Content recommendation method and content recommendation system

Abstract

The invention provides a content recommendation method and a content recommendation system, and relates to the technical field of data processing. The method comprises the following steps: acquiring a multi-modal behavior data stream; wherein the multi-modal behavior data stream comprises a plurality of pieces of multi-modal behavior data; the multi-modal behavior data comprises a timestamp and a scene label; according to the timestamp and the scene label, aggregating the multi-modal behavior data flow to generate a spatio-temporal behavior unit cluster; based on the spatio-temporal behavior unit cluster, updating the user historical behavior map to generate a user real-time behavior map; matching the user real-time behavior map with a preset candidate content set to generate a content recommendation result; wherein the content recommendation result is used for content recommendation. According to the method, the spatiotemporal behavior unit cluster is constructed based on the multi-modal user behaviors, and the user behavior graph is dynamically updated based on the spatiotemporal behavior unit cluster, so that the accuracy of content recommendation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a content recommendation method and a content recommendation system. BACKGROUND

[0002] Traditional content recommendation relies on user historical behavior records or static preference tags, and adopts a collaborative filtering, content tag matching or offline clustering based approach to push content. For example, some systems generate an interest profile based on user listening history, or recommend similar types of music and radio stations through keyword matching.

[0003] Such methods usually construct static user profiles and have basic personalization capabilities to some extent. However, they mainly rely on offline modeling and single modal input, lack deep perception of environmental, user current state, interaction mode and other dimensions, and are difficult to meet real-time adaptation requirements brought by user interest changes or scene switching. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a content recommendation method and a content recommendation system, which construct a spatio-temporal behavior unit cluster based on multi-modal user behavior, and dynamically update a user behavior graph based on the spatio-temporal behavior unit cluster, to improve the accuracy of content recommendation.

[0005] In a first aspect, the embodiments of the present application provide a content recommendation method, which includes: acquiring a multi-modal behavior data stream; wherein the multi-modal behavior data stream includes a plurality of multi-modal behavior data; the multi-modal behavior data includes a timestamp and a scene tag; aggregating the multi-modal behavior data stream according to the timestamp and the scene tag to generate a spatio-temporal behavior unit cluster; updating a user historical behavior graph based on the spatio-temporal behavior unit cluster to generate a user real-time behavior graph; matching the user real-time behavior graph with a preset candidate content set to generate a content recommendation result; wherein the content recommendation result is used for content recommendation.

[0006] The embodiments of the present application generate a spatio-temporal unit behavior cluster based on multi-modal behavior data, which represents the behavior characteristics of a user in a certain space-time. The user real-time behavior graph is generated by updating the user historical behavior graph using the spatio-temporal behavior unit cluster, and content recommendation is performed based on the user real-time behavior graph. In this process, since the user real-time behavior graph is dynamically updated according to the user's spatio-temporal behavior, the accuracy of content recommendation is improved.

[0007] In some embodiments, after generating the content recommendation result, the method further includes: receiving user response data; optimizing the user real-time behavior graph based on the response data to generate an optimized user real-time behavior graph; performing content recommendation based on the optimized user real-time behavior graph.

[0008] The embodiment of the application optimizes the user real-time behavior graph according to the user response data after content recommendation, and performs content recommendation according to the optimized user real-time behavior graph. In this process, the behavior graph is dynamically optimized through interaction behavior, and the accuracy of content recommendation is further improved.

[0009] In some embodiments, based on the spatio-temporal behavior unit cluster, the user historical behavior graph is updated to generate a user real-time behavior graph, including: determining a user behavior type according to the spatio-temporal behavior unit cluster; if it is determined that a target node matching the user behavior type exists in the user historical behavior graph, updating a current weight of the target node based on a preset weight update rule to generate an updated target node; updating the user historical behavior graph based on the updated target node to generate the user real-time behavior graph.

[0010] In the embodiment of the application, when the user historical behavior graph is updated based on the spatio-temporal behavior unit cluster, only the target node matching the user behavior type is updated, which not only reduces the computational complexity, but also improves the updating efficiency of the behavior graph, thereby improving the real-time recommendation. Further, since the weight reflects the importance of each node in the behavior graph for content recommendation, the user historical behavior graph is updated by updating the weight of the target node, thereby improving the accuracy of content recommendation.

[0011] In some embodiments, the current weight of the target node is updated based on the preset weight update rule to generate an updated target node, including: obtaining a current activity and a current weight of the target node; calculating a new weight of the target node based on the formula ; wherein, represents the new weight of the target node, represents the current weight of the target node, represents the current activity of the target node, represents the activity weight coefficient, ∈ (0, 1); and updating the current weight of the target node according to the new weight to generate the updated target node. In the embodiment of the application, the activity of the target node is also considered when calculating the weight, so that the weight update is more in line with user interest, and the accuracy of content recommendation is further improved.

[0012] In the embodiment of the application, the activity of the target node is also considered when calculating the weight, so that the weight update is more in line with user interest, and the accuracy of content recommendation is further improved.

[0013] ​​In some embodiments, updating the user's historical behavior graph based on spatiotemporal behavior unit clusters to generate a real-time user behavior graph further includes: if it is determined that all nodes in the user's historical behavior graph do not match the user's behavior type, then creating a new node based on the spatiotemporal behavior unit cluster; wherein the new node includes a node type; determining the weight of the new node based on the weights of all nodes in the user's historical behavior graph that match the node type; updating the user's historical behavior graph based on the new node and the weight of the new node to generate a real-time user behavior graph.

[0014] In this embodiment of the application, by creating new nodes, the graph is dynamically updated based on changes in user interests, increasing the types of behaviors that the behavior graph can represent, expanding the boundaries of the behavior graph, and improving the richness of the behavior graph, thereby improving the accuracy of content recommendation.

[0015] In some embodiments, updating the user's historical behavior graph based on spatiotemporal behavior unit clusters to generate a real-time user behavior graph further includes: if it is determined that there are static nodes in the user's historical behavior graph whose weights have not been updated within a preset time period, then based on a preset weight decay function... Update the current weight of the static node to generate the updated static node; where; Represents static nodes The updated weights; Represents static nodes The last time it was effectively activated The weight value; Indicates the current system time; Represents static nodes The time point of the most recent effective activation; λ represents the time decay coefficient, which is dynamically determined based on the scene label; e represents the natural constant; the user's historical behavior graph is updated based on the updated static nodes to generate the user's real-time behavior graph.

[0016] In this embodiment, by reducing the weight of nodes with low activity in the behavior graph, the behavior graph representation is made more biased towards nodes that users are interested in, thereby improving the accuracy of content recommendation.

[0017] In some embodiments, matching a user's real-time behavior graph with a preset set of candidate content to generate content recommendation results includes: generating a user preference representation vector based on the user's real-time behavior graph; calculating the similarity between the user preference representation vector and the vector corresponding to the candidate content set to obtain a basic recommendation score for the recommended content; determining target recommended content based on a preset number of historical spatiotemporal behavior unit clusters; wherein the target recommended content includes a trend-weighted score; and generating content recommendation results based on the basic recommendation score and the trend-weighted score.

[0018] In this embodiment of the application, when generating content recommendation results, not only are the user's real-time preferences considered to obtain the basic recommendation score of the recommended content, but also the changes in the user's historical behavior are considered to capture the evolution trend of user behavior and obtain a trend-weighted score. Finally, the basic recommendation score and the trend-weighted score are combined to determine the content recommendation result, thereby improving the accuracy of content recommendation.

[0019] In some embodiments, the multimodal behavior data stream is aggregated based on timestamps and scene tags to generate a spatiotemporal behavior unit cluster, including: dividing the multimodal behavior data stream according to a preset time window based on timestamps to obtain multiple multimodal temporal behavior sequences; for each multimodal temporal behavior sequence, aggregating the multimodal behavior data corresponding to scene tags with similarity greater than a preset similarity threshold to generate multiple behavior clusters; and generating a spatiotemporal behavior unit cluster based on the multiple behavior clusters.

[0020] In this embodiment of the application, when generating spatiotemporal unit clusters, clustering is performed through similarity to improve the accuracy of behavior representation, thereby improving the accuracy of content recommendation.

[0021] Secondly, embodiments of this application provide a content recommendation system, which is used to execute the method steps of any embodiment of the first aspect; the system includes a multi-source behavior data acquisition module, a behavior sequence spatiotemporal clustering module, a graph construction and dynamic evolution module, and a content adaptation module; wherein, the multi-source behavior data acquisition module is used to acquire multimodal behavior data streams; the behavior sequence spatiotemporal clustering module is used to generate spatiotemporal behavior unit clusters based on the multimodal behavior data streams; the graph construction and dynamic evolution module is used to update the user's historical behavior graph based on the spatiotemporal behavior unit clusters to generate the user's real-time behavior graph; the content adaptation module is used to match the user's real-time behavior graph with a preset set of candidate content to generate content recommendation results; wherein, the content recommendation results are used for content recommendation.

[0022] In some embodiments, the system further includes a feedback update module; wherein the feedback update module is used to receive user response data and optimize the real-time user behavior graph based on the user response data to generate an optimized real-time user behavior graph.

[0023] Thirdly, embodiments of this application provide an electronic device, including: a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the machine-readable instructions, when executed by the processor, are capable of performing the method steps of any embodiment of the first aspect.

[0024] Fourthly, embodiments of this application provide a computer-readable storage medium, including: computer program instructions stored on the computer-readable storage medium, wherein the computer program instructions are executed by a processor to perform the method steps of any embodiment of the first aspect.

[0025] Fifthly, embodiments of this application provide a computer program product, including: computer program instructions, which are executed by a processor to perform the method steps of any embodiment of the first aspect.

[0026] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing embodiments of this application. Attached Figure Description

[0027] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A flowchart illustrating a content recommendation method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a first content recommendation system provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a second content recommendation system provided in an embodiment of this application; Figure 4 A content recommendation timing diagram provided for an embodiment of this application; Figure 5 This is a schematic diagram of the electronic device structure provided in an embodiment of this application. Detailed Implementation

[0029] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0030] It should be noted that all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0031] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0032] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0033] Traditional content recommendation systems often rely on users' historical behavior records or static preference tags, using methods such as collaborative filtering, content tag matching, or offline clustering to push content. For example, some systems generate interest profiles based on users' listening history, or recommend similar types of music and radio stations through keyword matching.

[0034] These methods typically construct static user profiles, which possess basic personalization capabilities to some extent. However, they mainly rely on offline modeling and single-modal input, lacking a deep understanding of dimensions such as the environment, the user's current state, and interaction methods, making it difficult to cope with the real-time adaptation requirements brought about by changes in user interests or scene switching.

[0035] For example, in in-vehicle entertainment content recommendation scenarios, the lack of deep perception of dimensions such as the driving environment, the user's current state, and interaction methods makes it impossible to effectively perceive changes in the user's behavioral state during driving. This results in a rigid data graph structure and an inability to respond to shifts in interests. The difficulty in handling the real-time adaptation requirements brought about by changes in user interests or scene transitions leads to low accuracy in content recommendations.

[0036] Therefore, to address the above problems, this application provides a content recommendation method and a content recommendation system, which constructs a spatiotemporal behavior unit cluster based on multimodal user behavior and dynamically updates the user behavior graph based on the spatiotemporal behavior unit cluster, thereby improving the accuracy of content recommendation.

[0037] Figure 1This is a flowchart illustrating a content recommendation method provided in an embodiment of this application. It is understood that the content recommendation method provided in this application can be applied to terminal devices (also known as electronic devices) and servers; specifically, terminal devices can be smartphones, tablets, computers, personal digital assistants (PDAs), vehicle control terminals, etc.; servers can be application servers or web servers. To facilitate understanding of the technical solutions provided in this application, the application scenarios of the content recommendation method provided in this application are described below, using a terminal device as the execution subject.

[0038] like Figure 1 As shown, the method includes: Step S101: Obtain multimodal behavior data stream; wherein, the multimodal behavior data stream includes multiple multimodal behavior data; the multimodal behavior data includes timestamps and scene labels.

[0039] Multimodal behavioral data streams refer to heterogeneous and parallel sequences of behavioral data continuously generated during a period of interaction between users and digital systems (such as apps, websites, smart devices, etc.).

[0040] The multimodal behavior data stream comprises multiple multimodal behavior data sets, each of which includes multiple data sets of different types.

[0041] For example, in the context of in-vehicle entertainment, multimodal behavioral data refers to the behavioral data generated by the user during driving, which is collected in real time by various sensors on the vehicle in the in-vehicle environment.

[0042] In-vehicle entertainment refers to the human-computer interaction environment and set of functions that provide information, entertainment and leisure services to drivers and passengers in the mobile and enclosed space of a car, in order to enhance the driving experience and relieve travel fatigue.

[0043] In-vehicle entertainment includes audio entertainment (radio stations, music, podcasts, etc.), video entertainment, communication (Bluetooth phone, in-vehicle Wi-Fi, etc.), cabin environment (scene modes, intelligent ambient lighting, audio system, etc.), navigation and positioning, and vehicle driving status.

[0044] In in-vehicle entertainment scenarios, multimodal behavioral data can specifically include voice commands collected by the in-vehicle microphone. Touch operations collected by touch sensors Eye movement trajectory captured by vehicle-mounted camera and facial expressions and interactive behaviors collected through a combination of multiple vehicle sensors. and vehicle driving status data Among them, vehicle driving status data Including speed, geographic location information and driving status .

[0045] Based on the data collected above, multimodal behavioral data It can be represented as At the same time, a timestamp is marked for each multimodal behavioral data. and geographic location information This allows us to clearly identify the time period and geographical location from which the data was collected, thereby improving the data's representation capabilities.

[0046] A multimodal behavior data stream composed of multiple multimodal behavior data. It can be represented as .

[0047] It should be understood that multimodal behavioral data streams It can also represent multimodal behavioral data collected in other application scenarios.

[0048] For example, in a smart home scenario, user interaction with smart home devices and the user's geographical location are collected as multimodal behavioral data. .

[0049] Timestamp Used to represent the time point of multimodal behavioral data collection.

[0050] Scene tags are used to abstractly describe the context in which a target object (such as a user) is situated and exhibits a specific behavioral pattern within a specific spatiotemporal dimension.

[0051] Scene tags Through a pre-defined scene encoding function It is a structured, high-level semantic identifier generated by fusing and semantically computing multimodal behavioral data from a specific environment.

[0052] Therefore, in the specific implementation process, after acquiring the multimodal behavior data stream, a preset scene encoding function is used for each multimodal behavior data in the multimodal behavior data stream. Generate corresponding scene labels and mark the scene labels on the corresponding multimodal behavior data.

[0053] For example, in the context of in-vehicle entertainment, scene labels Based on interaction behavior and vehicle driving status data Geographical location information With driving status Generation, that is, Specifically, when For "highway section", "Constant speed driving" with no interactive behavior Then, the scene label "Highway Section - Constant Speed ​​- Silent" will be generated.

[0054] Step S102: Aggregate the multimodal behavior data streams based on timestamps and scene tags to generate spatiotemporal behavior unit clusters.

[0055] Since each timestamp corresponds to a single piece of multimodal behavioral data, behavioral data at a single point in time can only indicate a user's behavior at a specific moment and lacks universality. Therefore, by aggregating multimodal behavioral data to form spatiotemporal behavioral unit clusters, we can characterize the user's behavioral features over a certain period of time. A spatiotemporal behavioral unit cluster is a collection composed of multiple pieces of multimodal behavioral data.

[0056] In the specific implementation process, the acquired multimodal behavioral data streams are targeted. According to the preset time window Segmentation: .

[0057] After obtaining multiple multimodal temporal behavior sequences in segments, based on scene labels Clustering algorithms (such as DBSCAN or KMeans) are used to group the same time slices. Internal, with similar scene tags The data is aggregated to generate spatiotemporal behavioral unit clusters: ,in, This represents a cluster of behaviors with similar scene labels.

[0058] Therefore, in some embodiments, the multimodal behavior data stream is aggregated based on timestamps and scene tags to generate spatiotemporal behavior unit clusters, including: dividing the multimodal behavior data stream according to a preset time window based on timestamps to obtain multiple multimodal temporal behavior sequences; for each multimodal temporal behavior sequence, aggregating the multimodal behavior data corresponding to scene tags with similarity greater than a preset similarity threshold to generate multiple behavior clusters; and generating spatiotemporal behavior unit clusters based on multiple behavior clusters.

[0059] It should be noted that the size of the preset time window can be set according to the actual situation, or it can be dynamically determined based on the scene labels of the latest multimodal behavior data acquired at the current moment.

[0060] For example, in an in-vehicle entertainment scenario, if the scene label of the latest acquired multimodal behavior data indicates that the vehicle belongs to an urban congestion scenario, then the time window... The time window can be set to 1 minute to capture high-frequency operations; if the scene label of the latest acquired multimodal behavior data indicates that the vehicle belongs to a high-speed constant-speed scene, then the time window... The time window can be set to 5 minutes to aggregate stable behavior; if the scene label of the latest acquired multimodal behavior data indicates that the vehicle belongs to a parking scene, then the time window... It can be set to 3 minutes to balance the operation density.

[0061] Therefore, in the specific implementation process, if the preset time window is dynamically determined by the scene labels of the latest acquired multimodal behavior data at the current moment, then the preset scene encoding function can be used. After generating scene labels corresponding to multimodal behavioral data, a time window is first determined based on the scene labels. Then, the multimodal behavioral data stream is divided using the dynamically determined time window so that the behavioral clusters obtained by clustering can adapt to dynamically changing scenes, improve the accuracy of spatiotemporal behavioral unit clusters, and thus improve the accuracy of content recommendation.

[0062] The preset similarity threshold is also a value set in advance based on the actual situation.

[0063] Similarity calculation methods can include Euclidean distance, Manhattan distance, cosine similarity, etc.

[0064] Step S103: Based on the spatiotemporal behavior unit cluster, update the user's historical behavior map to generate the user's real-time behavior map.

[0065] Step S104: Match the user's real-time behavior graph with the preset candidate content set to generate content recommendation results; wherein, the content recommendation results are used for content recommendation.

[0066] In the specific implementation process, the user historical behavior map is generated in advance based on the multimodal behavior data stream collected in history.

[0067] Specifically, historical interaction behaviors are determined based on historically collected multimodal behavioral data streams. Interest tags and content records Historical interaction behavior Interest tags Content Records Mapped to the initial set of nodes in the user behavior graph Establish a graph structure The nodes in the user behavior graph are divided into: behavior nodes. Preference Nodes and content nodes ,side It indicates semantic associations, temporal shifts, or co-occurrence relationships between behaviors or preferences.

[0068] Historical interaction behavior refers to the user's operational and interactive behaviors with the in-vehicle entertainment system in a vehicle setting. For example, a user can say "play rock music" via voice command while driving, manually touch the central control screen to change radio channels when parked, or pause audio playback using steering wheel buttons while driving.

[0069] Interest tags are structured labels that represent users' entertainment preferences. They can be extracted from historical interaction behavior or set proactively by the user upon first use. For example, a "rock music" tag can be generated based on a user's frequent listening to rock music, and a "financial news" tag can be generated based on a user's repeated requests for financial radio stations.

[0070] Content records refer to the specific entertainment content that users have previously encountered in the in-vehicle entertainment system. For example: the song "Seven Mile Fragrance" and the novel "The Three-Body Problem".

[0071] After acquiring the latest multimodal behavior data stream, the user's historical behavior map is updated based on the spatiotemporal behavior unit clusters generated from the latest multimodal behavior data stream to generate the user's real-time behavior map, in order to adapt to dynamically changing scenarios.

[0072] After obtaining the user's real-time behavior graph, the user's real-time behavior graph is matched with a preset set of candidate content to generate content recommendation results, which are then used for content recommendation.

[0073] This application's embodiments generate spatiotemporal unit behavior clusters based on multimodal behavioral data, representing the user's behavioral characteristics in a specific spatiotemporal context. These spatiotemporal behavior unit clusters are used to update the user's historical behavior graph, generating a real-time user behavior graph, and content recommendation is performed based on this graph. In this process, because the real-time user behavior graph is dynamically updated according to the user's spatiotemporal behavior, the accuracy of content recommendation is improved.

[0074] In some embodiments, after generating content recommendation results, the method further includes: receiving user response data; optimizing the user real-time behavior graph based on the response data to generate an optimized user real-time behavior graph; and performing content recommendation based on the optimized user real-time behavior graph.

[0075] In the specific implementation process, after recommending content based on the content recommendation results, users usually respond to it. Therefore, in order to further improve the accuracy of content recommendation, user response data is received, and the real-time user behavior graph is optimized based on the response data to generate an optimized real-time user behavior graph for content recommendation.

[0076] Specifically, user responses to recommended content (such as clicks, dwell time, skipping, and facial expression changes) are captured to form a feedback set. Then use the feedback set Reverse-adjust the weights of nodes associated with recommended content in the real-time user behavior graph: ,in, This refers to the nodes in the behavior graph whose weights need to be adjusted. Feedback bias refers to the difference between a user's actual response to recommended content and the system's expectation, used to quantify the degree of deviation between the recommendation results and the user's true preferences. Specifically, it is calculated by the degree of match between the user's explicit behavior (such as clicking or skipping) or implicit behavior (such as dwell time or facial expression changes) and the system's prediction. For example, if a user skips a recommended item, the feedback bias for that node is negative; if the user plays for a long time or displays positive facial expressions, the feedback bias is positive. For learning rate, This represents the weight of nodes in the user's real-time behavior graph that are associated with the recommended content. This indicates the adjusted weight of the nodes associated with the recommended content.

[0077] For example, the nodes and weights in the real-time user behavior graph are as follows: video games (w=0.9), NBA basketball (w=0.8), science fiction movies (w=0.7), Chinese pop music (w=0.6), and popular science (w=0.5).

[0078] The system prioritizes and recommends high-weight content based on the user's current interest graph. For example, it might push a game video related to a video game. The user immediately clicks on it, watches the entire 15-minute video, and finally likes it. This captures and forms a feedback set. This includes r1 (click) = +1, r2 (extremely long dwell time) = +1, and r3 (positive interaction: like) = +1.

[0079] Then, the weights of the game nodes are recalculated using a weight correction formula, thereby updating the user's real-time behavior graph. Since the captured feedback set is all positive, the weights of the game nodes are increased, and game-related content is still recommended to the user.

[0080] If a node is frequently triggered by negative feedback, it may trigger structural pruning or interest drift prompts. Ultimately, this achieves a closed loop of system self-adjustment and preference expression, continuously improving the accuracy of content recommendations and user satisfaction.

[0081] It should be noted that the learning rate The learning rate can also be dynamically determined based on the scene label of the latest acquired multimodal behavior data. Specifically, if the scene label of the latest acquired multimodal behavior data is a parking scene, then the learning rate... The learning rate is set to 0.3 to enhance effective feedback. If the scene label of the latest acquired multimodal behavior data is a driving scene, then the learning rate is set to 0.3. The learning rate is set to 0.1 to filter out interfering feedback. If the scene label of the latest acquired multimodal behavior data is an abnormal scene such as sudden braking, then the learning rate is set to 0.1. Set to 0 to ignore invalid feedback.

[0082] In this embodiment, after content recommendation, user response data is received to optimize the user's real-time behavior graph, and content recommendation is performed based on the optimized user real-time behavior graph. In this process, the behavior graph is dynamically optimized through interactive behavior, further improving the accuracy of content recommendation.

[0083] In some embodiments, updating the user's historical behavior graph based on spatiotemporal behavior unit clusters to generate a real-time user behavior graph includes: determining the user behavior type based on the spatiotemporal behavior unit clusters; if it is determined that there is a target node in the user's historical behavior graph that matches the user behavior type, updating the current weight of the target node based on a preset weight update rule to generate an updated target node; and updating the user's historical behavior graph based on the updated target node to generate a real-time user behavior graph.

[0084] As can be seen from the above embodiments, the spatiotemporal behavior unit cluster includes multiple behavior clusters. In the specific implementation process, the spatiotemporal behavior unit cluster is treated as a whole to update the user's historical behavior map.

[0085] Specifically, the spatiotemporal behavioral unit clusters are transformed into a set of structured data, and user behavior types are determined. User behavior types are used to characterize the category to which user behavior belongs and serve as a key bridge connecting specific user operations with abstract nodes in the user behavior graph.

[0086] Specific examples of user behavior types are as follows: A user saying "Play soothing music" via voice is a "voice command type" user behavior, which will be captured by the multi-source behavior data acquisition module and tagged with a timestamp and location. A user tapping the "Radio Switch" button on the central control screen while driving is a "touch operation type" user behavior. The system detecting that a user is looking at the navigation interface for an extended period instead of the entertainment interface through eye-tracking devices is a "gaze-focused type" user behavior. A user smiling because they hear a favorite song is captured by the facial recognition module and is a "facial expression feedback type" user behavior. A user frequently scrolling through the entertainment content list after parking is an "interactive browsing type" user behavior.

[0087] These different user behavior types constitute a spatiotemporal behavior unit cluster. When updating the user's historical behavior graph, based on the specific behavior types contained in the spatiotemporal behavior unit cluster, these behavior types are associated with corresponding nodes in the user's historical behavior graph through predefined mapping rules. This allows for matching with corresponding "voice interaction nodes," "touch operation nodes," "preference feedback nodes," etc., in the user's historical behavior graph to determine whether there are target nodes in the user's historical behavior graph that match the user behavior type.

[0088] For example, voice command type → voice interaction node, touch operation type / interactive operation browsing type → touch operation node, facial expression feedback type / eye focus type → preference feedback node, where "→" indicates a mapping relationship.

[0089] If it is determined that there is a target node in the user's historical behavior graph that matches the user's behavior type, the current weight of the target node is updated based on the preset weight update rule to generate the updated target node; the user's historical behavior graph is updated based on the updated target node to generate the user's real-time behavior graph.

[0090] In some embodiments, updating the current weight of the target node based on a preset weight update rule to generate an updated target node includes: obtaining the current activity and current weight of the target node; and updating the target node based on a formula. Calculate the new weights of the target node; where, Represents the target node The updated weights Represents the target node The current weight, Represents the target node Current activity level This represents the activity level weighting coefficient. (∈0,1); update the current weight of the target node according to the new weight, and generate the updated target node.

[0091] Activity A dynamic indicator used to quantify the degree to which nodes in a user behavior graph are stimulated by spatiotemporal behavioral unit clusters within the current update cycle. Therefore, if a user behavior type in a spatiotemporal behavioral unit cluster is related to a node in the user's historical behavior graph, the activity of that related node will increase. Activity It is a weighted value determined by three dimensions: frequency of behavior, duration, and intensity of interaction, and directly reflects the intensity of the node's recent activation.

[0092] Activity ,in, To match the frequency of occurrence of behaviors in spatiotemporal behavioral unit clusters, The percentage of time the matching action lasts. The interaction strength for matching behaviors. , and These represent the weighting coefficients.

[0093] Specifically, frequency of occurrence For example, if a user actively switches songs 5 ​​times within 30 minutes using voice or touch, then the value of ffreq represents these 5 switching actions.

[0094] Duration percentage During this 30-minute drive, the car's music played for a total of 25 minutes. So... The value is 25 / 30 ≈ 0.83. This parameter measures the degree to which a user remains engaged in a particular behavior or content.

[0095] Interaction strength During the 25-minute music listening session, the user hummed along (detected via microphone), nodded or clapped to the rhythm (analyzed through facial expressions and micro-movements from the in-car camera). The overall intensity of these behaviors is... It measures the quality and depth of user interaction.

[0096] For example, during 25 minutes of music listening: Humming behavior: The user's humming is detected by the microphone. If the preset humming intensity coefficient is 0.6 and the humming duration is 10 minutes, then the score for this item is 0.6×(10÷25)=0.24. Rhythmic interactive behavior: The in-car camera detects the user nodding / clapping in rhythm. If the intensity coefficient of this behavior is preset to 0.4 and the interaction duration is 15 minutes, then the score for this item is 0.4×(15÷25)=0.24. The final interaction strength fint = 0.24 + 0.24 = 0.48 (the higher the value, the stronger the interaction quality and depth).

[0097] Therefore, in the specific implementation process, when determining whether there is a target node in the user's historical behavior graph that matches the user's behavior type, the current activity of the target node is calculated using the above activity calculation formula based on the matching result.

[0098] When updating the current weight of a target node based on a preset weight update rule, the current activity and current weight of the target node are obtained, and then the formula is used to update the target node's current weight. Calculate the new weight of the target node. Then update the current weight of the target node according to the new weight to generate the updated target node.

[0099] in, Represents the target node The weight value in "After Current Update" is the final output of the weight adjustment, which directly reflects the importance of the node (behavior node, preference node, or content node) in modeling user preferences after the update.

[0100] Represents the target node The weight value of the target node before this update. The weights determined based on historical behavioral data (such as the previous spatiotemporal behavioral unit cluster) are the basis for weight adjustment.

[0101] This represents the activity weighting coefficient, typically ranging from 0 to 1, used to balance the "current activity level". "and historical weight" "Regarding the updated weights" The impact. When When the value approaches 1, the updated weight depends more on the current activity level, meaning the system places greater emphasis on responding to the user's latest behavior: when When the value approaches 0, the updated weights rely more on historical weights, meaning the system is more inclined to retain users' long-term stable preferences.

[0102] In this embodiment, when updating the user's historical behavior graph based on spatiotemporal behavior unit clusters, only target nodes matching the user's behavior type are updated. This not only reduces computational complexity but also improves the efficiency of updating the behavior graph, thereby enhancing the real-time performance of recommendations. Furthermore, since the weights reflect the importance of each node in the behavior graph to content recommendations, updating the weights of target nodes to update the user's historical behavior graph improves the accuracy of content recommendations.

[0103] In some embodiments, updating the user's historical behavior graph based on spatiotemporal behavior unit clusters to generate a real-time user behavior graph further includes: if it is determined that all nodes in the user's historical behavior graph do not match the user's behavior type, then creating a new node based on the spatiotemporal behavior unit cluster; wherein the new node includes a node type; determining the weight of the new node based on the weights of all nodes in the user's historical behavior graph that match the node type; updating the user's historical behavior graph based on the new node and the weight of the new node to generate a real-time user behavior graph.

[0104] In the actual implementation process, since user interests and behaviors are constantly changing, the constructed user history behavior map may not have a target node that matches the user behavior type determined based on the current spatiotemporal behavior unit cluster. In order to include the user's latest behavior into the user history behavior map and expand the map boundary to cover new behavior and content dimensions, it is necessary to create new nodes based on the spatiotemporal behavior unit cluster.

[0105] The logic for establishing edge connections between new nodes and related nodes is as follows: First, clarify the category attribute of the new node, that is, determine whether it belongs to a behavior node, preference node, or content node. Then, use the average weight of this type of node in the user's historical behavior graph as the weight of the new node. Next, extract the spatiotemporal scene features (such as "parking scene"), interaction behavior type (such as "voice command"), and associated content tags (such as "music style tags corresponding to customized playlists") associated with the new node from the spatiotemporal behavior unit cluster, and match the nodes in the user's historical behavior graph that have these associated features as related nodes associated with the new node. Finally, establish edge connections between the new node and these related nodes, thereby incorporating the new node into the user's historical behavior graph and generating the user's real-time behavior graph.

[0106] In this embodiment of the application, by creating new nodes, the graph is dynamically updated based on changes in user interests, increasing the types of behaviors that the behavior graph can represent, expanding the boundaries of the behavior graph, and improving the richness of the behavior graph, thereby improving the accuracy of content recommendation.

[0107] In some embodiments, updating the user's historical behavior graph based on spatiotemporal behavior unit clusters to generate a real-time user behavior graph further includes: if it is determined that there are static nodes in the user's historical behavior graph whose weights have not been updated within a preset time period, then based on a preset weight decay function... Update the current weight of the static node to generate the updated static node; where; Represents static nodes The updated weights; Represents static nodes The last time it was effectively activated The weight value; Indicates the current system time; Represents static nodes The time point of the most recent effective activation; λ represents the time decay coefficient, which is dynamically determined based on the scene label; e represents the natural constant; the user's historical behavior graph is updated based on the updated static nodes to generate the user's real-time behavior graph.

[0108] As can be seen from the above embodiments, there are nodes with high activity in the user's historical behavior graph, and conversely, there are nodes with low activity. Therefore, for nodes that have not been activated for a long time, a weight decay is applied to the node to further reduce the weight value of the node, so that the user behavior graph is more inclined to user interests.

[0109] Therefore, if it is determined that there are static nodes in the user's historical behavior graph whose weights have not been updated within a preset time period, then a preset weight decay function will be used. The current weights of static nodes are updated to generate updated static nodes. Based on the updated static nodes, the user's historical behavior graph is updated to generate the user's real-time behavior graph.

[0110] in, Represents static nodes At the current time The weight value, that is, the real-time weight of the node after exponential decay, directly reflects the importance of the node in modeling user preferences at the current moment.

[0111] Represents static nodes "Time of last effective activation" The weight value is the initial baseline weight for this attenuation calculation.

[0112] This indicates the "current system time," which is the time when the node weight decay determination and calculation are triggered, such as the time when the system periodically checks the node activation status or the time when the graph is updated after a new behavioral unit cluster is input.

[0113] Represents static nodes The time of the most recent effective activation, that is, the time when the node was last associated with a cluster of effective behavioral units (behavioral association under non-accidental touch and non-abnormal scenarios), is recorded by the system in real time and stored in the node attributes.

[0114] λ: Represents the "time decay coefficient," which is a positive number used to control the rate at which node weights decay over time. It is dynamically adjusted based on the characteristics of the vehicle scenario. The larger the λ value, the faster the weight decays, and the more quickly the impact of long-term inactive nodes can be weakened; the smaller the λ value, the slower the weight decays, and the better the historical preference value of the node can be preserved.

[0115] It should be noted that for nodes that have been inactive for a long time, once their weight drops below a certain threshold, the node can be removed from the user behavior graph, and the edges associated with the node can be removed to achieve graph pruning, control graph complexity, and avoid invalid nodes affecting the accuracy of preference modeling.

[0116] In this embodiment, by reducing the weight of nodes with low activity in the behavior graph, the behavior graph representation is made more biased towards nodes that users are interested in, thereby improving the accuracy of content recommendation.

[0117] In some embodiments, during the construction of a real-time user behavior graph, or after the real-time user behavior graph has been constructed, if there are two or more nodes in the graph that are "semantically similar" (e.g., "voice request for music" and "voice selection of songs" are semantically similar) and "behavioral overlap" (the user achieves the purpose of playing music through both behaviors, and they are frequently associated in the behavior unit cluster), these nodes are merged into a unified node, and the edge weights of the original nodes are weighted and summed according to the association strength and then assigned to the new node, thereby simplifying the graph structure and avoiding semantically redundant nodes from occupying resources.

[0118] If a user experiences a "shift in interests" or a "sudden change in behavior path," such as switching from a long-standing preference for rock music to an obsession with crosstalk, the edge connections between the original interest-related nodes are disconnected, the edge relationships between the new interest nodes are re-established, the behavioral logic is reorganized to adapt to the new interests and behavioral paths, and the subgraph is reconstructed.

[0119] In some embodiments, matching a user's real-time behavior graph with a preset set of candidate content to generate content recommendation results includes: generating a user preference representation vector based on the user's real-time behavior graph; calculating the similarity between the user preference representation vector and the vector corresponding to the candidate content set to obtain a basic recommendation score for the recommended content; determining target recommended content based on a preset number of historical spatiotemporal behavior unit clusters; wherein the target recommended content includes a trend-weighted score; and generating content recommendation results based on the basic recommendation score and the trend-weighted score.

[0120] In the specific implementation process, each node in the real-time user behavior graph is embedded into a vector space using a graph embedding method, generating a node vector for each node. Based on the node vector of each node Weighted summation is performed to construct a user preference representation vector for the entire real-time user behavior graph. :

[0121] Candidate content set It is also vectorized as .

[0122] The cosine similarity function is used to calculate the similarity between the user preference representation vector and the corresponding vector in the candidate content set, thus obtaining a basic recommendation score list for the recommended content. The basic recommendation scores in the basic recommendation score list are sorted from largest to smallest.

[0123] It should be noted that similarity can also be calculated using other methods, such as Euclidean distance.

[0124] After obtaining the basic recommendation score list, the optimal recommendation list is output based on behavioral evolution trends. Further adjustments can be made based on the context.

[0125] Specifically, first, select the basic recommended score list. The recommended content corresponding to the top K basic recommendation scores forms an initial recommendation candidate pool.

[0126] Secondly, when adjusting the ranking based on behavioral evolution trends, let the weight changes of the historical spatiotemporal behavioral unit clusters for the preference nodes in the last N update cycles (e.g., N=3) be respectively... , , If the average weight change rate of preference nodes associated with a certain type of content... (θ is the trend threshold, which can be determined according to the actual situation), then a trend-weighted score is added to this type of content: = , This is the trend coefficient. This refers to the average weight change rate of user preference nodes, used to quantify the evolving trend of user interest in a certain type of content.

[0127] At this time, the real-time score for this type of content is: .

[0128] Based on real-time scores The recommended content in the initial recommendation candidate pool is reordered to generate content recommendation results.

[0129] Furthermore, the ranking can be calibrated in conjunction with the context, and a scene weight factor can be introduced. The content scores for different scenarios are adjusted accordingly. Finally according to Output the best recommendation list in descending order.

[0130] In this embodiment of the application, when generating content recommendation results, not only are the user's real-time preferences considered to obtain the basic recommendation score of the recommended content, but also the changes in the user's historical behavior are considered to capture the evolution trend of user behavior and obtain a trend-weighted score. Finally, the basic recommendation score and the trend-weighted score are combined to determine the content recommendation result, thereby improving the accuracy of content recommendation.

[0131] Figure 2 This is a schematic diagram of the structure of the first content recommendation system provided in the embodiments of this application, as shown below. Figure 2The system 20 is used to execute the method steps of any of the above embodiments; the system 20 includes a multi-source behavior data acquisition module 201, a behavior sequence spatiotemporal clustering module 202, a graph construction and dynamic evolution module 203, and a content adaptation module 204; wherein, the multi-source behavior data acquisition module 201 is used to acquire multimodal behavior data streams; the behavior sequence spatiotemporal clustering module 202 is used to generate spatiotemporal behavior unit clusters based on the multimodal behavior data streams; the graph construction and dynamic evolution module 203 is used to update the user's historical behavior graph based on the spatiotemporal behavior unit clusters to generate the user's real-time behavior graph; the content adaptation module 204 is used to match the user's real-time behavior graph with a preset set of candidate content to generate content recommendation results; wherein, the content recommendation results are used for content recommendation.

[0132] In some embodiments, the system further includes a feedback update module; wherein the feedback update module is used to receive user response data and optimize the real-time user behavior graph based on the user response data to generate an optimized real-time user behavior graph.

[0133] Figure 3 This is a schematic diagram of the structure of a second content recommendation system provided in an embodiment of this application. The schematic diagram illustrates the collaborative relationship between the modules when the content recommendation system performs content recommendation in an in-vehicle entertainment scenario. Figure 4 This is a content recommendation timing diagram provided for an embodiment of this application.

[0134] like Figure 3 and Figure 4 As shown, the multi-source behavioral data acquisition module is used to collect multimodal behavioral data of users during driving in the in-vehicle environment. Specifically, this includes: voice commands, touch operations, eye trajectory, facial expressions, interactive operation behaviors (such as menu browsing and playback control), and vehicle driving status (such as speed, location information, and driving status). The collected multimodal behavioral data is then constructed into a multimodal behavioral data stream and transmitted to the behavior sequence spatiotemporal clustering module.

[0135] The behavior sequence spatiotemporal clustering module includes a time window sliding unit, a spatial scene recognition unit, and a spatiotemporal cluster clusterer. The time window sliding unit divides the user's multimodal behavior data stream into temporal behavior sequences according to a preset time window ΔT, forming behavior temporal clusters. The spatial scene recognition unit uses a preset scene encoding function... The system generates scene labels based on user interaction behavior and geographical location information and driving status data from vehicle driving status data, such as "Highway - Call - Silent" and "City - Multiple Touches - Parking". A spatiotemporal clusterer is used to aggregate multimodal behavioral data under the same time period, geographical location, and scene label based on timestamps and scene labels, generating spatiotemporal behavioral unit clusters with spatiotemporal attributes. These spatiotemporal behavioral unit clusters serve as the basic input for the graph construction and dynamic evolution module.

[0136] The graph construction and dynamic evolution module 203 includes a graph node initialization unit, a node weight transfer unit, and a structural evolution engine. The graph node initialization unit initializes and generates a user historical behavior graph based on historically collected multimodal behavior data streams. The node weight transfer unit adjusts the weight values ​​of corresponding nodes based on newly input spatiotemporal behavior unit clusters; specifically, it increases the weight of nodes related to recent behavior, applies an exponential decay function to reduce the weight of nodes that have been inactive for a long time, and creates new nodes and connects edges if the behavior does not match existing nodes. The structural evolution engine performs node merging or subgraph reconstruction operations on the graph structure according to the changing trend when behavioral mutations (such as interest shifts) occur, realizing the structured transfer expression of behavioral preferences.

[0137] The content adaptation module is used to embed the real-time user behavior graph generated by the graph construction and dynamic evolution module into the vector space, and match it with the candidate content vectors. Based on the weight of user preference nodes and the behavior evolution trend, it generates recommendation ranking results.

[0138] The feedback update module is used to collect user response information to recommended content, including interactive behaviors and implicit reactions, and feeds the response results back to the graph construction and dynamic evolution module for subsequent graph structure optimization and preference correction.

[0139] In the in-vehicle environment, by fusing multi-source behavioral data (such as voice, touch, gaze, facial expressions, operational behaviors, and vehicle driving status), a user behavior graph with spatiotemporal semantics is constructed. Based on the dynamic evolution capability of the graph, personalized and real-time adaptive recommendations for in-vehicle entertainment content are achieved. The system extracts behavioral patterns through spatiotemporal clustering, and combines incremental updates of the graph structure, weight transfer, and behavioral mutation recognition to accurately reflect the changing trends of user interests. Then, through graph vector embedding and content matching, a deep match between content and the user's current preferences is achieved, thereby improving recommendation accuracy and user experience. At the same time, a feedback update mechanism optimizes the graph structure in a closed loop, enabling the system to have continuous self-learning and adaptive capabilities, effectively meeting the rapidly changing user preferences and diverse content needs in in-vehicle scenarios.

[0140] Figure 5 This is a schematic diagram of the electronic device structure provided in the embodiments of this application, such as... Figure 5As shown, the electronic device includes a processor 501, a memory 502, and a bus 503; wherein the processor 501 and the memory 502 communicate with each other via the bus 503. The processor 501 is used to call program instructions in the memory 502 to execute the methods provided in the above-described method embodiments.

[0141] Processor 501 can be an integrated circuit chip with signal processing capabilities. The processor 501 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor.

[0142] The memory 502 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0143] This application provides a computer program product, including: computer program instructions, which are executed by a processor to perform the methods provided in the above-described method embodiments.

[0144] This application provides a computer-readable storage medium, including: computer program instructions stored on the computer-readable storage medium, which, when executed by a processor, perform the methods provided in the above-described method embodiments.

[0145] Computer-readable storage media can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0146] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0147] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0148] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0149] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A content recommendation method, characterized in that, The method includes: Acquire a multimodal behavior data stream; wherein the multimodal behavior data stream includes multiple multimodal behavior data; the multimodal behavior data includes timestamps and scene tags; The multimodal behavior data stream is aggregated based on the timestamp and the scene label to generate a spatiotemporal behavior unit cluster; Based on the spatiotemporal behavior unit cluster, the user's historical behavior graph is updated to generate the user's real-time behavior graph. The user's real-time behavior graph is matched with a preset set of candidate content to generate content recommendation results; wherein, the content recommendation results are used for content recommendation.

2. The method according to claim 1, characterized in that, After generating the content recommendation results, the method further includes: Receive user response data; The user real-time behavior graph is optimized based on the response data to generate an optimized user real-time behavior graph. Content recommendation is based on the optimized real-time user behavior graph.

3. The method according to claim 1, characterized in that, The step of updating the user's historical behavior graph based on the spatiotemporal behavior unit cluster to generate the user's real-time behavior graph includes: The user behavior type is determined based on the spatiotemporal behavior unit cluster; If it is determined that there is a target node in the user's historical behavior graph that matches the user's behavior type, then the current weight of the target node is updated based on the preset weight update rule to generate the updated target node. The user's historical behavior graph is updated based on the updated target node to generate the user's real-time behavior graph.

4. The method according to claim 3, characterized in that, The step of updating the current weight of the target node based on a preset weight update rule to generate an updated target node includes: Obtain the current activity level and current weight of the target node; Based on formula Calculate the new weights of the target node; where, Indicates the target node The updated weights Indicates the target node The current weight, Indicates the target node Current activity level This represents the activity level weighting coefficient. ∈ (0, 1); The current weight of the target node is updated according to the new weight to generate the updated target node.

5. The method according to claim 3, characterized in that, The step of updating the user's historical behavior graph based on the spatiotemporal behavior unit cluster to generate the user's real-time behavior graph also includes: If it is determined that none of the nodes in the user's historical behavior graph match the user's behavior type, then a new node is created based on the spatiotemporal behavior unit cluster; wherein, the new node includes a node type; The weight of the new node is determined based on the weights of all nodes in the user's historical behavior graph that match the node type. The user's historical behavior graph is updated based on the new node and its weight to generate the user's real-time behavior graph.

6. The method according to claim 3, characterized in that, The step of updating the user's historical behavior graph based on the spatiotemporal behavior unit cluster to generate the user's real-time behavior graph also includes: If it is determined that there are static nodes in the user's historical behavior graph whose weights have not been updated within a preset time period, then based on the preset weight decay function... The current weight of the static node is updated to generate the updated static node; in; Represents the static node The updated weights; Represents the static node The last time it was effectively activated The weight value; Indicates the current system time; Represents the static node The time point of the most recent effective activation; λ represents the time decay coefficient, which is dynamically determined based on the scene label; e represents the natural constant. The user's historical behavior graph is updated based on the updated static nodes to generate the user's real-time behavior graph.

7. The method according to claim 1, characterized in that, The step of matching the user's real-time behavior graph with a preset set of candidate content to generate content recommendation results includes: Generate a user preference representation vector based on the real-time user behavior graph; Calculate the similarity between the user preference representation vector and the vector corresponding to the candidate content set to obtain the basic recommendation score of the recommended content; The target recommended content is determined based on a preset number of historical spatiotemporal behavior unit clusters; wherein, the target recommended content includes trend-weighted scores; The content recommendation result is generated based on the basic recommendation score and the trend-weighted score.

8. The method according to any one of claims 1-7, characterized in that, The step of aggregating the multimodal behavior data stream based on the timestamp and the scene label to generate a spatiotemporal behavior unit cluster includes: Based on the timestamp, the multimodal behavior data stream is divided into multiple multimodal temporal behavior sequences according to a preset time window; For each of the multimodal temporal behavior sequences, the multimodal behavior data corresponding to the scene labels with similarity greater than a preset similarity threshold are aggregated to generate multiple behavior clusters; Based on the multiple behavior clusters, the spatiotemporal behavior unit cluster is generated.

9. A content recommendation system, characterized in that, The system is used to perform the method described in any one of claims 1-8; the system includes a multi-source behavioral data acquisition module, a behavioral sequence spatiotemporal clustering module, a graph construction and dynamic evolution module, and a content adaptation module; The multi-source behavioral data acquisition module is used to acquire multimodal behavioral data streams. The behavior sequence spatiotemporal clustering module is used to generate spatiotemporal behavior unit clusters based on the multimodal behavior data stream; The graph construction and dynamic evolution module is used to update the user's historical behavior graph based on the spatiotemporal behavior unit cluster, and generate the user's real-time behavior graph. The content adaptation module is used to match the user's real-time behavior graph with a preset set of candidate content to generate content recommendation results; wherein, the content recommendation results are used for content recommendation.

10. The system according to claim 9, characterized in that, The system also includes a feedback update module; wherein the feedback update module is used to receive user response data and optimize the real-time user behavior graph based on the user response data to generate an optimized real-time user behavior graph.

Images