A method and system for intelligently generating MR user portraits based on big data

Abstract

The application belongs to the technical field of mixed reality, and specifically discloses a MR user portrait intelligent generation method and system based on big data. The method and system of the application comprise the following steps: acquiring MR behavior data of a current session and at least two historical sessions, and respectively constructing a cross-session global heterogeneous user portrait graph and a current session event graph; performing bidirectional coupling representation updating between the two graphs, transmitting the global user representation to the session graph to update the session user representation, and converging the session graph representation to form a session side convergence representation, and feeding back the global user representation; and generating a user portrait according to the updated global and session representations. By utilizing the historical session data and the current session data to form complementary representations, the user portrait representation can simultaneously reflect long-term information across sessions and session information in the current session; by weighting and converging the node and edge representations on the session side and feeding back the global side representation, the portrait representation has an incremental updating capability.

Description

Technical Field

[0001] This invention relates to the field of mixed reality technology, and in particular to a method and system for intelligent generation of MR user profiles based on big data. Background Technology

[0002] With the development of head-mounted display devices, spatial awareness, and natural interaction technologies, MR systems enable users to continuously interact with virtual objects, spatial anchors / areas, and scene content in the real environment through gestures, gaze / look, voice, controller operations, and pose movement. These interactions generate a large amount of multi-source, heterogeneous behavioral data, characterized by multimodality, strong temporality, strong context, and the coexistence of noise and uncertainty. In practical applications, this behavioral data typically originates from continuous collection from a large number of users across devices and applications, converging into high-scale, high-frequency behavioral data sets in the form of log streams and event streams, exhibiting typical big data characteristics. Therefore, efficiently and stably completing user profile modeling and updating under large-scale behavioral data conditions has become a crucial foundation for the implementation of personalized MR services. To achieve personalized capabilities such as content presentation, interactive feedback, task guidance, and service recommendations, it is usually necessary to automatically generate user profiles reflecting user preferences and behavioral patterns from this behavioral data. However, MR interaction data has significant semantic and contextual differences between different sessions and is affected by factors such as sensor tracking errors, false triggers and occasional behaviors, making it difficult for user profiles to balance cross-session stability and real-time performance in the current session.

[0003] Existing user modeling / personal profiling technologies typically utilize historical behavioral data to form user representations, which are then used for personalized decision-making on the application side. However, in MR scenarios, user behavior not only includes discrete actions such as object clicks or content browsing, but also involves complex relationships with multiple entities such as space, objects, and tasks. Modeling using only sequence or aggregate statistics methods often fails to express the interaction structure of event-object-space-context. Furthermore, session-oriented modeling methods often focus on learning temporal dependencies or local structures within a single session, making it difficult to effectively utilize long-term preference information across sessions; while modeling methods that focus on global history may ignore short-term intent changes within the current session, leading to delayed profile updates or insensitivity to the current session.

[0004] In recent years, graph structure modeling has been used to represent multiple entities and their relationships, and combined with sequence modeling to characterize temporal changes within a session, thereby capturing both long-term preferences and short-term intentions. However, in existing solutions, the global graph and session modeling often exhibit a one-way utilization relationship; for example, global history is used to assist in the initialization or prediction of the current session. When new high-confidence interaction evidence is generated on the session side, how to feed it back into the global profile structure in a controllable manner to achieve incremental updates still lacks a unified and clear closed-loop mechanism. Furthermore, facing the coexistence of session sparsity and noisy behavior, how to adaptively adjust the contributions of different nodes or edges during cross-graph information transmission and session convergence to improve the robustness and consistency of profile representation remains a key issue that needs to be addressed.

[0005] Therefore, under the conditions of multi-source heterogeneous, strong temporal and contextual, and noisy behavioral big data, it is still an urgent technical problem to be solved in this field to simultaneously take into account the stability of long-term information across sessions and the agility of current session representation, and to suppress the interference caused by noisy evidence during the profile update process. Summary of the Invention

[0006] The purpose of this invention is to overcome the difficulty of existing user profile generation methods in simultaneously ensuring the stability of long-term cross-session information and the agility of current session representation, and to provide an intelligent user profile generation scheme for MR big data. This scheme can simultaneously construct a global heterogeneous user profile graph across sessions and a current session event graph, and perform bidirectional coupled representation updates between the two, thereby improving cross-session consistency without sacrificing session real-time performance.

[0007] To achieve the above objectives, this invention provides a method for intelligently generating MR user profiles based on big data, which specifically includes: Acquire MR user behavior data in the current session and at least two historical sessions; Construct a global heterogeneous user profile graph across sessions using MR user behavior data from at least two historical sessions; Construct a session event graph using MR user behavior data from the current session; Perform bidirectional coupled representation update processing between the global heterogeneous user profile graph and the session event graph; User profile representations are generated based on the updated global user representation and the updated session user representation; The bidirectional coupling representation update process includes: passing the global user representation in the global heterogeneous user profile graph to the session event graph as the representation input for nodes and edges in the session event graph to obtain the updated session user representation; and aggregating the node and edge representations of the session event graph to obtain the session-side aggregated representation, and performing feedback update on the global user representation in the global heterogeneous user profile graph based on the session-side aggregated representation to obtain the updated global user representation.

[0008] More preferably, the global heterogeneous user profile graph includes at least: user nodes, entity nodes related to MR interaction, and multi-type relationship edges; The entity nodes include at least two of the following categories: session nodes, interactive event nodes, virtual object nodes, spatial anchor points or region nodes, scene nodes, content nodes, and task nodes; The multi-type relationship edge is used to describe the association relationship category between at least two selected entity nodes. Each relationship edge includes a relationship type identifier to characterize the association relationship category, and also includes at least one relationship attribute among timestamp, duration, confidence level, and pose information parameters.

[0009] More preferably, the session event graph includes: An event node and at least one of its associated object nodes and spatial nodes; At least one of the following: a temporal association edge used to characterize the order in which events occur; an interactive association edge used to characterize the relationship between an event and an object; and a contextual association edge used to characterize the relationship between an event and its spatial context.

[0010] More preferably, before performing bidirectional coupled representation update processing, the initial representations of nodes and edges in the session event graph are set based on the global representation vectors of user nodes in the global heterogeneous user profile graph and entity nodes corresponding to nodes in the session event graph, so as to obtain an initialized session event graph. The initial representation setting includes: generating initialization vectors for corresponding nodes and edges, and using the initialization vectors as the initial representations of the nodes and edges; The generation of the initialization vector includes: transforming the global representation vector through mapping to obtain the initialization vector in the session representation space, or fusing the global representation vector with the event type features, context features, and relation type features corresponding to the edges of the current session to generate the initialization vectors of the corresponding nodes and edges.

[0011] More preferably, in the bidirectional coupled representation update process, when the global user representation in the global heterogeneous user profile graph is passed to the session event graph, the corresponding injection weights are determined for the nodes and edges in the session event graph, which are used to determine the proportion of the global user representation participating in the corresponding node and edge representation update. The injected weights are generated based on at least one piece of evidence quality information from the following: confidence level of the interaction event, duration of the interaction event, time interval between adjacent interaction events, spatial stability index, pose tracking quality index, interaction event type identifier, and context category identifier.

[0012] More preferably, in the bidirectional coupled representation update process, the convergence weights of the node representations and edge representations in the session event graph are determined, and the node representations and edge representations are weighted and converged according to the convergence weights to obtain the session-side converged representation. The convergence weight is generated based on at least one piece of evidence quality information from the following: confidence level of interaction event, duration of interaction event, time interval between adjacent interaction events, spatial stability index, pose tracking quality index, interaction event type identifier, relationship type identifier, and context category identifier. The feedback update includes at least one of the following methods: incrementally updating the global user representation vector corresponding to the user node in the global heterogeneous user profile graph; incrementally updating the representation vector corresponding to the entity node related to the current session in the global heterogeneous user profile graph; updating the relation attributes or relation weights of multi-type relation edges in the global heterogeneous user profile graph; and incrementally adjusting the adjacency relations in the global heterogeneous user profile graph.

[0013] More preferably, the determination of the injection weight or the aggregation weight includes an adaptive adjustment process, which includes: constructing at least two perturbation views based on the session event graph; performing bidirectional coupled representation update processing on the at least two perturbation views to obtain corresponding session user representations and session-side aggregation representations; calculating the consistency differences of the session user representations and session-side aggregation representations corresponding to the at least two perturbation views; and adaptively correcting the injection weights and the aggregation weights based on the consistency differences of the session user representations. The perturbation view is obtained by applying a perturbation operation to the session event graph, which is either a structural perturbation or a feature perturbation.

[0014] More preferably, the generation of user profile representation includes: determining a fusion coefficient between the updated global user representation and the updated session user representation, and generating a profile vector based on the fusion coefficient; The process of generating the profile vector includes: performing representation space alignment processing on the updated global user representation and the updated session user representation respectively to obtain a global profile sub-vector and a session profile sub-vector; and performing a weighted combination of the global profile sub-vector and the session profile sub-vector based on the first fusion coefficient and the second fusion coefficient to obtain the profile vector. The fusion coefficients include a first fusion coefficient for adjusting the contribution of global profile sub-vectors and a second fusion coefficient for adjusting the contribution of session profile sub-vectors, wherein the first fusion coefficient and the second fusion coefficient satisfy normalization constraints.

[0015] More preferably, it also includes: writing the image vector into the image storage structure, and performing incremental updates on the historical image vectors in the image storage structure; The incremental update includes at least one of the following update strategies: weighting the historical profile vector according to a preset time decay rule; storing the historical profile vector according to a versioning rule and recording the current profile vector as the new version; maintaining user profile representations according to scenario categories and writing the current profile vector into the user profile representation of the scenario category corresponding to the current session.

[0016] In addition, the present invention also provides an intelligent generation system for MR user profiles based on big data, the system specifically comprising: The data acquisition module acquires MR user behavior data in the current session and at least two historical sessions; The global user profile construction module uses MR user behavior data from at least two historical sessions to construct a cross-session global heterogeneous user profile graph and obtains the global user representation in the global heterogeneous user profile graph. The session event graph construction module uses the MR user behavior data of the current session to construct a session event graph; The bidirectional coupling module performs bidirectional coupling representation update processing between the global heterogeneous user profile graph and the session event graph. The bidirectional coupling representation update processing includes: passing the global user representation to the session event graph as the representation input for nodes and edges in the session event graph to obtain the updated session user representation; and aggregating the node and edge representations of the session event graph to obtain the session-side aggregated representation, and performing feedback update on the global user representation based on the session-side aggregated representation to obtain the updated global user representation. The profile generation module is used to generate user profile representations based on the updated global user representation and the updated session user representation.

[0017] Compared with the prior art, the present invention has at least one or more of the following beneficial effects: By utilizing historical and current session data to form complementary representations, user profile representations can simultaneously reflect long-term information across sessions and session information of the current session. By weighting and aggregating node and edge representations on the session side and updating the global representations, the profile representation is made capable of incremental updates. By adaptively adjusting the weights based on the consistency differences of the perturbation view, the impact of low-reliability evidence on the portrait generation process can be reduced, and the stability and consistency of the portrait representation can be improved. By using fusion coefficients, a controllable combination of global representation and session representation can be achieved, which facilitates the generation of user profile vectors that are adapted to different application scenarios, and can further support profile storage and historical profile update management. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the operation of the intelligent generation method for MR user profiles in an embodiment of the present invention.

[0019] Figure 2 This is a simplified structural diagram of the MR user profile intelligent generation system in an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the present invention is not limited to the following embodiments. Equivalent substitutions or modifications made by those skilled in the art without departing from the spirit and substance of the present invention should be considered to fall within the protection scope of the present invention, which is defined by the appended claims.

[0021] To facilitate a unified description and understanding of the technical solution of this invention, the following conventions apply to some terms in the following embodiments: A continuous interaction process of a user in an MR environment can be defined as a session, where the current session refers to the current session data corresponding to the profile to be generated, and the historical session refers to previous session data that is different from the current session and can be traced; the global heterogeneous user profile graph is used to represent entities such as users, interaction events, virtual objects, spatial anchors / regions, scenes, content, and tasks and their relationships in at least two historical sessions; the session event graph is used to represent events within the current session and their relationships with objects and spatial context; global user representation and session user representation refer to the vectorized representations obtained from the global heterogeneous user profile graph and the session event graph, respectively. The above terms are only used to illustrate the descriptive convenience of the embodiments of this invention and should not be construed as limiting the scope of protection of this invention.

[0022] In MR applications, user interactions typically exhibit characteristics of multi-source heterogeneity, strong temporal sequence, and strong spatial context. On the one hand, the preferences and habits of the same user in different sessions have a certain stability; on the other hand, the current session often contains immediate intentions related to task objectives, scene content, or spatial location. If modeling relies solely on single session data, the representation is prone to instability due to session fluctuations; if modeling relies solely on historical data, it may be difficult to reflect changes in intent within the current session in a timely manner. Based on this, this invention, from the perspective of collaborative representation of long-term information across sessions and short-term evidence within sessions, organizes historical session data and current session data into computable graph structures, and establishes iterative association and update relationships between the two types of graph structures. This maintains the continuity of the user profile representation while also considering the characterization of the current conversation context. The following will describe the intelligent user profile generation process of this invention with specific embodiments, and in subsequent paragraphs, will elaborate on key aspects such as the construction of a global heterogeneous user profile graph, the construction of a session event graph, bidirectional coupled representation update processing, and the generation, storage, and update of user profile representations.

[0023] like Figure 1 As shown, in one embodiment, the present invention provides a method for intelligent generation of MR user profiles based on big data, the method specifically including the following steps: Acquire MR user behavior data in the current session and at least two historical sessions; Construct a global heterogeneous user profile graph across sessions using MR user behavior data from at least two historical sessions; Construct a session event graph using MR user behavior data from the current session; Perform bidirectional coupled representation update processing between the global heterogeneous user profile graph and the session event graph; User profile representations are generated based on the updated global user representation and the updated session user representation; The bidirectional coupling representation update process includes: passing the global user representation in the global heterogeneous user profile graph to the session event graph as the representation input for nodes and edges in the session event graph to obtain the updated session user representation; and aggregating the node and edge representations of the session event graph to obtain the session-side aggregated representation, and performing feedback update on the global user representation in the global heterogeneous user profile graph based on the session-side aggregated representation to obtain the updated global user representation.

[0024] MR user behavior data refers to recorded data that characterizes a user's interactions with virtual objects, spatial locations / areas, and scene content in an MR environment. This data can be collected by the MR terminal or platform and generated into logs or record entries. Each record typically includes at least the user identifier, interaction time, interaction object identifier, and its contextual information. Spatial pose information or interaction type information may also be included when necessary.

[0025] A cross-session global heterogeneous user profile graph is a graph structure representation used to organize user-related information from at least two historical sessions. The purpose of this graph structure is to link scattered interaction records from historical sessions in the form of entities and relationships, enabling historical behavior to be expressed and computed in a structured manner. By processing this global graph, a global user representation can be obtained, which can be understood as a vectorized representation of a user's long-term preferences, habits, or relationships across sessions.

[0026] A session event graph is a graph structure used to organize the interaction process within the current session. Unlike global graphs, which focus on long-term information across sessions, session event graphs characterize the interaction events and their relationships within the current session, enabling the immediate intent and context embodied in the current session to be expressed and computed. By processing the session event graph, a session user representation can be obtained, which can be understood as a vectorized representation of the user's behavioral characteristics and conversational context within the current session.

[0027] In the bidirectional coupling representation update process between the global heterogeneous user profile graph and the session event graph, the bidirectional coupling manifests as linkage in two directions. First, the global user representation in the global heterogeneous user profile graph is passed to the session event graph as the representation input for nodes and edges in the session event graph, enabling the session event graph to utilize long-term information reflected in historical sessions during representation calculation. Second, the node and edge representations of the session event graph are aggregated to form a session-side aggregated representation, and based on this session-side aggregated representation, the global user representation in the global heterogeneous user profile graph is updated. This allows the historical profile to absorb newly generated interaction evidence in the current session, achieving continuous evolution across session profiles. Through this bidirectional coupling process, this embodiment can simultaneously obtain updated session user representations and updated global user representations, and generate user profile representations based on these. The user profile representation can be specifically represented in vector form to support subsequent personalized services or user behavior analysis.

[0028] By organizing historical and current sessions into computable structured representations and introducing a global user representation as input on the session side, the generated session user representation not only reflects interaction evidence within the current session but also possesses cross-session information support, thereby improving the continuity of the user profile representation. Furthermore, by aggregating the session-side representation to perform feedback updates on the global user representation, the global user representation can be iteratively updated along with session evidence, avoiding the problem of the global profile remaining static and unchanging, leading to a disconnect from current behavior. Ultimately, the user profile representation generated based on the updated global user representation and the updated session user representation can take into account both long-term and session information, providing an achievable technical path for user profile generation in MR scenarios.

[0029] MR terminals or platforms typically record user interactions and system status information during application operation, forming raw behavior logs. These raw behavior logs include at least user identifiers, timestamps, interaction-related information (e.g., interaction trigger source / interaction type / interaction object identifier), and scene / application status information at the time of the interaction. In MR devices with spatial awareness capabilities, they may also include pose information (e.g., head-mounted display pose, hand / controller pose, gaze direction), spatial anchor point identifiers, or spatial region identifiers. To facilitate subsequent processing, these raw fields can be uniformly mapped to a preset data structure, such as storing them as log entries or record tuples, and ensuring consistency in timestamp units, coordinate systems, and object identifier encoding methods. For missing fields, null values ​​can be used to mark them, and the original record can be retained to avoid making unnecessary assumptions about the data.

[0030] In this embodiment, a session represents a continuous interaction process of a user in an MR environment. Sessions can be divided based on the time sequence of the original records. After sorting the behavior records of the same user by timestamp, the continuous records are divided into multiple session segments according to a preset session segmentation rule. The session segmentation rule can split sessions based on the following triggering conditions: the time interval between two adjacent records exceeds a preset threshold; the application scenario / task state changes, such as a change in scenario ID, a change in task stage, or a switch between foreground and background; the user enters / exits the MR application or the device enters / exits immersive mode. After segmentation, the most recent session segment can be defined as the current session, and the remaining session segments that meet the time sequence and are different from the current session can be defined as historical sessions, thus obtaining the data basis for the current session and at least two historical sessions.

[0031] Within a session, to form events that can be used for subsequent graph structure representation, the original records can be further organized into event units. An event unit can be understood as a recognizable interaction or state change process. Its construction can be as follows: when the original record itself is recorded according to interaction triggers, a single or a small number of related records can be merged into one event unit; when the original record is a continuously sampled signal, such as a pose trajectory or gaze trajectory, the start and end of the event can be determined based on trigger conditions (e.g., selection / grabbing / releasing / voice commands) or state change conditions (e.g., from gaze to departure, from contact to separation), thus forming an event unit. Each event unit includes at least: an event identifier, event occurrence time information (start and end times), the session identifier to which the event belongs, and the object identifier and context identifier associated with the event; if necessary, pose information or system state information at the time of the event can also be recorded. Through the above methods, an event set can be formed in the current session, providing a source of basic nodes for the construction of the subsequent session event graph.

[0032] To ensure events can be clearly associated with contexts such as objects, spaces, and scenes / tasks, the following elements can be obtained based on the asset management and spatial awareness capabilities of MR applications: virtual object identifiers, spatial anchor / region identifiers, and scene / content / task identifiers. Regarding virtual object identifiers, MR applications typically maintain unique identifiers for virtual objects, such as object IDs or resource IDs. When an event occurs, its associated object identifier can be recorded. For objects reused across sessions, persistent identifiers or unified identifiers mapped by application logic can be used. Regarding spatial anchor / region identifiers, MR devices or applications can provide identifiers for spatial anchors or spatial regions, such as spatial anchor IDs, room region IDs, grid / plane partition IDs, etc. When an event occurs, it can be associated with the spatial anchor / region based on the location of the event or the location of the object, forming the spatial context identifier of the event. Regarding scene / content / task identifiers, MR applications typically have corresponding scene IDs, content IDs, or task IDs, which can be directly obtained from event records or session states, used to characterize the application context in which the event occurs.

[0033] By performing the above preprocessing on the raw data, we obtain the session set and session boundary information of the current session and at least two historical sessions, the event set corresponding to the current session, and the entity identifier set associated with the events. It may also include attribute fields such as pose information, duration, and confidence level, along with their standardized storage format. This set of basic elements provides the data input source and definition basis for subsequent construction of a global heterogeneous user profile graph, a session event graph, and the execution of bidirectional coupled representation update processing.

[0034] In this embodiment, based on the aforementioned set of basic elements of the output, user-related information from at least two historical sessions is organized into a global heterogeneous user profile graph across sessions. The global heterogeneous user profile graph includes at least: user nodes, entity nodes related to MR interaction, and multi-type relationship edges. The entity nodes include at least two of the following categories: session nodes, interactive event nodes, virtual object nodes, spatial anchor points or region nodes, scene nodes, content nodes, and task nodes; The multi-type relationship edges are used to describe the association category between at least two selected entity nodes. Each relationship edge includes a relationship type identifier to represent the association category, and also includes at least one relationship attribute selected from timestamp, duration, confidence level, and pose information parameters. Specifically, the user to be modeled is instantiated as a user node; at least two types of entity identifiers are selected from session identifiers, event identifiers, object identifiers, spatial anchor point / region identifiers, and scene / content / task identifiers and instantiated as entity nodes related to MR interaction; based on the interaction records and context information in historical sessions, multi-type relationship edges are established between user nodes and entity nodes, and between entity nodes, to represent the association category between entities. The relationship edge carries at least a relationship type identifier and further carries at least one relationship attribute selected from timestamp, duration, confidence level, and pose information parameters.

[0035] In this embodiment, the interaction process in the current session is further organized into a session event graph to facilitate bidirectional coupled representation update processing with the global heterogeneous user profile graph. The session event graph includes: An event node and at least one of its associated object nodes and spatial nodes; At least one of the following: a temporal association edge used to characterize the order in which events occur; an interactive association edge used to characterize the relationship between an event and an object; and a contextual association edge used to characterize the relationship between an event and its spatial context.

[0036] The formation of the session event graph can be directly completed based on the current session event set and its associated information. Specifically, the implementation steps can be summarized as follows: 1) Forming nodes: Instantiate event units in the current session as event nodes, and select at least one of the object identifier and spatial context identifier associated with the event unit as an auxiliary node type, instantiating it as an object node or spatial node; 2) Forming edges: Establish temporal association edges between event nodes based on the temporal order of event units to represent the order of event occurrence; establish interactive association edges between event nodes and object nodes based on the correspondence between events and objects to represent the effect of events on objects; establish context association edges between event nodes and spatial nodes based on the spatial context at the time of event occurrence to represent the association between events and spatial context. At least one of the above edge types can be selected based on the availability of the current session data. Furthermore, to facilitate subsequent representation calculations, attribute fields such as time information, duration, confidence level, and spatial pose information related to the current session can be added to the nodes or edges. These attribute fields can be directly derived from the normalized output of the aforementioned data preprocessing stage.

[0037] The above processing yields a session event graph, which represents events within the current session and their relationships with objects and spatial context. Subsequently, a bidirectional coupled representation update process will be performed based on this session event graph and the global heterogeneous user profile graph to obtain updated session user representations and updated global user representations, and further generate user profile representations.

[0038] In this embodiment, after obtaining the session event graph, before performing the bidirectional coupled representation update process, the session event graph can be initialized to enable subsequent representation updates to be carried out in a unified representation space and to allow session-side node and edge representations to carry initial references across session history information. Specifically, before performing the bidirectional coupled representation update process, based on the global representation vectors of user nodes in the global heterogeneous user profile graph and the entity nodes corresponding to nodes in the session event graph, the initial representations of nodes and edges in the session event graph are set to obtain an initialized session event graph. The initial representation setting includes: generating initialization vectors for corresponding nodes and edges, and using the initialization vectors as the initial representations of the nodes and edges; The generation of the initialization vector includes: transforming the global representation vector through mapping to obtain the initialization vector in the session representation space, or fusing the global representation vector with the event type features, context features, and relation type features corresponding to the edges of the current session to generate the initialization vectors of the corresponding nodes and edges.

[0039] The above initialization process can be understood as follows: based on establishing entity correspondence between the global heterogeneous user profile graph and the session event graph, the global representation information from the historical side is injected into the node and edge representations of the session side in the form of initialization vectors, thereby obtaining the initialized session event graph. Further, based on the consistency of entity identifiers or preset mapping rules, the correspondence between some nodes in the session event graph and entity nodes in the global heterogeneous user profile graph can be determined. For example, object nodes, spatial nodes, and scene / content / task-related nodes in the current session can be considered as corresponding entity nodes to nodes with the same entity identifiers in the global heterogeneous user profile graph. For session nodes whose corresponding entity identifiers cannot be found in the global heterogeneous user profile graph, a default vector or a vector generated from the current session features can be used as its initialization vector to ensure that the representation of the session event graph can be completely defined.

[0040] After determining the correspondence, initialization vectors for nodes and edges are generated and used as the initial representations of nodes and edges in the session event graph. The generation of these initialization vectors includes two methods: First, performing a mapping transformation on the global representation vector obtained from the global heterogeneous user profile graph, converting it from the global representation space to the session representation space, to obtain initialization vectors for session-side nodes or edges; second, fusing the global representation vector with the event type features and context features of the current session to generate initialization vectors that better fit the current session context. For edge initialization, the relationship type features corresponding to the edge can be further fused so that the edge representation can reflect the type information of the association relationships in the current session. Through the above initialization process, an initialized session event graph can be obtained, providing initial representation input on the session side for the subsequent bidirectional coupled representation update process.

[0041] In this embodiment, after initializing the session event graph, the core step of the bidirectional coupling representation update process can be entered. The global user representation from the global heterogeneous user profile graph is transferred to the session event graph to achieve continuous support for the current session representation calculation from long-term cross-session information. It should be noted that if global information is injected indiscriminately during this transfer process, session elements corresponding to low-reliability or sporadic interactions may also receive strong global prior participation, thus introducing unnecessary offsets during session representation updates. Therefore, this embodiment introduces selective injection control for session-side nodes and edges when transferring global user representations to the session event graph, so that the degree of global information participation can vary with the quality of session evidence. In the bidirectional coupling representation update process, when transferring global user representations from the global heterogeneous user profile graph to the session event graph, corresponding injection weights are determined for nodes and edges in the session event graph to determine the proportion of global user representation participation in the corresponding node and edge representation updates. The injected weights are generated based on at least one piece of evidence quality information from the following: confidence level of the interaction event, duration of the interaction event, time interval between adjacent interaction events, spatial stability index, pose tracking quality index, interaction event type identifier, and context category identifier.

[0042] The injection weight can be understood as a coefficient determined for each node or edge in the session event graph, used to adjust the participation ratio of the global user representation in the representation update of that node or edge. The injection weight applies to node types such as event nodes, object nodes, and spatial nodes in the session event graph, as well as edge types such as temporally related edges, interactively related edges, and contextually related edges. Furthermore, the injection weight can be determined independently for both nodes and edges, allowing the intensity of the introduction of different types of session elements into global information to be controlled separately. Based on this, during subsequent session-side representation updates, the global user representation can be applied to the update process of the corresponding node or edge representation with different participation ratios according to the injection weight, thereby achieving controllable injection from the global representation to the session representation.

[0043] Furthermore, the generation of injection weights is based on evidence quality information. This evidence quality information can be obtained directly from the aforementioned normalized output or derived from it. For example, the confidence level of an interaction event can be derived from the event validity determination result of the interaction recognition or tracking module; the duration of an interaction event can be calculated from the event start and end times; the time interval between adjacent interaction events can be calculated from the timestamp difference between adjacent events; spatial stability indicators and pose tracking quality indicators can be derived from quality outputs such as spatial anchor stability assessment, tracking state quality assessment, or pose estimation residuals; and interaction event type identifiers and context category identifiers can be mapped from application-side event enumeration types, scene / task states, or context fields. Through the aforementioned evidence quality information, corresponding injection weights can be formed for each session node or edge, enabling the global information injection process to have an adjustment basis consistent with the session evidence quality.

[0044] In specific implementation, the generation method of injection weights is not limited. This embodiment can be implemented in at least one of the following ways: generating injection weights based on single evidence quality information, such as generating injection weights based only on the confidence of interaction events or only on the pose tracking quality index; generating injection weights based on the fusion of multiple evidence quality information, such as weighting and combining interaction event confidence, duration, and spatial stability index to generate injection weights; generating injection weights according to node / edge type, such as prioritizing interaction event confidence and duration for event nodes, prioritizing spatial stability index and pose tracking quality index for spatial nodes, and prioritizing context category identifier for context-related edges.

[0045] When using a single evidence object to generate injected weights, a type of evidence quality information can be selected as input for each session node or edge, and the injected weights can be output through preset mapping rules. Specifically, this can include: 1. Normalizing or intervalizing the selected evidence quality information to make it fall within a preset numerical range, such as setting a range of 0 to 1, so that data from different sources have a consistent numerical scale; 2. Mapping the normalized evidence value to the injected weight using a monotonic mapping method, such as using linear mapping, piecewise mapping, or threshold mapping, so that the higher the evidence value, the greater the injected weight; 3. Applying upper and lower bound restrictions or saturation constraints to the output injected weights to avoid extreme values ​​causing unstable updates; 4. When a session node or edge lacks the evidence quality field, a default weight, nearest neighbor substitution, or substitution based on the statistical value of elements of the same type can be used to ensure that the injected weight can be determined for all target elements.

[0046] When generating injection weights by combining multiple evidence objects, at least two pieces of evidence quality information can be selected as inputs for each session node or edge, and the injection weights can be output through evidence fusion rules. Specifically, this can include: 1. Combining at least two of the following as evidence input sets: interaction event confidence, duration, time interval between adjacent events, spatial stability index, pose tracking quality index, interaction event type identifier, and context category identifier. For discrete class identifiers, category encoding or embedding representation can be used as input; 2. Evidence fusion can employ weighted summation fusion, multiplicative fusion, gating fusion, etc., where the weighting coefficients or rule parameters can be given by preset configuration or obtained based on historical data statistics; 3. After fusion output, normalization and boundary constraints are also applied to the injection weights to obtain stable injection weights; 4. When some evidence is missing, the weights of existing evidence can be reallocated. When different pieces of evidence contradict each other in reliability assessment, such as high confidence but low pose quality, priority rules or penalty rules can be used to make the injection weights more consistent with engineering reliability judgments.

[0047] The above implementation method adapts the injected weights to the reliability, interaction nature, and context category of the MR session data, thereby providing controllable global information input for the session-side representation update in the subsequent bidirectional coupled representation update.

[0048] In this embodiment, after the controllable transfer of the global user representation to the session event graph is completed, another key step in the co-enhancement closed loop is taken: extracting the session-side converged representation that can represent the current session from the session event graph, and performing feedback updates on the global heterogeneous user profile graph based on this session-side converged representation. This process is used to enable the global long-term profile to absorb relatively reliable interaction evidence from the current session, thereby achieving feedback correction from the session to the global level and avoiding the global profile from remaining static and out of touch with the latest behavior. It should be further noted that different events, objects, and their associated edges within a session may differ in reliability and representativeness. Therefore, this embodiment introduces convergence weights during the convergence process to achieve the screening and emphasis of session evidence. Specifically, in the bidirectional coupled representation update process, convergence weights are determined for the node representations and edge representations in the session event graph, and the node representations and edge representations are weighted and converged according to the convergence weights to obtain the session-side converged representation. The convergence weight is generated based on at least one piece of evidence quality information from the following: confidence level of interaction event, duration of interaction event, time interval between adjacent interaction events, spatial stability index, pose tracking quality index, interaction event type identifier, relationship type identifier, and context category identifier. The feedback update includes at least one of the following methods: incrementally updating the global user representation vector corresponding to the user node in the global heterogeneous user profile graph; incrementally updating the representation vector corresponding to the entity node related to the current session in the global heterogeneous user profile graph; updating the relation attributes or relation weights of multi-type relation edges in the global heterogeneous user profile graph; and incrementally adjusting the adjacency relations in the global heterogeneous user profile graph.

[0049] The convergence weights can be understood as weighting coefficients determined for node or edge representations in the session event graph, used to adjust the contribution of different session elements in forming the converged representation of the session. The objects for determining the convergence weights include, but are not limited to, node types such as event nodes, object nodes, and spatial nodes, as well as edge types such as temporally related edges, interactively related edges, and contextually related edges. Convergence weights can be determined separately for nodes and edges to consider the representativeness of the event / object / space itself, as well as the representativeness of the relationships between events or between events and the context, when converging session evidence.

[0050] The aggregation weights are generated based on evidence quality information, which can be directly obtained from the normalized output of the data preprocessing stage or derived from it. For example, the confidence level, duration, and time interval between adjacent events of an interaction event can be calculated from event record fields or timestamps; spatial stability indicators and pose tracking quality indicators can be derived from quality outputs such as spatial anchor stability assessment, tracking quality assessment, or positioning residuals; interaction event type identifiers and context category identifiers can be mapped from event / scene / task status fields; and relationship type identifiers can be obtained from the type encoding of edges in the session event graph or from the semantic mapping of edges. Through the above evidence quality information, a corresponding aggregation weight can be formed for each node or edge in the session event graph, making the session-side aggregation representation more inclined to be composed of highly credible and representative session evidence.

[0051] The aforementioned session-side convergent representation is used to represent the overall evidence summary or session-level intent summary of the current session. Its formation process can be summarized as follows: weighted convergence of node representations and edge representations in the session event graph is performed based on convergence weights to obtain the session-side convergent representation. In practical applications, the session-side convergent representation can be formed using at least one of the following convergence methods: weighted convergence of node representations in the session event graph, where the weights are generated from the evidence quality information corresponding to the nodes; weighted convergence of edge representations in the session event graph, where the weights are generated from the evidence quality information corresponding to the edges; weighted convergence of node representations and edge representations separately, and combination of the two convergence results to form the session-side convergent representation, thereby simultaneously preserving the evidence contribution of key elements and key relationships within the session.

[0052] Through the aforementioned weighted aggregation, the session-side aggregation representation can better reflect the key and reliable behavioral evidence in the current session compared to simple average or indiscriminate aggregation, providing an input basis for subsequent feedback updates to the global heterogeneous user profile graph.

[0053] In the above embodiments, performing feedback updates on the global heterogeneous user profile graph based on the session-side converged representation can be understood as writing the session evidence extracted in the current session back to the global side in an incremental manner, so that the global user representation and its related structure are updated as the session evolves.

[0054] Optionally, the session-side aggregated representation can be used as the session evidence vector to incrementally update the global user representation vector corresponding to the user node in the global heterogeneous user profile graph. For example, a new user node representation can be formed by fusing historical representations with session evidence, so that the updated global user representation maintains long-term stability while absorbing new evidence from the current session. The fusion can adopt commonly used engineering methods such as weighted combination, gated fusion, or sliding window update.

[0055] Optionally, based on the object nodes, spatial nodes, and scene / content / task-related nodes appearing in the current session event graph, the set of entity nodes corresponding to them in the global heterogeneous user profile graph is determined, and the representation vectors of the entity nodes are incrementally updated. This approach enables the global side to update not only the overall user representation but also the representations of object / spatial / context entities strongly related to the current session, so that the association calculations of the same entities in subsequent sessions have the updated prior knowledge.

[0056] Optionally, the evidence of relationships reflected in the current session can be written back to the multi-type relationship edges on the global side. This includes updating relationship attributes, such as updating the most recent interaction timestamp, cumulative duration, and overall confidence level, and updating relationship weights, such as increasing the weight of relationships corresponding to frequent and high-confidence interactions and decreasing the weight of relationships corresponding to low-reliability or occasional interactions. This approach enables the relationship layer of the global graph to reflect the latest session evidence, supporting subsequent cross-session relationship inference and profile evolution.

[0057] Optionally, when new entity-entity or user-entity associations appear in the current session, corresponding relationship edges can be added or the adjacency connection strength can be adjusted in the global heterogeneous user profile graph; when certain associations are missing for a long time or the evidence quality remains low, the corresponding connection strength can be reduced or they can be deleted, thereby achieving incremental adjustment of global adjacency relationships. This approach allows the global graph structure to evolve incrementally with the session, rather than just updating node vectors.

[0058] To implement the above-mentioned multi-selection convergence weight strategy, the specific generation process of the convergence weight may include the following steps: 1. Extract at least one piece of evidence quality information for each session node or edge. The evidence quality information includes at least one of the following: interaction event confidence, duration, time interval, spatial stability, pose tracking quality, event type identifier, relationship type identifier, and context category identifier. The discrete category identifier can be represented by category encoding or embedding. 2. Normalize / intervalize continuous evidence and encode and map discrete evidence to ensure that different evidences have a scale basis for fusion. 3. The convergence weight can be generated by single evidence mapping or by multi-evidence fusion rules. The fusion rules can be weighted summation, multiplicative combination, gated fusion, etc. 4. Apply upper and lower bound restrictions to the output convergence weight. When some evidence is missing, default values, statistical values ​​of the same type can be used as substitutes, or the weights of existing evidence can be redistributed to ensure that all target nodes / edges can obtain convergence weights. 5. Perform weighted convergence on the node representation and edge representation based on the generated convergence weight to obtain the session-side convergence representation, and use the convergence representation in the above-mentioned feedback update process.

[0059] MR session data may be subject to uncertainties such as noise interaction, false triggering, and tracking drift during actual acquisition. If the injection weights or convergence weights remain fixed throughout the iteration process or are statically determined by a single piece of evidence, low-reliability evidence may be repeatedly amplified in the closed loop, leading to instability in the session representation or global representation. Therefore, this embodiment further introduces an adaptive weight adjustment process to automatically correct the injection weights or convergence weights based on the stability of the session data without changing the overall closed-loop structure, thereby improving the robustness of the closed-loop update. Specifically, the adaptive adjustment process for injection weights or convergence weights includes: constructing at least two perturbation views based on the session event graph; performing bidirectional coupled representation update processing under each of the at least two perturbation views to obtain corresponding session user representations and session-side convergence representations; calculating the consistency differences between the session user representations and session-side convergence representations corresponding to the at least two perturbation views; adaptively correcting the injection weights based on the consistency differences of the session user representations; and adaptively correcting the convergence weights based on the consistency differences of the session-side convergence representations. The perturbation view is obtained by applying a perturbation operation to the session event graph, which is either a structural perturbation or a feature perturbation.

[0060] In this embodiment, perturbation views are used to construct several minor variations of the same session event graph to test the sensitivity of session representations and convergent representations to local changes. Perturbation views include at least two, which may include a combination of the original view and perturbation views, or a combination of views generated by multiple different perturbation strategies. The perturbation operation belongs to at least one of the following: firstly, structural perturbation, which performs lightweight structural changes on the nodes or edges of the session event graph, such as randomly discarding some nodes / edges, removing edges with low evidence quality, disconnecting or reconnecting weakly related edges, sampling local subgraphs, etc.; secondly, feature perturbation, which performs lightweight perturbation on the node or edge features of the session event graph, such as masking some event features / context features, applying small noise perturbations to continuous quality indicators, replacing discrete type identifiers with "unknown / default" encodings, etc. It should be noted that the perturbation operation should maintain the nature of a slight perturbation, that is, it should not change the basic structural semantics of the current session, but only perturb local information that may have noise or unstable sources, so that the consistency difference can reflect the reliability of the evidence rather than the complete destruction of the session semantics.

[0061] After constructing at least two perturbation views, this embodiment performs the aforementioned bidirectional coupling representation update process under each perturbation view to obtain the session user representation and session-side converged representation corresponding to each perturbation view. This step can be understood as repeatedly traversing the session-side update path in the co-enhancement closed loop for the same session under different slight perturbation conditions to observe the stability of the output representation to the perturbation.

[0062] The consistency difference is used to measure the degree of consistency of the output representations under different perturbation views. Specifically, the consistency difference can be calculated separately for the session user representation and the session-side converged representation. When the session user representation changes significantly under different perturbation views, it indicates that the session representation is sensitive to perturbation, and there may be some node / edge evidence that has an unstable impact on the representation update. When the session-side converged representation changes significantly under different perturbation views, it indicates that the convergence of session evidence is sensitive to perturbation, and there may be some nodes / edges that contribute unreasonably or unstablely during the convergence stage. The consistency difference can be calculated using a metric based on vector distance or similarity. This invention does not limit this method, as long as it can reflect the degree of difference between the output representations of different views.

[0063] After obtaining the consistency differences, this embodiment adaptively adjusts the injection weights or aggregation weights based on these differences. For the adaptive adjustment of injection weights, when the consistency differences of session user representations under different perturbation views are large, it indicates that the session-side update is sensitive to perturbations. The injection weights corresponding to nodes / edges related to this sensitivity can be lowered or redistributed, reducing the participation ratio of global user representations on these unstable session elements. When the consistency differences are small, the corresponding injection weights can be maintained or moderately increased, allowing stable and representative session elements to make fuller use of global information. For the adaptive adjustment of aggregation weights, when the consistency differences of session-side aggregated representations under different perturbation views are large, the aggregation weights corresponding to nodes / edges that cause instability can be lowered or redistributed, reducing their contribution to session evidence aggregation. When the consistency differences are small, the corresponding aggregation weights can be maintained or moderately increased, thereby allowing stable evidence to account for a larger proportion in aggregation.

[0064] The aforementioned adaptive correction process can be executed in one or more rounds according to a preset number of iterations. In each round, a perturbation view is constructed, bidirectional coupling updates are performed, consistency differences are calculated, and weights are corrected, so that the weights gradually converge to a state consistent with the session stability performance. Through this process, the co-enhancement closed loop can maintain high representation consistency when facing MR interaction noise and uncertainty, thereby improving the stability and robustness of the portrait generation process.

[0065] In this embodiment, after the aforementioned co-enhancement closed-loop processing, updated global user representations and updated session user representations can be obtained respectively. The updated global user representation focuses more on representing the user's long-term preferences and stable associations across sessions, while the updated session user representation focuses more on representing the user's immediate intent and conversational context in the current session. To facilitate the output of these two types of representations in a unified form for subsequent profile storage, retrieval, or personalized services, this embodiment further fuses them to generate a directly usable user profile representation, which is then output in the form of a profile vector. The generation of the user profile representation includes: determining a fusion coefficient for the updated global user representation and the updated session user representation, and generating a profile vector based on the fusion coefficient; The process of generating the profile vector includes: performing representation space alignment processing on the updated global user representation and the updated session user representation respectively to obtain a global profile sub-vector and a session profile sub-vector; and performing a weighted combination of the global profile sub-vector and the session profile sub-vector based on the first fusion coefficient and the second fusion coefficient to obtain the profile vector. The fusion coefficients include a first fusion coefficient for adjusting the contribution of global profile sub-vectors and a second fusion coefficient for adjusting the contribution of session profile sub-vectors, wherein the first fusion coefficient and the second fusion coefficient satisfy normalization constraints.

[0066] The updated global user representation and the updated session user representation may originate from different graph structures or update paths, and their vector dimensions, numerical distributions, or semantic spaces may differ. To avoid semantic inconsistencies or scale mismatches caused by direct fusion, this step first performs representation space alignment processing on the two types of representations separately, transforming them into a unified portrait representation space. Specifically, the updated global user representation can be input into the global-side alignment module, which outputs global portrait sub-vectors; the updated session user representation can be input into the session-side alignment module, which outputs session portrait sub-vectors. The alignment module can be implemented using mapping transformation methods, such as linear mapping, nonlinear mapping, or normalization transformation, to project the original representations onto a portrait space of a unified dimension. This alignment processing ensures that the global portrait sub-vectors and session portrait sub-vectors are comparable in the same representation space, providing a foundation for subsequent fusion.

[0067] To achieve controllable adjustment of the contribution of long-term information and conversation information, a fusion coefficient is further determined, including a first fusion coefficient and a second fusion coefficient. The first fusion coefficient is used to adjust the contribution of the global profile sub-vector, and the second fusion coefficient is used to adjust the contribution of the conversation profile sub-vector. Both satisfy the normalization constraint. This invention does not limit the method of determining the fusion coefficient, and can adopt at least one of the following engineering implementation methods, including: assigning values ​​to the first fusion coefficient and the second fusion coefficient according to the system's preset strategy, for example, increasing the first fusion coefficient in business scenarios that need to highlight the stability of long-term preferences, and increasing the second fusion coefficient in business scenarios that need to highlight the immediate intent of the conversation; determining the fusion coefficient based on the overall state information of the current conversation, for example, adjusting the fusion coefficient based on the conversation duration, number of events, conversation stage, or scenario / task state; determining the fusion coefficient based on the numerical characteristics, similarity, or difference between the updated global user representation and the updated conversation user representation, so that the two types of representations adopt a more balanced fusion when the consistency is high, and adopt a fusion that is more biased towards one side when the difference is large; determining the fusion coefficient based on the overall statistical results of the conversation-side evidence quality, so that the second fusion coefficient is appropriately increased when the conversation evidence quality is higher, and the first fusion coefficient is appropriately increased when the conversation evidence quality is lower or unstable.

[0068] After obtaining the global profile sub-vector, the session profile sub-vector, and the fusion coefficients, this step performs a weighted combination of the two types of sub-vectors based on the first and second fusion coefficients to obtain the profile vector. The profile vector is used as a vectorized output representing the user profile and can serve as input for profile storage, profile retrieval, profile matching, or personalized service strategy selection, thereby enabling the output of the co-enhancement closed loop to be provided externally in a unified form.

[0069] Through the above-mentioned fusion mechanism of representation space alignment and fusion coefficient control, this embodiment can comprehensively reflect the user's long-term preference information across sessions and the real-time intent information of the current session within the same portrait representation space, avoiding portrait offset or response lag caused by relying solely on a single source representation.

[0070] In this embodiment, after obtaining the profile vector, to facilitate continuous maintenance and subsequent invocation of the profile, such as cross-session profile inheritance, long-term preference tracking, and personalized service triggering, this step further writes the current profile vector into the profile storage structure and performs incremental updates on the historical profile vectors in the profile storage structure. This allows the user profile to maintain timeliness and traceability as the session data continues to evolve. The incremental update includes at least one of the following update strategies: weighting the historical profile vector according to a preset time decay rule; storing historical profile vectors according to a versioning rule and recording the current profile vector as the new version; maintaining user profile representations according to scenario categories and writing the current profile vector into the user profile representation of the scenario category corresponding to the current session.

[0071] In this embodiment, the profile storage structure is used to store the profile vector corresponding to the user and its necessary index information to support subsequent reading, updating, and backtracking. The profile storage structure can be implemented using a key-value structure, a table structure, a vector index structure, or a combination thereof, and includes at least a user identifier and its corresponding profile vector field; and optionally includes metadata fields such as session identifier, timestamp, scene / content / task identifier, and version number to support subsequent profile maintenance strategies based on time or scene.

[0072] When writing the current profile vector to the profile storage structure, insert or overwrite operations can be performed. For example, the current profile vector can be written as the latest profile vector to the user's master record; or, while retaining historical records, the current profile vector can be written to a new record entry for subsequent version management or scenario-based maintenance.

[0073] Optionally, to ensure that user profiles reflect the time-related characteristics of recent behavior being more important and long-term behavior gradually fading, historical profile vectors can be updated with weights according to a preset time decay rule. Specifically, the decay weight can be determined based on the interval between the timestamp corresponding to the historical profile vector and the current session time, and the historical profile vector is decayed accordingly. This decayed vector is then combined with the current profile vector to obtain an updated historical profile vector or an updated long-term profile vector. Through this time decay mechanism, the sensitivity of the profile to recent behavioral changes can be improved without losing historical information.

[0074] Optionally, to support the traceability and rollback requirements of user profiles, historical profile vectors can be stored according to versioning rules, and the current profile vector can be recorded as the new version. Specifically, a version number or version time series can be maintained for the same user. After the current profile vector is generated in each session, it is written as the new version into the profile storage structure; and a preset number of historical versions can be retained or historical versions can be retained according to time windows. Through versioned storage, comparative analysis of profile status at different points in time can be achieved, and restoration to historical versions can be supported in the event of abnormal updates.

[0075] Optionally, considering the potential differences in user behavior preferences across different scenarios in MR applications, user profile representations can be maintained separately for each scenario category, and the current profile vector can be written into the user profile representation of the scenario category corresponding to the current session. Specifically, a mapping structure between scenario category and profile vector can be established for each user. When the current session is identified as belonging to a certain scenario category, the current profile vector is written into the profile record corresponding to that scenario category. When the profile is subsequently retrieved, the profile vector matching the target scenario category can be read first to improve the profile's adaptability in specific scenarios. The scenario category can be obtained by mapping scenario identifiers, content identifiers, task identifiers, or combinations thereof in the session. This embodiment does not limit the specific classification method.

[0076] The above incremental update strategies can be implemented independently or in combination. For example, historical profile vectors can be weighted and updated first according to time decay rules, and then the updated results can be recorded as the new version according to versioning rules; or, based on maintaining them separately by scenario category, time decay updates and versioned storage can be performed separately for the profile vectors under each scenario category. By combining strategies, the manageability of the profile storage structure in specific application deployments can be improved while satisfying the timeliness, traceability, and scenario adaptability of profiles.

[0077] In one system embodiment, the present invention also provides a big data-based intelligent generation system for MR user profiles, used to implement the intelligent generation method for MR user profiles in any of the foregoing method embodiments. This system can be deployed on MR terminals, edge computing nodes, servers, or cloud platforms, and can also be implemented using a terminal-server collaborative approach. Figure 2 The system structure shown, the MR user profile intelligent generation system includes: The data acquisition module acquires MR user behavior data in the current session and at least two historical sessions; The global user profile construction module uses MR user behavior data from at least two historical sessions to construct a cross-session global heterogeneous user profile graph and obtains the global user representation in the global heterogeneous user profile graph. The session event graph construction module uses the MR user behavior data of the current session to construct a session event graph; The bidirectional coupling module performs bidirectional coupling representation update processing between the global heterogeneous user profile graph and the session event graph. The bidirectional coupling representation update processing includes: passing the global user representation to the session event graph as the representation input for nodes and edges in the session event graph to obtain the updated session user representation; and aggregating the node and edge representations of the session event graph to obtain the session-side aggregated representation, and performing feedback update on the global user representation based on the session-side aggregated representation to obtain the updated global user representation. The profile generation module is used to generate user profile representations based on the updated global user representation and the updated session user representation.

[0078] In some embodiments, the present invention can also be implemented as a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, causes the processor to perform the steps described in any of the foregoing method embodiments. The computer-readable storage medium can be any tangible medium capable of storing program code, including but not limited to: read-only memory, random access memory, disk, optical disk, flash memory, solid-state drive, memory card, etc. The term "module / unit" in this specification can be understood as a software or hardware component or combination thereof that implements a corresponding function, and the specific implementation should not constitute a limitation on the present invention.

[0079] In some embodiments, the present invention may also be implemented as a computer program product, the computer program product including a computer program that may be stored in a computer-readable storage medium or carried in the form of a signal and transmitted over a network; when the computer program is executed by a processor, the processor performs the steps described in any of the foregoing method embodiments.

[0080] In some embodiments, the present invention can also be implemented as an electronic device, comprising: one or more processors, a memory, and a computer program stored in the memory and executable by the one or more processors; the one or more processors execute the computer program to implement the steps described in any of the foregoing method embodiments. The electronic device may be an MR terminal, an edge computing node, a server, or a cloud computing device, or a combination thereof.

[0081] The technical solution of the present invention has been described above with reference to the accompanying drawings and specific embodiments. It should be understood that the above embodiments are only used to explain the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions or variations made by those skilled in the art to the above embodiments without departing from the spirit and substance of the present invention should fall within the protection scope of the present invention, and the protection scope is determined by the appended claims.

Claims

1. A method for intelligently generating MR user profiles based on big data, characterized in that, The method includes: Acquire MR user behavior data in the current session and at least two historical sessions; Construct a global heterogeneous user profile graph across sessions using MR user behavior data from at least two historical sessions; Construct a session event graph using MR user behavior data from the current session; Perform bidirectional coupled representation update processing between the global heterogeneous user profile graph and the session event graph; User profile representations are generated based on the updated global user representation and the updated session user representation; The bidirectional coupling representation update process includes: passing the global user representation in the global heterogeneous user profile graph to the session event graph as the representation input for nodes and edges in the session event graph to obtain the updated session user representation; and aggregating the node and edge representations of the session event graph to obtain the session-side aggregated representation, and performing feedback update on the global user representation in the global heterogeneous user profile graph based on the session-side aggregated representation to obtain the updated global user representation.

2. The method for intelligent generation of MR user profiles based on big data according to claim 1, characterized in that, The global heterogeneous user profile graph includes at least: user nodes, entity nodes related to MR interaction, and multi-type relationship edges; The entity nodes include at least two of the following categories: session nodes, interactive event nodes, virtual object nodes, spatial anchor points or region nodes, scene nodes, content nodes, and task nodes; The multi-type relationship edge is used to describe the association relationship category between at least two selected entity nodes. Each relationship edge includes a relationship type identifier to characterize the association relationship category, and also includes at least one relationship attribute among timestamp, duration, confidence level, and pose information parameters.

3. The intelligent generation method for MR user profiles based on big data according to claim 1, characterized in that, The session event graph includes: An event node and at least one of its associated object nodes and spatial nodes; At least one of the following: a temporal association edge used to characterize the order in which events occur; an interactive association edge used to characterize the relationship between an event and an object; and a contextual association edge used to characterize the relationship between an event and its spatial context.

4. The method for intelligent generation of MR user profiles based on big data according to any one of claims 1 to 3, characterized in that, Before performing bidirectional coupled representation update processing, the initial representations of nodes and edges in the session event graph are set based on the global representation vectors of user nodes in the global heterogeneous user profile graph and entity nodes corresponding to nodes in the session event graph, so as to obtain an initialized session event graph. The initial representation setting includes: generating initialization vectors for corresponding nodes and edges, and using the initialization vectors as the initial representations of the nodes and edges; The generation of the initialization vector includes: transforming the global representation vector through mapping to obtain the initialization vector in the session representation space, or fusing the global representation vector with the event type features, context features, and relation type features corresponding to the edges of the current session to generate the initialization vectors of the corresponding nodes and edges.

5. The method for intelligent generation of MR user profiles based on big data according to any one of claims 1 to 4, characterized in that, In the bidirectional coupled representation update process, when the global user representation in the global heterogeneous user profile graph is passed to the session event graph, the corresponding injection weights are determined for the nodes and edges in the session event graph, which are used to determine the proportion of the global user representation participating in the corresponding node and edge representation update. The injected weights are generated based on at least one piece of evidence quality information from the following: confidence level of the interaction event, duration of the interaction event, time interval between adjacent interaction events, spatial stability index, pose tracking quality index, interaction event type identifier, and context category identifier.

6. The method for intelligent generation of MR user profiles based on big data according to any one of claims 1 to 5, characterized in that, In the bidirectional coupled representation update process, the convergence weights of the node representations and edge representations in the session event graph are determined, and the node representations and edge representations are weighted and converged according to the convergence weights to obtain the session-side converged representation. The convergence weight is generated based on at least one piece of evidence quality information from the following: confidence level of interaction event, duration of interaction event, time interval between adjacent interaction events, spatial stability index, pose tracking quality index, interaction event type identifier, relationship type identifier, and context category identifier. The feedback update includes at least one of the following methods: incrementally updating the global user representation vector corresponding to the user node in the global heterogeneous user profile graph; Incrementally update the representation vectors of entity nodes related to the current session in the global heterogeneous user profile graph; Update the relation attributes or relation weights of multi-type relation edges in the global heterogeneous user profile graph; Incremental adjustments are made to the adjacency relationships of the global heterogeneous user profile graph.

7. The method for intelligent generation of MR user profiles based on big data according to claim 5 or 6, characterized in that, The determination of the injection weight or the aggregation weight includes an adaptive adjustment process, which includes: constructing at least two perturbation views based on the session event graph; performing bidirectional coupled representation update processing under the at least two perturbation views to obtain the corresponding session user representation and session-side aggregation representation; calculating the consistency difference of the session user representation and session-side aggregation representation corresponding to the at least two perturbation views; and adaptively adjusting the injection weight and the aggregation weight based on the consistency difference of the session user representation. The perturbation view is obtained by applying a perturbation operation to the session event graph, which is either a structural perturbation or a feature perturbation.

8. The method for intelligent generation of MR user profiles based on big data according to any one of claims 1 to 7, characterized in that, The process of generating user profile representations includes: determining a fusion coefficient between the updated global user representation and the updated session user representation, and generating a profile vector based on the fusion coefficient; The process of generating the profile vector includes: performing representation space alignment processing on the updated global user representation and the updated session user representation respectively to obtain a global profile sub-vector and a session profile sub-vector; and performing a weighted combination of the global profile sub-vector and the session profile sub-vector based on the first fusion coefficient and the second fusion coefficient to obtain the profile vector. The fusion coefficients include a first fusion coefficient for adjusting the contribution of global profile sub-vectors and a second fusion coefficient for adjusting the contribution of session profile sub-vectors, wherein the first fusion coefficient and the second fusion coefficient satisfy normalization constraints.

9. The method for intelligent generation of MR user profiles based on big data according to claim 8, characterized in that, Also includes: Write the image vector into the image storage structure and perform incremental updates on the historical image vectors in the image storage structure; The incremental update includes at least one of the following update strategies: weighting the historical image vector according to a preset time decay rule; storing the historical image vector according to a versioning rule, and recording the current image vector as the new version; Maintain user profile representations for each scenario category, and write the current profile vector into the user profile representation of the scenario category corresponding to the current session.

10. A big data-based intelligent generation system for MR user profiles, characterized in that, The system includes: The data acquisition module acquires MR user behavior data in the current session and at least two historical sessions; The global user profile construction module uses MR user behavior data from at least two historical sessions to construct a cross-session global heterogeneous user profile graph and obtains the global user representation in the global heterogeneous user profile graph. The session event graph construction module uses the MR user behavior data of the current session to construct a session event graph; The bidirectional coupling module performs bidirectional coupling representation update processing between the global heterogeneous user profile graph and the session event graph. The bidirectional coupling representation update processing includes: passing the global user representation to the session event graph as the representation input for nodes and edges in the session event graph to obtain the updated session user representation; and aggregating the node and edge representations of the session event graph to obtain the session-side aggregated representation, and performing feedback update on the global user representation based on the session-side aggregated representation to obtain the updated global user representation. The profile generation module is used to generate user profile representations based on the updated global user representation and the updated session user representation.

Images