Incremental thinking chain-based museum visitor interest evolution recommendation system and method

Abstract

This paper presents a museum visitor interest evolution recommendation system and its methodology based on incremental thinking chains, involving the interdisciplinary fields of intelligent recommendation systems and edge computing. The system obtains the actual interaction time and associated basic cognitive time of the target visitor with the first exhibit, and derives a relative interest bias parameter based on the ratio of their differences. If the bias exceeds a threshold, it is converted into a graph update weight, updating the state attributes of associated nodes in the behavior-related state graph and cascading updates to adjacent nodes; otherwise, the cache is cleared. Finally, nodes whose state attributes satisfy the ranking criteria are extracted from the reconstructed graph, and their associated objects are used as target recommendation objects. This system improves the accuracy of museum visitor interest evolution recommendations and the utilization efficiency of edge computing resources, ensuring that the recommendation results are highly consistent with the visitor's actual interest state.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of intelligent recommendation systems and edge computing, specifically a museum visitor interest evolution recommendation system and method based on incremental thinking chains. Background Technology

[0002] Existing museum recommendation systems use edge computing nodes to collect the absolute dwell time of visitors in front of exhibits as the sole criterion for determining the degree of interest. When the dwell time exceeds a fixed threshold, evidence is injected into the Bayesian graph network and the recommendation graph is updated. The implicit assumption is that dwell time and degree of interest are linearly positively correlated.

[0003] However, the visual structural complexity and textual description load of different exhibits vary significantly, and a considerable portion of the time visitors spend in front of complex exhibits is not due to subjective interest. Existing technologies fail to separate decoding time from interest response at the cognitive mechanism level, causing dwell time to fail to reflect true attention. Decoding time is misjudged as a sudden surge in interest, resulting in a large number of false interest nodes being injected into the network, and edge computing nodes frequently performing invalid incremental operations.

[0004] Therefore, how to break through the inherent assumption of a linear mapping between absolute dwell time and audience interest, and how to strip away the undesirable impact of cognitive decoding time on audience behavior by starting from the visual information entropy of the displayed object itself, in order to obtain interest characteristics that can truly reflect the relative attention of the audience, has become an urgent technical problem to be solved. Summary of the Invention

[0005] The purpose of this application is to provide a museum visitor interest evolution recommendation system and method based on incremental thinking chain, aiming to overcome the defects of existing technologies, such as the waste of computing power and deviation of reasoning logic caused by ignoring the inherent cognitive decoding time of the exhibits and the incorrect extraction of absolute dwell time as pseudo-interest features, thereby improving the accuracy of museum visitor interest evolution recommendation and the utilization efficiency of edge computing resources.

[0006] The objective of this application can be achieved through the following technical solution: Firstly, a museum visitor interest evolution recommendation method based on incremental thinking chain, including the following steps:

[0007] Obtain the actual interaction time of the target audience with the first displayed object, and obtain the basic cognitive time associated with the first displayed object;

[0008] Based on the difference ratio between the actual interaction time and the basic cognition time, a relative interest bias parameter representing the relative attention level of the target object is obtained.

[0009] When the relative interest bias parameter is greater than the preset bias threshold, the relative interest bias parameter is converted into a graph update weight parameter based on the preset weight conversion coefficient; otherwise, the process cache containing the actual interaction duration is cleared.

[0010] Based on the preset behavior association state graph, the first graph node associated with the first display object is extracted, and the current state attribute value of the first graph node is fused and updated based on the graph update weight parameter to obtain the updated state attribute value.

[0011] Based on the updated state attribute values ​​of the first graph node, the state attribute values ​​of the second graph nodes connected to it are updated to obtain a reconstructed behavior association state graph. Based on the reconstructed behavior association state graph, the second graph nodes whose state attribute values ​​meet the preset sorting conditions are extracted, and the second display objects associated with them are used as target recommendation objects.

[0012] Secondly, the museum visitor interest evolution recommendation system based on incremental thinking chain includes the following modules:

[0013] The data acquisition module is used to obtain the actual interaction time of the target audience with the first display object, and to obtain the basic cognitive time bound to the first display object;

[0014] The feature extraction module is used to obtain a relative interest bias parameter that represents the relative attention level of the target object based on the difference ratio between the actual interaction time and the basic cognition time.

[0015] The bias determination module is used to convert the relative interest bias parameter into a graph update weight parameter based on a preset weight conversion coefficient when the relative interest bias parameter is greater than a preset bias threshold; otherwise, it clears the process cache containing the actual interaction duration.

[0016] The state update module is used to extract the first graph node associated with the first display object based on the preset behavior association state graph, and to fuse and update the current state attribute value of the first graph node based on the graph update weight parameter to obtain the updated state attribute value.

[0017] The result output module is used to update the state attribute values ​​of the second graph nodes connected to the first graph node based on the updated state attribute values ​​of the first graph node to obtain a reconstructed behavior association state graph. Based on the reconstructed behavior association state graph, the module extracts the second graph nodes whose state attribute values ​​meet the preset sorting conditions and uses the second display object associated with them as the target recommendation object.

[0018] Thirdly, a computer storage medium stores computer-executable instructions that, when executed, implement the museum visitor interest evolution recommendation method based on incremental thought chain described in the first aspect.

[0019] Compared with the prior art, the beneficial effects of this application are:

[0020] This application constructs a cognitive baseline time by introducing the visual information entropy of the displayed object itself, thus eliminating the interfering influence of cognitive decoding time on the audience's dwell behavior and significantly reducing the injection ratio of pseudo-interest features. Through a local graph reconstruction mechanism within the neighborhood range, the peak concurrent processing of edge computing nodes is significantly reduced without increasing computing power resources. By strictly constraining the graph update range with damping coefficients and transmission truncation thresholds, the consistency between recommendation results and the audience's true psychological state is improved. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the steps of the museum visitor interest evolution recommendation method based on incremental thinking chain in this application;

[0022] Figure 2 This is a schematic diagram of the modules of the museum visitor interest evolution recommendation system based on incremental thinking chain in this application. Detailed Implementation

[0023] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only to illustrate selected embodiments of this application.

[0024] Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item has been defined in one figure, it does not need to be further defined and explained in subsequent figures. The terms first, second, etc. are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0025] In the typical operation of incremental recommendation in museums, existing edge computing nodes usually use the absolute dwell time of visitors in front of exhibits as the basis for judging the intensity of interest during the behavior collection and encoding stage. When the dwell time exceeds a certain fixed threshold, strong interest evidence is injected into the gated chain editor, triggering a local incremental update of the Bayesian mind chain graph.

[0026] This approach may be barely applicable in scenarios where the information density of the exhibits is relatively small, but the visual composition complexity and textual load of exhibits in museum exhibition halls often vary greatly: the physiological time required for viewers to complete basic visual decoding is orders of magnitude different between a simple porcelain piece with minimalist design and a bronze artifact with intricate patterns and lengthy explanatory text.

[0027] Specifically, when viewers linger in front of an object with dense patterns for tens of seconds, a large portion of that time is actually spent visually scanning and semantically decoding the complex patterns and explanatory text, rather than being driven by subjective interest. However, existing systems, because they do not distinguish between the decoding time and the interest time, mistakenly extract this part of the dwell time, which is actually a physiological burden, as a feature of sudden interest surge, thus frequently triggering the gated chain editor to start the Bayesian prior network for re-inference.

[0028] Under normal high-traffic conditions, the high-frequency triggering of such pseudo-interest features not only causes a large number of invalid computing power peaks in edge computing nodes, but also continuously pollutes the node states of the Bayesian prior network. This leads to a significant deviation between the recommendation reasoning chain presented to the audience and the audience's actual psychological activities, and even logically inconsistent jump recommendations. In the long run, this will also weaken the audience's trust in the overall recommendation system.

[0029] Furthermore, existing technologies typically employ a global recalculation model after triggering a graph update: whenever a displayed object generates new behavioral data, the backend needs to reorder all nodes in the recommendation graph. This model is unsustainable given the extremely limited computing power of edge computing nodes in museums. The core technical problem this invention aims to solve, and the key difference between it and existing technologies, lies in both eliminating the interference of the inherent cognitive load of displayed objects on interest determination and strictly confining the computational scope of graph updates within a local topology that matches the current interaction.

[0030] Therefore, such as Figure 1 As shown, this application provides a museum visitor interest evolution recommendation method based on incremental thinking chain, including the following steps:

[0031] Obtain the actual interaction time of the target audience with the first displayed object, and obtain the basic cognitive time associated with the first displayed object;

[0032] Based on the difference ratio between the actual interaction time and the basic cognition time, a relative interest bias parameter representing the relative attention level of the target object is obtained.

[0033] When the relative interest bias parameter is greater than the preset bias threshold, the relative interest bias parameter is converted into a graph update weight parameter based on the preset weight conversion coefficient; otherwise, the process cache containing the actual interaction duration is cleared.

[0034] Based on the preset behavior association state graph, the first graph node associated with the first display object is extracted, and the current state attribute value of the first graph node is fused and updated based on the graph update weight parameter to obtain the updated state attribute value.

[0035] Based on the updated state attribute values ​​of the first graph node, the state attribute values ​​of the second graph nodes connected to it are updated to obtain a reconstructed behavior association state graph. Based on the reconstructed behavior association state graph, the second graph nodes whose state attribute values ​​meet the preset sorting conditions are extracted, and the second display objects associated with them are used as target recommendation objects.

[0036] In another implementation, such as Figure 2 As shown, this application also provides a museum visitor interest evolution recommendation system based on incremental thinking chain, including the following modules:

[0037] The data acquisition module is used to obtain the actual interaction time of the target audience with the first display object, and to obtain the basic cognitive time bound to the first display object;

[0038] The feature extraction module is used to obtain a relative interest bias parameter that represents the relative attention level of the target object based on the difference ratio between the actual interaction time and the basic cognition time.

[0039] The bias determination module is used to convert the relative interest bias parameter into a graph update weight parameter based on a preset weight conversion coefficient when the relative interest bias parameter is greater than a preset bias threshold; otherwise, it clears the process cache containing the actual interaction duration.

[0040] The state update module is used to extract the first graph node associated with the first display object based on the preset behavior association state graph, and to fuse and update the current state attribute value of the first graph node based on the graph update weight parameter to obtain the updated state attribute value.

[0041] The result output module is used to update the state attribute values ​​of the second graph nodes connected to the first graph node based on the updated state attribute values ​​of the first graph node to obtain a reconstructed behavior association state graph. Based on the reconstructed behavior association state graph, the module extracts the second graph nodes whose state attribute values ​​meet the preset sorting conditions and uses the second display object associated with them as the target recommendation object.

[0042] I. Acquisition of actual interaction time and basic cognitive time;

[0043] The inputs to this step are a continuous video data stream covering the observation area of ​​the first display object and unobstructed standard image data of the first display object collected during non-business hours. The outputs are the actual interaction time of the target audience with the first display object and the basic cognitive time associated with the first display object.

[0044] By pre-quantifying the visual information entropy of the displayed object as an independent feature, this invention enables the system to obtain an objective benchmark for measuring the inherent cognitive load of the displayed object while acquiring audience behavior data. This lays a data foundation for subsequently removing the interference of decoding time on interest determination. This design breaks away from the existing approach of mining interest features solely from the audience's own behavioral sequences, and instead incorporates the physical attributes of the displayed object itself into the computational basis of interest determination.

[0045] Regarding the acquisition of basic cognitive duration, edge cameras are used during non-business hours to collect unobstructed standard image data of the first display object under preset standard lighting conditions and at preset standard shooting distances. The data is then normalized to a preset standard resolution and processed into grayscale. The uniform lighting conditions, shooting distance, and resolution normalization process eliminate the interference of different on-site imaging conditions on subsequent feature extraction values, ensuring the comparability of basic cognitive duration between different display objects and different collection batches.

[0046] The Laplacian operator is applied to the grayscale image to calculate its second-order derivative of pixels, and the variance of the second-order derivative matrix is ​​calculated. This variance value is used as the texture complexity feature value to characterize the spatial structural complexity of the first display object. Furthermore, a preset text region detection algorithm is invoked to identify the bounding boxes of all connected text regions in the standard image data. The ratio of the sum of the pixel areas of each bounding box to the total pixel area of ​​the standard image data is calculated, and this ratio is used as the proportion of the text semantic domain to characterize the text reading load of the first display object.

[0047] Specifically, standard image data is first converted into a single-channel grayscale image. Then, the Laplacian operator is used to perform convolution filtering on the grayscale image to obtain a response matrix that characterizes the intensity of local grayscale abrupt changes in the image. Furthermore, the variance of all elements in the response matrix is ​​calculated and used as a texture complexity feature value. This index has been proven in the field of computer vision to stably characterize the spatial structural complexity of an image and is often used as the core basis for camera autofocus evaluation functions. Moreover, the calculation process is extremely lightweight and can be completed in real time on edge computing terminals without relying on any deep neural network model, thus possessing good engineering feasibility.

[0048] Regarding the acquisition of the text semantic domain proportion, a preset text region detection algorithm is invoked to identify the connected components of text in the standard image data one by one, and the bounding boxes of each connected component are merged and deduplicated to avoid the same text region being counted repeatedly due to broken character strokes, thereby ensuring the accuracy of the text semantic domain proportion calculation. The visual saccade conversion constant, text reading conversion constant, and basic dwell time constant are predetermined by statistical regression fitting on multiple display object samples with known real basic cognitive durations.

[0049] Specifically, the process involves: pre-collecting multiple sets of display object sample data with different texture complexity feature values ​​and text semantic domain proportions, and whose actual basic cognitive duration is known. Using the following cognitive baseline time mapping relationship as the regression equation, the least squares method is used to fit this set of sample data to obtain the combination of visual saccade conversion constant, text reading conversion constant, and basic dwell time constant values ​​that minimize the fitting error. These combinations are then used as preset constants to calculate the basic cognitive duration of all subsequent display objects. Combining the aforementioned predetermined preset constants, the basic cognitive duration of the first display object is calculated according to the following cognitive baseline time mapping relationship:

[0050] ;

[0051] in, This indicates the basic cognitive duration of the first displayed object, in seconds; The texture complexity feature value is represented by the variance of the second derivative of pixels after processing standard image data with the Laplacian operator, and is dimensionless. This represents the proportion of the semantic domain of the text, and its value is a dimensionless ratio between zero and one.

[0052] This represents a preset visual scanning conversion constant, used to map texture complexity feature values ​​to the corresponding visual scanning time. This represents a preset text reading conversion constant, used to map the proportion of the text's semantic domain to the corresponding semantic decoding time. This represents the preset basic dwell time constant, in seconds, which is used to characterize the minimum time required for an audience to complete a basic gait dwell in front of any displayed object.

[0053] In embodiments of the present invention, a set of visual saccade transition constants determined by the aforementioned statistical regression fitting method is provided. The value is 0.15, a text reading conversion constant. The value is 18, which is the basic residence time constant. The value is set to 3 seconds as a reference.

[0054] The basic cognitive duration is calculated only once for each displayed object during system initialization and bound to the corresponding displayed object identifier, stored in a static baseline mapping table for subsequent real-time retrieval. The construction of this cognitive baseline time draws on a physiological model of human visual search and fixation: visual complexity corresponds to the time required for a viewer to complete an overall scan of an image, while text area corresponds to the time required for a viewer to fixate on each word of the text. These two are converted to time units using preset conversion constants and then linearly superimposed, thereby transforming the originally abstract visual information entropy into a physically meaningful time benchmark. This provides an objective quantitative foundation for distinguishing between physiological decoding and subjective interest.

[0055] Regarding the acquisition of actual interaction time, the edge acquisition node acquires a continuous video data stream covering the physical exhibition area where the first displayed object is located. It extracts video frames one by one at a preset frame rate and calls a preset target detection model (such as a YOLO series lightweight network) to identify the target audience in the frame, outputting the corresponding two-dimensional human bounding box. The coordinates of the top-left and bottom-right corners of this bounding box are extracted, and the midpoint of its diagonal vertices is calculated. This midpoint is used as the real-time centroid coordinates of the target audience. The calculation relationship is as follows:

[0056] ;

[0057] ;

[0058] in, , The x and y coordinates represent the real-time centroid coordinates of the target audience. , This represents the coordinates of the top-left corner of the bounding box; , This represents the coordinates of the lower right corner of the bounding box.

[0059] The system pre-stores a preset observation area bound to the first display object. This area is a closed polygonal region composed of multiple vertices in the video frame coordinate system. For the centroid coordinates extracted frame by frame, the Ray-Casting Algorithm is used to determine in real time whether the centroid coordinates are located inside the closed polygonal region and outputs the corresponding instantaneous Boolean state of being inside or outside the domain.

[0060] When the centroid coordinates enter the preset observation area bound to the first display object, a start timestamp is recorded; when the centroid coordinates leave the preset observation area, a stop timestamp is recorded, and the actual interaction duration is obtained according to the following calculation formula:

[0061] ;

[0062] in, This indicates the actual interaction time of the target audience with the first displayed object, in seconds; Indicates the stop timestamp; This indicates the start timestamp.

[0063] To eliminate visual tracking jitter noise caused by crowd occlusion or brief offset, a state machine with temporal tolerance is maintained for the target audience: a valid entry event is triggered and the start timestamp is recorded only when the centroid coordinates enter the interior of the preset observation area from outside the area and remain in the state within the area for a preset number of consecutive frames.

[0064] A valid detachment event is triggered only when the centroid coordinates leave the area from the outside and do not return to the inside of the area within a preset tolerance time window, and the last time it was within the area is traced back to the stop timestamp. By introducing a dual de-jitter mechanism of continuous frame count confirmation and tolerance time window, this invention effectively avoids the problem of incorrect truncation or repeated counting of actual interaction time due to brief audience occlusion or boundary wandering, thus improving the robustness of the actual interaction time acquisition process.

[0065] II. Calculation of relative interest bias parameters;

[0066] The inputs to this step are the actual interaction time and basic cognitive time of the above outputs, and the output is a relative interest bias parameter that characterizes the relative attention level of the target audience to the first displayed object. By performing a normalized difference calculation on the absolute dwell time and the inherent cognitive load of the displayed object, this invention can mathematically separate the decoding time and interest time that were originally mixed together, so that the obtained bias parameter only reflects the intensity of the audience's subjective attention beyond basic cognitive needs.

[0067] The difference between the actual interaction time and the basic cognitive time is obtained, and the sum of the basic cognitive time and a preset smoothing constant is obtained. The difference is divided by the sum to obtain the relative deviation ratio. When the relative deviation ratio is less than or equal to zero, it is determined that the audience has not exceeded the basic decoding time of the displayed object and has not generated excessive interest, so the value of the relative interest bias parameter is set to zero. When the relative deviation ratio is greater than zero, the value of the relative deviation ratio is determined as the relative interest bias parameter, and its calculation relationship is as follows:

[0068] ;

[0069] in, This represents the relative interest bias parameter, which is dimensionless. Indicates the actual interaction duration; Indicates the duration of basic cognitive processing; This represents a preset smoothing constant, which is a positive number close to zero, used to prevent division by zero errors when the basic cognitive duration is zero.

[0070] Because the denominator uses the basic cognitive duration instead of a fixed constant, this relative interest bias parameter achieves scale uniformity between display objects with different information densities. This allows the equivalent excess dwell time generated in front of high-cognitive-load display objects and low-cognitive-load display objects to be converted into comparable relative attention intensities.

[0071] The difference in the intensity of subjective interest reflected by spending ten more seconds in front of an object that can be cognitively decoded in just ten seconds versus spending ten more seconds in front of an object that takes one minute to cognitively decode is clearly orders of magnitude different. After normalization of basic cognitive time, the former will obtain a much higher relative interest bias parameter than the latter, thus accurately reflecting this difference.

[0072] Furthermore, the maximum value function ensures that the parameter will not have a negative value, thus clearly distinguishing between the two distinct behaviors of viewers leaving before finishing watching and viewers staying after finishing watching. This avoids unnecessary interference from negative values ​​in subsequent graph calculations and provides input parameters with a clear range of values ​​and clear physical meaning for the weight conversion and threshold determination steps in the later stages.

[0073] III. Generation of map update weight parameters and clearing of process cache;

[0074] The inputs for this step are the relative interest bias parameters and preset bias thresholds from the above outputs, and the outputs are the map update weight parameters or a clear command.

[0075] By setting a bias threshold as a rigid judgment checkpoint, this invention cuts off the slight fluctuations caused by the information load of the displayed object itself at the source, allowing only truly statistically significant evidence of interest to enter the subsequent graph calculation stage, thereby protecting the stability of the behavioral association state graph and avoiding a large amount of low-confidence behavioral data from causing continuous invalid disturbances to the graph node state.

[0076] The relative interest bias parameter is compared with a preset bias threshold: when the relative interest bias parameter is greater than the preset bias threshold, a preset weight conversion coefficient is invoked to convert the relative interest bias parameter into a map update weight parameter, and an upper limit is set for the conversion result. The calculation relationship is as follows:

[0077]

[0078] in, This indicates that the graph updates the weight parameters; This represents the preset weight conversion coefficient, used to compress and map the dimensionless relative interest bias parameter to the legal value range allowed by the graph node state attribute values. This represents the relative interest bias parameter; This represents the preset weight upper limit, which is set to 1 in this embodiment. This is used to ensure that no matter how large the relative interest bias parameter is, the generated graph update weight parameter will not exceed the legal value range of the graph node state attribute value (i.e., probability value), thereby ensuring that the subsequent update posterior probability value is always between zero and one.

[0079] When the relative interest bias parameter is not greater than the preset bias threshold, it is determined that the current interaction is a normal behavior within the inherent decoding time range of the displayed object and has no statistical significance. Then, the cache of this process, including the actual interaction time, is cleared, the graph update weight parameter is not generated, and no subsequent graph operation is triggered.

[0080] This hard-blocking mechanism is equivalent to setting up a filter valve at the entry point of the graph operation: only when the audience's excessive dwell time reaches a statistically significant level is the behavioral data carried by this interaction allowed to be converted into graph evidence, while the interaction data that does not reach this significance level is regarded as background noise and directly discarded without leaving any trace in the system. This avoids the storage and computing overhead caused by repeatedly maintaining the process cache for a large number of invalid interactions, and also protects the behavioral association state graph from being contaminated by a large amount of low-confidence data from the root.

[0081] IV. Updating the status attribute values ​​of the first graph node;

[0082] The inputs to this step are the graph update weight parameters output above and the historical states associated with the first displayed object in the behavior-related state graph. The output is the updated state attribute value of the first graph node. By controlling the fusion of newly added interest evidence with historically accumulated state values, this invention enables the state of graph nodes to dynamically reflect the latest evolution trend of audience interest, while avoiding the complete overwriting of historical data by a single interaction, thereby achieving a reasonable balance between stability and sensitivity of node states.

[0083] In the preset behavior association state graph, find the network node corresponding to the entity identifier of the first displayed object and use it as the first graph node. Extract the initial prior probability value currently stored in the first graph node. By combining the initial prior probability value, the graph update weight parameters generated above, and the preset dynamic forgetting constant, obtain the updated posterior probability value of the first graph node according to the following calculation relationship, and use the updated posterior probability value to cover its initial prior probability value:

[0084] ;

[0085] in, This represents the updated state attribute value of the first graph node, i.e., the updated posterior probability value, with a value range of zero to one. This represents the state attribute value of the first graph node before it is updated, i.e., the initial prior probability value. This indicates that the graph updates the weight parameters;

[0086] In an embodiment of the present invention, This represents the preset dynamic forgetting constant, which ranges from zero to one. It is used to control the fusion ratio between historical states and new evidence. The larger the value, the more significant the immediate impact of the current interaction on the node state. The smaller the value, the higher the dependence of the node state on historical accumulation.

[0087] This momentum decay update method ensures that the state attribute values ​​of the first graph node will not change drastically due to a single interaction, causing numerical saturation, nor will it fail to reflect the latest evolution trend of audience interest due to the rigidity of historical data. This is more in line with the objective law that audience interest dynamically drifts with the progress of the tour in the museum setting.

[0088] It should be noted that since the updated posterior probability value is obtained by weighted fusion of the initial prior probability value and the map update weight parameter, when the map update weight parameter is less than the initial prior probability value, the fusion result will be lower than the initial prior probability value before the update, that is, the difference in probability values ​​may be negative.

[0089] This situation indicates that although the relative interest bias parameter corresponding to this interaction has reached a statistically significant level and is sufficient to trigger this graph update process, the graph update weight parameter obtained by its conversion is itself lower than the historical accumulated state level of the graph node. Therefore, the fused state value shows a slight pullback. This pullback is a true reflection of the relative strength of the audience's current relative attention and the historical state of the node, rather than a denial of the interest attribute of this interaction, and does not affect the record of the physical meaning of the positive and negative signs of the probability value difference in this invention.

[0090] This invention sets an upper limit on the value of the map update weight parameter. This ensures that regardless of the relative magnitude of the map update weight parameter and the initial prior probability value, the updated posterior probability value is strictly within the legal range of zero to one, and will not exceed the legal range of probability values.

[0091] V. Local Reconstruction of the Behavioral Association State Graph;

[0092] The input for this step is the state attribute values ​​of the first graph node before and after the update, as well as the topological relationships connected to the first graph node in the behavior association state graph. The output is the reconstructed behavior association state graph.

[0093] By strictly limiting the transmission of node state changes along directed connections to the range of first-order neighbors, this invention completely abandons the traditional recommendation algorithm's operation mode of updating one place and recalculating the entire graph. This allows the system to complete the local evolution of the graph with only a minimal amount of computing power matching the current interaction. This design is particularly crucial for museum edge computing nodes with extremely limited computing resources.

[0094] The difference between the updated posterior probability value and the initial prior probability value corresponding to the first graph node is obtained, and the calculation relationship is as follows:

[0095] ;

[0096] in, It represents the probability difference, and its positive or negative sign has a definite physical meaning: when it is positive, it indicates that the first displayed object has aroused positive interest in the audience, and when it is negative, it indicates that the first displayed object has aroused negative interest in the audience.

[0097] In the preset behavior association state graph, each adjacent node directly pointed to by the first graph node through a one-way connection edge is extracted and used as the corresponding second graph node. Based on the preset transmission weight of each one-way connection edge, the probability value difference is weighted and calculated to generate the transmission difference of each second graph node.

[0098] The preset transfer weight is configured in advance based on prior statistical information such as the thematic relevance and spatial proximity between the displayed objects. It is used to characterize the degree of connection between two displayed objects. The larger the weight, the closer the connection between the first displayed object and the corresponding second displayed object, and the easier it is for the audience's interest to shift.

[0099] Based on a preset mapping function, the relative interest bias parameter is converted into a corresponding dynamic damping coefficient. The mapping function is configured such that the larger the relative interest bias parameter is, the closer the generated dynamic damping coefficient is to a preset minimum threshold.

[0100] Specifically, the preset mapping function described in this embodiment adopts In the form of, The preset attenuation coefficient is used to control the rate at which the dynamic damping coefficient decreases as the relative interest bias parameter increases; when the value calculated according to the preset mapping function is not greater than a preset minimum threshold... ( When, the preset minimum threshold is used. The value of the dynamic damping coefficient is used to avoid the target connection weight from being divided into zero or increasing infinitely when the dynamic damping coefficient approaches zero.

[0101] Divide the preset transfer weight of each unidirectional connection edge by the dynamic damping coefficient to obtain the target connection weight. Then, multiply the target connection weight by the difference between the probability values ​​as the transfer difference of the corresponding second graph node. The calculation relationship is as follows:

[0102] ;

[0103] ;

[0104] in, Indicates the target connection weight; This indicates the preset transfer weight of the corresponding unidirectional connection edge; This indicates that the dynamic damping coefficient has a value greater than a preset minimum threshold. And less than or equal to 1; This represents the transmission difference corresponding to the second graph node, and its sign is the difference between the probability value and the value. The signs of the positive and negative signs must remain consistent; This represents the difference in probability values.

[0105] When the absolute value of the transmission difference is less than or equal to the preset transmission truncation threshold, the initial prior probability value currently stored in the corresponding second graph node remains unchanged, and no superposition operation is performed; when the absolute value of the transmission difference is greater than the preset transmission truncation threshold, the transmission difference is superimposed on the initial prior probability value currently stored in the corresponding second graph node according to its positive or negative sign to obtain the updated posterior probability value of the second graph node, and its initial prior probability value is overwritten, thereby obtaining the reconstructed behavior association state graph.

[0106] Therefore, regardless of whether the interaction generates positive interest stimulation or negative interest suppression for the first graph node, as long as the absolute value of its corresponding transmission difference exceeds the preset transmission cutoff threshold, it can be propagated truthfully along the unidirectional connection edge to the adjacent second graph node, avoiding the situation where negative suppression evidence is blocked by the unilateral threshold judgment mechanism and cannot be propagated.

[0107] Specifically, when the absolute value of the transmission difference is determined to be greater than a preset transmission truncation threshold, the following restricted algebraic sum operation is performed to obtain the updated posterior probability value of the second graph node, and this value is used to overwrite its initial prior probability value:

[0108] ;

[0109] in, This represents the updated state attribute value of the second graph node, i.e., the updated posterior probability value of the second graph node. This represents the value of the state attribute currently stored before the second graph node is updated, that is, the initial prior probability value of the second graph node;

[0110] This indicates the corresponding difference, whose positive or negative sign is fully preserved and substituted into the calculation in this superposition operation. Regardless of whether its value is positive or negative, it is directly substituted into the calculation according to the aforementioned restricted algebraic sum and relation, without any additional processing. After completing the above superposition operation, the obtained... The value overwrites the original value stored in the second graph node. The numerical value enables the state attribute value of the second graph node in the behavior association state graph to be updated in real time.

[0111] Once all second graph nodes in the preset behavior association state graph that satisfy the condition that the absolute value of the transmission difference is greater than the preset transmission truncation threshold have completed the above-mentioned coverage operation, the entire graph will change from the topological state before the update to the topological state after the update, thereby obtaining the reconstructed behavior association state graph.

[0112] VI. Extraction of target recommendation objects;

[0113] The input to this step is the state attribute value of each second graph node in the reconstructed behavior association state graph obtained above, and the output is the target recommendation object. By uniformly sorting the state attribute values ​​of candidate nodes, this invention converges the final result of the graph operation into a clear recommendation decision, completing a complete closed loop from the collection of original behavior data to the output of the final recommendation result. This ensures that each recommendation result can be traced back to a specific behavior trigger node, possessing clear causal interpretability.

[0114] In the reconstructed behavior association state graph, the state attribute values ​​of each second graph node are sorted in descending order to generate a corresponding candidate queue. The second graph node that is ranked first in the candidate queue is extracted. The display object corresponding to the second graph node is queried and extracted in the preset mapping relationship table and confirmed as the target recommendation object. A recommendation instruction containing the target recommendation object is sent to the mobile terminal corresponding to the target audience.

[0115] The preset mapping table stores the correspondence between the map node identifier and the physical attributes of the physical display object in the form of a hash table, including the display object name, physical coordinates within the venue, and multimedia guide resource index. By performing a hash query on the target second map node identifier, the mapping extraction from the abstract map node to the specific display object can be completed within a constant time overhead, ensuring the real-time performance of the target recommended object extraction process.

[0116] Furthermore, to support the stable operation of the above steps in the context of high visitor traffic in museums, and to ensure that the various calculations carried out can be executed smoothly under limited edge computing power, the physical deployment of this invention includes two parts: edge acquisition nodes deployed in each exhibition area and a cloud-based graph computing cluster.

[0117] Edge acquisition nodes complete the collection of actual interaction time, comparison of basic cognition time, and calculation of relative interest bias parameters nearby. Only when the relative interest bias parameter is determined to be greater than the preset bias threshold, the map update weight parameter is encapsulated into an update instruction and uploaded to the cloud map computing cluster, thereby cutting off the communication bandwidth occupied by the original video data transmission to the cloud.

[0118] After receiving an update command, the cloud-based graph computing cluster reads the current concurrent traffic parameters and the load parameters of the graph computing nodes. When the load parameters exceed the preset warning load threshold, it prioritizes allocating independent computing resources to the update command to ensure that the status attribute values ​​of the first and second graph nodes can be updated within a limited time. When the load parameters do not exceed the threshold, the update command is queued and processed in the normal order.

[0119] The behavior association state graph is organized in the form of an adjacency list in terms of storage structure. Each graph node stores its unique identifier, current state attribute value, and unidirectional connection edges pointing to each second graph node and the corresponding preset transfer weights. This storage structure enables the extraction of second graph nodes to be completed with constant-level lookup overhead, further ensuring the real-time performance of local graph reconstruction.

[0120] Upon receiving the recommendation instruction, the mobile terminal corresponding to the target audience unpacks the instruction and renders the name, thumbnail, and navigation path of the target recommended object for the audience's reference. This allows the audience to intuitively understand the evolutionary process of being recommended a second object after focusing on the first one. Through collaborative scheduling between the edge and cloud, this invention ensures real-time map reconstruction while keeping the computational overhead under high-concurrency scenarios within a reasonable range.

[0121] Through the above technical solutions, the present invention can achieve the following technical effects:

[0122] I. In view of the shortcomings of existing technologies that extract the inherent visual decoding time-consuming errors of the displayed object as interest surge features, this invention constructs a cognitive baseline time by introducing the visual information entropy of the displayed object itself, and uses the relative overflow ratio of the actual interaction time exceeding the baseline as the basis for interest determination. This physically removes the undesirable influence of cognitive load on audience behavior, significantly reducing the proportion of pseudo-interest features. It can significantly pre-filter pseudo-interest triggered by reading long texts and viewing complex patterns, and this filtering process can be completed in real time on edge computing terminals without relying on any deep learning model.

[0123] Second, in view of the shortcomings of existing technologies that waste computing power due to the triggering of full graph re-inference by a single behavior update, the present invention limits the transmission of changes in state attribute values ​​to the range of second graph nodes and blocks them by combining transmission cutoff thresholds, so that the peak concurrent processing of edge computing nodes is significantly reduced. Without increasing computing power resources, the efficient local evolution of the recommendation graph is achieved, which significantly alleviates the resource pressure on museum edge computing infrastructure under high visitor flow conditions.

[0124] Third, in response to the shortcomings of existing technologies, such as the deviation between the reasoning chain and the audience's actual psychological state and the frequent logical jumps, this invention couples the strength of the audience's interest with the spread of the graph's influence through a dynamic damping coefficient. This makes the causal path followed by the recommendation results highly consistent with the audience's actual relative attention level, eliminating cognitive jumps and logical biases in the incremental recommendation process and improving the interpretability and credibility of the recommendation results for the audience.

[0125] In another embodiment, this application also provides a computer storage medium storing computer-executable instructions that, when executed, implement the aforementioned museum visitor interest evolution recommendation method based on incremental thought chain.

[0126] 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 museum visitor interest evolution recommendation method based on incremental thinking chain, characterized in that, Includes the following steps: Obtain the actual interaction time of the target audience with the first displayed object, and obtain the basic cognitive time associated with the first displayed object; Based on the difference ratio between the actual interaction time and the basic cognition time, a relative interest bias parameter representing the relative attention level of the target object is obtained. When the relative interest bias parameter is greater than the preset bias threshold, the relative interest bias parameter is converted into a graph update weight parameter based on the preset weight conversion coefficient; otherwise, the process cache containing the actual interaction duration is cleared. Based on the preset behavior association state graph, the first graph node associated with the first display object is extracted, and the current state attribute value of the first graph node is fused and updated based on the graph update weight parameter to obtain the updated state attribute value. Based on the updated state attribute values ​​of the first graph node, the state attribute values ​​of the second graph nodes connected to it are updated to obtain a reconstructed behavior association state graph. Based on the reconstructed behavior association state graph, the second graph nodes whose state attribute values ​​meet the preset sorting conditions are extracted, and the second display objects associated with them are used as target recommendation objects.

2. The museum visitor interest evolution recommendation method based on incremental thinking chain according to claim 1, characterized in that, The process of acquiring basic cognitive time includes: Obtain the standard image data of the first display object in an unobstructed state, and extract the variance of the second-order derivative of the pixels in the standard image data as the texture complexity. ; Based on a preset text region detection algorithm, the ratio of the total pixel area of ​​the bounding boxes of all connected text regions in the standard image data to the total pixel area of ​​the standard image data is extracted as the proportion of the text semantic domain. ; Combined with preset visual saccade conversion constants Text reading conversion constant Basic residence time constant Obtain the basic cognitive duration of the first displayed object. ; 。 3. The museum visitor interest evolution recommendation method based on incremental thinking chain according to claim 2, characterized in that, The process of obtaining the relative interest bias parameters includes: Obtain the actual interaction duration With the aforementioned basic cognitive duration The difference is used to obtain the basic cognitive duration. With preset smoothing constant The sum of; When the ratio of the difference to the sum is less than or equal to zero, the relative interest bias parameter is... The value is set to zero. When the ratio result is greater than zero, the value of the ratio result is determined as the relative interest bias parameter. ; 。 4. The museum visitor interest evolution recommendation method based on incremental thinking chain according to claim 1, characterized in that, The process of updating the state attribute value of the first graph node includes: When the relative interest bias parameter is greater than a preset bias threshold, the preset weight is converted into coefficients. With the relative interest bias parameter The product and the preset weight limit The smaller of the two values ​​is used as the graph update weight parameter for the first graph node. ; ； In the preset behavior association state graph, find the network node corresponding to the entity identifier of the first displayed object, and use it as the first graph node. Extract the initial prior probability value currently stored in the first graph node. ; Based on the initial prior probability value, the map update weight parameters, and the preset dynamic forgetting constant. Obtain the updated posterior probability value of the first graph node. And use the updated posterior probability value to cover its initial prior probability value; 。 5. The museum visitor interest evolution recommendation method based on incremental thinking chain according to claim 4, characterized in that, The process of obtaining the reconstructed behavioral association state graph includes: The probability difference between the updated posterior probability value and the initial prior probability value corresponding to the first graph node is obtained. In the preset behavior association state graph, each adjacent node directly pointed to by the first graph node through a one-way connection edge is extracted as the corresponding second graph node. Based on the preset transmission weights of each of the unidirectional connection edges, the probability value differences are weighted and calculated to generate the transmission difference of each of the second graph nodes. The transmission difference is then superimposed on the initial prior probability value currently stored in each of the second graph nodes to obtain the reconstructed behavior association state graph.

6. The museum visitor interest evolution recommendation method based on incremental thinking chain according to claim 5, characterized in that, The process of generating the propagation difference of the second graph node includes: The relative interest bias parameter is converted into a corresponding dynamic damping coefficient based on a preset mapping function, and the preset transmission weight of each unidirectional connection edge is divided by the dynamic damping coefficient to obtain the target connection weight. The target connection weight is used as the weight for weighting the probability value difference to obtain the transmission difference of the corresponding second graph node. When the absolute value of the transmission difference is less than or equal to the preset transmission truncation threshold, the initial prior probability value currently stored in the corresponding second graph node remains unchanged. When the absolute value of the transmission difference is greater than the preset transmission truncation threshold, the transmission difference is superimposed on the initial prior probability value currently stored in the corresponding second graph node according to its positive or negative sign to obtain the updated posterior probability value of the second graph node, and its initial prior probability value is overwritten.

7. The museum visitor interest evolution recommendation method based on incremental thinking chain according to claim 1, characterized in that, The process of obtaining target recommendations includes: In the reconstructed behavior association state graph, each second graph node is sorted in descending order according to the state attribute values ​​to generate a corresponding candidate queue. The second graph node that is ranked first in the candidate queue is extracted, and the display object corresponding to the second graph node is queried and extracted in the preset mapping relationship table and used as the target recommendation object.

8. The museum visitor interest evolution recommendation method based on incremental thinking chain according to claim 1, characterized in that, The process of obtaining the actual interaction duration includes: Acquire a continuous video data stream containing the target audience, extract the centroid coordinates of the target audience in the continuous video data stream, record a start timestamp when the centroid coordinates enter the preset observation area bound to the first display object, and record a stop timestamp when the centroid coordinates leave the preset observation area, and use the time difference between the stop timestamp and the start timestamp as the actual interaction duration.

9. A museum visitor interest evolution recommendation system based on incremental thinking chain, characterized in that, Includes the following modules: The data acquisition module is used to obtain the actual interaction time of the target audience with the first display object, and to obtain the basic cognitive time bound to the first display object; The feature extraction module is used to obtain a relative interest bias parameter that represents the relative attention level of the target object based on the difference ratio between the actual interaction time and the basic cognition time. The bias determination module is used to convert the relative interest bias parameter into a graph update weight parameter based on a preset weight conversion coefficient when the relative interest bias parameter is greater than a preset bias threshold; otherwise, it clears the process cache containing the actual interaction duration. The state update module is used to extract the first graph node associated with the first display object based on the preset behavior association state graph, and to fuse and update the current state attribute value of the first graph node based on the graph update weight parameter to obtain the updated state attribute value. The result output module is used to update the state attribute values ​​of the second graph nodes connected to the first graph node based on the updated state attribute values ​​of the first graph node to obtain a reconstructed behavior association state graph. Based on the reconstructed behavior association state graph, the module extracts the second graph nodes whose state attribute values ​​meet the preset sorting conditions and uses the second display object associated with them as the target recommendation object.

10. A computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, they implement the museum visitor interest evolution recommendation method based on incremental thinking chain as described in any one of claims 1-8.

Images