determine the importance of branches.

[0055] ;

[0056] ;

[0057] in, This represents the total number of branches under the current node. Let i be the relevance weight of the i-th branch. This is the normalized probability value of the weight. This is the information entropy of the correlation weight distribution of each branch under the current node. Its magnitude reflects the degree of dispersion or concentration of the correlation weight distribution of each branch.

[0058] ;

[0059] in, This is the preset dynamic entropy threshold as the basic quantity. for The maximum entropy value corresponding to the uniform distribution of branches. The normalized entropy value. This is the dynamic entropy threshold for the current node. When the correlation weights of each branch are relatively concentrated ( When it is smaller, Approaching The judgment criteria become stricter; when the correlation weights of each branch are relatively dispersed ( When it is relatively large, The corresponding reduction loosens the judgment criteria, avoiding over-pruning when the importance of branches is difficult to distinguish.

[0060] ,like ;

[0061] in, This is the index path weight offset generated when the relevance weight of the i-th branch is lower than the dynamic entropy threshold. Its value is negative, indicating the degree of deviation of the branch from the current judgment benchmark. It is passed to the closed-loop evolution coordination module as an additional input in the aforementioned spatial evolution gradient correlation regression operation, and is used to reflect the influence of the pruned branches on the adjustment direction of the scene state tensor and index path bias mapping matrix during the gradient calculation process.

[0062] Specifically, a certain node has a total of There are 1 branch, with relevance weights of 0.5, 0.3, 0.15, and 0.05 respectively. The calculated... , ,Pick Substituting into the formula, we get The branch with a relevance weight of 0.05 is lower than... This triggers asymmetric pruning, generating a weight offset of 0.007, which is then passed to the closed-loop evolutionary coordination module to participate in the calculation of the spatial evolution gradient.

[0063] By using the above technical solution, a dynamic entropy threshold that changes with the distribution of branch relevance weights is used instead of a fixed pruning benchmark. This reduces candidate set bias caused by overly strict or lenient pruning of fixed thresholds in scenarios where the importance of each branch is difficult to distinguish. When combined with the aforementioned process of generating spatial evolution gradients, the index path weight offset corresponding to the pruned branch is used as an additional input to participate in the association regression calculation. This allows the pruning results of the retrieval path to be further reflected in the update direction of the scene state tensor and the index path bias mapping matrix, providing a more complete input source for parameter correction of the closed-loop evolutionary coordination module.

[0064] This application further proposes the following process for obtaining the spatial evolution gradient:

[0065] Construct the objective function ,in For the predicted interactive feedback calculated based on the current bias weights, For the actual human-computer interaction feedback collected, The regularization coefficient is . The mapping matrix parameters are used; the spatial evolution gradient is the objective function. Partial derivatives with respect to the bias weights. Predicting interactive feedback. It refers to the estimated value of user interaction effect obtained by estimating the bias weights in the current index path bias mapping matrix and the high-dimensional scene state tensor. Specifically, it can be obtained by nonlinear mapping through the inner product of the bias weights and the scene state tensor.

[0066] ;

[0067] in, Let m be the index of the m-th video in the target candidate video set. This is the bias weight vector corresponding to the branch to which the video belongs in the index path bias mapping matrix. This is a unified high-dimensional scene state tensor for this round of retrieval (for each candidate video within the same user and retrieval loop). The values ​​are the same; the difference lies in the bias weight of the branch to which each candidate video belongs. different), It is a non-linear mapping function (such as the Sigmoid function). The first one calculated based on this The predicted interactive feedback value corresponding to each candidate video, its value range is determined by... After function mapping, it falls into The interval is used for numerical comparison with the actual collected human-computer interaction feedback.

[0068] ;

[0069] in, This refers to the number of target candidate videos that have generated human-computer interaction feedback, collected by the closed-loop evolutionary collaborative module in this round. To iterate through the indices of each video in the target candidate video set, , The first The predicted interactive feedback and the actual human-computer interaction feedback (after normalization) for each candidate video fall within the same numerical range. The regularization coefficient is . This refers to the index path bias mapping matrix and the adjustable parameters in the matrix mapping function. This term is used to constrain the range of parameter values, preventing the bias weights from diverging after multiple updates. The summation sign represents the sum of values ​​that have already received feedback. The prediction bias is accumulated for each candidate video, making the objective function... This comprehensively reflects the overall prediction accuracy of all candidate videos in this round of retrieval.

[0070] ;

[0071] in, for The derivative of a function That is, the first The spatial evolution gradient of the bias weights corresponding to the branches to which each candidate video belongs, along with its direction and magnitude, characterizes the direction and magnitude of the adjustment of the bias weights for that branch: when the deviation between the predicted and actual interaction feedback corresponding to that branch is large, the gradient value increases, and the adjustment magnitude of the bias weights for that branch in the corresponding round also increases accordingly; when the deviation is small, the gradient value approaches zero, and the adjustment of the bias weights for that branch tends to stabilize. The spatial evolution gradient of the bias weights for each branch... By writing back the gradient matrix according to its row and column positions in the index path bias mapping matrix, we obtain the complete spatial evolution gradient matrix for this round. This matrix is ​​used to recursively update the state tensor of the high-dimensional scene and to perform nonlinear correction on the bias weights in the index path bias mapping matrix.

[0072] Specifically, in the target candidate video set, the first Bias weights of the branches to which the videos belong With scene state tensor The inner product is used to obtain the predicted interactive feedback through the Sigmoid mapping. The closed-loop evolutionary collaboration module collects the actual human-computer interaction feedback (after normalization) corresponding to this video. Take the regularization coefficient Substituting into the gradient formula, we obtain the spatial evolution gradient corresponding to this branch. A positive value indicates that the bias weight of that branch should be increased in the next update to make the subsequent predicted interaction feedback closer to the actual feedback level; the branches to which the remaining candidate videos belong are determined according to their respective... and Calculate the gradients separately in the same way.

[0073] The above technical solution uses correlation regression to obtain the spatial evolution gradient by analyzing the deviation between predicted and actual interactive feedback. This reduces the bias caused by the lack of quantitative basis and reliance on empirical settings in the bias weight adjustment process. When combined with the process of collecting human-computer interaction feedback data by the aforementioned closed-loop evolution collaboration module, the spatial evolution gradient is used to recursively update the high-dimensional scene state tensor and to nonlinearly correct the bias weights in the index path bias mapping matrix. This provides a unified numerical calculation basis for the adjustment process between scene perception, path constraints, and retrieval effect feedback.

[0074] This application further proposes that the closed-loop evolutionary coordination module is also used for:

[0075] After the human-computer interaction feedback data undergoes nonlinear activation mapping processing, it is simultaneously injected into the feature fusion process of the multi-dimensional perception collaboration module and the matrix mapping operation process of the dynamic retrieval decoupling module. The multi-dimensional perception collaboration module and the dynamic retrieval decoupling module synchronously update their internal mapping parameters before the next retrieval cycle to construct a nonlinear data processing closed loop. Nonlinear activation mapping processing refers to the computational process of converting the original human-computer interaction feedback data into mapping quantities that can be directly used for parameter updates. Specifically, it can be implemented using a hyperbolic tangent mapping with trainable weights.

[0076] ;

[0077] in, This refers to the actual human-computer interaction feedback data collected by the closed-loop evolutionary collaboration module. This is the trainable weight matrix corresponding to the mapping process. For the corresponding bias term, The hyperbolic tangent function is used to constrain the mapping result to... interval, This is the injection amount obtained after nonlinear activation mapping.

[0078] ;

[0079] ;

[0080] ;

[0081] in, This refers to the weight matrix of the fully connected layer used in the t-th round during the feature fusion process of the multi-dimensional perception collaborative module. This is the index path bias mapping matrix used in the t-th round of the dynamic retrieval decoupling module matrix mapping function. The partial derivative of the objective function J with respect to the high-dimensional scene state tensor (from the aforementioned prediction interaction feedback formula) (obtained by chain rule differentiation) The partial derivative of the high-dimensional scene state tensor with respect to the weight matrix of the fully connected layer (as described in the aforementioned fusion formula). (and layer normalization operations are obtained through chain rule differentiation). That is, the update direction obtained by multiplying the two; ∇M is the spatial evolution gradient corresponding to each branch. The gradient matrix itself is obtained by summing its positions in the matrix, which is the output of the aforementioned spatial evolution gradient process. This indicates element-wise weighting. , These are the corresponding update step size coefficients for the two. , For injection The parameters for the next round are obtained after synchronization. Because... and Both can be traced back to the same objective function J and the same batch of human-computer interaction feedback data, and they also share the same nonlinear activation mapping injection amount. The multi-dimensional perception collaboration module and the dynamic retrieval decoupling module always adjust parameters based on the same batch of feedback information before entering the next retrieval cycle, thus avoiding misalignment in update direction and update rhythm between the two modules due to using different batches of feedback data or different gradient sources.

[0082] Specifically, the actual human-computer interaction feedback data collected in a certain search cycle through , Map and retrieve Later obtained Take the step size coefficient , ,Will After element-wise weighting of the update directions corresponding to the feature fusion process and the matrix mapping process, respectively, the results are obtained synchronously. , The updated results before the next round of retrieval are used for the next generation of high-dimensional scene state tensors and dynamic path constraint vector mapping.

[0083] By using the above technical solution, the same nonlinear activation mapping result is synchronously injected into the feature fusion process and the matrix mapping operation process, reducing the time difference that occurs when the two modules respond to the same batch of feedback data due to their independent parameter updates. After being combined with the aforementioned spatial evolution gradient generation process, the gradient information is synchronously used to update the feature fusion weight and index path bias mapping matrix after nonlinear activation mapping. This ensures that the multidimensional perception collaboration module and the dynamic retrieval decoupling module always evolve based on consistent feedback information in multiple retrieval cycles, providing a synchronous update implementation method for constructing a closed loop of iterative nonlinear data processing.

[0084] This application further proposes that the multidimensional perception collaboration module is also used to calculate the state transition probability distribution of the high-dimensional scene state tensor within a future preset time window based on the multidimensional evolutionary increment matrix output by the closed-loop evolution collaboration module, and to use the probability distribution as the prediction regularization coefficient for the matrix mapping operation performed by the dynamic retrieval decoupling module. The multidimensional evolutionary increment matrix refers to the matrix output by the closed-loop evolution collaboration module during the recursive update of the high-dimensional scene state tensor, which characterizes the changing trend of the scene state in each feature dimension. Specifically, it can be obtained by statistically analyzing the differences in the scene state tensor during multiple rounds of updates. The state transition probability distribution is based on... A typical scenario state category is formed by pre-clustering high-dimensional scenario state tensor samples accumulated in the platform's history (e.g., -Mean clustering) partitioning There are 10 categories, each corresponding to a category center tensor, used to represent a class of scene states with similar spatiotemporal and equipment safety characteristics; the number of categories is... The central tensors of each category are predetermined before system deployment and can be periodically re-clustered and updated as historical data continues to accumulate.

[0085] ;

[0086] in, For the current high-dimensional scene state tensor, The multidimensional evolutionary increment matrix output by the closed-loop evolutionary coordination module. For the reasons mentioned above A matrix formed by concatenating the class center tensors row by row is used to project the computation result onto the matrix. Each typical scenario state category corresponds to one 3D space The preset future time window length, This is the bias term corresponding to the operation. The function is used to convert the projection result into a normalized probability distribution. That is, the state tensor of the high-dimensional scene is transferred to this state tensor within a preset time window in the future. The probability distribution of each category in a typical scenario state category, and the number of components and categories. Consistent.

[0087] ;

[0088] in, This is a pre-defined base quantity for the regularization coefficients in the matrix mapping operation of the dynamic retrieval decoupling module. This represents the maximum probability value in the state transition probability distribution. This is the prediction regularization coefficient after correction for the state transition probability distribution. When When the value is relatively large, meaning the direction of future scenario state transitions is relatively certain. Correspondingly reduced, matrix mapping operations place a higher degree of trust in the mapping results of the current scene state tensor; when When the value is relatively small, meaning there is significant uncertainty regarding the direction of future scenario state transitions. The matrix mapping operation is increased accordingly, and it imposes stronger constraints on the current mapping result, avoiding the output of overly absolute path constraint vectors when the scene state is about to undergo significant fluctuations.

[0089] Specifically, take a future preset time window Minutes, the current scene state tensor With multidimensional evolutionary increment matrix Perform calculations and through After processing, a probability distribution covering typical scenario states is obtained, where the maximum probability value is... Take the basic value of regularization coefficient. Substituting into the formula, we get This value serves as the prediction regularization coefficient and participates in the matrix mapping operation of the dynamic retrieval decoupling module in this round.

[0090] By using the above technical solution, the state transition probability distribution of the high-dimensional scene state tensor within a future preset time window is calculated based on the multidimensional evolution increment matrix. This reduces the path constraint bias caused by matrix mapping operations that only consider the scene state at the current moment without taking into account its subsequent change trend. After being combined with the aforementioned nonlinear data processing closed loop, the state transition probability distribution is transformed into prediction regularization coefficients and fed back to the matrix mapping operation process. This allows the generation of dynamic path constraint vectors to reflect not only the current scene state but also, to a certain extent, the possible subsequent change trend of the scene state.

[0091] This application further proposes that the closed-loop evolutionary collaboration module is configured to perform feature comparison between the human-computer interaction feedback data and the high-dimensional scene state tensor; the multi-dimensional perception collaboration module receives update parameters sent by the closed-loop evolutionary collaboration module, and uses the human-computer interaction feedback data to recursively weight and correct the viewing window vector and the risk state vector to update the feature weights of the high-dimensional scene state tensor. Feature comparison refers to the process of comparing the difference or correlation between the human-computer interaction feedback data and the high-dimensional scene state tensor in the feature dimension, used to determine the contribution of the viewing window vector and the risk state vector to the current interaction effect.

[0092] ;

[0093] ;

[0094] in, , The first The wheel's viewing window vector and risk state vector, , These are the actual and predicted human-computer interaction feedback, respectively. , These are the partial derivatives of the high-dimensional scene state tensor with respect to the viewing window vector and the risk state vector, respectively (obtained by chain rule from the aforementioned fusion formula). , These are the corresponding correction step size coefficients for the two. , This is the next round vector obtained after recursive weighted correction. This correction process adjusts the viewing window vector and the risk state vector by different magnitudes according to their respective contributions to the prediction bias, rather than applying the same magnitude of adjustment to both.

[0095] Specifically, in a certain round and The deviation is 0.18. Feature comparison determines that the viewing window vector contributes more to this deviation than the risk state vector. , After substituting into the formula, the correction magnitude of the viewing window vector is greater than that of the risk state vector. After both are updated together, they re-participate in the next round of fusion calculation to obtain the updated high-dimensional scene state tensor.

[0096] The above technical solution determines the correction magnitude of the viewing window vector and the risk state vector by comparing the features of the human-computer interaction feedback data and the high-dimensional scene state tensor. This reduces the update bias caused by applying the same correction magnitude to both types of vectors without distinguishing their respective contributions to the prediction bias. When combined with the aforementioned specific process of generating the high-dimensional scene state tensor, the viewing window vector and the risk state vector re-participate in the fusion calculation after recursively weighted correction, enabling the feature weights of the high-dimensional scene state tensor to be updated differently according to the actual contribution of each component to the interaction effect.

[0097] This application further proposes that the closed-loop evolutionary coordination module is also used to detect the risk state vector component in the high-dimensional scene state tensor; when the risk level represented by the risk state vector component is updated to below a preset safety threshold, a task clearing instruction is sent to the dynamic retrieval decoupling module, and the dynamic retrieval decoupling module removes the corresponding bias weight data from the index path bias mapping matrix according to the task clearing instruction. The risk level refers to the scalar value obtained by projection operation on the risk state vector component, used to characterize the severity of security risks in the current device access environment.

[0098] ;

[0099] in, For the first The risk state vector after each round of updates The preset risk projection weight vector is obtained by pre-training through logistic regression fitting of historically labeled risk event samples (i.e., risk state vectors corresponding to devices confirmed to have abnormal access). It is used to project the multi-dimensional risk state vector into a single risk level scalar. Dimensions with larger projection weights correspond to topological anomaly features that have a greater impact on the risk level. This is the current risk level scalar value obtained through projection calculation. The trigger condition for the task reset instruction is... ,in This is a preset security threshold. When this condition is met, it indicates that the security risks previously recorded in the device access environment have subsided to an acceptable range. At this time, the closed-loop evolution collaboration module sends a task clearing instruction to the dynamic retrieval decoupling module. The instruction carries the device access branch identifier corresponding to the original gateway log that triggered the current risk status escalation. After responding to the instruction, the dynamic retrieval decoupling module locates the corresponding row and column position in the index path bias mapping matrix based on the branch identifier (this position is the position written when the risk status vector was projected by the matrix mapping function to generate the bias weight). The bias weight value at this position is set to the preset initial baseline value to complete the removal operation. The remaining bias weight data in the matrix that are unrelated to the current risk status are unaffected.

[0100] Specifically, preset security threshold The risk level of a device, calculated by projection, was 0.35 for a risk state vector component previously generated due to accessing an abnormal gateway. After several rounds of closed-loop updates, the risk level decreased to 0.15, satisfying the requirement. The closed-loop evolution coordination module sends a task clearing command, and the dynamic retrieval decoupling module locates the bias weight data previously generated due to the risk state in the index path bias mapping matrix and removes it to prevent the risk state that has subsided from continuing to affect the constraint results of subsequent retrieval paths.

[0101] By using the above technical solution, the risk level obtained by projection calculation of the risk state vector components is used as the judgment criterion, which reduces the path constraint deviation caused by the lack of a clearing mechanism to continuously affect the index path bias mapping matrix due to the elimination of safety hazards. After being combined with the aforementioned generation and correction process of the index path bias mapping matrix, the introduction of the task zeroing instruction enables the bias weight data in the index path bias mapping matrix to be removed or retained as the actual safety status of the device changes, providing a specific execution method for keeping the index path bias mapping matrix consistent with the current safety status of the device after multiple rounds of closed-loop updates.

[0102] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0103] A user of a short video platform watched videos on the platform while riding the subway in the evening. Their device experienced multiple base station and gateway switching along the subway line. The multi-dimensional perception and collaboration module extracted the latitude and longitude coordinates and timestamps corresponding to 50 clicks made by the user within the last 10 minutes, obtaining a 32-dimensional viewing window vector through discrete Fourier transform. Simultaneously, it extracted 8 gateway security logs associated with the user's device MAC address during this access process, obtaining a 32-dimensional risk state vector through graph attention network encoding. Due to the frequent gateway switching along the subway line, the risk state vector exhibited high topological fluctuation characteristics. The two vectors were fused through a deep fully connected network and layered normalization to obtain a 256-dimensional high-dimensional scene state tensor. The dynamic retrieval decoupling module previously used the database storage engine cache hit rate. Query response latency The milliseconds, after feedback correction, yielded the adjusted feature fusion weighted value, which has already participated in the generation of the scene state tensor in this round.

[0104] After the high-dimensional scene state tensor is processed by a preset 256×128-dimensional index path bias mapping matrix and a matrix mapping function, the depth offset parameter is output. Attribute filtering threshold numerical components The dynamic index retrieval module traverses an inverted index tree constructed from approximately 1.2 million video metadata entries for that region's subject matter. Upon reaching a node at level 6, it calculates the semantic vector of that node's attributes. The cosine similarity is 0.6. Since 6 is greater than 5 and 0.6 is less than 0.75, asymmetric pruning is triggered. The relevance weights of the four branches under a certain node are 0.5, 0.3, 0.15, and 0.05, respectively. After dynamic entropy threshold calculation, the result is 0.043. The branch with a weight of 0.05 is judged to be below the threshold, generating a weight offset of 0.007, which is then passed to the closed-loop evolutionary coordination module. Finally, the dynamic index retrieval module outputs a total of 860 target candidate videos.

[0105] After users interact with 860 videos in the candidate video set by watching or swiping, the closed-loop evolutionary collaboration module collects the normalized actual human-computer interaction feedback corresponding to all videos that have generated feedback. Interactive feedback with the predictions calculated based on the bias weights of their respective branches. Compare each item one by one, with the 17th video corresponding to... , An objective function J is constructed, and the spatial evolution gradients corresponding to each branch are calculated. These gradients are used to recursively update the high-dimensional scene state tensor and to nonlinearly correct the bias weights in the index path bias mapping matrix. Simultaneously, the aforementioned weight offset of 0.007 is used as an additional input in this round of regression calculation. The human-computer interaction feedback data is processed by nonlinear activation mapping to obtain the injection volume. The feature fusion process of the synchronously injected multidimensional perception collaboration module and the matrix mapping operation process of the dynamic retrieval decoupling module enable the two modules to complete synchronous updates before the next retrieval cycle.

[0106] Based on the multidimensional evolution increment matrix output by the closed-loop evolution collaboration module, the multidimensional perception collaboration module calculates the state transition probability distribution of the high-dimensional scene state tensor within the next 5 minutes, where the maximum probability value is 0.72. The prediction regularization coefficient is then obtained from this distribution. The closed-loop evolutionary collaboration module further compares the human-computer interaction feedback data with the high-dimensional scene state tensor to determine that the contribution of the viewing window vector to the prediction deviation in this round is greater than that of the risk state vector. Based on this, the module adjusts the two parameters by step size coefficients. , A recursive weighted correction is performed, and the updated vector participates in the next round of fusion calculation. After several rounds of closed-loop updates, the user has left the densely signal-switching section along the subway line, and the risk level of the risk state vector component, obtained through projection calculation, has decreased from 0.35 to 0.15, which is below the preset safety threshold. The closed-loop evolution coordination module sends a task clearing instruction to the dynamic retrieval decoupling module, and the corresponding bias weight data previously generated due to the switching of metro gateways is removed from the index path bias mapping matrix.

[0107] The scene perception, path constraint, index traversal, and feedback correction processes involved in the multi-dimensional perception collaboration module, dynamic retrieval decoupling module, dynamic index retrieval module, and closed-loop evolution collaboration module are interconnected, forming a reciprocating adjustment process from the user's spatiotemporal and device safety status input to the target candidate video set output, and then to feedback correction. Based on the aforementioned technical contents, the collaborative change relationship of the parameters of each module within the same retrieval cycle is further demonstrated, providing a complete example for those skilled in the art to understand and implement the technical solution of this application.

[0108] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A precise audience matching system for content delivery on short video platforms, characterized in that, Includes the following modules: The multi-dimensional perception collaboration module is used to obtain a viewing window vector containing the user's spatiotemporal state distribution characteristics and a risk state vector containing the device's security topology association, and to perform multi-dimensional coupling processing on the viewing window vector and the risk state vector to generate a high-dimensional scene state tensor representing the scene state. The dynamic retrieval decoupling module is used to store a preset index path bias mapping matrix. It projects the high-dimensional scene state tensor onto the index path bias mapping matrix through a preset matrix mapping function to perform nonlinear spatial mapping operations, generating a dynamic path constraint vector including depth offset parameters, attribute filtering thresholds and bias weights. The dynamic index retrieval module is used to obtain video content metadata stored in the database storage engine within the short video platform, and based on the depth offset parameter extracted by the dynamic path constraint vector, control the traversal depth of the database storage engine on the inverted index tree, and restrict the attribute filtering logic of the video content metadata to a specific content attribute subspace defined by the dynamic path constraint vector, and output a target candidate video set. The closed-loop evolutionary coordination module is used to collect human-computer interaction feedback data corresponding to the target candidate video set, and perform correlation regression calculation on the human-computer interaction feedback data, the high-dimensional scene state tensor and the dynamic path constraint vector to obtain the spatial evolution gradient. The high-dimensional scene state tensor is recursively updated using the spatial evolution gradient, and the bias weights in the index path bias mapping matrix are nonlinearly corrected.

2. The content delivery audience matching system for short video platforms according to claim 1, characterized in that, The process of generating a high-dimensional scene state tensor is as follows: The latitude and longitude coordinates and timestamps of N user clicks are extracted by a sliding time window, and a viewing window vector is generated by discrete Fourier transform; a risk state vector is generated by encoding the gateway security log associated with the device MAC address using a graph attention network and topological adjacency matrix. The viewing window vector and the risk state vector are connected and fused along the feature dimension using a deep fully connected neural network, and a high-dimensional scene state tensor is output through layer normalization processing.

3. The content delivery audience matching system for short video platforms according to claim 1, characterized in that, The process of controlling the traversal depth of the inverted index tree by the database storage engine includes: Traversing to the index tree When layering, calculate the current node's attributes and the attribute filtering threshold. The semantic cosine similarity; if the semantic cosine similarity is less than And the current depth Greater than the depth offset parameter If the similarity is greater than or equal to a certain value, then perform asymmetric pruning, discarding all child branches of the current node; only perform pruning on nodes with similarity greater than or equal to a certain value. And depth Less than or equal to The branches are recursively traversed to extract video content metadata from the candidate space.

4. The content delivery audience matching system for short video platforms according to claim 1, characterized in that, The dynamic retrieval decoupling module also monitors the cache hit rate and query response latency of the database storage engine in real time; it feeds back the cache hit rate and query response latency to the multi-dimensional perception and collaboration module as a dynamic correction benchmark for adjusting the feature fusion weighting value in the high-dimensional scene state tensor, so as to realize the adaptive linkage between retrieval resource allocation and scene state.

5. The content delivery audience matching system for short video platforms according to claim 1, characterized in that, When the dynamic index retrieval module performs a retrieval path bias traversal, it performs asymmetric pruning on different branches of the inverted index tree according to the depth offset parameter in the dynamic path constraint vector. If the relevance weight of the branch is lower than the preset dynamic entropy threshold, the index path weight offset of the branch is passed to the closed-loop evolution coordination module to adjust the spatial evolution gradient.

6. The content delivery audience matching system for short video platforms according to claim 1, characterized in that, The process of obtaining the spatial evolution gradient is as follows: Construct the objective function ; in, For the predicted interactive feedback calculated based on the current bias weights, For the actual human-computer interaction feedback collected, The regularization coefficient is . The mapping matrix parameters are used; the spatial evolution gradient is the objective function. Partial derivatives with respect to the bias weights.

7. The content delivery audience matching system for short video platforms according to claim 1, characterized in that, The closed-loop evolutionary coordination module is also used to: after the human-computer interaction feedback data has undergone nonlinear activation mapping processing, simultaneously inject it into the feature fusion process of the multidimensional perception coordination module and the matrix mapping operation process of the dynamic retrieval decoupling module, and synchronously update the internal mapping parameters of the multidimensional perception coordination module and the dynamic retrieval decoupling module before the next retrieval cycle, so as to construct a nonlinear data processing closed loop.

8. The content delivery audience matching system for short video platforms according to claim 1, characterized in that, The multidimensional perception collaboration module is also used to calculate the state transition probability distribution of the high-dimensional scene state tensor within a future preset time window based on the multidimensional evolution increment matrix output by the closed-loop evolution collaboration module, and use the probability distribution as the prediction regularization coefficient for matrix mapping operation by the dynamic retrieval decoupling module.

9. The content delivery audience matching system for short video platforms according to claim 1, characterized in that, The closed-loop evolutionary collaboration module is also configured to perform feature comparison between the human-computer interaction feedback data and the high-dimensional scene state tensor; The multidimensional perception collaboration module receives the update parameters sent by the closed-loop evolution collaboration module, and uses the human-computer interaction feedback data to recursively weight and correct the viewing window vector and the risk state vector in order to update the feature weights of the high-dimensional scene state tensor.

10. The content delivery audience matching system for short video platforms according to claim 1, characterized in that, The closed-loop evolution collaboration module is also used to detect the risk state vector component in the high-dimensional scene state tensor; when the risk level represented by the risk state vector component is updated to below the preset safety threshold, a task clearing instruction is sent to the dynamic retrieval decoupling module, and the dynamic retrieval decoupling module removes the corresponding bias weight data from the index path bias mapping matrix according to the task clearing instruction.