Optimization Method for First-Promoted Items of Exhibits Combining User Portrait and User Location

Abstract

This application relates to the technical field of exhibit recommendation. Specifically, it discloses an optimized method for the first recommended item of exhibits by combining user portraits and user positioning. It extracts user positioning data from the indoor positioning system, determines the target exhibit group based on the user's real-time location and orientation information, then introduces deep learning technology to perform semantic parsing on the user portrait data to extract user portrait features, and through semantic matching analysis of the user portrait features with the feature labels of each exhibit in the target exhibit group, determines the first recommended item of exhibits from the target exhibit group. In this way, it can comprehensively consider the user's behavioral preferences and real-time physical environment information, provide more accurate and personalized exhibit recommendation services for users, thereby improving the visiting experience of visitors and optimizing the resource utilization efficiency of cultural exhibition venues such as museums and exhibition halls.

Description

Technical Field

[0001] This application relates to the technical field of exhibit recommendation, and more specifically, to an optimization method for the first recommended item of exhibits that combines user portraits and user positioning. Background Art

[0002] In cultural exhibition venues such as museums and exhibition halls, how to provide accurate and personalized exhibit recommendation services for visitors has always been a key issue in enhancing the visiting experience and optimizing resource utilization. Traditional exhibit recommendation methods are mostly based on the popularity of exhibits, historical visit data, or fixed exhibition route planning. This "one-size-fits-all" recommendation model fails to fully consider the unique interest preferences and real-time visiting status of each visitor.

[0003] With the continuous development of technology, user portrait technology has gradually matured. It can construct a detailed user interest model by collecting and analyzing multi-dimensional data of users, such as age, gender, occupation, historical browsing records, search preferences, etc., so as to accurately describe the characteristics and needs of users. However, currently in the field of exhibit recommendation, the recommendation algorithm based on collaborative filtering focuses on mining the common behaviors of user groups, but it is difficult to handle the immediate needs of individuals in specific scenarios; the recommendation method based on content similarity can match according to the attribute information of exhibits, but it ignores the influence mechanism of the user's current physical environment on their attention.

[0004] Therefore, an optimization method for the first recommended item of exhibits that combines user portraits and user positioning is expected. Summary of the Invention

[0005] To solve the above technical problems, this application is proposed. An embodiment of this application provides an optimization method for the first recommended item of exhibits that combines user portraits and user positioning. It extracts user positioning data from an indoor positioning system, determines a target exhibit group based on the user's real-time location and orientation information, then introduces deep learning technology to perform semantic parsing on the user portrait data to extract user portrait features, and determines the first recommended item of exhibits from the target exhibit group by performing semantic matching analysis between the user portrait features and the feature tags of each exhibit in the target exhibit group. In this way, it can comprehensively consider the user's behavioral preferences and real-time physical environment information, provide more accurate and personalized exhibit recommendation services for users, thereby improving the visiting experience of visitors and optimizing the resource utilization efficiency of cultural exhibition venues such as museums and exhibition halls.

[0006] Correspondingly, according to one aspect of this application, there is provided an optimization method for the first recommended item of exhibits that combines user portraits and user positioning, which includes:

[0007] Extract user positioning data from an indoor positioning system, where the user positioning data includes the user's real-time location and the user's orientation information;

[0008] Based on the user location data, determine a target exhibit group;

[0009] Extract the feature tags of each exhibit from the target exhibit group to obtain a set of exhibit label feature vectors;

[0010] Obtain user portrait data and encode the user portrait data to obtain a user portrait feature vector;

[0011] Based on the semantic matching analysis between the user portrait feature vector and the set of exhibit label feature vectors, determine the first recommended exhibits from the target exhibit group.

[0012] Compared with the prior art, the method for optimizing the first recommended exhibits by combining user portraits and user locations provided by this application extracts user location data from an indoor positioning system, determines a target exhibit group based on the real-time location and orientation information of the user, and then introduces deep learning technology to semantically analyze the user portrait data to extract user portrait features. By performing semantic matching analysis on the user portrait features and the feature tags of each exhibit in the target exhibit group, the first recommended exhibits are determined from the target exhibit group. In this way, it is possible to comprehensively consider the user's behavioral preferences and real-time physical environment information, provide a more accurate and personalized exhibit recommendation service for the user, thereby improving the visitor experience and optimizing the resource utilization efficiency of cultural exhibition venues such as museums and exhibition halls. Description of the Drawings

[0013] By describing the embodiments of the present application in more detail with reference to the accompanying drawings, the above and other objects, features, and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. They are used together with the embodiments of the present application to explain the present application and do not constitute a limitation to the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0014] Figure 1 It is a flowchart of a method for optimizing the first recommended exhibits by combining user portraits and user locations according to an embodiment of the present application.

[0015] Figure 2 It is a schematic diagram of data flow of a method for optimizing the first recommended exhibits by combining user portraits and user locations according to an embodiment of the present application.

[0016] Figure 3 It is a flowchart of step S5 in a method for optimizing the first recommended exhibits by combining user portraits and user locations according to Embodiment 2 of the present application.

[0017] Figure 4It is a flowchart of step S51 in the optimized method for the first recommended item of exhibits combining user portraits and user positioning according to Embodiment 2 of the present application. Detailed implementation manners

[0018] Next, exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all embodiments of the present application. It should be understood that the present application is not limited by the exemplary embodiments described herein.

[0019] Embodiment 1

[0020] Figure 1 It is a flowchart of the optimized method for the first recommended item of exhibits combining user portraits and user positioning according to the embodiment of the present application. Figure 2 It is a schematic diagram of data flow of the optimized method for the first recommended item of exhibits combining user portraits and user positioning according to the embodiment of the present application. As Figure 1 and Figure 2 shown, the optimized method for the first recommended item of exhibits combining user portraits and user positioning according to the embodiment of the present application includes the steps: S1, extracting user positioning data from an indoor positioning system, where the user positioning data includes user real-time location and user orientation information; S2, determining a target exhibit group based on the user positioning data; S3, extracting feature tags of each exhibit from the target exhibit group to obtain a set of exhibit label feature vectors; S4, obtaining user portrait data and encoding the user portrait data to obtain a user portrait feature vector; S5, determining the first recommended item of exhibits from the target exhibit group based on semantic matching analysis between the user portrait feature vector and the set of exhibit label feature vectors.

[0021] In the above optimized method for the first recommended item of exhibits combining user portraits and user positioning, in step S1, user positioning data is extracted from an indoor positioning system, and the user positioning data includes user real-time location and user orientation information. It should be understood that traditional exhibit recommendation means often ignore the immediate needs of individuals in specific visiting scenarios - in a museum, the focus of the same visitor may vary greatly at different positions. Given that in a large exhibition hall, the position and orientation of a user can intuitively reflect their immediate focus and movement trajectory, which is crucial for achieving accurate recommendations. Therefore, the present application uses an indoor positioning system to obtain user real-time location and orientation information, so that subsequent exhibit recommendations can closely fit the actual visiting scenario of the user and provide more immediate and personalized recommendation services.

[0022] Specifically, indoor positioning technology aims to determine the precise location of users in indoor spaces and the orientation of devices through various means, which can be roughly classified into several categories such as signal strength-based, time-based, angle-based, and hybrid technology-based. Currently, signal strength-based positioning technology is a widely used method. Among them, Bluetooth Low Energy (BLE) positioning technology occupies an important position in indoor positioning systems. In specific implementation, multiple Bluetooth beacons are pre-deployed in the venue, and these beacons periodically broadcast signals containing information such as their own identifiers. The Bluetooth module built into the mobile devices (such as smartphones) carried by users can receive these signals and estimate the distance from the beacons based on the Received Signal Strength Indication (RSSI). Generally speaking, there is a certain functional relationship between signal strength and distance. By measuring the signal strength at different positions in advance and establishing a mapping model between signal strength and distance, the distance between the device and the beacon can be deduced according to the received signal strength. For example, in a museum exhibition hall, a Bluetooth beacon is deployed every certain distance (such as 5 meters). When a user enters the exhibition hall, their mobile phone receives the signals from nearby beacons. By measuring the signal strength and referring to the pre-established model, the distance between the mobile phone and each beacon can be calculated. To determine the real-time position of the user, at least three beacons with known positions are required. Through the principle of triangulation, three distance values are used to determine the coordinate position of the user device on a two-dimensional plane. In practical applications, to improve the positioning accuracy, more beacons are often deployed, and optimization algorithms such as the weighted centroid algorithm are used to comprehensively consider the weights of the signal strengths of multiple beacons to calculate the user position more accurately.

[0023] In addition to Bluetooth, Wi-Fi positioning technology is also a common signal strength-based positioning method. Wi-Fi access points (APs) widely existing in indoor environments can be used as positioning reference points. User devices will scan the visible Wi-Fi access points around and obtain information such as the signal strength of each access point. Similar to Bluetooth positioning, by pre-collecting the signal strength data of each Wi-Fi access point at different positions, a fingerprint database is constructed. When the user device moves indoors, the real-time scanned Wi-Fi signal strength information is matched with the data in the fingerprint database, and matching algorithms such as the nearest neighbor algorithm and probability algorithm are used to find the fingerprint data closest to the current signal strength pattern, thereby determining the user's position. In a large exhibition center, hundreds of Wi-Fi access points may be deployed. User devices can scan the signals of multiple access points. By comparing with a large amount of pre-collected data in the fingerprint database, even in a complex indoor environment, the real-time position of the user can be determined relatively accurately.

[0024] Time-based positioning technologies mainly include two methods: Time of Arrival (TOA) and Time Difference of Arrival (TDOA). The principle of TOA positioning is to measure the propagation time of a signal from the transmitter (such as a base station) to the receiver (user equipment), and combine it with the signal propagation speed (usually the speed of light or sound) to calculate the distance between the two. In an indoor environment, if ultrasonic signals are used, multiple ultrasonic transmitting base stations need to be deployed indoors, and the user equipment is equipped with an ultrasonic receiver. The base station emits ultrasonic signals, and after the user equipment receives the signal, it records the reception time. By calculating the difference between the reception time and the transmission time of the base station, the signal propagation time can be obtained, and then the distance can be calculated. However, this method requires extremely high time synchronization accuracy, and a small time error will lead to a large distance calculation error. TDOA positioning is relatively more practical. It determines the position of the signal source (user equipment) by measuring the time difference of the signal arriving at two or more receivers. For example, when deploying multiple receiving base stations indoors, when the user equipment emits a signal, there are differences in the time when different base stations receive the signal. Based on these time differences and the known position relationship between the base stations, the position of the user equipment can be calculated using the hyperbola positioning principle. This method reduces the requirement for time synchronization accuracy to a certain extent and can achieve high-precision positioning with fewer base stations.

[0025] Angle-based positioning technology is represented by Angle of Arrival (AOA) positioning. This technology determines the position of the signal source by measuring the angle at which the signal arrives at the receiver. In indoor positioning implementation, receiving devices with array antennas, such as smart antenna arrays, are usually installed at fixed positions. When the user equipment emits a signal, the array antenna of the receiving device can measure the phase difference or signal strength difference of the signal on different antenna elements, and use signal processing algorithms to calculate the angle of arrival of the signal. For example, when installing multiple smart antenna arrays on the ceiling of an exhibition hall, when a user carrying a signal-emitting device moves in the exhibition hall, the antenna array can measure the angle of arrival of the signal in real time. Combining the known position and direction information of the antenna array, the position of the user equipment can be calculated using the triangulation principle. At the same time, by continuously tracking the change of the angle of arrival of the signal, the orientation information of the user equipment can also be obtained. For example, if the user equipment continuously emits a signal, as the user moves and turns, the angle of arrival of the signal measured by the antenna array will change accordingly. By analyzing the trend and amplitude of these angle changes, the orientation of the user equipment can be inferred.

[0026] Hybrid positioning technology combines the advantages of multiple positioning technologies to improve the accuracy and reliability of positioning. Common hybrid positioning methods include combining Bluetooth with Wi-Fi positioning technology, or combining signal strength-based positioning technology with inertial navigation technology. Taking the Bluetooth and Wi-Fi hybrid positioning as an example, in an indoor environment, first, Bluetooth beacons are used for preliminary positioning to determine the general area where the user is located, and then Wi-Fi positioning technology is used to perform more accurate position calculations within this area. Since Bluetooth positioning has high accuracy in the short-distance range, while Wi-Fi positioning can cover a larger area and has better stability in complex environments, the combination of the two can complement each other's advantages. In actual implementation, the user device simultaneously receives the signals of Bluetooth beacons and Wi-Fi access points, and the positioning system flexibly switches between using Bluetooth and Wi-Fi positioning algorithms to calculate the user's position according to the accuracy requirements at different stages. Another example is the combination of signal strength-based positioning technology and inertial navigation technology. Inertial navigation technology uses sensors such as accelerometers and gyroscopes built into the user device. By integrating the acceleration and angular velocity during the movement of the device, the displacement and direction changes of the device are estimated. During indoor positioning, when the signal strength-based positioning technology (such as Bluetooth or Wi-Fi positioning) experiences a decrease in positioning accuracy or signal loss due to signal occlusion or other reasons, inertial navigation technology can be used as a supplement to continue tracking the position and orientation changes of the user device based on sensor data until the signal returns to normal or the positioning system switches to other more reliable positioning methods.

[0027] In terms of obtaining the orientation information of the user device, in addition to the relevant methods in angle-based positioning technology, the sensors built into the device also play an important role. Mobile devices such as smartphones are usually equipped with gyroscopes and accelerometers. The gyroscope can measure the rotational angular velocity of the device around each axis. By integrating over time, the change in the rotation angle of the device can be calculated, thereby determining the change in the device's orientation. The accelerometer can measure the acceleration of the device in each direction. By analyzing the acceleration data, the posture of the device can be judged, such as whether the device is placed horizontally or tilted. Combining the data of the gyroscope can more accurately determine the absolute orientation of the device. For example, when a user holds a mobile phone to visit exhibits in a museum hall, the gyroscope and accelerometer built into the mobile phone collect data in real time. Through a specific sensor fusion algorithm, the rotation angle data measured by the gyroscope and the posture data measured by the accelerometer are comprehensively processed, and the orientation direction of the mobile phone (i.e., the user) can be accurately determined, providing key information for subsequent recommendation of relevant exhibits based on the user's orientation.

[0028] In the above-mentioned method for optimizing the first recommended item of exhibits by combining user portraits and user positioning, in step S2, based on the user positioning data, a target exhibit group is determined. It should be understood that since there are a large number of exhibits in the exhibition hall space and the user's attention resources are limited, it is impossible to pay attention to all exhibits at the same time. Usually, users are only interested in the exhibits within a certain range around them and within their line of sight. Therefore, in this application, the target exhibit group is further clarified according to the user's positioning data to narrow the recommendation scope, so as to concentrate resources on analyzing the valuable exhibits around the user, thereby improving the recommendation efficiency and accuracy. Specifically, first, according to the user's real-time position, a reasonable distance range is set centered on this position (for example, a circular area with a radius of 5 meters is set). Within this area, all exhibits within this area are determined by querying the exhibit database. Then, further screening is carried out in combination with the user's orientation information. Assuming that the user's orientation angle is θ, a fan-shaped area centered on the user's orientation is set (for example, the fan angle is 120°). By retaining the exhibits within this fan-shaped area and filtering out the exhibits that the user is facing away from, the target exhibit group is further streamlined to improve the efficiency and pertinence of exhibit recommendation.

[0029] In the above-mentioned method for optimizing the first recommended item of exhibits by combining user portraits and user positioning, in step S3, the feature labels of each exhibit are extracted from the target exhibit group to obtain a set of exhibit label feature vectors. It should be understood that in order to achieve precise matching recommendations for each exhibit in the target exhibit group, it is necessary to deeply understand the characteristics of each exhibit in the target exhibit group. Therefore, in this application, for each exhibit in the target exhibit group, multiple feature labels such as its title, description, theme classification, exhibit type, relevant artist or creation information are extracted, and the word vector technology in natural language processing (such as the Bert model) is used to perform semantic embedding encoding on the feature label information of each exhibit to realize the vector representation of the attribute information of each exhibit, thereby obtaining a set of exhibit label feature vectors. Those of ordinary skill in the art should know that the word embedding encoding technology can reflect the semantic similarity between texts based on the distance metric in the vector space by converting text data into numerical representations in the vector space, thus providing a data basis for subsequent user portrait matching analysis.

[0030] In the above method for optimizing the first-recommended items of exhibits by combining user portraits and user positioning, in step S4, user portrait data is obtained and the user portrait data is encoded to obtain a user portrait feature vector. Specifically, the sources of user portrait data are extensive, covering multiple dimensions such as the basic attribute information, behavior data, and preference data of users. Among them, the basic attribute information includes the age, gender, occupation, region, etc. of the user, and this information can be collected through the required items in the user registration process. In the exhibition venue or related applications, when the user registers an account, the system will require the user to fill in this basic information. For example, in the registration interface of the official APP of the museum, there are set required age ranges, gender selection boxes, and occupation drop-down menus, etc. After the user fills in, this data will be directly stored in the user information database, providing basic information for subsequent construction of the user portrait.

[0031] Behavior data is data that reflects the actual behavior trajectories and operation records of users and is crucial for accurately depicting user portraits. In the exhibition venue scenario, behaviors such as the user's visiting path in the venue, staying time, and the number of times of viewing different exhibits can be collected. Through the sensor network deployed in the exhibition hall, such as the indoor positioning system mentioned above, the position information of the user can be tracked in real time, and then the user's visiting path can be analyzed. For example, in an art exhibition venue, by using Bluetooth positioning technology combined with video surveillance analysis, the staying time of the user in each exhibition hall and the staying duration in front of each piece of art can be recorded. At the same time, the operation behaviors of the user on the relevant application, such as search records, browsing history, likes, comments, etc. are also recorded in detail. Taking the museum APP as an example, the system will automatically record the keywords of the exhibits searched by the user, the exhibition introduction pages browsed, and the like and comment contents on specific exhibits. These behavior data can intuitively reflect the user's interest points and attention areas, providing important clues for in-depth understanding of user needs.

[0032] The acquisition of preference data focuses on exploring the user's deep-seated interest preferences and value orientations. For example, the museum can push questionnaires through the APP to ask the user about their preference levels for art styles such as Impressionism and Classicism, and their preference tendencies for exhibit types such as ancient cultural relics and modern artworks. User feedback is also an important way to obtain preference data. Encourage users to submit feedback after visiting, including evaluations of the exhibition content and the exhibit types they hope to see. At the same time, through in-depth analysis of the user's behavior data, the user's potential preferences can also be mined. For example, if the user frequently views exhibits related to a specific historical period and stays for a long time, it can be inferred that the user has a high interest preference for the exhibits of that historical period.

[0033] To effectively obtain multi-source user profile data, a complete data collection system needs to be constructed. During the data collection process, it is first necessary to ensure the accuracy and integrity of the data. For the basic user attribute information, strict data verification rules are set to prevent users from filling in incorrect information by mistake or maliciously. For behavior data and preference data, multiple data collection technologies are used for cross-verification to ensure the reliability of the data. For example, when recording the user's visit path, the indoor positioning system and video surveillance data are combined for comparative analysis to correct possible positioning errors. At the same time, attention should be paid to the timeliness of the data, especially in the collection of user behavior data, to promptly capture the user's latest operations and behavior changes, so as to update the user profile in a timely manner and provide services that better meet the user's current needs.

[0034] In terms of data security and privacy protection, the process of obtaining user profile data must strictly comply with relevant privacy requirements. Before user registration and data collection, clearly inform users of the purpose, scope, and usage method of data collection, and obtain the user's explicit authorization. At the same time, adopt advanced data encryption technology to encrypt and store and transmit the collected user data to prevent data leakage. For example, sensitive user information such as ID numbers and contact information is encrypted and stored in the database, and only specific authorized programs can decrypt and use it. At the same time, establish a perfect data access permission management mechanism to ensure that only authorized personnel and programs can access and process user profile data, and safeguard the security and privacy of user data.

[0035] Similarly, since user profile data comes from a wide range of sources and contains a large amount of multi-dimensional information such as age, gender, occupation, historical browsing records, and search preferences, in order to achieve an effective match between user profile data and exhibit label features, it is necessary to further convert user profile data into a unified and comparable numerical vector form, so as to realize personalized exhibit recommendations based on the similarity calculation between vectors. In a specific example of this application, the Bert model is also used to perform semantic embedding encoding on user profile data, in order to utilize the powerful semantic understanding and representation ability of the Bert model to convert the multi-dimensional information in the user profile into a numerical representation in the vector space, so as to capture key information such as the user's interest preferences and historical behaviors, obtain the user profile feature vector, and thus provide strong data support for subsequent exhibit feature matching.

[0036] In the above-mentioned method for optimizing the first-recommended item of exhibits by combining user portraits and user positioning, in step S5, based on the semantic matching analysis between the user portrait feature vector and the set of exhibit label feature vectors, the first-recommended item of exhibits is determined from the target exhibit group. More specifically, step S5 includes: calculating the correlation between the user portrait feature vector and each exhibit label feature vector in the set of exhibit label feature vectors to obtain a set of correlations; arranging the target exhibit group in the order of the magnitudes of the correlations to obtain the first-recommended item of exhibits. In a specific example of the present application, the correlation is the value of the cosine function. That is, the present application uses cosine similarity as the matching index between the user portrait and the exhibit label, and measures the degree of correlation between the two by calculating the cosine function values between the user portrait feature vector and each exhibit label feature vector. Specifically, the value range of the cosine function is [-1, 1], where 1 indicates complete similarity, -1 indicates complete opposition, and 0 indicates no relation. The closer the cosine function value between the user portrait feature vector and each exhibit label feature vector is to 1, the higher the semantic correlation between the user portrait and the corresponding exhibit label, that is, the higher the matching degree between the exhibit and the user's interest preference. Therefore, the present application takes the exhibit with the highest cosine function value as the first-recommended item, so as to be able to provide the exhibits recommended to the user that best meet their interests and needs. In this way, not only the user's historical behaviors and interest preferences are considered, but also the user's real-time location and orientation information are combined, making the recommendation result closely fit the user's actual visit scenario and improving the timeliness and pertinence of the recommendation.

[0037] Embodiment 2

[0038] Figure 3 It is a flowchart of step S5 in the method for optimizing the first-recommended item of exhibits by combining user portraits and user positioning according to Embodiment 2 of the present application. As Figure 3 shown, step S5 includes: S51, performing portrait-exhibit feature dynamic query analysis on the user portrait feature vector and the set of exhibit label feature vectors to obtain a portrait-exhibit feature dynamic query response coding vector; S52, inputting the portrait-exhibit feature dynamic query response coding vector into a classifier-based personalized recommendation engine to obtain the first-recommended item of exhibits.

[0039] Specifically, in step S51, a portrait-exhibit feature dynamic query analysis is performed on the set of the user portrait feature vectors and the exhibit label feature vectors to obtain a portrait-exhibit feature dynamic query response coding vector. In particular, considering that in the actual exhibition scenario, the matching method based on cosine similarity is essentially a measurement of the angle in the vector space, and its calculation premise is to assume that all dimension information has the same importance. However, in the actual scenario, information in different dimensions often carries different weights and has different degrees of influence on the recommendation result, which may cause the matching method based on cosine similarity to over-respond to the secondary interest features of the user, resulting in recommendation bias. Therefore, in order to more accurately capture the potential connection between the user's interests and the exhibit features, this application proposes a portrait-exhibit feature dynamic query analysis method, which performs a non-linear semantic response analysis on the user portrait feature vectors and each exhibit label feature vector to deeply capture the implicit association patterns between the user's interest preferences and each exhibit feature, and through a spectral aggregation operation, realizes the integration and refinement of the implicit association information between the user's interest preferences and the exhibit features, and excavates the key semantic connection between the user portrait and the exhibit features, so as to obtain a portrait-exhibit feature dynamic query response coding vector that fuses the deep connection between the user's main interest points and the core features of the exhibits.

[0040] Figure 4 FIG. is a flowchart of step S51 in the exhibit top recommendation item optimization method combining user portrait and user location according to Embodiment 2 of the present application. As Figure 4 shown, step S51 includes: S511, performing a non-linear decision interaction response coding on each exhibit label feature vector in the set of the user portrait feature vectors and the exhibit label feature vectors to obtain a set of portrait-exhibit feature implicit interaction coding vectors; S512, performing a graph structure modeling on the set of the portrait-exhibit feature implicit interaction coding vectors to obtain a portrait-exhibit implicit interaction decision point state Laplacian matrix; S513, performing a core component extraction based on spectral decomposition on the portrait-exhibit implicit interaction decision point state Laplacian matrix to obtain the portrait-exhibit feature dynamic query response coding vector.

[0041] More specifically, step S511 is represented by the formula:

[0042]

[0043]

[0044]

[0045] where represents the user portrait feature vector, represents the set of exhibit label feature vectors, 、 、 and Respectively represent the first, second, and third in the set of exhibit label feature vectors and The feature vector of exhibit labels, Indicates the number of exhibit label feature vectors, represents the weight matrix, represents the bias term, represents matrix multiplication, represents the sigmoid function, represents the set of implicit interaction coding vectors of portrait-exhibit features, 、 、 and They represent the first, second, and third vectors in the set of implicit interaction coding vectors of portrait-exhibit features. and A vector encoding the implicit interaction between portrait and exhibit features.

[0046] Here, because the interaction pattern between user portraits and exhibit label features exhibits highly nonlinear characteristics, traditional matching methods based on cosine similarity are limited by the linear assumption of Euclidean space and have difficulty capturing the implicit associations between exhibit attributes and user interests (such as the cross-modal implicit association between "Bronze Research" and "Inscription Patterns"). To this end, this application constructs a nonlinear decision response coding unit based on a deep neural network, leveraging its powerful nonlinear mapping capabilities to perform deep interactive response analysis on user portrait feature vectors and each exhibit label feature vector in a high-dimensional nonlinear semantic latent space, in order to explore the potential nonlinear semantic associations between user portraits and exhibit label features, and generate a set of portrait-exhibit feature implicit interaction coding vectors.

[0047] More specifically, step S512 includes: first, calculating a state-class neighborhood matrix of the portrait-exhibit implicit interaction decision point based on the set of portrait-exhibit feature implicit interaction coding vectors. In a specific example of the present application, the Poincare distance between every two portrait-exhibit feature implicit interaction coding vectors in the set of portrait-exhibit feature implicit interaction coding vectors is calculated to obtain the portrait-exhibit implicit interaction decision point state-class neighborhood matrix composed of multiple Poincare distances, which can be expressed as follows:

[0048]

[0049]

[0050] in, Represents the neighborhood matrix of the state class of the implicit interaction decision points of the image-exhibit, , , and respectively represent the element values at the (1,1) position, the (n,1) position, the (1,n) position, and the (n,n) position in the neighborhood matrix of the state class of the implicit interaction decision points of the image-exhibit. Represents the th image-exhibit feature implicit interaction coding vector in the set of image-exhibit feature implicit interaction coding vectors. Represents the L1 norm. Represents the inverse hyperbolic cosine function. Represents the element value at the (i,j) position in the neighborhood matrix of the state class of the implicit interaction decision points of the image-exhibit, that is, the Poincaré norm between and .

[0051] That is, taking the interaction response information between the user image and each exhibit label as the decision point, further capturing the hierarchical structure of the set of the image-exhibit feature implicit interaction coding vectors through hyperbolic space metrics, and constructing the neighborhood matrix of the state class of the implicit interaction decision points of the image-exhibit. For example, in the East Asian cultural relics exhibition area, the implicit interaction features of the user image information with respect to the "Oracle Bone Inscriptions exhibit" and the "Bronze Ware exhibit" may form a close local neighborhood, while the implicit interaction features with the "Modern Calligraphy exhibit" show a relatively long distance. This geometric representation can effectively model the topological distribution of the user's interests in the cultural context of the exhibits, avoiding the flattening of the cultural correlation degree by the traditional Euclidean distance.

[0052] Next, based on the set of the image-exhibit feature implicit interaction coding vectors, calculate the degree matrix of the state class of the implicit interaction decision points of the image-exhibit. In a specific example of the present application, calculate the feature offset degrees of each image-exhibit feature implicit interaction coding vector in the set of the image-exhibit feature implicit interaction coding vectors with respect to the set of the image-exhibit feature implicit interaction coding vectors, and construct a diagonal matrix based on the feature offset degrees of each image-exhibit feature implicit interaction coding vector to obtain the degree matrix of the state class of the implicit interaction decision points of the image-exhibit, which is expressed by the formula:

[0053]

[0054]

[0055] Among them, Represents the degree matrix of the state class of the implicit interaction decision points of the image-exhibit, and respectively represent the eigenvalues at the (1,1) position and the (n,n) position in the image-exhibit implicit interaction decision point state degree matrix, represent the eigenvalue at the (i,i) position in the image-exhibit implicit interaction decision point state degree matrix, that is, the semantic offset metric value relative to the set of the image-exhibit feature implicit interaction coding vectors

[0056] Here, the present application captures the relative importance of each user image-exhibit interaction decision point by further performing feature offset metrics on each image-exhibit feature implicit interaction coding vector. It should be understood that the feature offset degree of each image-exhibit feature implicit interaction coding vector reflects the degree of its distribution difference from the overall set. For example, an image-exhibit feature implicit interaction coding vector with a larger feature offset degree may represent a relatively unique association pattern between the user image and a specific exhibit, and may mean that it is in a hub position connecting different clusters in the graph structure. By converting the feature offset degree into the diagonal elements of the image-exhibit implicit interaction decision point state degree matrix, it helps to dynamically adjust the weights of each decision point in the graph spectrum, thereby more accurately reflecting the key decision nodes in the user-exhibit interaction, and helping to simplify and effectively represent the complex relationships between high-dimensional data through the adjacency relationship of the graph.

[0057] Then, calculate the difference matrix between the image-exhibit implicit interaction decision point state degree matrix and the image-exhibit implicit interaction decision point state neighborhood matrix to obtain the image-exhibit implicit interaction decision point state Laplacian matrix, which is expressed by the formula:

[0058]

[0059] wherein, represents the image-exhibit implicit interaction decision point state Laplacian matrix.

[0060] Here, by performing a difference operation on the image-exhibit implicit interaction decision point state neighborhood matrix and the image-exhibit implicit interaction decision point state degree matrix, the local connection information of the neighborhood matrix and the node importance information of the degree matrix are fused, so that the spectral characteristics of the generated image-exhibit implicit interaction decision point state Laplacian matrix directly reflect the internal structure of the data and effectively reveal the clustering characteristics of the user-exhibit interaction pattern.

[0061] More specifically, in step S513, since the Laplacian matrix of the portrait-exhibit implicit interaction decision point states is a union representation of all portrait-exhibit implicit interaction decision point states in the topological space, in this case, a topological closure mechanism can be applied to perform a more robust topological feature representation on the Laplacian matrix of the portrait-exhibit implicit interaction decision point states. Based on this, in a preferred example of the present application, the Laplacian matrix of the portrait-exhibit implicit interaction decision point states is constrained and optimized based on the cut space and the cycle space to obtain an optimized Laplacian matrix of the portrait-exhibit implicit interaction decision point states, which is expressed by the formula:

[0062]

[0063]

[0064]

[0065] Wherein, and respectively represent the cut space and the cycle space of the Laplacian matrix of the portrait-exhibit implicit interaction decision point states The cut space and the cycle space of the Laplacian matrix of the portrait-exhibit implicit interaction decision point states, is the exponential function operation with e as the base, represents the optimized Laplacian matrix of the portrait-exhibit implicit interaction decision point states.

[0066] That is, the connected component information of the portrait-exhibit implicit interaction decision point state class neighborhood matrix and the portrait-exhibit implicit interaction decision point state class degree matrix is extracted through the pseudo-inverse operation, and thus the Laplacian matrix of the portrait-exhibit implicit interaction decision point states is optimized through the cut space closure and the cycle space closure. In this way, the global structure invariants in the structural information are extracted based on the closure operation of the pseudo-inverse, making the spectral topological information representation of the Laplacian matrix of the portrait-exhibit implicit interaction decision point states more robust. For example, when there is sparsity or noise interference in the exhibit labels, the cut space closure can strengthen the separation boundary between clusters, while the cycle space closure can maintain the circular association of exhibits within the same cluster (such as a series of works by related artists), so that the optimized Laplacian matrix can more stably characterize the complex user-exhibit interaction topology.

[0067] Next, the optimized Laplacian matrix of the portrait-exhibit implicit interaction decision point states is spectrally decomposed to obtain a set of portrait-exhibit implicit interaction core component coding vectors, which is expressed by the formula:

[0068]

[0069] Wherein, represents the spectral decomposition function, Denote as the matrix composed of the set arrangement of the portrait-exhibit implicit interaction core component coding vectors obtained by spectral decomposition of , and respectively represent the 1st, 2nd, and k-th portrait-exhibit implicit interaction core component coding vectors in the set of portrait-exhibit implicit interaction core component coding vectors, Denote as the diagonal matrix composed of the portrait-exhibit implicit interaction core component eigenvalues obtained by spectral decomposition of , and respectively represent the 1st, 2nd, and k -th portrait-exhibit implicit interaction core component eigenvalues in the diagonal matrix, Denote matrix diagonalization

[0070] Here, by decomposing the optimized portrait-exhibit implicit interaction decision point state Laplacian matrix into a combination of eigenvalues and eigenvectors, nonlinear dimensionality reduction is achieved based on spectral domain transformation to capture the main structural directions of the user-exhibit interaction pattern. It should be understood that each portrait-exhibit implicit interaction core component coding vector obtained after decomposition is essentially the projection representation of the user-exhibit core interaction pattern in the nonlinear dimensionality reduction space. It not only retains the most critical interaction information between the user and the exhibit but also effectively reduces the data dimension, providing a more concise and effective feature representation for subsequent exhibit recommendations.

[0071] Finally, perform attention-driven adaptive fusion on the set of the portrait-exhibit implicit interaction core component coding vectors to obtain the portrait-exhibit feature dynamic query response coding vector, which is expressed by the formula:

[0072]

[0073] where denotes the adaptive fusion network, denotes the weight parameter matrix of the adaptive fusion network, denotes the bias term of the adaptive fusion network, denotes the core component feature significance scoring conversion vector of the adaptive fusion network, denotes the normalized exponential function, denotes the normalized core component feature significance scoring factor, denotes the gating threshold, denotes the masking operation, denotes the corresponding feature fusion weight factor, denotes the portrait-exhibit feature dynamic query response coding vector.

[0074] That is, the contributions of each core component are dynamically integrated using an attention-driven adaptive fusion mechanism. In different exhibition recommendation scenarios, the importance of each user-exhibit core interaction component may change. For example, when the user's browsing history shows their attention to a specific artist, the weight of the core component associated with the artist increases; while when the user is in an exploration mode, the component reflecting style diversity may dominate. This application introduces an attention mechanism to evaluate the importance of each user-exhibit core interaction component through learnable weight parameters, so as to achieve context-related feature fusion, enabling the finally generated portrait-exhibit feature dynamic query response encoding vector to adaptively focus on the most relevant feature combination, not only enhancing the model's response ability to diverse query intents, but also improving the accuracy of the recommendation results by suppressing noise components.

[0075] Specifically, in step S52, the portrait-exhibit feature dynamic query response encoding vector is input into a classifier-based personalized recommendation engine to obtain the top recommended exhibit. Specifically, the classifier-based personalized recommendation engine is pre-trained on a large amount of user portraits and exhibit data, and by learning the mapping relationship between the association pattern of user portraits and exhibit features and the actual user's selection of exhibits, it can accurately predict and recommend the exhibit that best matches the user's current interests. During the recommendation process, after receiving the portrait-exhibit feature dynamic query response encoding vector as input, the classifier-based personalized recommendation engine performs feature parsing on it through the trained classification model, and based on the user portrait-exhibit feature association pattern learned from the vector, judges the potential matching degree between the user and each exhibit, so as to query the top recommended exhibit that best matches the user's current interests from the target exhibit group.

[0076] In a specific example of this application, step S52 includes: performing fully connected encoding on the portrait-exhibit feature dynamic query response encoding vector using the fully connected layer of the personalized recommendation engine to obtain a portrait-exhibit feature dynamic query response fully connected encoding vector; inputting the portrait-exhibit feature dynamic query response fully connected encoding vector into the Softmax classification function of the personalized recommendation engine to obtain the probability values of the portrait-exhibit feature dynamic query response encoding vector belonging to each exhibit; and determining the exhibit corresponding to the largest probability value as the top recommended exhibit.

[0077] In summary, the optimization method for the first-recommended exhibit items combining user portraits and user positioning according to the embodiments of the present application is elucidated. It extracts user positioning data from the indoor positioning system, determines the target exhibit group based on the real-time position and orientation information of the user, then introduces deep learning technology to semantically analyze the user portrait data to extract user portrait features, and determines the first-recommended exhibit items from the target exhibit group by performing semantic matching analysis between the user portrait features and the feature tags of each exhibit in the target exhibit group. In this way, it is possible to comprehensively consider the user's behavioral preferences and real-time physical environment information, provide a more accurate and personalized exhibit recommendation service for users, thereby improving the visiting experience of visitors and optimizing the resource utilization efficiency of cultural exhibition venues such as museums and exhibition halls.

Claims

1. An optimization method for the first recommended item of an exhibit by combining user portraits and user positioning, characterized in that Including: Extracting user location data from an indoor positioning system, where the user location data includes the user's real-time location and the user's orientation information; Determining a target exhibit group based on the user location data; Extracting the feature tags of each exhibit from the target exhibit group to obtain a set of exhibit label feature vectors; Obtaining user portrait data and encoding the user portrait data to obtain a user portrait feature vector; Determining the first-recommended exhibits from the target exhibit group based on semantic matching analysis between the user portrait feature vector and the set of exhibit label feature vectors; Among them, determining the first-recommended exhibits from the target exhibit group based on semantic matching analysis between the user portrait feature vector and the set of exhibit label feature vectors includes: Performing portrait-exhibit feature dynamic query analysis on the user portrait feature vector and the set of exhibit label feature vectors to obtain a portrait-exhibit feature dynamic query response coding vector, including: performing non-linear decision interaction response coding on each exhibit label feature vector in the user portrait feature vector and the set of exhibit label feature vectors to obtain a set of portrait-exhibit feature implicit interaction coding vectors; performing graph structure modeling on the set of portrait-exhibit feature implicit interaction coding vectors to obtain a portrait-exhibit implicit interaction decision point state Laplacian matrix; performing core component extraction based on spectral decomposition on the portrait-exhibit implicit interaction decision point state Laplacian matrix to obtain the portrait-exhibit feature dynamic query response coding vector; Inputting the portrait-exhibit feature dynamic query response coding vector into a personalized recommendation engine based on a classifier to obtain the first-recommended exhibits.

2. The optimized method for the first recommended item of the exhibit combining user portraits and user positioning according to claim 1, wherein Performing graph structure modeling on the set of portrait-exhibit feature implicit interaction coding vectors to obtain a portrait-exhibit implicit interaction decision point state Laplacian matrix, including: Calculating a portrait-exhibit implicit interaction decision point state class neighborhood matrix based on the set of portrait-exhibit feature implicit interaction coding vectors; Calculating a portrait-exhibit implicit interaction decision point state degree matrix based on the set of portrait-exhibit feature implicit interaction coding vectors; Calculating the difference matrix between the portrait-exhibit implicit interaction decision point state degree matrix and the portrait-exhibit implicit interaction decision point state class neighborhood matrix to obtain the portrait-exhibit implicit interaction decision point state Laplacian matrix.

3. The optimized method for the first recommended item of an exhibit by combining a user profile and user positioning according to claim 2, wherein Calculating a portrait-exhibit implicit interaction decision point state class neighborhood matrix based on the set of portrait-exhibit feature implicit interaction coding vectors, including: Calculating the Poincaré distance between every two portrait-exhibit feature implicit interaction coding vectors in the set of portrait-exhibit feature implicit interaction coding vectors to obtain the portrait-exhibit implicit interaction decision point state class neighborhood matrix composed of multiple Poincaré distances.

4. The optimized method for the first recommended item of the exhibit combining user portraits and user positioning according to claim 3, characterized in that Calculating a portrait-exhibit implicit interaction decision point state degree matrix based on the set of portrait-exhibit feature implicit interaction coding vectors, including: Calculate the feature offset degrees of each image-exhibit feature implicit interaction coding vector in the set of image-exhibit feature implicit interaction coding vectors with respect to the set of image-exhibit feature implicit interaction coding vectors, and construct a diagonal matrix based on the feature offset degrees of each image-exhibit feature implicit interaction coding vector to obtain the image-exhibit implicit interaction decision point state degree matrix.

5. The optimized method for the first recommended item of exhibits combining user portraits and user positioning according to claim 4, characterized in that, Perform core component extraction based on spectral decomposition on the image-exhibit implicit interaction decision point state Laplacian matrix to obtain the image-exhibit feature dynamic query response coding vector, including: Perform constraint optimization based on the cut space and the cycle space on the image-exhibit implicit interaction decision point state Laplacian matrix to obtain an optimized image-exhibit implicit interaction decision point state Laplacian matrix; Perform spectral decomposition on the optimized image-exhibit implicit interaction decision point state Laplacian matrix to obtain a set of image-exhibit implicit interaction core component coding vectors; Perform attention-driven adaptive fusion on the set of image-exhibit implicit interaction core component coding vectors to obtain the image-exhibit feature dynamic query response coding vector.

6. The optimized method for the first recommended item of an exhibit by combining user portraits and user positioning according to claim 5, characterized in that Input the image-exhibit feature dynamic query response coding vector into a personalized recommendation engine based on a classifier to obtain the top recommended exhibit item, including: Use the fully connected layer of the personalized recommendation engine to perform fully connected coding on the image-exhibit feature dynamic query response coding vector to obtain an image-exhibit feature dynamic query response fully connected coding vector; Input the image-exhibit feature dynamic query response fully connected coding vector into the Softmax classification function of the personalized recommendation engine to obtain the probability values of the image-exhibit feature dynamic query response coding vector belonging to each exhibit; Determine the exhibit corresponding to the maximum value among the probability values as the top recommended exhibit item.

Images