An end-side system recommendation method based on semantic attenuation and memory fusion

Abstract

The application discloses an end-side system recommendation method based on semantic attenuation and memory fusion, adopts a double-mode architecture: in the real-time recommendation mode, a short-term semantic scene memory module is constructed to perform semantic density screening and coding on user behavior to generate a short-term interest vector, and a long-term interest vector extracted from a long-term privacy interest memory bank is combined to generate a basic interest mask; the desensitization-processed long-term memory generalization representation is sent to the cloud for rough recall, a system-level recommendation mask is generated after privacy reordering at the end side, and the system IPC interface is broadcasted for third-party application subscription and calling. In the background resident maintenance mode, the end-side semantic attenuation engine is started, a lightweight large model is used for semantic logic reasoning on the short-term memory in a period, a semantic probability vector is output, and the semantic probability vector is mapped into an attenuation factor through a deterministic rule, and the weight of the long-term memory node is updated and new semantics are fused.

Description

Technical Field

[0001] This invention belongs to the fields of computer technology, artificial intelligence and recommendation system technology, and in particular relates to a recommendation method for edge systems based on semantic decay and memory fusion. Background Technology

[0002] With the development of artificial intelligence technology, large-scale edge models are gradually being deployed and applied in terminal devices such as mobile operating systems, in-vehicle systems, smart wearable devices, and lightweight AI assistants. Edge devices continuously generate a large amount of user behavior data during daily use, which can be applied to more personalized recommendation services, such as news feed recommendations, smart assistant prompts, contextualized service pushes, and commercial services. Compared to traditional application-level recommendations, edge system-level recommendations directly impact the core entry point of the operating system, placing higher demands on the real-time performance, stability, security, and user experience of the recommendation results.

[0003] In today's trend of application ecosystem development, recommendation systems have become the core carrier for efficient information distribution within the ecosystem. However, existing cloud-based recommendation algorithms have the following prominent problems when deployed in edge-side system-level scenarios:

[0004] 1. The problem of difficulty in distinguishing between highly reproducible and low-value behaviors in edge-side interest modeling.

[0005] User behavior on the device side is characterized by high frequency, high repetition, and low trigger cost. A large amount of data only reflects immediate inertia and has limited practical value. Existing algorithms are mostly based on frequency or short-window statistical modeling, lacking a mechanism to distinguish between the effectiveness of behavior and sustainable reference value. They are prone to overlearning low-value, highly reproducible redundant behaviors, thereby interfering with the accurate characterization of the real interest structure.

[0006] 2. The problem of lacking fine-grained decay and stability control over the evolution of interest representation over time.

[0007] Edge-side interests exhibit significant phased and contextual characteristics, displaying both short-term fluctuations and long-term stable demands. Existing algorithms mostly employ uniform time decay or sliding window strategies, making it difficult to differentiate management based on the formation path and activity patterns of interests. This leads to problems such as rapid forgetting or ineffective retention during interest evolution, severely impacting the accuracy and controllability of recommendation decisions.

[0008] 3. The problem that edge-side interest results are difficult to form a unified expression that can be called at the system level.

[0009] Existing algorithms primarily output content-based ranking, with their interest modeling results existing in a closed internal state, lacking standardized external expression. This closed nature makes it difficult for operating systems or third-party applications to securely and flexibly invoke user intent, hindering the formation of a unified system-level interest interface and limiting the expansion and application of edge recommendation capabilities in cross-application collaboration and differentiated services. Summary of the Invention

[0010] In view of this, the present invention aims to overcome the shortcomings of the above-mentioned problems in the prior art and proposes an edge system recommendation method based on semantic decay and memory fusion. This method overcomes the problems of lack of edge context awareness, rigid forgetting mechanism and difficulty in utilizing privacy data in the cloud in the existing recommendation system. This improves the accuracy of user interest perception, the context adaptability of cross-application recommendation, and thus enhances the personalized experience and system-level privacy protection capabilities throughout the user's life cycle.

[0011] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0012] A recommendation method for edge systems based on semantic decay and memory fusion includes:

[0013] Real-time recommendation modalities include:

[0014] A short-term semantic context memory module is built on the device side. A sliding window mechanism based on semantic density is used to perform capacity control and noise reduction on the real-time collected user interaction behavior sequences to generate short-term interest vectors.

[0015] Long-term interest vectors are generated by reading active interest nodes from the long-term privacy interest memory on the edge and fusing the short-term interest vectors with the long-term interest vectors to generate a basic interest mask.

[0016] The desensitized long-term memory generalized representation is sent to the cloud for coarse recall, and the candidate content returned by the cloud is reordered for privacy using the basic interest mask on the device side to generate a system-level recommendation mask.

[0017] The system-level recommendation mask is broadcast as a standard intent signal of the operating system for third-party applications to subscribe to in order to achieve cross-application scheduling.

[0018] Background maintenance modes include:

[0019] When the device is in a background running state with sufficient resources, start the edge-side semantic decay engine;

[0020] Multiple short-term interest vectors accumulated within a period are aggregated, and the logical relationship between the aggregated vectors and long-term memory nodes is analyzed using a large edge model to output a semantic probability vector.

[0021] The semantic probability vector is mapped to a decay factor using deterministic rules, and the weights of long-term memory nodes are updated according to the decay factor. New semantics are then fused, and short-term caches are cleared.

[0022] Furthermore, the short-term semantic context memory module controls its capacity through a sliding window mechanism based on semantic density, specifically including:

[0023] When a new behavior causes the queue to overflow, the semantic information entropy of each behavior in the queue is calculated. Behavior nodes with low information entropy are removed first, while heterogeneous behaviors with high information content are retained.

[0024] The retained behavioral sequences are vectorized and aggregated with real-time contextual information to generate short-term interest vectors.

[0025] Furthermore, the interest nodes in the long-term privacy interest memory are represented by a triple structure, wherein the triple includes: an interest semantic vector, an interest weight, and a semantic state identifier; wherein the semantic state identifier includes an active state, a frozen state, and an archived state.

[0026] Furthermore, the step of sending the desensitized long-term memory generalization representation to the cloud for coarse recall specifically includes:

[0027] Long-term interest vectors are clustered and mapped to preset public interest cluster centers in the cloud, and noise that satisfies differential privacy constraints is added to generate a generalized representation.

[0028] The generalized representation is sent to the cloud, and the cloud retrieves the content library based on the generalized representation and returns a set of candidate content.

[0029] Furthermore, the generation of the system-level recommendation mask specifically includes:

[0030] On the edge, the cosine similarity between the candidate content vector and the basic interest mask is calculated, and the Top-K content with the highest scores after rearrangement is selected.

[0031] The selected content is weighted and aggregated to generate a system-level recommendation mask that includes ranking weights.

[0032] Furthermore, the broadcasting of the system-level recommended mask as a standard intent signal of the operating system specifically includes:

[0033] The operating system encapsulates and exposes standardized interfaces at the framework layer for authorized third-party applications to call;

[0034] Third-party applications obtain the current system-level recommendation mask through the interface and calculate the similarity between their internal candidate content vectors and the mask to achieve alignment with system-level intent.

[0035] Furthermore, the edge semantic decay engine is built on a lightweight large language model and is used to perform logical reasoning on the semantic state of user interests and output a semantic probability vector. The semantic probability vector includes the probability of task completion, the probability of context suspension, the probability of interest reinforcement, and the probability of natural decay.

[0036] Furthermore, the mapping of semantic probability vectors to attenuation factors through deterministic rules specifically includes:

[0037] A confidence threshold is preset for each semantic state; when the probability of a certain state in the semantic probability vector exceeds the corresponding threshold, it is mapped to the corresponding decay factor according to the preset deterministic rule.

[0038] Furthermore, the decay factor is used to update the weights of long-term memory nodes, including blocking, freezing protection, dynamic enhancement, or natural decay.

[0039] Furthermore, the updating of the weights of long-term memory nodes and the integration of new semantics specifically includes:

[0040] The weights of the non-frozen nodes are updated according to the decay factor.

[0041] The periodic aggregated vector is fused with the original semantic vector of interest according to a preset update rate to generate a new semantic vector of interest.

[0042] Clear the short-term behavior sequence and short-term vector cache, and enter the real-time monitoring state of the next cycle.

[0043] Compared with existing technologies, the edge system recommendation method based on semantic decay and memory fusion described in this invention has the following advantages:

[0044] This invention constructs a long-term memory and lifecycle management model based on triples, and introduces a "frozen state" and continuous weight mechanism, which effectively solves the defect of traditional algorithms that cannot distinguish between "temporary suspension" and "complete forgetting", and realizes intelligent dynamic management of periodic or phased interests, such as exam preparation and holidays.

[0045] This invention proposes an LSDA semantic decay engine based on probability mapping, which transforms the fuzzy reasoning of large models into a deterministic control process of "probability evaluation-rule mapping", avoiding the risk of uninterpretability brought by black box models and improving the robustness of the algorithm and the suitability of the patented technology.

[0046] This invention establishes a system-level recommendation mask generation and broadcasting mechanism, which is no longer limited to recommendations within a single app, but transforms user intent into standardized interface signals at the operating system level, realizing cross-application context awareness and strategy coordination, and improving the intelligence level of the operating system.

[0047] This invention adopts a dual-modal architecture of "real-time lightweight computing + background resident deep updates" and combines it with short-term memory sliding window capacity control. While ensuring the real-time performance and accuracy of recommendations, it greatly reduces the power consumption and memory usage of edge devices, achieving a balance between privacy protection and computing efficiency. Attached Figure Description

[0048] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0049] Figure 1 This invention provides a real-time recommendation modal flowchart for an edge system recommendation method based on semantic decay and memory fusion. Detailed Implementation

[0055] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0056] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0057] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0058] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0059] The present invention provides a recommendation method for edge systems based on semantic decay and memory fusion, specifically including the following steps:

[0060] 1. Short-Term Semantic Encoding and Capacity Control: A Short-Term Semantic Context Memory (S-SCM) module is constructed on the device side to collect user interaction behavior sequences within the current time window in real time. Capacity control is performed based on a sliding window mechanism of semantic density. When the queue overflows, the semantic information entropy of the behavior sequence is calculated, and behavior nodes with low information entropy, i.e. semantic repetition and redundancy, are removed first, while heterogeneous behaviors with high information content are retained. The retained behavior sequence is then combined with real-time context information and mapped into a short-term interest vector.

[0061] 2. Edge-Cloud Collaboration and System-Level Mask Generation: Active interest nodes are read from the edge's Long-term Privacy Interest Memory (LPIM) database to generate long-term interest vectors. These vectors are then weighted and fused with the short-term interest vectors to generate a basic interest mask. The de-identified generalized representation of the long-term memory is sent to the cloud for coarse recall. On the edge, the candidate content vectors returned from the cloud are re-ranked for privacy using the basic interest mask. High-matching content is filtered by calculating cosine similarity, generating a system-level recommendation mask that includes the final ranking weights.

[0062] 3. System-level intent broadcasting and cross-application scheduling: The generated system-level recommendation mask is used as the standard intent signal of the operating system and broadcast through the inter-process communication (IPC) interface. The operating system provides a standardized interface for third-party applications to subscribe. After receiving the mask, the third-party application does not need to perform complex end-side modeling. It directly calculates the similarity between its internal candidate content vector and the system-level recommendation mask, thereby achieving lightweight alignment and strategy adjustment with the user's current intent.

[0063] 4. Background Resident Semantic State Determination and Update: When the device is in a background resident state with sufficient resources (such as charging and screen off), the edge-side semantic decay engine (Large-model Semantic Decay Analysis Engine, abbreviated as LSDA) is started; the short-term memory vectors accumulated within the period are aggregated, and the logical relationship between the aggregated vectors and long-term memory nodes is analyzed using the edge-side large model, and semantic probability vectors (complete / suspended / enhanced / decayed) are output; the probability is converted into decay factors through deterministic rule mapping, and the weights of long-term memory nodes are updated accordingly, and new semantics are integrated, and finally the short-term cache is cleared.

[0064] The present invention will now be described in detail.

[0065] This invention discloses a recommendation method for edge systems based on semantic decay and memory fusion. The overall logical architecture is divided into a "real-time recommendation mode" and a "backend resident maintenance mode." The real-time recommendation mode process is roughly as follows: Figure 1 As shown.

[0066] 1. Construct the S-SCM module (real-time recommendation modality)

[0067] To address the common issues of "high frequency but low efficiency" and "noise interference" in edge-side user behavior data, a short-term memory structure with adaptive filtering capabilities is first required. In this embodiment, the system constructs a short-term semantic contextual memory module (S-SCM), which uses a semantic density-based sliding window mechanism to clean and encode real-time collected behavior sequences, generating a short-term interest feature vector V_short that incorporates real-time contextual information and has undergone noise reduction.

[0068] Specifically, let the short-term behavioral sequence generated by the user within the current real-time window be: When a new action causes the queue to overflow, the system does not simply follow the first-in, first-out (FIFO) principle. Instead, it calculates the semantic information entropy of each action in the queue and prioritizes removing action nodes with low information entropy, i.e. semantic repetition, redundancy, or low weight, to ensure that the maximum amount of contextual information is retained within a limited window.

[0069] After completing the filtering, the system retains the relevant behaviors. Vectorization and aggregation are performed. First, the behavior is vectorized and aggregated using a lightweight encoder on the edge. Mapped to initial semantic vector ,in This refers to a lightweight embedding layer on the edge, used to transform discrete behavior IDs or text descriptions into high-dimensional dense vectors. Subsequently, it incorporates current contextual information. The vector is modulated by factors such as application context, sentiment tags, and time status, and finally weighted and aggregated to obtain a short-term interest vector. :

[0070]

[0071] in, For context modulation function, For the normalized time weights, strong time decay is not introduced here to fully preserve the contextual features within the window.

[0072] 2. Edge-cloud collaboration and system-level mask generation (real-time recommendation modality)

[0073] This step aims to generate a system-level mask that incorporates both long-term user preferences and intent correction via cloud-based algorithms, utilizing short-term interest vectors and existing long-term memory profiles. To address the cold start problem, an initial base long-term memory profile can be configured for each system during system initialization. To achieve this, the system first needs to maintain an edge-side Long-Term Privacy Interest Memory (LPIM). This memory contains several interest nodes. Its structure is defined as follows:

[0074]

[0075] in: For interest semantic vectors; Weighted by interest; It serves as a semantic state identifier, and its possible values ​​include: Where Active indicates that the interest is currently active and can directly participate in the recommendation calculation; Frozen indicates that the interest is temporarily suspended, retains its weight but does not participate in the current recommendation; Archived indicates that the interest's goal has been completed or it has not been activated for a long time and does not participate in the recommendation.

[0076] When generating the mask, the system reads from the LPIM. For nodes in the Active state, generate long-term interest vectors. Subsequently, based on the long-term and short-term equilibrium coefficients Its value ranges from [0,1], and it is used to adjust whether the system is more biased towards long-term habits or short-term intentions:

[0077]

[0078] After obtaining the basic mask, perform edge-cloud collaborative operations: send the de-identified long-term memory to the cloud for coarse recall and return the candidate set. Regarding the anonymization process, this embodiment employs differential privacy or clustering generalization techniques. The system reads long-term memory, adds noise to it, and performs generalization processing to generate a generalized representation of the long-term memory. .

[0079] Specifically, this generalized representation no longer points to specific user feature points, but rather to the "center of the interest cluster" to which that feature belongs.

[0080]

[0081] Where Cluster(⋅): represents the clustering mapping operation, which groups user-specific clusters... Mapped to the nearest neighbor "Cluster Centroid" in the cloud, thus hiding individual characteristics; : indicates that the condition is satisfied - Laplace noise or Gaussian noise with differential privacy constraints.

[0082] This generalized representation is sent to the cloud. The cloud can only know "what kind of user group the user belongs to," such as a basketball enthusiast, but cannot know the "specific user profile" or which team the user likes. The cloud leverages its computing power advantage, based on... Retrieve from a massive content library and return a large candidate set. .

[0083] On the endpoint, each candidate content is calculated. Content vector and The cosine similarity is used, and the top-K items with the highest scores after rearrangement are selected to generate the final system-level recommendation mask:

[0084]

[0085] Among them Normalized sorting weights. This mask. This represents "the user's strongest current intent after extensive cloud-based retrieval and edge-side precision correction".

[0086] 3. System-level intent broadcasting and cross-application scheduling (real-time recommendation modality)

[0087] To address the issues of "data silos" and "semantic fragmentation" between apps in traditional recommendation systems, this invention generates... This signal is elevated to an operating system-level standard. In this embodiment, the operating system kernel layer is configured to support a standardized intent distribution protocol. The generated... Broadcasting is performed through the operating system's internal inter-process communication (IPC) mechanism:

[0088] (1) Interface definition: The operating system encapsulates and exposes standard API interfaces (e.g., named getSystemInterestMask()) at the framework layer.

[0089] (2) Calling logic: This interface does not require the operating system to refactor its core code, but is an extension based on existing IPC mechanisms (such as Binder or Intent). Third-party applications (with authorization) can obtain the current system-level mask vector by calling this interface.

[0090] (3) Lightweight alignment: After receiving the mask, the third-party application does not need to perform complex end-side modeling, but only needs to calculate its internal candidate content vector. and The cosine similarity can be used to adjust the internal recommendation list based on the similarity level, thereby achieving alignment with the system-level intent.

[0091] 4. Background Resident Semantic State Determination and Update (Backend Maintenance Modality)

[0092] To address the logical flaw in existing technologies that rely solely on time decay, resulting in tasks being completed but interest not being cleared, this step is executed when the device is in a background resident state with sufficient resources and meets the specified cycle time, such as when the device is charging and the screen is off in the early morning.

[0093] First, the system calculates the short-term vectors of multiple time slices accumulated within the previous complete cycle, such as 24 hours or other defined cycles. Processing is performed to calculate the periodic aggregation vector:

[0094]

[0095] in, This represents the total number of time slices in the previous period.

[0096] Then, the LSDA-Engine is started, as detailed below:

[0097] (1)LSDA-Engine is an inference engine that runs on the edge NPU neural network processor. Its core carrier is a specially optimized lightweight large language model.

[0098] (2) This model is not a general text generation model, but a lightweight version based on a large language model, using knowledge distillation and quantization techniques, such as compressing the number of parameters to 3B or 7B and using int4 precision quantization.

[0099] (3) The model was pre-trained using large-scale “user behavior logic chain” data and was fine-tuned specifically for the “interest state determination” task, enabling it to judge complex logics such as “purchase ends” or “viewing enhances”.

[0100] The core function of this engine is no longer text generation, but semantic and logical reasoning. To ensure the interpretability of the decisions, a two-tier architecture of "probability evaluation - rule mapping" is adopted. The large-scale model on the edge receives... and The concatenated input is output as a standardized semantic probability vector. .

[0101]

[0102] in, : Indicates the probability of task completion; : Indicates the probability of the situation being suspended; : Represents the probability of interest reinforcement; : Represents the probability of natural decay.

[0103] Next, state mapping is introduced. The function transforms a probability vector into discrete decay / update factors. This process possesses mathematical determinism:

[0104]

[0105] Where τ is the preset confidence threshold for each state; : This is the interest enhancement coefficient, used to increase the weight of positive interests; : is the natural time decay constant; Δt: is the time interval since the last activation.

[0106] According to the calculation The system performs long-term memory self-forgetting and updating on non-frozen nodes, including the following two dimensions of operations:

[0107] (1) Weight update:

[0108]

[0109] (2) Semantic fusion (overlay of new interests):

[0110]

[0111] in The memory update rate is used to control the speed at which new knowledge is integrated.

[0112] Finally, after updating all long-term memory nodes, perform a short-term memory cleanup operation:

[0113] (1) Clear the original behavior sequence in the S-SCM module ;

[0114] (2) Clear the short-term vector cache within the cycle ;

[0115] (3) The system is reset and enters the real-time monitoring state of the next cycle.

[0116] By leveraging the semantic understanding capabilities of the large-scale model on the device and the deterministic state machine logic, dynamic management of the entire lifecycle of user interests is achieved. This significantly reduces the information cocoon and invalid recommendation problems caused by the lack of cloud-based perception, effectively improves the context sensitivity and user privacy security of the recommendation system, reduces the resource consumption of real-time computing on the device, and can be adapted to the operating systems of various mobile terminals and IoT devices, better adapting to complex application environments with high privacy requirements and multiple scenario switching.

[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A recommendation method for edge systems based on semantic decay and memory fusion, characterized in that: include: Real-time recommendation modalities include: A short-term semantic context memory module is built on the device side. A sliding window mechanism based on semantic density is used to perform capacity control and noise reduction on the real-time collected user interaction behavior sequences to generate short-term interest vectors. Long-term interest vectors are generated by reading active interest nodes from the long-term privacy interest memory on the edge and fusing the short-term interest vectors with the long-term interest vectors to generate a basic interest mask. The desensitized long-term memory generalized representation is sent to the cloud for coarse recall, and the candidate content returned by the cloud is reordered for privacy using the basic interest mask on the device side to generate a system-level recommendation mask. The system-level recommendation mask is broadcast as a standard intent signal of the operating system for third-party applications to subscribe to in order to achieve cross-application scheduling. Background maintenance modes include: When the device is in a background running state with sufficient resources, start the edge-side semantic decay engine; Multiple short-term interest vectors accumulated within a period are aggregated, and the logical relationship between the aggregated vectors and long-term memory nodes is analyzed using a large edge model to output a semantic probability vector. The semantic probability vector is mapped to a decay factor using deterministic rules, and the weights of long-term memory nodes are updated according to the decay factor. New semantics are then fused, and short-term caches are cleared.

2. The edge system recommendation method based on semantic decay and memory fusion according to claim 1, characterized in that: The short-term semantic context memory module controls its capacity through a sliding window mechanism based on semantic density, specifically including: When a new behavior causes the queue to overflow, the semantic information entropy of each behavior in the queue is calculated. Behavior nodes with low information entropy are removed first, while heterogeneous behaviors with high information content are retained. The retained behavioral sequences are vectorized and aggregated with real-time contextual information to generate short-term interest vectors.

3. The edge system recommendation method based on semantic decay and memory fusion according to claim 1, characterized in that: Interest nodes in the long-term privacy interest memory are represented by a triplet structure, wherein the triplet includes: an interest semantic vector, an interest weight, and a semantic state identifier; wherein the semantic state identifier includes an active state, a frozen state, and an archived state.

4. The edge system recommendation method based on semantic decay and memory fusion according to claim 1, characterized in that: The step of sending the desensitized long-term memory generalization representation to the cloud for coarse recall specifically includes: Long-term interest vectors are clustered and mapped to preset public interest cluster centers in the cloud, and noise that satisfies differential privacy constraints is added to generate a generalized representation. The generalized representation is sent to the cloud, and the cloud retrieves the content library based on the generalized representation and returns a set of candidate content.

5. The edge system recommendation method based on semantic decay and memory fusion according to claim 1, characterized in that: The generation of the system-level recommendation mask specifically includes: On the edge, the cosine similarity between the candidate content vector and the basic interest mask is calculated, and the Top-K content with the highest scores after rearrangement is selected. The selected content is weighted and aggregated to generate a system-level recommendation mask that includes ranking weights.

6. The edge system recommendation method based on semantic decay and memory fusion according to claim 1, characterized in that: The broadcasting of the system-level recommended mask as a standard intent signal of the operating system specifically includes: The operating system encapsulates and exposes standardized interfaces at the framework layer for authorized third-party applications to call; Third-party applications obtain the current system-level recommendation mask through the interface and calculate the similarity between their internal candidate content vectors and the mask to achieve alignment with system-level intent.

7. The edge system recommendation method based on semantic decay and memory fusion according to claim 1, characterized in that: The edge semantic decay engine is built on a lightweight large language model and is used to perform logical reasoning on the semantic state of user interests and output a semantic probability vector. The semantic probability vector includes the probability of task completion, the probability of context suspension, the probability of interest reinforcement, and the probability of natural decay.

8. The edge system recommendation method based on semantic decay and memory fusion according to claim 1, characterized in that: The process of mapping semantic probability vectors to attenuation factors using deterministic rules specifically includes: A confidence threshold is preset for each semantic state; when the probability of a certain state in the semantic probability vector exceeds the corresponding threshold, it is mapped to the corresponding decay factor according to the preset deterministic rule.

9. The edge system recommendation method based on semantic decay and memory fusion according to claim 1, characterized in that: The decay factor is used to update the weights of long-term memory nodes, including blocking, freezing protection, dynamic enhancement, or natural decay.

10. The method according to claim 1, characterized in that: The updating of the weights of long-term memory nodes and the integration of new semantics specifically includes: The weights of the non-frozen nodes are updated according to the decay factor. The periodic aggregated vector is fused with the original semantic vector of interest according to a preset update rate to generate a new semantic vector of interest. Clear the short-term behavior sequence and short-term vector cache, and enter the real-time monitoring state of the next cycle.

Images