Personalized recommendation system and method for intelligent terminal

Abstract

The invention provides a personalized recommendation system and method for an intelligent terminal, and belongs to the technical field of artificial intelligence and big data, and the system comprises a five-element multi-mode dynamic perception module, a space-time attention feature fusion unit, a hierarchical federal transfer learning framework, a context perception enhanced recommendation engine and an edge-cloud co-evolution mechanism. The five-element multi-mode dynamic sensing module is used for synchronously collecting physiological features, environmental parameters, behavior data, spatio-temporal information and a social relation graph; and the space-time attention feature fusion unit is used for fusing the multi-modal data by adopting an ST-Transform model. According to the system, on the premise of ensuring the privacy of the user, the recommendation accuracy and real-time performance in a complex scene are remarkably improved, and a new technical normal form is provided for the personalized service of the intelligent terminal; according to the personalized recommendation method, cross-device transfer learning enables the time consumption of new user feature mapping to be greatly shortened, environment-driven brightness adjustment effectively reduces the visual fatigue of the user, and incremental learning obviously reduces the data volume updating demand of the model.

Description

Technical Field

[0001] The present invention relates to the technical fields of artificial intelligence and big data, and specifically relates to a personalized recommendation system and method for intelligent terminals. Background Art

[0002] Generally speaking, an intelligent terminal is a type of embedded computer system device, so its architecture framework is consistent with that of an embedded system; at the same time, as an application direction of the embedded system, the application scenario of the intelligent terminal is relatively clear. Therefore, its architecture is clearer, finer-grained than that of a general embedded system, and has some of its own characteristics.

[0003] With the popularization of intelligent terminal devices, the personalized recommendation system has become one of the core technologies to improve the user experience. Although traditional recommendation systems (such as collaborative filtering, content-based recommendation, etc.) perform well in specific scenarios, they have significant limitations in multi-modal data fusion, real-time capture of user intentions, privacy protection, and edge computing efficiency. Specifically, they are manifested as follows:

[0004] 1. Single data dimension

[0005] Existing systems mainly rely on user behavior logs or explicit feedback (such as ratings, clicks), lacking the deep fusion of multi-dimensional data such as physiological characteristics (heart rate, body temperature), environmental parameters (light, noise), and spatio-temporal information (GPS, timestamp), resulting in insufficient accuracy in user state modeling and a disconnection between recommendation strategies and real needs.

[0006] 2. Weak privacy protection mechanism

[0007] Traditional centralized recommendation architectures need to upload user data to the cloud, posing a risk of data leakage. Although federated learning can protect privacy, existing methods (such as federated averaging) are difficult to handle cross-device feature differences, and the model update cycle is long, unable to meet the requirements of dynamic scenarios.

[0008] 3. Prominent cold start problem

[0009] New users or long-tail items lack historical interaction data. Traditional transfer learning methods rely on a large number of labeled samples, and their generalization ability is limited, resulting in low recommendation coverage and decreased click-through rate.

[0010] 4. High computational resource consumption

[0011] When deploying deep models (such as Transformer, BERT) on terminal devices, there are problems such as insufficient computing power and high energy consumption. Centralized inference in the cloud leads to increased latency and is difficult to meet real-time requirements.

[0012] Therefore, a personalized recommendation system and method for intelligent terminals are proposed. Summary of the Invention

[0013] The present invention aims to solve the problems raised in the background technology and provides a personalized recommendation system and method for intelligent terminals.

[0014] The specific technical solutions are as follows:

[0015] A personalized recommendation system for intelligent terminals, comprising:

[0016] A five - element multi - modal dynamic perception module, used for synchronously collecting physiological characteristics, environmental parameters, behavior data, spatio - temporal information, and social relationship graphs;

[0017] A spatio - temporal attention feature fusion unit, used for fusing multi - modal data by using the ST - Transformer model and detecting feature drift through the dynamic time warping (DTW) algorithm;

[0018] A hierarchical federated transfer learning framework, including a terminal lightweight model, an edge differential privacy aggregation node, and a Multi - TaskBERT model enhanced by a cloud knowledge graph;

[0019] A context - aware reinforcement recommendation engine, generating dynamic recommendation strategies based on a dual - channel reinforcement learning network LSTM + CNN;

[0020] An edge - cloud co - evolution mechanism, used for realizing continuous model optimization through the incremental knowledge distillation OTF - Distill algorithm.

[0021] In the above - mentioned personalized recommendation system for intelligent terminals, the five - element multi - modal dynamic perception module includes:

[0022] A biometric sensor with an accuracy of ±0.5°C and ±2 bpm;

[0023] A multi - modal data synchronizer, synchronously collecting multi - source data at a frequency of 100 Hz and using timestamp alignment technology to ensure data consistency;

[0024] An environmental parameter acquisition unit, covering three - axis sensing of light intensity from 0 to 100 klux, noise decibels from 30 to 120 dB, and air pressure from 300 to 1100 hPa.

[0025] In the above - mentioned personalized recommendation system for intelligent terminals, the spatio - temporal attention feature fusion unit includes:

[0026] An ST - Transformer model: composed of a spatio - temporal position encoding layer, a cross - modal attention matrix, and a feature importance weight calculation module;

[0027] A dynamic interest decay function: where t is the time variable; λ(t) is the dynamic decay factor; T(u) is the personalized period, and cos(πt / T(u)) is the period adjustment term;

[0028] Feature drift detection module: Calculate the similarity of the feature sequence based on the dynamic time warping (DTW) algorithm, and trigger model update when the offset exceeds the threshold of 0.15.

[0029] The above personalized recommendation system for intelligent terminals, wherein the hierarchical federated transfer learning framework includes:

[0030] Terminal layer: Process visual data using quantized MobileViT and text data using TinyBERT, and add Laplace noise ∈ = 1.0 to the output feature vector;

[0031] Edge layer: Used to deploy the secure multi-party computation (SMPC) protocol to aggregate feature vectors and support cross-device feature mapping migration;

[0032] Cloud layer: Used to build a knowledge graph containing more than 3 million entities and achieve rapid adaptation to new scenarios through meta-learning algorithms.

[0033] The above personalized recommendation system for intelligent terminals, wherein the context-aware reinforcement recommendation engine includes a dual-channel reinforcement learning network, a dynamic environment factor matrix, and a policy generator, where:

[0034] The dual-channel reinforcement learning network includes a long-term preference branch and a short-term intention branch. The long-term preference branch uses an LSTM network to model the user's historical behavior sequence; the short-term intention branch uses a CNN network to process the real-time environment parameter matrix;

[0035] The dynamic environment factor matrix is used to sense 8-dimensional parameters in real time. The 8-dimensional parameters include light intensity, noise decibel, social scene, device type, network status, time category, location type, and user activity status;

[0036] The policy generator is used to dynamically adjust the recommended content attributes according to the environment parameters, including brightness adjustment, information density, and interaction method.

[0037] The above personalized recommendation system for intelligent terminals, wherein the edge-cloud co-evolution mechanism includes:

[0038] Incremental knowledge distillation mechanism: Use the OTF-Distill algorithm to compress the cloud GNN model to the terminal, and the knowledge distillation loss function is:

[0039] L KD = KL(p s ||p t ) + αL2(W s ,Wt ) + β·PrivacyLoss(D sens );

[0040] Where KL(p s || p t ) is the KL divergence term; Where p s , p t are the softmax output probabilities; L2(W s , W t ) is the L2 regularization term; α is the regularization coefficient; PrivacyLoss(D sens ) is the privacy loss term; β is the privacy-performance balance coefficient;

[0041] Model warm-up cache mechanism, which predicts the user behavior pattern based on the Markov chain and pre-loads the recommendation model of the regional edge server in advance;

[0042] Federated learning update mechanism, which uses the Paillier homomorphic encryption technology to transmit the model gradient, and the update period is triggered every 2000 user interactions.

[0043] The present invention also provides a personalized recommendation method for a personalized recommendation system based on an intelligent terminal, including the following steps:

[0044] S1. Multi-modal data collection: Synchronously obtain five-element data through biosensors, environmental sensors and terminal interaction interfaces;

[0045] S2. Feature fusion and modeling: Perform Z-score normalization on the physiological data and generate the user state vector ut through ST-Transformer;

[0046] S3. Hierarchical federated inference: The terminal generates a Top-20 candidate set, the response time < 150ms, and the cloud complements the long-tail items through the knowledge graph, and the inference time < 80ms;

[0047] S4. Dynamic policy generation: The dual-channel reinforcement learning network outputs the recommendation action space A, and adjusts the candidate set weights in combination with the interest decay function w(t);

[0048] S5. Online evolution: When the feature drift detection module is triggered, perform incremental knowledge distillation, and the user's real-time feedback updates the local policy network through the TD(λ) algorithm.

[0049] The above-mentioned personalized recommendation method for the personalized recommendation system of the intelligent terminal, wherein the hierarchical federated inference in step S3 includes:

[0050] S31. Cross-device transfer learning: Construct the device feature similarity matrix M ∈ R n×n, new users obtain the initial feature vectors through k = 5 nearest neighbor mapping;

[0051] S32. Differential privacy protection: Add noise to the terminal feature vectors where ∈ = 1.0;

[0052] S33. Knowledge graph completion: Use the graph attention network GAT to mine potential association paths for long-tail items.

[0053] The personalized recommendation method of the personalized recommendation system of the intelligent terminal described above, wherein, the dynamic policy generation in the step S4 includes:

[0054] S41. Environmental parameter coupling: Automatically adjust the brightness of the recommended content according to the light intensity, and the mapping function is Brightness = 0.7×e 0.015L , where L is the light value;

[0055] S42. Social scenario adaptation: When a multi-person conversation is detected, filter sensitive content and increase the public recommendation weight by 30%.

[0056] The personalized recommendation method of the personalized recommendation system of the intelligent terminal described above, wherein, the online evolution in the step S5 includes:

[0057] S51. Model update trigger conditions: KL divergence threshold 0.1 and DTW feature offset > 0.15;

[0058] S52. Federal parameter aggregation: Adopt the weighted average algorithm:

[0059] where Di is the data volume of device i;

[0060] D is the total global data volume; TrustScore(i) is the device trustworthiness score; W i is the device local model parameter, and TrustScore(j) is the multi-dimensional dynamic weighting coefficient with a value range of [0, 1].

[0061] The personalized recommendation system and method of the intelligent terminal provided by the present invention have the following beneficial effects:

[0062] 1. Multimodal dynamic perception: Synchronously collect five-element data such as physiology, environment, and behavior, and combine the ST-Transformer model to achieve spatio-temporal feature fusion, improving the accuracy of user intention recognition.

[0063] 2. Privacy-efficient federal architecture: Adopt a hierarchical federal transfer learning framework, and enhance the model through edge differential privacy aggregation nodes and cloud knowledge graphs to balance privacy protection and recommendation efficiency.

[0064] 3. Cold start optimization: Introduce cross-device transfer learning and knowledge graph completion technologies to reduce dependence on historical data and improve the recommendation coverage rate for new users and long-tail items.

[0065] 4. Edge-cloud co-evolution: Achieve lightweight terminal models through incremental knowledge distillation (OTF-Distill algorithm), combined with the Markov chain warm-up strategy, reduce the cloud computing load, and shorten the end-to-end latency.

[0066] This system significantly improves the recommendation accuracy and real-time performance in complex scenarios while ensuring user privacy, providing a new technical paradigm for personalized services on intelligent terminals;

[0067] For this personalized recommendation method, cross-device transfer learning significantly shortens the time-consuming of new user feature mapping, environment-driven brightness adjustment effectively reduces user visual fatigue, and incremental learning significantly reduces the data volume requirement for model updates. Brief Description of the Drawings

[0068] Figure 1 It is a schematic diagram of the architecture of the personalized recommendation system for intelligent terminals provided by the embodiments of the present invention;

[0069] Figure 2 It is a module relationship diagram of the personalized recommendation system for intelligent terminals provided by the embodiments of the present invention;

[0070] Figure 3 It is a working flow chart of the personalized recommendation system for intelligent terminals provided by the embodiments of the present invention. Detailed Embodiments

[0071] The technical solutions of the present invention will be further described below with reference to the drawings and through specific embodiments.

[0072] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and cannot be construed as a limitation of this patent; for better illustration of the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, which do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0073] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if terms such as "upper", "lower", "left", "right", "inner", "outer", etc. are used to indicate the orientation or positional relationship, it is based on the orientation or positional relationship shown in the accompanying drawings. This is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the accompanying drawings are only for illustrative purposes and cannot be construed as a limitation of this patent. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0074] In the description of the present invention, unless otherwise clearly specified and defined, if terms such as "connection" are used to indicate the connection relationship between components, this term should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection, an electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, and can be the communication between the interiors of two components or the interaction relationship between two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0075] Embodiment 1. The personalized recommendation system of the intelligent terminal provided in this embodiment is as Figures 1-3 shown, and includes: a five - element multi - modal dynamic perception module, a spatio - temporal attention feature fusion unit, a hierarchical federated transfer learning framework, a context - aware reinforcement recommendation engine, and an edge - cloud co - evolution mechanism.

[0076] Among them, the five - element multi - modal dynamic perception module is communicatively connected to the spatio - temporal attention feature fusion unit, the spatio - temporal attention feature fusion unit is communicatively connected to the hierarchical federated transfer learning framework, the hierarchical federated transfer learning framework is communicatively connected to the context - aware reinforcement recommendation engine, the context - aware reinforcement recommendation engine is communicatively connected to the edge - cloud co - evolution mechanism, and the edge - cloud co - evolution mechanism is also communicatively connected to the hierarchical federated transfer learning framework;

[0077] Among them, the five - element multi - modal dynamic perception module is used to synchronously collect physiological characteristics (heart rate / temperature), environmental parameters (light / noise / air pressure), behavior data (gesture / voice), spatio - temporal information (GPS coordinates / timestamp), and social relationship graphs;

[0078] Among them, the spatio - temporal attention feature fusion unit is used to fuse multi - modal data using the ST - Transformer model and detect feature drift through the dynamic time warping (DTW) algorithm;

[0079] Among them, the hierarchical federated transfer learning framework includes a lightweight terminal model, an edge differential privacy aggregation node, and a knowledge graph enhanced Multi-TaskBERT model in the cloud;

[0080] Among them, the context-aware reinforcement recommendation engine generates dynamic recommendation strategies based on a dual-channel reinforcement learning network (LSTM + CNN);

[0081] Among them, the edge-cloud co-evolution mechanism is used to continuously optimize the model through incremental knowledge distillation (OTF-Distill algorithm).

[0082] The personalized recommendation system of this intelligent terminal comprehensively captures multi-dimensional user behavior data, greatly improving the recommendation accuracy. The hierarchical federated architecture balances privacy and efficiency, significantly reducing the risk of data leakage. The edge-cloud collaboration greatly reduces the cloud computing load and effectively extends the terminal battery life.

[0083] Specifically, in this embodiment, the five-element multi-modal dynamic perception module includes a biometric sensor, a multi-modal data synchronizer, and an environmental parameter acquisition unit, where:

[0084] The accuracy of the biometric sensor is ±0.5°C (body temperature) and ±2 bpm (heart rate);

[0085] The multi-modal data synchronizer synchronizes and collects multi-source data at a frequency of 100 Hz, and uses timestamp alignment technology to ensure data consistency;

[0086] The environmental parameter acquisition unit covers three-axis sensing of light intensity (0 - 100 klux), noise decibel (30 - 120 dB), and air pressure (300 - 1100 hPa).

[0087] High-precision physiological data effectively improves the accuracy of health-related recommendations. Millisecond-level data synchronization ensures multi-modal temporal consistency, and significantly reduces the feature fusion error.

[0088] Specifically, in this embodiment, the spatio-temporal attention feature fusion unit includes an ST-Transformer model, a dynamic interest decay function, and a feature drift detection module, where:

[0089] The ST-Transformer model consists of a spatio-temporal position encoding layer, a cross-modal attention matrix, and a feature importance weight calculation module;

[0090] Dynamic interest decay function: Among them, the time variable t represents the time interval (unit: minute) after the occurrence of the user behavior event, which is calculated in real time and starts timing from the user's last interaction; the dynamic decay factor λ(t) represents the control of the interest decay rate, which is adaptively adjusted according to the user's recent interaction frequency. λ(t) = λ0 × (1 + log(1 + N click )), where λ0 = 0.001 / min (basic decay rate), Nclick is the number of clicks within the last hour, and the personalized period T(u) is the personalized period parameter of the user interest change, which reflects the periodic law of the user behavior (such as the preference difference between weekdays / weekends), and is generated by K-means clustering analysis of the user's historical behavior data (time window = 7 days). Typical value: T(u) ∈ [24, 168] hours; the periodic adjustment term cos(πt / T(u) introduces the periodic characteristics of interest fluctuations to solve the problem that the static decay model cannot capture the user's long-term habits. The amplitude of the periodic term is fixed at 1, and the phase is controlled by T(u); through the adaptive adjustment of λ(t) and T(u), the model can distinguish high-active users (slow decay) from low-frequency users (fast decay), and adapt to the personalized period, and the retention rate is greatly improved compared with the traditional static model (such as λ = 0.001);

[0091] The feature drift detection module calculates the similarity of the feature sequences based on the DTW algorithm, and triggers model update when the offset exceeds the threshold of 0.15.

[0092] ST-Transformer spatio-temporal modeling greatly improves the short-term interest capture speed, the periodic decay function effectively improves the user retention rate, the DTW detection triggers model update in real time, and the interest drift adaptation speed is greatly accelerated.

[0093] Specifically, in this embodiment, the hierarchical federated transfer learning framework includes a terminal layer, an edge layer, and a cloud layer, where:

[0094] The terminal layer uses quantized MobileViT (8-bit integer) to process visual data and TinyBERT to process text data, and adds Laplace noise (∈ = 1.0) to the output feature vector;

[0095] The edge layer is used to deploy the secure multi-party computing (SMPC) protocol to aggregate the feature vectors and support cross-device feature mapping migration;

[0096] The cloud layer is used to build a knowledge graph containing more than 3 million entities and achieve rapid adaptation to new scenarios through the meta-learning algorithm.

[0097] The 8-bit quantization model can greatly reduce the terminal computing overhead, the SMPC protocol ensures zero data leakage in the federated learning process, and the knowledge graph completion greatly improves the coverage rate of long-tail item recommendations.

[0098] Specifically, in this embodiment, the context-aware reinforcement recommendation engine includes a dual-channel reinforcement learning network, a dynamic environmental factor matrix, and a policy generator, where:

[0099] The dual-channel reinforcement learning network includes a long-term preference branch and a short-term intention branch. The long-term preference branch uses an LSTM network to model the user's historical behavior sequence; the short-term intention branch uses a CNN network to process the real-time environmental parameter matrix;

[0100] The dynamic environmental factor matrix is used to sense 8-dimensional parameters in real time (light intensity, noise decibel, social scene, device type, network status, time category, location type, user activity status);

[0101] The policy generator is used to dynamically adjust the recommended content attributes according to the environmental parameters, including brightness adjustment (±30%), information density (concise / detailed mode), and interaction method (voice / touch priority).

[0102] The LSTM-CNN hybrid network effectively improves the recommendation accuracy in complex scenarios. The coupling of environmental parameters significantly increases the click-through rate of night mode recommendations, and the response time of dynamic policy generation is effectively shortened.

[0103] Specifically, in this embodiment, the edge-cloud co-evolution mechanism includes an incremental knowledge distillation mechanism, a model warm-up cache mechanism, and a federated learning update mechanism, where:

[0104] The incremental knowledge distillation mechanism uses the OTF-Distill algorithm to compress the cloud GNN model to the terminal, and the knowledge distillation loss function is

[0105] L KD =KL(p s ||p t )+αL2(W s ,W t )+β·PrivacyLoss(D sens ), where the KL divergence term KL(p s ||p t ) is used to measure the difference in the output probability distributions between the large cloud model (teacher) and the small terminal model (student), ensuring the effectiveness of knowledge transfer, where p s , p t are the softmax output probabilities; the L2 regularization term L2(W s ,W t ) is used to constrain the difference between the parameters Wt of the student model and the parameters Ws of the teacher model to prevent overfitting, The regularization coefficient α represents the importance of balancing the KL divergence and the L2 regularization term to prevent excessive distortion in model compression. The empirical value of α is 0.7, which can be optimized through grid search; the privacy loss term PrivacyLoss(D sens ) represents the risk of privacy leakage of sensitive data (such as heart rate, location) during the distillation process, based on the differential privacy mechanism: PrivacyLoss = ∈·||Dsens||1, where ∈ = 0.1 (privacy budget); the privacy-performance balance coefficient β is used to control the trade-off between privacy protection intensity and model performance. The larger the value, the stronger the privacy protection, but the model accuracy may decrease. It is dynamically adjusted as: β = 0.3×SensitivityScore(D sens ). The higher the sensitive data score, the larger β. Through the PrivacyLoss term, the risk of sensitive data leakage is significantly reduced, while the model compression rate remains at a high level;

[0106] The model warm-up cache mechanism predicts the user behavior pattern based on the Markov chain and pre-loads the recommendation model of the regional edge server in advance;

[0107] The federated learning update mechanism uses the Paillier homomorphic encryption technology to transmit the model gradient, and the update period is triggered every 2000 user interactions.

[0108] OTF-Distill significantly compresses the model volume, greatly reduces the memory occupancy, the warm-up strategy significantly reduces the cold start latency, the homomorphic encryption protects the gradient transmission, and the success rate of the model inversion attack is significantly reduced.

[0109] Embodiment 2 provides a personalized recommendation method for a personalized recommendation system of an intelligent terminal based on Embodiment 1, including the following steps:

[0110] S1. Multi-modal data collection: Synchronously obtain five-element data through biosensors, environmental sensors, and terminal interaction interfaces;

[0111] S2. Feature fusion and modeling: Perform Z-score standardization on the physiological data and generate the user state vector ut through ST-Transformer;

[0112] S3. Hierarchical federated inference: The terminal generates a Top-20 candidate set (response time < 150ms), and the cloud complements the long-tail items through the knowledge graph (inference time < 80ms);

[0113] S4. Dynamic policy generation: The dual-channel reinforcement learning network outputs the recommendation action space A, and adjusts the candidate set weights in combination with the interest decay function w(t);

[0114] S5. Online Evolution: When the feature drift detection module is triggered, incremental knowledge distillation is performed, and the user's real-time feedback is used to update the local policy network through the TD(λ) algorithm.

[0115] Among them, the hierarchical federated inference in step S3 includes:

[0116] S31. Cross-device Transfer Learning: Construct a device feature similarity matrix M ∈ R n×n , and the new user obtains the initial feature vector through k = 5 nearest neighbor mapping;

[0117] S32. Differential Privacy Protection: Add noise to the terminal feature vector where ∈ = 1.0;

[0118] S33. Knowledge Graph Completion: Use the Graph Attention Network (GAT) to mine potential association paths for long-tail items.

[0119] Among them, the dynamic policy generation in step S4 includes:

[0120] S41. Environmental Parameter Coupling: Automatically adjust the brightness of the recommended content according to the light intensity, and the mapping function is Brightness = 0.7×e 0.015L , where L is the light value;

[0121] S42. Social Scenario Adaptation: When a multi-person conversation is detected, filter sensitive content and increase the public recommendation weight by 30%.

[0122] Among them, the online evolution in step S5 includes:

[0123] S51. Model Update Trigger Conditions: KL divergence threshold 0.1 and DTW feature offset > 0.15;

[0124] S52. Federated Parameter Aggregation: Adopt the weighted average algorithm:

[0125] where Di is the data volume of device i, the size of the local dataset of device i, reflecting the data contribution amount, and directly count the number of samples of device i (such as Di = 1000 pieces); D is the total global data volume, the total data volume of all devices participating in federated learning, used to normalize the weights, TrustScore(i) is the device trust score, which evaluates the data quality and reliability of device i to prevent malicious devices or low-quality data from interfering with the global model. TrustScore(i) = ω1·DTW(Hi, Havg)+ω2Freq(i), where: DTW(H i ,H avg): DTW similarity between device historical data and the global average sequence (value range [0,1]); Freq(i): Frequency of the device participating in federated updates (normalized to [0,1]); ω1 = 0.6, ω2 = 0.4 (weights); W i is the device local model parameter, the model parameter after training of device i, which is used to aggregate and generate the global model, and is updated through local training (such as SGD optimizer). High - contribution devices are screened by TrustScore, and the model convergence speed is greatly accelerated in the cold - start scenario, and the interference of malicious devices is significantly reduced. TrustScore(j) is a multi - dimensional dynamic weighting coefficient with a value range of [0,1], and the closer the value is to 1, the higher the trust level of device j.

[0126] For this personalized recommendation method, cross - device transfer learning significantly shortens the time consumed for new user feature mapping, environment - driven brightness adjustment effectively reduces the user's visual fatigue, and incremental learning significantly reduces the demand for model update data volume.

[0127] In summary, the personalized recommendation system and method for intelligent terminals provided in this embodiment have the following advantages:

[0128] This personalized recommendation system for intelligent terminals combines ST - Transformer with DTW drift detection to solve the problem of multi - modal time - series alignment. It integrates knowledge graphs and meta - learning in federated learning, breaks through the data dependence of traditional transfer learning, significantly improves the click - through rate in the cold - start scenario, and significantly reduces the end - to - end delay. By fusing five - element data through ST - Transformer, the accuracy of user intention recognition is greatly improved. Cross - device transfer learning significantly improves the click - through rate of new user recommendations. Edge differential aggregation effectively reduces the risk of privacy leakage. Incremental knowledge distillation (OTF - Distill) realizes continuous optimization of the terminal model, and the model iteration cycle is significantly shortened to real - time update. The dynamic environment factor drives policy adjustment, and the click - through rate of recommended content in a noisy environment is greatly improved.

[0129] For this personalized recommendation method, cross - device transfer learning significantly shortens the time consumed for new user feature mapping, environment - driven brightness adjustment effectively reduces the user's visual fatigue, and incremental learning significantly reduces the demand for model update data volume.

[0130] The above are only the preferred embodiments of the present invention, and do not limit the implementation manners and protection scope of the present invention. For those skilled in the art, it should be able to realize that all equivalent replacements and obvious changes made by using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A personalized recommendation system for intelligent terminals, characterized in that: include: A five-element multimodal dynamic perception module for synchronously collecting physiological characteristics, environmental parameters, behavioral data, spatiotemporal information, and social relationship maps; The spatiotemporal attention feature fusion unit is used to fuse multimodal data using the ST-Transformer model and detect feature drift through the dynamic time warping DTW algorithm; A hierarchical federated transfer learning framework, including a lightweight terminal model, edge differential privacy aggregation nodes, and a Multi-TaskBERT model enhanced by a cloud knowledge graph; Context-aware reinforcement recommendation engine, which generates dynamic recommendation strategies based on the dual-channel reinforcement learning network LSTM+CNN; Edge-cloud co-evolution mechanism is used to achieve continuous model optimization through incremental knowledge distillation OTF-Distill algorithm.

2. The personalized recommendation system for smart terminals according to claim 1, characterized in that: The five-element multimodal dynamic perception module includes: Biometric sensor with an accuracy of ±0.5°C and ±2bpm; Multimodal data synchronizer, synchronously collects multi-source data at 100Hz frequency, and uses timestamp alignment technology to ensure data consistency; The environmental parameter collection unit covers three-axis sensing of light intensity 0-100klux, noise decibel 30-120dB and air pressure 300-1100hPa.

3. The personalized recommendation system for smart terminals according to claim 1, characterized in that: The spatiotemporal attention feature fusion unit comprises: ST-Transformer model: It consists of a spatiotemporal position encoding layer, a cross-modal attention matrix, and a feature importance weight calculation module; Dynamic interest decay function: Where t is the time variable; λ(t) is the dynamic attenuation factor; T(u) is the personalized period, and cos(πt / T(u)) is the period adjustment term; Feature drift detection module: Calculates feature sequence similarity based on the dynamic time warping (DTW) algorithm and triggers model update when the offset exceeds the threshold of 0.

15.

4. The personalized recommendation system for smart terminals according to claim 1, characterized in that: The hierarchical federated transfer learning framework includes: At the terminal layer, quantized MobileViT is used to process visual data, TinyBERT is used to process text data, and Laplace noise ∈ = 1.0 is added to the output feature vector; The edge layer is used to deploy the secure multi-party computing (SMPC) protocol to aggregate feature vectors and support cross-device feature mapping migration; The cloud layer is used to build a knowledge graph containing more than 3 million entities and achieve rapid adaptation to new scenarios through meta-learning algorithms.

5. The personalized recommendation system for smart terminals according to claim 1, characterized in that: The context-aware reinforcement recommendation engine includes a dual-channel reinforcement learning network, a dynamic environmental factor matrix, and a strategy generator, wherein: The dual-channel reinforcement learning network includes a long-term preference branch and a short-term intention branch. The long-term preference branch uses an LSTM network to model the user's historical behavior sequence; the short-term intention branch uses a CNN network to process the real-time environment parameter matrix; The dynamic environmental factor matrix is ​​used to perceive 8-dimensional parameters in real time, including light intensity, noise decibels, social scene, device type, network status, time category, location type, and user activity status; The strategy generator is used to dynamically adjust the recommended content attributes according to environmental parameters, including brightness adjustment, information density and interaction mode.

6. The personalized recommendation system for smart terminals according to claim 1, characterized in that: The edge-cloud co-evolution mechanism includes: The incremental knowledge distillation mechanism uses the OTF-Distill algorithm to compress the cloud GNN model to the terminal. The knowledge distillation loss function is: L KD =KL(p s ||p t )+αL2(W s ,W t )+β·PrivacyLoss(D sens ); Where KL(p s ||p t ) is the KL divergence term; where p s , p t is the softmax output probability; L2(W s ,W t ) is the L2 regularization term; α is the regularization coefficient; PrivacyLoss(D sens ) is the privacy loss term; β is the privacy-performance balance coefficient; Model preheating cache mechanism, based on Markov chain prediction of user behavior patterns, pre-loads the recommendation model of the regional edge server; The federated learning update mechanism uses Paillier homomorphic encryption technology to transmit model gradients, and the update cycle is triggered every 2,000 user interactions.

7. A personalized recommendation method based on the personalized recommendation system of the intelligent terminal according to any one of claims 1 to 6, characterized in that: The following steps are involved: S1, multimodal data acquisition: synchronously obtain five-dimensional data through biosensors, environmental sensors and terminal interaction interfaces; S2, feature fusion and modeling: perform Z-score normalization on physiological data and generate user state vector ut through ST-Transformer; S3, hierarchical federated reasoning: The terminal generates the top-20 candidate set with a response time of <150ms, and the cloud completes the long-tail items through the knowledge graph with a reasoning time of <80ms; S4, dynamic strategy generation: the dual-channel reinforcement learning network outputs the recommended action space A, and adjusts the candidate set weights in combination with the interest decay function w(t); S5, online evolution: When the feature drift detection module is triggered, incremental knowledge distillation is performed, and real-time user feedback is used to update the local policy network through the TD(λ) algorithm.

8. The personalized recommendation method of the personalized recommendation system of the intelligent terminal according to claim 7, characterized in that: The hierarchical federated reasoning in step S3 includes: S31. Cross-device transfer learning: Constructing the device feature similarity matrix M∈R n×n , the new user obtains the initial feature vector through k=5 nearest neighbor mapping; S32. Differential privacy protection: adding noise η to the terminal feature vector Laplace(0,1 / ∈), where ∈=1.0; S33. Knowledge graph completion: Use graph attention network GAT to mine potential association paths for long-tail items.

9. The personalized recommendation method of the personalized recommendation system of the intelligent terminal according to claim 7, characterized in that: The dynamic strategy generation in step S4 includes: S41, Environmental parameter coupling: Automatically adjust the brightness of recommended content according to the light intensity, the mapping function is Brightness = 0.7 × e 0.015L , where L is the light value; S42, social scenario adaptation: When a multi-person conversation is detected, filter sensitive content and increase the public recommendation weight by 30%.

10. The personalized recommendation method of the personalized recommendation system of the intelligent terminal according to claim 7, characterized in that: The online evolution in step S5 includes: S51, model update triggering conditions: KL divergence threshold 0.1 and DTW feature offset > 0.15; S52, federated parameter aggregation: using weighted average algorithm: Where Di is the amount of data of device i; D is the total amount of global data; TrustScore(i) is the trust score of the device; Wi is the local model parameter of the device, and TrustScore(j) is a multi-dimensional dynamic weighting coefficient with a value range of [0,1].

Images