Sensing data intelligent recommendation system

Abstract

The invention relates to the field of artificial intelligence and personalized recommendation, and discloses a sensing data intelligent recommendation system. The system comprises a perception data acquisition module, a multi-dimensional feature analysis module, a user portrait construction module, an intelligent recommendation decision module and an enhanced feedback optimization module. The method comprises the following steps: acquiring original data through a multi-source sensing device, and extracting multi-dimensional statistical characteristics after standardization processing; a user portrait model is constructed in combination with user historical behaviors, candidate recommendation vectors are generated, a recommendation strategy is dynamically optimized based on user feedback, efficient, accurate and personalized recommendation results are achieved, and the method is suitable for complex application environments such as intelligent interaction, mobile terminals and intelligent scenes.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and personalized recommendation, and in particular to a perceptual data-driven intelligent recommendation system. Background Technology

[0002] With the rapid development of IoT, big data, and AI technologies, personalized recommendation systems based on multi-source sensing data have been widely applied in smart homes, smart cities, and personalized content delivery. Multimodal sensors can collect user behavior, environmental, and device usage data in real time, providing rich contextual information for recommendation systems. However, existing systems still have shortcomings in data collection consistency and format fusion capabilities, leading to incomplete or inconsistent data input to downstream recommendation models, affecting the accuracy and real-time performance of recommendations.

[0003] In terms of feature extraction and user modeling, existing technologies mostly employ shallow feature combinations and static labeling systems, making it difficult to fully explore deep feature relationships in multidimensional time-series data. This is especially true when dealing with dynamic changes in user behavior, where a flexible label update mechanism is lacking. Furthermore, some recommendation systems fail to effectively utilize feedback mechanisms, unable to optimize models and adjust strategies in real time based on user feedback behavior across different devices and times, leading to a deviation between recommended content and actual user preferences.

[0004] Furthermore, existing intelligent recommendation systems often overlook the complex interplay between contextual factors and multi-objective business needs during the candidate selection and ranking stages. For example, they lack a balancing mechanism between accuracy, response speed, and system load, easily leading to problems such as delayed recommendation results and low resource utilization. Therefore, there is an urgent need for an intelligent recommendation system with high adaptability, multi-dimensional modeling capabilities, and feedback-driven optimization mechanisms to improve recommendation accuracy, response efficiency, and cross-scenario consistency. Summary of the Invention

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution: a perceptual data intelligent recommendation system, comprising the following modules:

[0006] The sensing data acquisition module is used to acquire raw sensing data D through multi-source sensing devices. raw The perceived data is then formatted and standardized to form standard input data D. std ;

[0007] The multidimensional feature parsing module is used to analyze the standard input data D. std Temporal correlation and multi-dimensional feature extraction are performed, and the extracted feature vector is represented as F. vec ;

[0008] The user profile building module is used to construct user profiles based on the feature vector F. vec Combine user historical behavior data U beh Constructing a multi-level user tagging model U tag ;

[0009] The intelligent recommendation decision-making module is used to make decisions based on the user tag model U. tag With external context factor C ext Generate candidate recommendation vectors R can ;

[0010] The enhanced feedback optimization module is used to optimize user feedback based on actual user feedback. usr For the candidate recommendation vector R acn The strategy is updated and dynamically adjusted to obtain the final recommendation result R. fin .

[0011] Preferably, the sensing data acquisition module includes:

[0012] The raw data acquisition unit is used to acquire raw sensing data D through a multimodal sensor network consisting of a temperature sensor, a motion detector, and an ambient light sensor. raw And uploads via a multi-channel concurrent interface;

[0013] The format standardization processing unit is used to process the raw sensor data D. raw Normalization is performed, and a weighted interpolation algorithm is used to generate standard input data D. std The weighted interpolation algorithm is defined as follows:

[0014]

[0015] in:

[0016] D std This represents the standard input data after processing;

[0017] D raw This represents the raw sensing data acquired by the sensor;

[0018] Let represent the weight coefficient of the j-th type of sensor, and satisfy . n represents the total number of sensors.

[0019] Preferably, the multidimensional feature parsing module includes:

[0020] Feature construction unit, used for processing standard input data D based on sliding window mechanism std Time-series slicing is performed to calculate multiple statistical feature vectors within each window segment. These statistical feature vectors include the mean vector μ. seg , standard deviation vector σseg ;

[0021] The feature fusion unit is used to construct a fusion expression based on the statistical feature vector to extract the final feature vector F. vec The fusion expression is:

[0022]

[0023] in:

[0024] μ seg It is the mean vector;

[0025] σ seg The standard deviation vector;

[0026] This represents the rate of change of the mean vector over time.

[0027] This represents the rate of change of the standard deviation vector over time.

[0028] β1, β2, β3, and β4 are fusion weighting coefficients, and β1 must satisfy the following condition: 2 +β2 2 +β3 2 +β4 2 =1.

[0029] Preferably, the user profile building module includes:

[0030] The user historical data integration unit is used to extract user historical behavior data U beh The user's historical behavior data U beh This includes visit frequency, click path, and dwell time;

[0031] The label generation unit employs a multi-layer graph neural network model to generate feature vectors F. vec With user historical behavior data U beh Mapped to the label space, the output is the user label model U. tag The mapping process formula is as follows:

[0032]

[0033] in:

[0034] α k It is a label weight factor, and satisfies

[0035] F vec *U beh Represents the eigenvector F vec With user historical behavior data U beh Perform convolution operations;

[0036] K represents user historical behavior data. beh The quantity, where K = 3.

[0037] The user profile building module further includes a dynamic user interest adjustment mechanism, used to adjust the user tag model U based on the latest behavioral characteristics during the recommendation period. tag It performs real-time updates to adapt to dynamic changes in user interests, and stores snapshots of user profiles at different time periods through a caching mechanism for subsequent comparison and optimization of recommendation models.

[0038] Preferably, the intelligent recommendation decision module includes:

[0039] Relevance scoring unit, used to calculate user label model U tag With each target resource T in the candidate resource pool res The multidimensional similarity score between them S sim The scoring algorithm is as follows:

[0040]

[0041] in:

[0042] u tag T User tagging model U tag The transpose of the matrix;

[0043] T res (i) For the i-th candidate resource target resource, determined by user level V meta User tag V label User activity V score Composition, denoted as T res (i) ={V meta (i) V label (i) V score (i)};

[0044] ||U tag ||and||T res (i) ||For the user tagging model U tag The modulus of the matrix of the i-th candidate resource and the target resource;

[0045] The candidate recommendation generation unit adopts a multi-objective ranking optimization strategy, combined with the external context factor C. ext Based on the business objective weights, a candidate recommendation vector R is generated. can :

[0046]

[0047] The candidate recommendation generation unit further integrates a multi-objective ranking and control mechanism, which is used to dynamically adjust the candidate resource screening threshold under multiple business objectives such as recommendation accuracy, response latency, and system load, so as to improve the overall recommendation efficiency of the system and support the adaptive balance between personalization and global optimization.

[0048] Preferably, the enhanced feedback optimization module includes:

[0049] The user feedback collection unit is used to receive actual user feedback F usr The actual user feedback includes click behavior, rating results, and dwell time metrics;

[0050] The strategy update unit is used to update the strategy based on actual user feedback. usr With the current candidate recommendation vector R can The differences between the recommendations are used to update and dynamically adjust the strategy, and the final recommendation result R is obtained based on the following expression. fin :

[0051] R fin (i) =R can (i) +λ·(F usr (i) -R can (i) );

[0052] in:

[0053] R fin (i) This represents the final score of the i-th recommendation result after policy optimization;

[0054] λ is the feedback response adjustment coefficient.

[0055] The enhanced feedback optimization module further supports continuous feedback fusion strategies, enabling unified modeling and analysis of feedback information from the same user across multiple time periods and devices. This improves the recommendation model's adaptability to cross-terminal behavioral characteristics in complex scenarios and maintains the consistency and coherence of recommended content across different interactive terminals.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0057] This invention introduces a multimodal perception fusion module, which supports the synchronous collection and normalization fusion of behavioral data, environmental data, and physiological data. This effectively improves the consistency and multidimensional linkage capabilities of the perception data, providing high-quality input for subsequent user modeling and personalized recommendations.

[0058] This invention constructs a user modeling mechanism based on a deep dynamic tag graph network, which can depict the evolution of user interests in real time and automatically update the relationship between tag nodes and edge weights, thereby significantly improving the matching degree between recommended content and users' current real needs.

[0059] This invention introduces a multi-objective reinforcement learning algorithm into the feedback-driven recommendation optimization module, which not only supports data adaptive learning based on multi-round feedback, but also achieves a dynamic balance between accuracy, response speed and resource efficiency, thereby enhancing the system's generalization ability and stability in multiple scenarios. Attached Figure Description

[0060] Figure 1 The system module flowchart provided for this application;

[0061] Figure 2 A schematic diagram of the sensing data acquisition module provided in this application;

[0062] Figure 3 A schematic diagram of the multidimensional feature parsing module provided in this application;

[0063] Figure 4 A schematic diagram of the user profile building module provided for this application;

[0064] Figure 5 A schematic diagram of the intelligent recommendation decision-making module provided in this application;

[0065] Figure 6 A schematic diagram of the enhanced feedback optimization module provided in this application. Detailed Implementation

[0066] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0067] refer to Figures 1-6 This invention provides a perceptual data-driven intelligent recommendation system, comprising the following modules:

[0068] The sensing data acquisition module is used to acquire raw sensing data D through multi-source sensing devices. raw The perceived data is then formatted and standardized to form standard input data D. std .

[0069] In the sensing data acquisition module, which is designed as the core front-end module of the system, the module is responsible for the efficient, stable, and scalable acquisition and preliminary processing of various sensor information. This module mainly consists of two sub-units: a raw data acquisition unit and a format standardization processing unit. These units work together to complete the entire process from physical data acquisition to structured data output, ensuring that downstream modules can perform further analysis and modeling based on accurate and consistent data.

[0070] First, in the raw data acquisition unit, the system integrates a multimodal sensor network, which consists of temperature sensors, motion detectors, and ambient light sensors. Each type of sensor communicates with the system's main control platform through a corresponding microcontroller unit, supporting SPI and I / O. 2 The system utilizes multiple communication protocols, including C and UART, to ensure the concurrency and stability of data transmission. A multi-channel concurrent interface architecture is designed, capable of simultaneously acquiring sensor data streams from multiple sensors. raw Furthermore, the data streams from each sensor are timestamped to ensure that the data remains synchronized in time and spatially relevant before entering the next stage.

[0071] Secondly, the format standardization processing unit is responsible for processing the acquired raw sensor data D. raw The task is to perform unified processing. Because the signals output by different sensors have different physical quantity dimensions, inconsistent sampling frequencies, and data distribution differences, directly inputting them into the subsequent modeling module would lead to decreased accuracy and system instability. Therefore, in this sub-unit, the system employs a weighted interpolation algorithm to normalize and weightedly fuse the data streams output by all sensors, generating standardized input data D. std The processing algorithm sets a weight coefficient for each sensor type, using sensor type as the dimension. The weighting coefficients reflect the contribution and reliability of this type of sensor in the overall sensing task, and satisfy... Specifically, the system performs linear interpolation and unit-scale normalization on the data collected by all sensors, and then performs weighted summation according to the aforementioned weighting coefficients to obtain a unified D. std The format and algorithm are as follows:

[0072]

[0073] To improve processing accuracy, this embodiment adjusts the weighting coefficients. The determination of the weights employs a dynamic weight adjustment mechanism based on sample validity assessment. Specifically, within each fixed time window (every 10 seconds), the system evaluates the impact of various sensor data on label modeling or prediction results during that time period, and adjusts the weights of each sensor type using a combination of moving average and entropy weighting. This allows the weight configuration to adapt to changes in different environmental conditions and user states. This effectively avoids the performance degradation problem of traditional fixed-weight mechanisms in dynamic environments.

[0074] Furthermore, the format standardization processing unit also integrates data integrity detection and outlier removal mechanisms. For example, when a sensor's data output is detected to be a fixed value or far from the mean range for a short period, the system will automatically trigger an interpolation reconstruction or alarm mechanism to ensure the final D... std Data continuity and reliability. All processed data D std It will be encapsulated into a structured input frame and uploaded to the user modeling module for further processing.

[0075] The multidimensional feature parsing module is used to analyze the standard input data D. std Temporal correlation and multi-dimensional feature extraction are performed, and the extracted feature vector is represented as F. vec .

[0076] In the multidimensional feature parsing module, which serves as the core functional unit connecting the preceding and following steps in the system, the main function is to process the formatted standard input data D. std This module performs in-depth feature extraction and representation transformation to provide discriminative and dynamically adaptive feature inputs for personalized modeling and recommendation. It consists of a feature construction unit and a feature fusion unit, relying on a sliding window mechanism and dynamic statistical feature modeling algorithms to achieve multi-dimensional evolutionary extraction from raw statistics to fused feature vectors.

[0077] First, the main task of the feature construction unit is to extract continuous standard input data D. std In this process, statistical features of time-series segments are extracted. In this embodiment, the system employs an equally spaced sliding window mechanism to divide D... std The system divides the data into several overlapping or non-overlapping time periods based on a preset time window length (5 seconds) and sliding step size (1 second). Each period is denoted as a window segment. For the data within each window segment, the system calculates its mean vector μ. seg and standard deviation vector σ seg This reflects the overall trend and fluctuations within the time period. The calculation is performed independently on the data dimension corresponding to each type of sensor, forming a corresponding multidimensional statistical vector. This effectively captures the changing patterns of data at micro-timescales, laying the foundation for subsequent dynamic modeling.

[0078] Secondly, in the feature fusion unit, the system introduces a fusion expression to linearly combine the aforementioned static statistics with their time derivatives, thereby obtaining a more dynamically sensitive final feature vector F. vecSpecifically, the system calculates the first-order time derivative of the mean vector and standard deviation vector within each of the aforementioned window segments to obtain the rate of change of the mean. and standard deviation change rate These reflect the rate of change of the data's mean and volatility over time, respectively. These first derivatives are calculated using the finite difference method, which involves taking the difference between two adjacent sliding windows and dividing by the window step size, ensuring both computational simplicity and real-time performance.

[0079] Next, the system integrates the above four types of feature terms according to the following fusion expression:

[0080]

[0081] Wherein, β1, β2, β3, and β4 are feature fusion weighting coefficients, used to balance the influence weights of static and dynamic features in the final feature vector. To ensure the stability and comparability of the vector across scales, the fusion coefficients must satisfy the constraint β1. 2 +β2 2 +β3 2 +β4 2 =1, and this constraint is solved using the Lagrange multiplier method during system initialization. In the initial stage, the system sets a set of default coefficients based on the requirements for stability, sensitivity, and responsiveness in different recommendation scenarios, and gradually adjusts and optimizes them according to performance evaluation indicators during operation.

[0082] Furthermore, to enhance the robustness of feature representation and noise suppression capabilities, this embodiment introduces a normalization and filtering mechanism for features within a window before calculating the fusion expression. Specifically, Z-score normalization is first performed on each feature term to make its mean 0 and standard deviation 1. Then, a one-dimensional Gaussian filter is used to smooth the rate of change, suppressing instantaneous fluctuations caused by sensor jitter or occasional events, and improving the final feature vector F. vec The stability and discriminative ability of expression.

[0083] The user profile building module is used to construct user profiles based on the feature vector F. vec Combine user historical behavior data U beh Constructing a multi-level user tagging model U tag .

[0084] The user profile building module serves as a bridge connecting perceived data features and personalized recommendation strategies, undertaking the crucial task of fusing and modeling environmental perception features acquired by the system with user behavior information. This module primarily comprises two sub-units: a user historical data integration unit and a tag generation unit. These units work synergistically to perform functions such as user interest identification, personalized tag generation, and time-series management of user profiles, thereby achieving dynamic adaptation and refined control of the recommendation strategy.

[0085] First, the user historical data integration unit connects to the user interaction log database interface in the system and periodically retrieves user historical behavior data. beh The data covers multiple dimensions of information, including user access frequency, click paths, and page dwell time on the platform. To ensure data quality, this unit performs structured processing and cleaning on the collected behavioral data, using window statistics to generate behavioral summary features while removing low-frequency noise behaviors and discontinuous jump paths to ensure temporal consistency and behavioral stability of subsequent modeling data. After processing, the system uniformly encodes the behavioral summaries and transforms them into tensor form, which is then used as part of the input to the graph neural network, providing computable behavioral basis for the label generation process.

[0086] Subsequently, the label generation unit uses the standard input feature vector F vec and user historical behavior data U beh As input, a multi-layer graph neural network is used for feature fusion and user label generation. This multi-layer graph neural network constructs a heterogeneous graph structure, where nodes represent the user's historical behavior state and the system's current perception state, and the edge weights are determined by the system's current feature vector F. vec The interaction strength between the interaction and historical behavior is represented by the inner product operation F based on the convolution mechanism. vec *U beh This is to capture the coupling pattern between behavioral habits and current perceived conditions. During the network's forward propagation, each graph convolutional layer constructs and updates user labels according to the following formula:

[0087]

[0088] Where K represents the type of behavioral data (in this system, it is three types: access frequency, click path, and dwell time), α k A fusion weighting factor is assigned to each behavior type to balance the influence of various behaviors on user interests and preferences, and to satisfy the following conditions: These weight coefficients can be optimized by the system based on user historical recommendation feedback, or initial values ​​can be set and updated online through strategy guidance, thereby improving the personalization and accuracy of model label generation.

[0089] To further enhance the responsiveness of user profiles to changes in interests, this embodiment introduces a dynamic user interest adjustment mechanism. This mechanism comprises two components: a behavior update detector and a profile snapshot cache. The behavior update detector monitors new user behavior data in real time within the current recommendation period to determine if significant interest drift exists. When a trend change is detected in the frequency of a user's access to a certain type of content or their click path, the system immediately triggers the user tag model U. tagThe system features online fine-tuning updates. This update mechanism uses a sliding window approach, initiating incremental graph neural network iterations only when the latest perceptual and behavioral features significantly deviate from the historical model, thereby reducing system overhead and improving model response efficiency.

[0090] The intelligent recommendation decision-making module is used to make decisions based on the user tag model U. tag With external context factor C ext Generate candidate recommendation vectors R can .

[0091] The intelligent recommendation decision-making module, as the core execution component of the perceptual data intelligent recommendation system, undertakes the crucial task of selecting the optimal recommendation results from candidate resources. This module mainly includes two functional units: a relevance scoring unit and a candidate recommendation generation unit. It further integrates a multi-objective ranking and control mechanism to achieve dynamic synergistic optimization between high accuracy, high response efficiency, and high system robustness.

[0092] First, the relevance scoring unit is used to construct a high-dimensional similarity evaluation model between user profiles and candidate resources. The system obtains the user tag model U generated by the user profile construction module. tag Next, retrieve all target resources T to be recommended from the current candidate resource library. res (i) Each resource contains user level information V meta (i) Label feature vector V label (i) And user activity score V score (i) The three together constitute the feature vector set representation of the target resource:

[0093] T res (i) ={V meta (i) V label (i) V score (i)};

[0094] During the scoring process, the system performs the following similarity calculation for each resource:

[0095]

[0096] This formula is based on the cosine similarity metric, calculating the angular similarity between the transpose of the user tag model and the target resource vector in vector space by taking the inner product of their inner products. This method considers both the matching degree between user interest direction and resource feature space, exhibiting good stability and discriminative power, and can adapt to performance requirements under high-dimensional sparse feature conditions. To improve processing efficiency, the system employs a vectorized parallel computing approach to perform batch scoring operations on all candidate resources, ensuring millisecond-level response time for recommendation decisions.

[0097] After completing the initial similarity assessment, the candidate recommendation generation unit is responsible for combining the similarity score with the context-aware factor C. ext Candidate resources are ranked, optimized, and filtered. Context factors include dynamic parameters such as time window, device terminal, geographical location, and environmental state. These exist in the system as a weight matrix and are used to adjust the sensitivity and acceptability of recommendations under different scenarios. This unit generates candidate recommendation vectors R in the following manner. can :

[0098]

[0099] Where n is the number of candidate resources. In this formula, the system constructs the most suitable set of recommended resources for the user's current scenario by weighted fusion of similarity scores and contextual regulation factors. In its implementation, to avoid overfitting, the system introduces an entropy regularization strategy to limit the extreme weights of contextual factors, thereby ensuring the stability and universality of the recommendation results.

[0100] To achieve a higher level of comprehensive performance optimization, this embodiment further integrates a multi-objective ranking and control mechanism into the candidate recommendation generation unit. This mechanism introduces three performance metrics as ranking objectives into the system's business logic: recommendation accuracy, response latency, and system load. It also globally optimizes the candidate resource selection process by constructing a dynamic threshold control model. Based on a reinforcement learning framework, the control model perceives the current system operating state in each round of recommendation tasks and automatically adjusts the resource selection threshold based on the objective function feedback. Recommendation accuracy is measured by user click-through rate or satisfaction index; response latency is dynamically calculated from system processing time; and system load is estimated through resource call frequency and server occupancy.

[0101] The enhanced feedback optimization module is used to optimize user feedback based on actual user feedback. usr For the candidate recommendation vector R can The strategy is updated and dynamically adjusted to obtain the final recommendation result R. fin .

[0102] The enhanced feedback optimization module, a key component of the perceptual data intelligent recommendation system for achieving closed-loop tuning and continuous learning, primarily functions to collect, model, and fuse user behavioral feedback during actual use in real time. This dynamically updates the recommendation strategy and outputs a final recommendation result that better reflects the user's personalized preferences and the current context. This module comprises two basic sub-units: a user feedback collection unit and a strategy update unit. It further integrates a continuous feedback fusion strategy to enhance the consistency of behavioral modeling across different terminals and time periods.

[0103] Specifically, the user feedback collection unit is responsible for continuously monitoring and acquiring various feedback data generated during user interaction with recommended content, including but not limited to explicit or implicit behavioral indicators such as click behavior, rating results, and page dwell time. This feedback data is uniformly labeled as the user actual feedback vector F. usr The data is encoded in a high-dimensional feature format within the system, incorporating multiple metadata dimensions such as timestamps, device identifiers, and content identifiers, thus ensuring traceability and attributability in subsequent processing. The system employs a combination of edge acquisition and central aggregation to ensure both low latency and global integrity in the feedback acquisition process. Regarding data quality assurance, the system introduces an outlier detection mechanism and a weighted reliability evaluation strategy to mitigate the impact of extreme behaviors on strategy updates.

[0104] After obtaining actual user feedback, the policy update unit updates the current candidate recommendation vector R. can Based on actual user feedback F usr The differences between the information are used to fine-tune and optimize the recommendation strategy in real time. The recommendation strategy update adopts the classic prediction-feedback correction mechanism in reinforcement learning, and the final recommendation vector is adjusted according to the following formula:

[0105] R fin (i) =R can (o) +λ·(F usr (i) -R can (o) ).

[0106] Among them, R fin (i)Let λ represent the final score of the i-th recommendation result after strategy optimization, and λ be the feedback response adjustment coefficient, used to control the degree of influence of feedback on the current strategy. This formula reflects the system's linear weighted correction of the confidence level of each candidate result based on feedback; the greater the deviation of the feedback information, the higher the magnitude of the recommendation strategy correction. In actual implementation, the system adopts a dynamic adaptive setting mechanism for λ, adjusting it through the model convergence rate and the effectiveness of historical feedback, so that the strategy can both quickly respond to changes in user preferences and avoid overfitting caused by short-term behavioral noise.

[0107] Furthermore, to enhance the stability and continuity of the recommendation system across multiple scenarios, this embodiment introduces a continuous feedback fusion strategy in the enhanced feedback optimization module. This strategy constructs a more comprehensive and stable dynamic profile of user interests by uniformly modeling and analyzing user feedback data across different time periods and devices. The system first normalizes cross-terminal feedback data, including eliminating device differences and correcting consistency in behavioral feature mapping, ensuring the comparability of feedback from different terminals. Subsequently, the system uses a multi-granularity time series model and a sequence attention mechanism to weighted aggregate historical feedback, extracting long-term preference trends and short-term interest changes in user behavior. Through this mechanism, the system can effectively learn the evolutionary patterns of user behavior across intraday, interday, and even periodic dimensions, providing more accurate contextual support for the recommendation strategy.

[0108] This continuous feedback fusion strategy also supports a cross-terminal consistency guarantee mechanism, ensuring that users receive a consistent recommended content presentation experience across different interactive terminals such as mobile devices, web pages, and smart TVs. The system establishes a unique user identifier bound to a terminal and utilizes a distributed recommendation caching and synchronization mechanism to share intermediate results of the recommendation strategy across terminals, achieving real-time synchronization and semantic coherence of recommended content. Through these mechanisms, the system significantly improves the accuracy and stability of user interest modeling in complex interactive environments, enhancing the overall adaptability of the recommendation system and the consistency of user experience.

[0109] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the present invention can be combined with each other.

[0110] Obviously, the embodiments described above are merely some embodiments of the present invention, not all embodiments. The accompanying drawings show preferred embodiments of the present invention, but do not limit the scope of the patent. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this invention.

Claims

1. A sensory data-driven intelligent recommendation system, characterized in that, Includes the following modules: The sensing data acquisition module is used to acquire raw sensing data D through multi-source sensing devices. raw The perceived data is then formatted and standardized to form standard input data D. std ; The multidimensional feature parsing module is used to analyze the standard input data D. std Temporal correlation and multi-dimensional feature extraction are performed, and the extracted feature vector is represented as F. vec ; The user profile building module is used to construct user profiles based on the feature vector F. vec Combine user historical behavior data U beh Constructing a multi-level user tagging model U tag ; The intelligent recommendation decision-making module is used to make decisions based on the user tag model U. tag With external context factor C ext Generate candidate recommendation vectors R can ; The enhanced feedback optimization module is used to optimize user feedback based on actual user feedback. usr For the candidate recommendation vector R can The strategy is updated and dynamically adjusted to obtain the final recommendation result R. fin .

2. The intelligent recommendation system based on sensory data according to claim 1, characterized in that, The sensing data acquisition module includes: The raw data acquisition unit is used to acquire raw sensing data D through a multimodal sensor network consisting of a temperature sensor, a motion detector, and an ambient light sensor. raw And uploads via a multi-channel concurrent interface; The format standardization processing unit is used to process the raw sensor data D. raw Normalization is performed, and a weighted interpolation algorithm is used to generate standard input data D. std The weighted interpolation algorithm is defined as follows: in: D std This represents the standard input data after processing; D raw This represents the raw sensing data acquired by the sensor; Let represent the weight coefficient of the j-th type of sensor, and satisfy . n represents the total number of sensors.

3. The intelligent recommendation system based on sensory data according to claim 1, characterized in that, The multidimensional feature parsing module includes: Feature construction unit, used for processing standard input data D based on sliding window mechanism std Time-series slicing is performed to calculate multiple statistical feature vectors within each window segment. These statistical feature vectors include the mean vector μ. seg , standard deviation vector σ seg ; The feature fusion unit is used to construct a fusion expression based on the statistical feature vector to extract the final feature vector F. vec The fusion expression is: in: μ seg It is the mean vector; σ seg The standard deviation vector; This represents the rate of change of the mean vector over time. This represents the rate of change of the standard deviation vector over time. β1, β2, β3, and β4 are fusion weighting coefficients, and β1 must satisfy the following condition: 2 +β2 2 +β3 2 +β4 2 =1.

4. The intelligent recommendation system based on sensory data according to claim 1, characterized in that, The user profile building module includes: The user historical data integration unit is used to extract user historical behavior data U beh The user's historical behavior data U beh This includes visit frequency, click path, and dwell time; The label generation unit employs a multi-layer graph neural network model to generate feature vectors F. vec With user historical behavior data U beh Mapped to the label space, the output is the user label model U. tag The mapping process formula is as follows: in: α k It is a label weight factor, and satisfies F vec *U beh Represents the eigenvector F vec With user historical behavior data U beh Perform convolution operations; K represents user historical behavior data. beh The quantity, where K = 3.

5. The intelligent recommendation system based on sensory data according to claim 1, characterized in that, The intelligent recommendation decision-making module includes: Relevance scoring unit, used to calculate user label model U tag With each target resource T in the candidate resource pool res The multidimensional similarity score between them S sim The scoring algorithm is as follows: in: U tag T User tagging model U tag The transpose of the matrix; T res (i) For the i-th candidate resource target resource, determined by user level V meta User tag V label User activity V score Composition, denoted as T res (i) ={V meta (i) V label (i) V score (i) }; ||U tag ||and||T res (i) ||For the user tagging model U tag The modulus of the matrix of the i-th candidate resource and the target resource; The candidate recommendation generation unit adopts a multi-objective ranking optimization strategy, combined with the external context factor C. ext Based on the business objective weights, a candidate recommendation vector R is generated. can :

6. The intelligent recommendation system based on sensory data according to claim 1, characterized in that, The enhanced feedback optimization module includes: The user feedback collection unit is used to receive actual user feedback F usr The actual user feedback includes click behavior S click Score results S rate Dwell time S time Among them, after the user clicks, S click =1, if the user does not click, S click =0; S = 0 when the user's rating is greater than 80. rate =1, when the score is less than or equal to 80, S rate =0; when the user stays for more than 1 minute, S time =1, otherwise, S time =0; The strategy update unit is used to update the strategy based on actual user feedback. usr With the current candidate recommendation vector R can The differences between the recommendations are used to update and dynamically adjust the strategy, and the final recommendation result R is obtained based on the following expression. fin : R fin (i) =R can (i) +λ·(F usr (i) -R can (i) ); in: R fin (i) This represents the final score of the i-th recommendation result after policy optimization; λ is the feedback response adjustment coefficient.

7. The intelligent recommendation system based on perceptual data according to claim 4, characterized in that, The user profile building module further includes a dynamic user interest adjustment mechanism, used to adjust the user tag model U based on the latest behavioral features during the recommendation period. tag It performs real-time updates to adapt to dynamic changes in user interests, and stores snapshots of user profiles at different time periods through a caching mechanism for subsequent comparison and optimization of recommendation models.

8. The intelligent recommendation system based on sensory data according to claim 5, characterized in that, The candidate recommendation generation unit further integrates a multi-objective ranking and control mechanism, which is used to dynamically adjust the candidate resource screening threshold under multiple business objectives such as recommendation accuracy, response latency, and system load, so as to improve the overall recommendation efficiency of the system and support the adaptive balance between personalization and global optimization.

9. The intelligent recommendation system based on sensory data according to claim 8, characterized in that, The enhanced feedback optimization module further supports continuous feedback fusion strategy, which can uniformly model and analyze the feedback information of the same user on multiple time periods and multiple devices, improve the adaptability of the recommendation model to cross-terminal behavioral characteristics in complex scenarios, and maintain the consistency and coherence of recommended content across different interactive terminals.

Images