MR content recommendation method fusing preference modeling

Abstract

The invention relates to the technical field of personalized recommendation, in particular to a preference modeling fused MR content recommendation method, which comprises the following steps of: constructing a multi-modal cross weight by collecting eye movement level data and gesture track intensity and combining voice space mapping, aggregating user preference intentions according to a behavior direction, and performing recommendation on the user preference intentions according to the behavior direction; and associating the fixation frequency with the content semantic label in the space grid, and outputting content mapping configuration. According to the method, a spatial hierarchical gaze structure and action intensity expression are constructed, the gaze trend and operation intention of a user in a three-dimensional space are converted into spatialization preference features, the corresponding relation between voice content and a scene target is achieved through semantic anaphora mapping of a voice instruction, and coincidence density analysis of a multi-modal point location is combined, so that the voice content and the scene target are obtained. The accuracy of preference expression is enhanced, the behavior intention is identified in combination with the aggregation trend, the binding relationship between the fixation frequency and the semantic tag is used, the spatial selection of content presentation is optimized, and the consistency of recommended content and the real attention path of the user is enhanced.

Description

Technical Field

[0001] This invention relates to the field of personalized recommendation technology, and in particular to a method for recommending content based on MR (Mixed Context) content that incorporates preference modeling. Background Technology

[0002] Personalized recommendation technology involves analyzing various data sources such as user behavior, interests, historical records, and contextual information to build user interest models, and then personalized push information, content, or services. Core aspects include user profile construction, recommendation algorithm design, recommended content filtering and ranking, and feedback mechanism updates. It is widely used in e-commerce, social platforms, video streaming, news information, education and training, and other scenarios. It relies on collaborative filtering, content filtering, deep learning, knowledge graphs, and other means to achieve accurate recommendations, improve user experience and information matching efficiency. Among them, traditional MR content recommendation methods refer to the processing methods that recommend suitable digital content or services based on the user's current context and interests in a mixed reality environment. It usually relies on the user's interaction records in the MR scene, such as gaze trajectory, voice input, gesture recognition, and other behavioral data, combined with preset content feature tags, and uses recommendation mechanisms based on collaborative filtering or content matching to filter candidate content and generate recommendation lists. In terms of user preference modeling, traditional methods mostly use static user attributes and simple behavior frequency statistics as the basis for preferences, lacking the means to model changes in users' dynamic interests. The recommendation methods are mainly based on fixed rules or similarity calculations, lacking in-depth integration of high-dimensional feature correlation and semantic preference expression.

[0003] Existing technologies for content recommendation in mixed reality environments rely on user gaze patterns, gesture triggers, and voice commands as behavioral records. Based on discrete behavior collection and static preference statistics, they cannot effectively characterize the continuous gaze trends and dwell depth of users in three-dimensional space. They struggle to identify the true interest orientation towards different spatial layers. When target references appear in voice commands, there is often a lack of a corresponding mechanism with scene objects, resulting in a break between semantic input and spatial targets. Hand movement paths are often simplified to trigger events rather than continuous spatial behaviors in existing models, making it impossible to accurately capture dynamic action intentions. Multimodal information lacks a unified spatial and temporal coordination process, leading to ambiguous hotspot judgments, distorted preference expressions, and a deviation between recommended content and users' actual attention, affecting the targeting and continuity of content delivery. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a content recommendation method based on fusion preference modeling.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for recommending MR content by incorporating preference modeling, comprising the following steps: S1: Obtain the user's eye movement trajectory sequence, analyze the spatial coordinates and depth distribution of continuous points, calculate the density and dwell trend of fixation points in each spatial layer, construct a fixation concentration distribution model, and perform label classification based on spatial hierarchy to generate spatial fixation hierarchy structure data. S2: Based on the spatial gaze hierarchy structure data, obtain the user's hand trajectory point sequence, analyze the direction deviation and spatial dispersion of the trajectory, determine the continuity and change pattern of the action path, construct the action feature expression structure, classify the operation intensity, and generate interactive trajectory intensity feature information. S3: Combining the spatial gaze hierarchy structure data and interaction trajectory intensity feature information, analyze user voice data, determine the overlap density of multimodal points through spatial location mapping and time synchronization, construct a modal cross model, initialize weight labels for semantic content matching, and generate spatial fusion weight configuration. S4: Using the spatial fusion weight configuration, analyze the directional changes and path extension trends of the operational behavior trajectory, construct a directional vector cluster, determine the frequency of directional offset and path continuity characteristics, form a preference aggregation direction, and generate behavioral guidance intention data; S5: Based on the spatial gaze hierarchy structure data and behavioral guidance intent data, analyze the spatial distribution of user gaze focus, calculate the gaze intensity index by recording the number of gaze points and focus range of each spatial unit, establish regional priority ranking, bind regions in combination with content semantic tags, match content semantic tags according to gaze activity, and generate gaze hotspot mapping configuration.

[0006] As a further embodiment of the present invention, the spatial gaze hierarchy structure data includes a gaze region coordinate group, a hierarchy label set, and a depth-of-field continuous distribution feature; the interaction trajectory intensity feature information includes a trajectory coherence segment set, a direction offset feature group, and a motion amplitude classification label; the spatial fusion weight configuration includes a modality mapping coordinate group, a modality overlap ratio group, and a weight allocation label; the behavior guidance intent data specifically includes a trajectory direction vector set, a behavior offset path group, and a preference aggregation direction label; and the gaze hotspot mapping configuration specifically refers to a priority region index, a gaze activity index set, and a semantic label binding group.

[0007] As a further aspect of the present invention, the step of acquiring the spatial gaze hierarchy structure data specifically includes: S111: Obtain the user's eye movement trajectory sequence, analyze the spatial coordinates and corresponding depth distribution of the user's continuous eye movement points in the MR scene, calculate the continuous density and dwell duration of the fixation point in each spatial layer, and obtain spatial fixation focusing data; S112: Based on the spatial gaze focusing data, segment the eye movement trajectory segments generated by the user in the MR scene, calculate the focusing stability of each gaze path, and combine cumulative trend modeling to obtain the gaze stability distribution coefficient. S113: Call the gaze stability distribution coefficient, divide the range of numerical variation into intervals, construct the focus segment mapping region in three-dimensional space, and map the located focus interval markers to the hierarchical label group of the MR content region model, establish the gaze distribution association structure under the spatial hierarchy, and generate spatial gaze hierarchy structure data.

[0008] As a further aspect of the present invention, the step of obtaining the interaction trajectory intensity feature information specifically includes: S211: Based on the spatial gaze hierarchy data, obtain the user's hand trajectory point sequence, analyze the directional offset change of adjacent trajectory segments in continuous segments, calculate the directional vector change rate, determine the directional change trend of each segment, and obtain the directional coherence index. S212: Based on the directional coherence index, analyze the degree of dispersion of continuous segments in the spatial coordinate distribution of the trajectory path, calculate the discrete density of the three-dimensional coordinate distribution within the segment, and generate a stable trajectory distribution feature. S213: Based on the directional coherence index and trajectory stability distribution characteristics, and according to the combination relationship between trajectory continuity and directional change trend, construct an action feature expression structure, and calculate the trajectory operation intensity coefficient value based on the density of trajectory feature combinations in the structure, classify the operation intensity, and obtain interactive trajectory intensity feature information.

[0009] As a further aspect of the present invention, the step of obtaining the spatial fusion weight configuration specifically includes: S311: Obtain the spatial gaze hierarchy structure data and interaction trajectory intensity feature information, obtain and analyze the user's voice data, aggregate the coordinate information of multiple modal points in each spatial unit through spatial location mapping and time synchronization, count the number of multimodal points in each region, and establish a multimodal spatial distribution density group. S312: Based on the multimodal spatial distribution density group, by analyzing the spatial landing point distribution of multiple modes in the MR scene, and according to the degree of overlap density of multiple modal points in the same spatial location, a spatial crossover rate set is generated; S313: Based on the spatial overlap rate set and the distribution density of each modality in the intersection region, initialize weight labels for semantic input content and generate spatial fusion weight configuration.

[0010] As a further aspect of the present invention, the process of aggregating the coordinate information of multiple modal points within each spatial unit specifically comprises: Acquire user voice data, convert it into text information through speech recognition, perform semantic analysis on the converted text, identify verbs in the instructions and the target entities they refer to, combine the spatial distribution information of objects in the MR system, map the target entities referred to in the text to entities in the MR space, and identify the spatial location of the target entities in the speech based on the spatial coordinate data of the entities, and output the spatial coordinate data of the speech referent entities at multiple times. By combining the spatial gaze hierarchy structure data and the interaction trajectory intensity feature information, spatial coordinate data of each modality within each spatial unit are obtained, including eye movement trajectory, hand movement trajectory, and spatial coordinate data of the voice-referencing entity at multiple times. The modal point list of each spatial unit is classified and statistically analyzed according to the modal type. The number of points, time sequence range and coordinate centroid data of each mode in each spatial unit are recorded to establish the multimodal spatial distribution density group.

[0011] As a further aspect of the present invention, the step of obtaining the behavior-oriented intent data specifically includes: S411: Obtain the spatial fusion weight configuration, extract the start and end points of the trajectory of each behavior of the user during the MR content operation, combine the direction change frequency and path extension trend between trajectory points, construct a set of directional trajectories, and generate a trajectory direction change sequence. S412: Based on the trajectory direction change sequence, calculate the direction offset frequency of each trajectory set, and count the continuity state in the path extension. Compare the direction change characteristics in the continuous segments, evaluate the stability of each trajectory, and obtain the trajectory stability parameter set. S413: Based on the trajectory stability parameter set, construct a direction vector cluster, analyze the intersection and change trend of each vector cluster in space, form a preference aggregation direction, identify the user's preference intention, and generate behavior-oriented intention data.

[0012] As a further aspect of the present invention, the step of obtaining the gaze hotspot mapping configuration specifically includes: S511: Based on the spatial gaze hierarchy structure data and behavioral guidance intention data, analyze the spatial distribution of the user's gaze focus in the MR scene, and generate a gaze intensity index set by recording the number of gaze points and the focus range in each spatial unit; S512: Calculate the gaze activity level of each spatial unit according to the gaze activity level set, and arrange the spatial units according to the gaze activity level to obtain a spatial priority sequence; S513: Combining the spatial priority sequence with the content semantic classification tags, multiple priority regions are bound together. The content semantic tags are matched according to the gaze activity level to form a structured semantic association result and generate a gaze hotspot mapping configuration.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by constructing a spatial hierarchical gaze structure and action intensity expression, the user's gaze trend and operation intention in three-dimensional space are transformed into spatialized preference features. The semantic reference mapping of voice commands realizes the correspondence between voice content and scene targets. Combined with the overlap density analysis of multimodal points, the accuracy of preference expression is enhanced. Combined with the aggregation trend to identify behavioral intentions, the binding relationship between gaze intensity and semantic tags is used to optimize the spatial selection of content presentation and enhance the consistency between recommended content and the user's actual attention path. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the main steps of the present invention; Figure 2 This is a flowchart of the spatial gaze hierarchy structure data acquisition process of the present invention; Figure 3 This is a flowchart of the process for obtaining interactive trajectory intensity feature information according to the present invention; Figure 4 This is a flowchart of the spatial fusion weight configuration acquisition process of the present invention; Figure 5 This is a flowchart of the behavior-oriented intent data acquisition process of the present invention; Figure 6 A flowchart for obtaining the gaze hotspot mapping configuration of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0017] Please see Figure 1 This invention provides a technical solution, a method for recommending MR content by incorporating preference modeling, comprising the following steps: S1: Obtain the user's eye movement trajectory sequence, analyze the spatial coordinates and depth distribution of continuous points, calculate the density and dwell trend of fixation points in each spatial layer, construct a fixation concentration distribution model, and perform label classification based on spatial hierarchy to generate spatial fixation hierarchy structure data. S2: Based on spatial gaze hierarchy data, obtain the user's hand trajectory point sequence, analyze the trajectory direction offset and spatial dispersion, determine the continuity and change pattern of the action path, construct the action feature expression structure, classify the operation intensity, and generate interactive trajectory intensity feature information. S3: Combining spatial gaze hierarchy structure data and interaction trajectory intensity feature information, analyze user voice data, determine the overlap density of multimodal points through spatial location mapping and time synchronization, construct a modality cross model, initialize weight labels for semantic content matching, and generate spatial fusion weight configuration; S4: Utilize spatial fusion weight configuration to analyze the directional changes and path extension trends of operational behavior trajectories, construct directional vector clusters, determine the frequency of directional offsets and path continuity features, form preference aggregation directions, and generate behavioral guidance intent data; S5: Based on spatial gaze hierarchy data and behavioral guidance intent data, analyze the spatial distribution of user gaze focus. By recording the number of gaze points and focus range of each spatial unit, calculate the gaze intensity index, establish regional priority ranking, bind regions in combination with content semantic tags, match content semantic tags according to gaze activity, and generate gaze hotspot mapping configuration.

[0018] Spatial gaze hierarchy data includes gaze region coordinate set, hierarchy label set, and depth of field continuous distribution features. Interaction trajectory intensity feature information includes trajectory coherence segment set, direction offset feature set, and action amplitude classification label. Spatial fusion weight configuration includes modality mapping coordinate set, modality overlap ratio group, and weight allocation label. Behavior guidance intent data specifically includes trajectory direction vector set, behavior offset path set, and preference aggregation direction label. Gaze hotspot mapping configuration specifically refers to priority region index, gaze activity index set, and semantic label binding group.

[0019] Please see Figure 2 The specific steps for obtaining spatial gaze hierarchy data are as follows: S111: Obtain the user's eye movement trajectory sequence, analyze the spatial coordinates and corresponding depth distribution of the user's continuous eye movement points in the MR scene, calculate the continuous density and dwell duration of the fixation point in each spatial layer, and obtain spatial fixation focusing data; In the initial stage of building the MR (Mixed Reality) scene, the system first activates the infrared eye-tracking sensor built into the head-mounted display device to continuously capture the user's gaze data at a sampling frequency of 120Hz, obtaining a sequence of the user's eye movement trajectory. This sequence consists of a series of three-dimensional coordinate points with timestamps. The system consists of several components. It synchronously loads the 3D spatial mesh data of the MR scene, maps eye-tracking coordinates to the virtual spatial coordinate system, and analyzes the spatial coordinates and corresponding depth distribution of continuous eye-tracking points in the MR scene. Specifically, the system calculates the spatial coordinates of each gaze point using the camera position as the origin. The Euclidean distance to the origin is used as the depth of field value. The system divides the scene along the Z-axis into three discrete spatial layers: a near-field interaction layer (0.3m-1.0m), a mid-field focus layer (1.0m-3.0m), and a far-field background layer (>3.0m). The system counts the number of gaze points falling into each spatial layer within a unit time window (e.g., 200ms) and calculates the continuous density and dwell duration of gaze points in each layer. For example, if there are 240 sampling points within 2 seconds, and 180 of them fall within the bounding box of the virtual car model in the mid-field focus layer, then the gaze density of that layer is 0.75, and the dwell duration is 1.5 seconds. The system further detects the transition frequency of gaze points between different layers in the time series to determine the dwell trend of continuous gaze points in the spatial hierarchy. If a gaze point remains within the same depth of field in adjacent time slices and the spatial coordinate variance is less than 0.05m², the system determines it to be in a "depth focus" state; if the gaze point switches rapidly between the three layers, it is determined to be in a "spatial browsing" state.

[0020] Table 1. Sampling table of eye-tracking gaze levels in MR scenes. ; Referring to Table 1, the system recorded spatial hierarchy gaze data in a certain session. Based on the spatial continuity of the focused area and the depth-of-field change path, the system identified object A, which is the user's main focus area, as the mid-field layer. It constructed a gaze concentration distribution model and performed label classification based on the spatial hierarchy regions already divided in the model (such as the mid-field layer - object A region). The region was marked as "high attention - detailed viewing", generating spatial gaze hierarchy structure data containing hierarchy ID, region coordinate range, and gaze status label.

[0021] S112: Based on spatial gaze focusing data, segmented localization is performed on the eye movement trajectory segments generated by the user in the MR scene, using the following formula: ; The focus stability of each gaze path is calculated, and the gaze stability distribution coefficient is obtained by combining cumulative trend modeling. in, For the first The focusing stability of an eye-tracking path is obtained by jointly calculating the gaze density shift and spatial perturbation terms within that path. For the first The eye movement path in the first The normalized fixation density in a spatial segment is obtained by counting the number of fixation points within that segment and normalizing the spatial sampling range. For the first The average normalized fixation density of each eye movement path is calculated by analyzing all the fixations within the path. The average was calculated to obtain the result. For the first The dimensionless depth factor of each spatial segment is obtained by performing normalization on the depth change sequence of that segment. For the first The eye movement path in the first The normalized lateral offset in a spatial segment is obtained by comparing and normalizing the lateral coordinate differences between adjacent points in that segment. For the first The normalized jump amplitude of a spatial segment during inter-layer transition is obtained by detecting and normalizing the spatial layer change amplitude between the segment and its adjacent segments. This is the index of the spatial segment within the path, used to identify discrete sampling locations within the path. This is a sequence index for the eye-tracking path, used to distinguish multiple discontinuous gaze trajectory segments. The number of spatial segments divided in a single eye-tracking path is obtained by performing segmentation processing on the continuous point sequence of the path according to the spatial layer change pattern and the depth change sequence; Based on spatial gaze focusing data, the system first segments the continuous long eye movement trajectory into segments according to the spatial clustering characteristics of the gaze points. Each segment is a separate discrete space segment, and each segment is numbered and indexed. Subsequently, the system targeted the first... A complete eye-tracking path, calling the formula Calculate its focusing stability. In the formula, the numerator part... The fluctuations in gaze density are accumulated and weighted by depth of field, where the summation sign... This indicates all paths within the path. The weighted deviations of each spatial segment are accumulated, and then subtraction is performed. The aim is to quantify the deviation of the current segment fixation density from the average level; the absolute value sign ensures that the magnitude of the deviation is positive; multiplication operations are used. Its function is to introduce depth weights, making gaze fluctuations in deep space have a greater impact on the results; the denominator part Addition operations are used to measure physical offsets and hierarchical jumps in space. The perturbations, which combine lateral movement and vertical hierarchy switching, are calculated using square and square root operations to determine the geometric norm of this combined perturbation. Regarding parameter acquisition: By statistical analysis of the first The number of fixations within a segment is divided by the area of ​​the segment's bounding box, and the result is mapped to the [0,1] interval. For all in the path The arithmetic mean; Take the ratio of the depth of field value at the center point of the segment to the maximum visible depth of the scene; Calculate and normalize the sum of Euclidean distances between adjacent fixation points within the segment; The result is obtained by dividing the spatial hierarchy ID difference between the current segment and the previous segment (e.g., a difference of 2 if jumping from level 1 to level 3) by the maximum hierarchy number (3). A practical example is used for calculation: Suppose an eye-tracking path is divided into 3 segments ( The parameters were measured as follows: Then the average density The density deviations are respectively ; Set depth factor ; Set lateral offset ; Set up hierarchical jump (A layer change occurs in the middle). Substitute into the formula to calculate the numerator: ; Calculate the denominator: .

[0022] Finally obtained .

[0023] The result The value is in a lower range (the reference range is set to [0,1]), which indicates that the gaze density fluctuation within the path is small and the spatial jump is small. Combined with cumulative trend modeling, the gaze stability distribution coefficient is obtained, and it is determined that the user's gaze behavior has high stability.

[0024] S113: Call the gaze stability distribution coefficient, divide the range of numerical variation into intervals, construct the focus segment mapping region in three-dimensional space, and map the located focus interval markers to the hierarchical label group of the MR content region model, establish the gaze distribution association structure under the spatial hierarchy, and generate spatial gaze hierarchy structure data. System call gaze stability distribution coefficient The system divides the range of numerical variation into intervals based on a preset quantization standard. Specifically, the system sets a first stability threshold. Second stability threshold The threshold setting process is as follows: The system collects 100 sets of standard "gaze" behavior data. The mean is 0.15, the variance is 0.05, and the set value is... Collect 100 sets of "rapid scanning" data. The mean is 0.8, set .like It is determined to be a "highly stable focusing range"; if This is determined to be an "exploratory observation period"; if This is determined to be a "divergent browsing range". Based on the calculations in S112... The system categorizes this area as a "high-stability focus region," constructs a focus fragment mapping region in 3D space, and defines this region as a spherical space with a radius of 0.5m centered on the user's line of sight. The system then retrieves the hierarchical label group of the MR content region model, which pre-assigns labels such as "structural layer," "functional layer," and "appearance layer" to virtual objects in the scene. Since this focus fragment is located on the vehicle's hood (belonging to the appearance layer), the system maps the located focus region label to the "appearance layer" label, establishing a spatial hierarchy of gaze distribution association structure, i.e., the association relationship is {User ID: U001, Spatial Region: Region_Hood, Stability: High, Semantic Layer: Appearance}, generating spatial gaze hierarchy structure data.

[0025] Please see Figure 3 The specific steps for obtaining interaction trajectory intensity feature information are as follows: S211: Based on spatial gaze hierarchy data, obtain the user's hand trajectory point sequence, analyze the directional offset changes of adjacent trajectory segments in continuous segments, calculate the directional vector change rate, determine the directional change trend of each segment, and obtain the directional coherence index. Based on the user's attention region determined by the spatial gaze hierarchy data, the system simultaneously activates the gesture recognition module to acquire the user's hand trajectory point sequence through a data glove or visual sensor. The system selects the hand trajectory that overlaps with the time of the highly stable gaze zone and analyzes the directional shift changes of adjacent trajectory segments within continuous segments. In practice, the system fits a micro-vector to every three points in the trajectory point sequence. And calculate adjacent vectors and cosine value of the angle between Using the inverse cosine function to obtain the included angle. Define the rate of change of the direction vector For example, during a 0.5-second "swipe right" operation, five vectors are extracted, with adjacent vectors having angles of [missing information]. The corresponding average rate of change is low, indicating that the direction is stable; however, during a period of "random aerial drawing," the included angle may vary. to The frequency fluctuates drastically. The system calculates the directional change trend of each segment, if... If the variance of the sequence is less than 0.1, it is judged as "unidirectional linear motion"; if the variance is greater than 0.5, it is judged as "multidirectional complex motion", thus obtaining the directional coherence index.

[0026] S212: Based on the directional coherence index, analyze the degree of dispersion of continuous segments in the spatial coordinate distribution of the trajectory path, calculate the dispersion density of the three-dimensional coordinate distribution within the segment, and generate the trajectory stable distribution characteristics. Based on the directional coherence index, the system further analyzes the dispersion of continuous segments in the trajectory path in terms of spatial coordinate distribution. The system constructs a minimum bounding box (OBB) containing all sampling points of the current trajectory segment and calculates the root mean square error (RMSE) of the perpendicular distance from all points within the trajectory segment to the trajectory fitting centerline as a measure of dispersion. Taking a set of hand hovering operations as an example, the sampling point set is... These points are closely distributed around the center point, and the calculated RMSE is 0.02m, indicating low dispersion. However, for large-amplitude waving movements, the sampling points are sparsely distributed and cover a wide area, and the RMSE can reach 0.2m. The system defines the ratio of the calculated RMSE value to the total trajectory length as the discrete density. ,like It is labeled as a "compact trajectory"; if It is identified as a "loose trajectory".

[0027] Table 2. Spatial Discrete Feature Analysis of Hand Interaction Trajectory ; Referring to Table 2, the system records the discrete data of different action types and generates stable trajectory distribution characteristics.

[0028] S213: Based on the directional coherence index and the stable distribution characteristics of the trajectory, and according to the combination relationship between the continuity of the trajectory and the trend of directional change, a motion feature expression structure is constructed, and based on the density of the trajectory feature combination in the structure, the following formula is used: ; Calculate the trajectory operation intensity coefficient value, classify the operation intensity, and obtain interactive trajectory intensity feature information; in, The trajectory operation intensity coefficient is calculated by combining the frequency of direction change and the ratio of stable distribution. For the first The normalized value of the rate of change of each directional vector is obtained by normalizing the difference in the angle between adjacent vectors in a continuous segment of the trajectory. The mean of the normalized rates of change of all direction vectors is obtained by applying the normalized values ​​of all direction vectors. We obtain the arithmetic mean. The index of the rate of change of the direction vector in the trajectory segment represents the sequence number of the current trajectory point. The total number of normalized values ​​for the rate of change of the direction vector is obtained by counting the number of valid vectors in a continuous trajectory segment. This is a dimensionless normalized value representing the overall shape of the trajectory path, obtained by normalizing after calculation of path length and spatial extensibility. The proportion of stable segments in the overall trajectory is obtained by dividing the total length of trajectory segments that satisfy directional continuity and low spatial discreteness by the total trajectory length. This is generated by sequentially numbering adjacent trajectory point pairs within consecutive segments of the trajectory path, and is used to identify the segment. The positional order of the directional vectors; Based on the directional coherence index and the stable distribution characteristics of the trajectory, the system constructs a motion feature expression structure, and uses a formula based on the density of trajectory feature combinations in the structure. Calculate the trajectory operation intensity coefficient value. In the formula, the summation symbol... Division operation The mean absolute deviation used to calculate the rate of change of direction reflects the degree or complexity of jitter in the direction of motion. The trajectory shape normalization value is incorporated into the intensity consideration through square root operation, representing the magnitude of the movement amplitude; multiplication operation... The overall intensity is then weighted and corrected using the proportion of stable segments. If the trajectory contains a large number of invalid or unstable swaying movements, A smaller value will reduce the overall strength value. The parameters are obtained as follows: For the first The rate of change of each directional vector is normalized to a value in [0,1]. for The mean of the sequence; The total number of vectors; It is the normalized value of the ratio of the actual path length of the trajectory to the straight-line distance between the beginning and end (reflecting the tortuosity of the path); To satisfy the criteria of "directional coherence index > 0.8 and discrete density < 0.1" for the proportion of trajectory segment length to the total length. Assume the user performs a selection action of "forcefully and rapidly drawing a circle". Detection Rate of change of a vector, sequence The fluctuations are large, and the calculated average deviation is... The trajectory is circular, the path is longer than the displacement, and the normalized morphological value is... , The movements were fluid and powerful throughout, with stable proportions. Substitute the values ​​into the calculation: ; The system sets the strength baseline to 0.5, because... The system categorizes the intensity of the operation, classifying the action as "high-intensity intentional interaction" rather than unconscious hand movements. This result indicates that the user has a strong desire to manipulate or select the current holographic object, thus obtaining information on the intensity characteristics of the interaction trajectory.

[0029] Please see Figure 4 The specific steps for obtaining the spatial fusion weight configuration are as follows: S311: Acquire spatial gaze hierarchy structure data and interaction trajectory intensity feature information, acquire and analyze user voice data, aggregate coordinate information of multiple modal points in each spatial unit through spatial location mapping and time synchronization, count the number of multimodal points in each region, and establish a multimodal spatial distribution density group. After acquiring spatial gaze hierarchy data and interaction trajectory intensity features, the system simultaneously activates a microphone array to acquire the user's voice data. The system first converts the audio stream "Open the hood and take a look" into text using an ASR (Automatic Speech Recognition) engine, and then uses Natural Language Processing (NLP) technology to extract the verb "open" and the entity noun "hood". The system then queries the MR scene database to obtain the center coordinates of the virtual object "hood". And its bounding box range. Through spatial location mapping and time synchronization, the system aligns the time window of the voice command occurrence (e.g., T=2.5s to 3.0s) with the eye-tracking fixation points and hand trajectory points within that time period. The system divides the MR scene into The voxel grid (spatial cell) aggregates the coordinate information of multiple modal points within each spatial cell. For example, in the coordinate... Within the spatial units of the system and its neighborhood, the system identified: 1 speech mapping point (based on entity coordinates), 15 eye-tracking fixation points, and 8 hand trajectory points. The system recorded the number of points, time sequence range (e.g., [T250, T300]), and coordinate centroid data for each modality in each spatial unit, establishing a multimodal spatial distribution density group.

[0030] S312: Based on the multimodal spatial distribution density group, by analyzing the spatial landing point distribution of multiple modes in the MR scene, a spatial crossover rate set is generated according to the degree of overlap density of multiple modal points in the same spatial location; Based on a multimodal spatial distribution density set, the system traverses all spatial cells containing data to analyze the spatial distribution of various modalities in the MR scene. The system defines "overlap" as heterogeneous modal points within the same or adjacent spatial cells with a timestamp difference of less than 200ms. The system calculates the modal overlap index for each region. The numerator is the frequency of the simultaneous occurrence of the three modes, and the denominator is the total frequency of the occurrence of any mode.

[0031] Table 3. Example of Multimodal Spatial Crossover Rate Analysis ; Referring to Table 3, the system generates a spatial overlap rate set based on the degree of overlap of multiple modal points in the same spatial location. For example, in the "hood" area, the overlap of the three is extremely high; while in the "wheel" area, only eye movement scans are performed, and the overlap is 0.

[0032] S313: Based on the spatial overlap rate set and the distribution density of each modality in the intersection region, initialize weight labels for semantic input content matching and generate spatial fusion weight configuration; Based on the spatial overlap rate set, the system combines the distribution density of each modality in the intersection region and, according to the intersection density ratio, initializes weight labels for semantic input content matching. The system sets the weight calculation rules: initial weights... ,in Trust coefficients for each modality are set. The coefficient setting process is as follows: Based on historical interaction data statistics, the system found that voice commands have the highest clarity, followed by gestures, while eye movements are most easily interfered with. Through calculation examples, the following coefficients are set: For the "hood" area in Table 3, the normalized density... (If a voice command exists, it is 1), calculate the weight. The value of 0.92 is significantly higher than the baseline threshold of 0.6. Based on this, the system resets the matching weight of semantic tags such as "view internal structure" or "disassemble" with the content of this area to 0.92, and generates a spatial fusion weight configuration containing {Target:Hood,Intent:Open,Weight:0.92}.

[0033] Please see Figure 5 The specific steps for obtaining behavior-oriented intent data are as follows: S411: Obtain spatial fusion weight configuration, extract the start and end points of the trajectory of each behavior of the user in the MR content operation process, combine the frequency of directional changes between trajectory points and the path extension trend, construct a set of directional trajectories, and generate a trajectory direction change sequence. By leveraging spatial fusion weighting to configure high-weighted interactive events, the system extracts the start and end points of each user action during MR content interaction. The system then backtracks on previously acquired hand trajectory data to lock in the weights. Extract the starting coordinates of the trajectory within the given time window. and termination coordinates Simultaneously, the system discretizes the trajectory into a series of unit vectors by combining the frequency of directional changes between trajectory points and the path extension trend. The system calculates the cross product and dot product of every two adjacent unit vectors to construct a set of directional trajectories. For example, for a waving gesture from left to right, the generated vector sequence is consistently positive on the X-axis component, while it is close to zero on the Y-axis and Z-axis components. The system packages these vectors to generate a sequence of trajectory direction changes, which not only contains spatial direction information but also implicitly contains temporal characteristics of velocity and acceleration.

[0034] S412: Based on the trajectory direction change sequence, calculate the direction offset frequency of each trajectory set, and count the continuity state in the path extension. Compare the direction change characteristics in the continuous segments, evaluate the stability of each trajectory, and obtain the trajectory stability parameter set. Based on the trajectory direction change sequence, the system calculates the direction offset frequency for each trajectory set, that is, the frequency of direction changes exceeding a certain threshold in the statistical sequence. The system counts the number of times the path extends. Simultaneously, it statistically analyzes the continuity of the path extension and calculates the sequence of curvature radii. The system compares the directional change characteristics within continuous segments. If the first 50% of the trajectory has a low frequency of directional deviation and a large curvature radius, while the latter 50% has a high frequency and a small curvature radius, it is classified as a "stopping operation." If the deviation frequency is low throughout the entire path, it is classified as a "penetrating operation." The system evaluates the stability of each trajectory and defines stability parameters. ,in For the number of large-angle offset points, This represents the total number of points. For example, in a trajectory containing 100 points, only 5 points experience abrupt changes. Thus, the trajectory stability parameter set is obtained.

[0035] S413: Based on the trajectory stability parameter set, construct a direction vector cluster, analyze the intersection and change trend of each vector cluster in space, form a preference aggregation direction, identify the user's preference intention, and generate behavior-oriented intention data; Based on the trajectory stability parameter set, the system selects stability parameters. The system constructs directional vector clusters from high-quality trajectories. Using the K-means clustering algorithm, the system projects the principal direction vectors of all selected trajectories onto a unit sphere for clustering. Assuming the clustering results show that 80% of the stable trajectory vectors point to the "right front" sector in the MR scene (corresponding to the location of the virtual display shelf), the system analyzes the spatial convergence and change trends of each vector cluster to determine this sector as the user's "operation hotspot." Through the convergence and change trajectories of each directional cluster, the system forms a preference aggregation direction, such as "going deeper to the right front," and combines this with the obtained semantic weights to identify the user's preference intent as "tendent to explore sports products on the right side of the display shelf," generating behavioral guidance intent data.

[0036] Please see Figure 6 The specific steps to obtain the focus hotspot mapping configuration are as follows: S511: Based on spatial gaze hierarchy data and behavioral guidance intent data, analyze the spatial distribution of user gaze focus in MR scenes, and generate a set of gaze intensity indicators by recording the number of gaze points and focus range in each spatial unit. Based on spatial gaze hierarchy data and behavioral guidance intent data, the system begins to construct the final heatmap. The system first subdivides the MR scene space into high-resolution 3D meshes (e.g., The system analyzes the spatial distribution of user gaze focus in MR scenes. It iterates through all historical time slices, projecting gaze points marked as "highly stable focus" onto a grid. By recording the number of gaze points and the focus range (i.e., the dispersion volume of gaze points within the grid) in each spatial unit, a gaze intensity index is generated. For example, if 500 highly stable gaze points fall within grid Grid_1024, and these points are concentrated within a 2cm radius of the grid center, then the gaze point count index for this grid is 500, the focus range index is "highly concentrated," and a gaze intensity index set is generated.

[0037] S512: Based on the gaze intensity index set, calculate the gaze activity level of each spatial unit, and arrange the spatial units according to the gaze activity level to obtain the spatial priority sequence; Based on the gaze frequency index set, the system calculates the gaze activity level of each spatial unit. The calculation logic is as follows: ,in For the number of fixation points, This indicates whether the cell is located on the "preference aggregation direction" identified by S4 (1 if yes, 0 otherwise). and Preset coefficients. Example of coefficient setting: Setting... , If a grid has 500 fixation points located in the preferred direction, then The system calculates the values ​​of all grids. The spatial priority sequence is obtained by arranging the spatial units from high to low according to the degree of gaze activity and selecting the top 10% of high-scoring grids.

[0038] S513: Combine spatial priority sequences and content semantic classification tags to bind multiple priority regions, match content semantic tags according to gaze activity, form structured semantic association results, and generate gaze hotspot mapping configuration; Combining spatial priority sequences, the system extracts the virtual object IDs (e.g., "red sports car model") corresponding to high-priority regions. The system queries the content database to obtain the semantic classification tags for the object (e.g., "vehicle," "sports," "red," "high price"). The system binds multiple priority regions based on these content semantic classification tags; for example, it binds the highest priority region with the tag "sports vehicle." Subsequently, the system matches content semantic tags based on gaze activity levels. If the user's gaze activity is extremely high (…), the system will match the content semantic tags accordingly. If the activity level is moderate, a more granular tag (such as "aerodynamic components") will be matched; if the activity level is moderate, a generalized tag (such as "car") will be matched. The system generates a structured semantic association result, such as {User_Preference:[Sports_Car,Red,Aerodynamics]}, and adjusts the recommendation strategy accordingly, prioritizing the push of MR content with similar attributes in subsequent displays, and generating attention hotspot mapping configurations.

[0039] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method of MR content recommendation with fusion preference modeling, characterized in that, Includes the following steps: S1: Obtain the user's eye movement trajectory sequence, analyze the spatial coordinates and depth distribution of continuous points, calculate the density and dwell trend of fixation points in each spatial layer, construct a fixation concentration distribution model, and perform label classification based on spatial hierarchy to generate spatial fixation hierarchy structure data. S2: Based on the spatial gaze hierarchy structure data, obtain the user's hand trajectory point sequence, analyze the direction deviation and spatial dispersion of the trajectory, determine the continuity and change pattern of the action path, construct the action feature expression structure, classify the operation intensity, and generate interactive trajectory intensity feature information. S3: Combining the spatial gaze hierarchy structure data and interaction trajectory intensity feature information, analyze user voice data, determine the overlap density of multimodal points through spatial location mapping and time synchronization, construct a modal cross model, initialize weight labels for semantic content matching, and generate spatial fusion weight configuration. S4: Using the spatial fusion weight configuration, analyze the directional changes and path extension trends of the operation behavior trajectory, construct a directional vector cluster, determine the frequency of directional offset and path continuity characteristics, form a preference aggregation direction, and generate behavior guidance intention data. 2.The MR content recommendation method of fusion preference modeling according to claim 1, characterized in that, The spatial gaze hierarchy structure data includes a gaze region coordinate set, a hierarchy label set, and a depth-of-field continuous distribution feature. The interaction trajectory intensity feature information includes a trajectory coherence segment set, a direction offset feature set, and a motion amplitude classification label. The spatial fusion weight configuration includes a modality mapping coordinate set, a modality overlap ratio group, and a weight allocation label. The behavior guidance intent data specifically includes a trajectory direction vector set, a behavior offset path set, and a preference aggregation direction label. 3.The MR content recommendation method of fusion preference modeling according to claim 1, characterized in that, The specific steps for obtaining the spatial gaze hierarchy structure data are as follows: S111: Obtain the user's eye movement trajectory sequence, analyze the spatial coordinates and corresponding depth distribution of the user's continuous eye movement points in the MR scene, calculate the continuous density and dwell duration of the fixation point in each spatial layer, and obtain spatial fixation focusing data; S112: Based on the aforementioned spatial gaze focusing data, segment and locate the eye movement trajectory segments generated by the user in the MR scene, using the formula: ； The focus stability of each gaze path is calculated, and the gaze stability distribution coefficient is obtained by combining cumulative trend modeling. in, For the first The focusing stability of the eye movement path, For the first The eye movement path in the first Normalized gaze density in a spatial segment For the first The average normalized fixation density of each eye movement path For the first The dimensionless depth factor of a spatial segment For the first The eye movement path in the first Normalized lateral offset in a spatial segment For the first Normalized jump amplitude of a spatial segment during interlayer transition This is the index of the spatial segment in the path. This is the index of the eye-tracking path. This represents the number of spatial segments divided within a single eye-tracking path. S113: Call the gaze stability distribution coefficient, divide the range of numerical variation into intervals, construct the focus segment mapping region in three-dimensional space, and map the located focus interval markers to the hierarchical label group of the MR content region model, establish the gaze distribution association structure under the spatial hierarchy, and generate spatial gaze hierarchy structure data.

4. The MR content recommendation method based on fusion preference modeling according to claim 3, characterized in that, The specific steps for obtaining the interaction trajectory intensity feature information are as follows: S211: Based on the spatial gaze hierarchy data, obtain the user's hand trajectory point sequence, analyze the directional offset change of adjacent trajectory segments in continuous segments, calculate the directional vector change rate, determine the directional change trend of each segment, and obtain the directional coherence index. S212: Based on the directional coherence index, analyze the degree of dispersion of continuous segments in the spatial coordinate distribution of the trajectory path, calculate the discrete density of the three-dimensional coordinate distribution within the segment, and generate a stable trajectory distribution feature. S213: Based on the directional coherence index and trajectory stability distribution characteristics, and according to the combination relationship between trajectory continuity and directional change trend, construct an action feature expression structure, and calculate the trajectory operation intensity coefficient value based on the density of trajectory feature combinations in the structure, classify the operation intensity, and obtain interactive trajectory intensity feature information.

5. The MR content recommendation method based on fusion preference modeling according to claim 4, characterized in that, The specific steps for obtaining the spatial fusion weight configuration are as follows: S311: Obtain the spatial gaze hierarchy structure data and interaction trajectory intensity feature information, obtain and analyze the user's voice data, aggregate the coordinate information of multiple modal points in each spatial unit through spatial location mapping and time synchronization, count the number of multimodal points in each region, and establish a multimodal spatial distribution density group. S312: Based on the multimodal spatial distribution density group, by analyzing the spatial landing point distribution of multiple modes in the MR scene, and according to the degree of overlap density of multiple modal points in the same spatial location, a spatial crossover rate set is generated; S313: Based on the spatial overlap rate set and the distribution density of each modality in the intersection region, initialize weight labels for semantic input content and generate spatial fusion weight configuration.

6. The MR content recommendation method based on fusion preference modeling according to claim 5, characterized in that, The process of aggregating the coordinate information of multiple modal points within each spatial unit is specifically as follows: Acquire user voice data, convert it into text information through speech recognition, perform semantic analysis on the converted text, identify verbs in the instructions and the target entities they refer to, combine the spatial distribution information of objects in the MR system, map the target entities referred to in the text to entities in the MR space, and identify the spatial location of the target entities in the speech based on the spatial coordinate data of the entities, and output the spatial coordinate data of the entities referred to in the speech at multiple times. By combining the spatial gaze hierarchy structure data and the interaction trajectory intensity feature information, spatial coordinate data of each modality within each spatial unit are obtained, including eye movement trajectory, hand movement trajectory, and spatial coordinate data of the voice-referencing entity at multiple times. The modal point list of each spatial unit is classified and statistically analyzed according to the modal type. The number of points, time sequence range and coordinate centroid data of each mode in each spatial unit are recorded to establish the multimodal spatial distribution density group.

7. The MR content recommendation method based on fusion preference modeling according to claim 5, characterized in that, The specific steps for obtaining the behavior-oriented intent data are as follows: S411: Obtain the spatial fusion weight configuration, extract the start and end points of the trajectory of each behavior of the user during the MR content operation, combine the direction change frequency and path extension trend between trajectory points, construct a set of directional trajectories, and generate a trajectory direction change sequence. S412: Based on the trajectory direction change sequence, calculate the direction offset frequency of each trajectory set, and count the continuity state in the path extension. Compare the direction change characteristics in the continuous segments, evaluate the stability of each trajectory, and obtain the trajectory stability parameter set. S413: Based on the trajectory stability parameter set, construct a direction vector cluster, analyze the intersection and change trend of each vector cluster in space, form a preference aggregation direction, identify the user's preference intention, and generate behavior-oriented intention data.

8. The MR content recommendation method based on fusion preference modeling according to claim 1, characterized in that, The method further includes: S5: Based on the spatial gaze hierarchy structure data and behavioral guidance intent data, analyze the spatial distribution of user gaze focus, calculate the gaze intensity index by recording the number of gaze points and focus range of each spatial unit, establish regional priority ranking, bind regions in combination with content semantic tags, match content semantic tags according to gaze activity, and generate gaze hotspot mapping configuration. The gaze hotspot mapping configuration specifically refers to the priority region index, gaze activity indicator set, and semantic tag binding group.

9. The MR content recommendation method based on fusion preference modeling according to claim 8, characterized in that, The specific steps for obtaining the gaze hotspot mapping configuration are as follows: S511: Based on the spatial gaze hierarchy structure data and behavioral guidance intention data, analyze the spatial distribution of the user's gaze focus in the MR scene, and generate a gaze intensity index set by recording the number of gaze points and the focus range in each spatial unit; S512: Calculate the gaze activity level of each spatial unit according to the gaze activity level set, and arrange the spatial units according to the gaze activity level to obtain a spatial priority sequence; S513: Combining the spatial priority sequence with the content semantic classification tags, multiple priority regions are bound together. The content semantic tags are matched according to the gaze activity level to form a structured semantic association result and generate a gaze hotspot mapping configuration.

Images