House resource intelligent recommendation and matching method based on cloud computing and user behaviors
By constructing a multimodal property recommendation method, monitoring the status of cloud and edge nodes in real time, and introducing a drift suppression control strategy, the problem of the cloud model and local cache losing synchronization is solved, realizing the high efficiency, stability and intelligence of the property recommendation system, and improving user satisfaction and matching efficiency.
Patent Information
- Application Number
- CN202511496322.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-20
- Publication Date
- 2025-12-16
AI Technical Summary
In cloud computing and user behavior-based intelligent housing recommendation systems, the time misalignment between the cloud model and the local cache leads to model desynchronization, resulting in inconsistent recommendation results, system misjudgment of user preferences, unstable recommendation feedback, and impact on user experience and system stability.
By collecting multimodal behavior vectors, a model synchronization metric function and a resource balance function are constructed. The status of cloud and edge nodes is monitored in real time. Drift suppression control equations and bidirectional control strategies are introduced to achieve model synchronization and resource balance, thereby reducing the risk of recommendation feedback instability.
It improves the robustness and adaptability of the housing recommendation system in high-concurrency environments, reduces model desynchronization and inference bias, enhances the accuracy and continuity of recommendation results, and reduces the risk of fragmented user experience.
Smart Images
Figure CN121144618A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent housing recommendation and matching technology, and more specifically, to a method for intelligent housing recommendation and matching based on cloud computing and user behavior. Background Technology
[0002] In cloud-based and user-behavior-driven intelligent housing recommendation and matching systems, recommendation models typically operate with centralized cloud training and real-time inference at edge nodes to achieve dynamic housing matching and personalized recommendations under high concurrency. However, in this distributed architecture, there is often a time misalignment between the asynchronous updates of the cloud model and the local caching mechanism. When the cloud performs incremental parameter updates under high-concurrency tasks, local recommendation nodes may still perform inference based on outdated model weights or feature vectors, resulting in a typical "model desynchronization" phenomenon.
[0003] This desynchronization phenomenon is not isolated but closely related to sudden congestion and imbalanced allocation of cloud computing resources. When the system simultaneously handles multi-tenant model training, parameter synchronization, and recommendation requests during peak business periods, the scheduling priority of computing resources is dynamically adjusted. Model update tasks are delayed due to resource occupation, further postponing the time for parameter distribution to edge nodes. Consequently, the time difference between the cloud model and the local cache continues to widen, causing some nodes to remain running in an old model state for extended periods. Meanwhile, before receiving delayed parameter updates, local nodes continue to perform adaptive adjustments based on the current input data, resulting in a directional shift between local feature weights and the main cloud model.
[0004] This interaction between model update delays caused by fluctuations in computing resources and cache desynchronization creates a magnifying effect: resource congestion delays model synchronization, while model desynchronization leads to the accumulation of inference errors, causing the system to misjudge the direction of user preference evolution and triggering erroneous adaptive retraining. As the feedback loop continues to operate, inconsistencies between models are periodically amplified, manifesting as oscillations in recommendation results and fragmented user experiences. Ultimately, this mutual influence mechanism leads to cross-layer model state drift and recommendation feedback instability within the system, disrupting the consistency between the cloud and local systems, and in severe cases, even causing cyclical deviations in recommendation logic and global performance degradation. Summary of the Invention
[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for intelligent housing recommendation and matching based on cloud computing and user behavior, in order to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: The intelligent property recommendation and matching method based on cloud computing and user behavior includes the following steps: Collect user operation logs, text search content, geographical location preferences, and page dwell time information on the housing platform to form a multimodal behavior vector set; The multimodal behavior vector set is used as the input benchmark for the cloud recommendation model to train the cloud recommendation model. After training, the model parameters are distributed to the edge nodes, and the current model version number and timestamp are recorded to form a model status registration table. By comparing the timestamps and parameter differences between the cloud and local models, a model synchronization metric function is constructed to calculate the model synchronization offset, and the rate of change of the model synchronization offset is recorded. Real-time monitoring of cloud CPU, GPU, memory, and I / O queue utilization; constructing a resource balance function to calculate resource balance. The cross-layer state drift risk index is calculated based on the model synchronization offset and resource balance, and the recommended feedback instability index is calculated based on the rate of change of the model synchronization offset. Based on the cross-level state drift risk index and the recommendation feedback instability index, we can identify whether the housing recommendation system has a combined risk of cross-level state drift and feedback instability. To address the combined risks of cross-level state drift and feedback instability, a two-way control strategy is initiated, and a drift suppression control equation is introduced. After risk control, a multi-objective comprehensive scoring is performed on the candidate housing set, and the recommendation results are finally output in descending order of the multi-objective comprehensive scores.
[0007] In a preferred embodiment, the cloud recommendation model adopts a two-layer time-series fusion structure, including a long-term preference sub-model and a short-term interest sub-model.
[0008] In a preferred embodiment, for quantifying the cloud recommendation model Local caching model The degree of difference between them is determined by the model out-of-sync metric function, which calculates the model synchronization offset as follows: ,in This is the model synchronization offset. This represents the parameter matrix of the cloud-based recommendation model at time t. For the local caching model at time The parameter matrix, Synchronization delay time represents the time difference between the local model and the cloud version. It is an L2 norm.
[0009] In a preferred embodiment, the rate of change of the recording model synchronization offset is specifically as follows: ,in This represents the rate of change of the model synchronization offset.
[0010] In a preferred embodiment, the resource balance is calculated by constructing a resource balance function based on real-time monitoring of cloud CPU, GPU, memory, and I / O queue utilization, and according to the standard deviation and average utilization of resource utilization. ,in For resource balance, This is to prevent division by zero by a very small constant. For the standard deviation of resource usage, This represents the average occupancy rate.
[0011] In a preferred embodiment, the calculation of the cross-layer state drift risk index based on the model synchronization offset and resource balance is as follows: ,in This is a cross-layer state drift risk index. These represent the preset weighting coefficients for model synchronization offset and resource balance, respectively.
[0012] In a preferred embodiment, the calculation of the recommended feedback instability index based on the rate of change of the model synchronization offset is as follows: ,in To provide feedback on the instability index, Let V be the variance of the confidence scores of the past k recommendation results. These are the preset weighting coefficients representing the rate of change of the model's synchronous offset and the variance of the confidence level of the recommendation results, respectively.
[0013] In a preferred embodiment, the identification logic for whether the housing recommendation system has a combined risk of cross-level state drift and feedback instability is as follows: If the cross-level state drift risk index is greater than the preset cross-level state drift risk index threshold and the recommendation feedback instability index is greater than the preset recommendation feedback instability index threshold, then the housing recommendation system will be marked as having a combined risk of cross-level state drift and feedback instability.
[0014] In a preferred embodiment, the bidirectional control strategy includes uplink synchronous calibration and downlink adaptive feedback; a drift suppression control equation is introduced: ,in This is the model parameter matrix after drift suppression. This is the local model parameter matrix. For dynamic learning rate, This is the gradient operator.
[0015] In a preferred embodiment, the multi-objective comprehensive scoring of the candidate housing set is performed as follows: ,in This is a comprehensive score based on multiple objectives. The multimodal behavior vector of the current user With candidate housing The semantic similarity between feature vectors These represent the preset weight coefficients for semantic similarity and recommendation feedback instability index, respectively.
[0016] The technical effects and advantages of this invention are as follows: 1. This invention constructs a multimodal housing recommendation and matching method based on cloud computing and user behavior. It achieves comprehensive perception and dynamic control of cross-layer model state drift and recommendation feedback instability under a centralized cloud training and edge real-time inference architecture. By introducing real-time measurement of model synchronization offset and rate of change, resource balance monitoring, and cross-layer risk index calculation, this invention can accurately identify the combined risks caused by asynchronous cloud updates, local cache out-of-sync issues, and sudden congestion of computing resources. Simultaneously, through a bidirectional collaborative control strategy and drift suppression control equation, it achieves self-stabilizing synchronization and feedback adjustment between cloud and edge nodes, significantly reducing the accumulation of model out-of-sync and inference bias. Combined with a multi-objective comprehensive scoring mechanism, this invention can dynamically balance user interest matching, overall housing quality, and system stability in a steady state, thereby improving the accuracy and continuity of recommendation results and reducing the risk of recommendation oscillations and user experience fragmentation. Overall, this invention significantly enhances the robustness and adaptability of the housing recommendation system in high-concurrency, multi-tenant environments, achieving efficient, stable, and intelligent recommendation processes, and effectively improving user satisfaction and housing matching efficiency. Attached Figure Description
[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.
[0019] Example: Figure 1 This invention presents a method for intelligent housing recommendation and matching based on cloud computing and user behavior, comprising the following steps: Collect user activity logs (browsing, favorites, interactions), text search content, geographic location preferences, and page dwell time information on the property listing platform to form a multimodal behavior vector set; The multimodal behavior vector set is used as the input benchmark for training the cloud recommendation model. After training, the model parameters are distributed to edge nodes (housing matching agents), and the current model version number and timestamp are recorded to form a model status registration table. By comparing the timestamps and parameter differences between the cloud and local models, a model synchronization metric function is constructed to calculate the model synchronization offset, and the rate of change of the model synchronization offset is recorded. Real-time monitoring of cloud CPU, GPU, memory, and I / O queue utilization; constructing a resource balance function to calculate resource balance. The cross-layer state drift risk index is calculated based on the model synchronization offset and resource balance, and the recommended feedback instability index is calculated based on the rate of change of the model synchronization offset. Based on the cross-level state drift risk index and the recommendation feedback instability index, we can identify whether the housing recommendation system has a combined risk of cross-level state drift and feedback instability. To address the combined risks of cross-level state drift and feedback instability, a two-way control strategy is initiated, and a drift suppression control equation is introduced. After risk control, a multi-objective comprehensive scoring is performed on the candidate housing set, and the recommendation results are finally output in descending order of the multi-objective comprehensive scores.
[0020] Collect user activity logs (browsing, favorites, interaction feedback), text search content, geographic location preferences, and page dwell time information on the housing platform to form a multimodal behavior vector set; In this embodiment of the invention, operation logs (browsing, favorites, interaction feedback), text search content, geographic location preferences, and page dwell time information are first collected in real time from the housing platform to form an original multi-source dataset. ,in The data is visual modality data, derived from the duration of user browsing of property images and videos, zoom-in / zoom-out operations, and image region attention heatmaps. This is semantic modality data, derived from users' text search content, input keywords, and the distribution of semantically similar words; This is geographic modal data, derived from users' geographic location, route planning, and map click preferences; It is temporal modal data, derived from the time sequence of user access and interaction at different time periods; Standardization, denoising, and outlier removal operations are performed on the original data to obtain a time-consistent, multimodal standardized dataset: Extracting a set of multimodal behavior vectors from a standardized dataset: ,in As a visual preference vector, visual features (such as image content style, color composition, and spatial layout) can be extracted through a convolutional feature extraction network. The semantic retrieval vector is obtained by inputting the user's search query at time t into the BERT model (Natural Language Processing model). This semantic embedding vector reflects the semantic topics related to the properties the user is interested in. As a geographic preference vector, geographic interest point embedding features can be extracted using a geocoding model. As a time-based behavior vector, the Fast Fourier Transform (FFT) can be used to extract periodic preferences (such as frequent browsing on weekends) from the user's time-seriesd access frequency. The multimodal behavior vector set is used as the input benchmark for training the cloud recommendation model. After training, the model parameters are distributed to edge nodes (housing matching agents), and the current model version number and timestamp are recorded to form a model status registration table. The cloud recommendation model described in this embodiment of the invention The two-layer timing fusion structure consists of the following two sub-models: ,in This is the long-term preference model. This is the Short-term Preference Model. The preset weight coefficients for the long-term preference model and the short-term interest model satisfy... ; It should be noted that, The settings should be tailored to the specific circumstances. For example, an expert weighting method can be used, which involves inviting experts in relevant fields to determine the pre-defined weighting coefficients for each indicator through professional opinion surveys and comprehensive evaluations. The initial value can be 0.5 or 0.5.
[0021] The Long-Term Preference Model (LPM) is primarily used to capture users' stable housing interest trends and historical decision-making patterns, employing a gated recurrent unit (GRU) structure to extract long-term sequence features. , ,in This is a hidden state of long-term preference (long-term memory of accumulated historical behavior). and These represent the weight matrix and bias term of the long-term preference mapping layer, respectively. The hyperbolic tangent activation function is used to limit the output range to [-1, 1], thus suppressing overfitting. The long-term preference embedding vector is used to capture long-term behavioral features; the long-term preference prediction probability is output. ,in The long-term preference prediction probability represents the long-term prediction result of interest in matching housing properties. This is the Sigmoid function, used to generate the recommendation probability output; The Short-Term Interest Sub-Model (SPM) is primarily used to model recent user behavior fluctuations and short-term hotspot shifts, employing an Attention-CNN structure. ,in Embed vectors for short-term interests to capture recent behavioral features. , , , These represent the query, key, and value matrices of the behavior sequence, respectively, where d is the feature dimension; the output is the short-term interest prediction probability: ,in This represents the short-term interest prediction probability, indicating a short-term prediction of interest in matching housing listings. This is the Sigmoid function, used to generate the recommendation probability output; The model is trained using a joint loss function. Perform dynamic optimization: ,in The loss function for the long-term preference sub-model is... The loss function for the short-term interest sub-model is... represents the preset weight coefficients of the long-term preference sub-model loss function and the short-term interest sub-model loss function, respectively; parameter updates employ the momentum adaptive optimization algorithm (AdamW). ,in This represents the parameter matrix of the cloud-based recommendation model at time t. For learning rate, , These represent the first-order and second-order momentum estimates, respectively. This is to prevent division by zero by a very small constant (generally taken as...). After training is complete, the latest parameter version of the cloud-based recommendation model will be provided. Distribute to edge nodes (property matching agents) and record the current model version number and timestamp to form a model status registration table.
[0022] It should be noted that the above formulas are all dimensionless calculations. Commonly used methods for removing dimensions include Min-Max normalization and Z-Score standardization, which will not be elaborated here. The settings should be tailored to the specific circumstances. For example, an expert weighting method can be used, which involves inviting experts in relevant fields to determine the pre-defined weighting coefficients for each indicator through professional opinion surveys and comprehensive evaluations. The initial value can be 0.5 or 0.5.
[0023] By comparing the timestamps and parameter differences between the cloud and local models, a model synchronization metric function is constructed to calculate the model synchronization offset, and the rate of change of the model synchronization offset is recorded. In this embodiment of the invention, a quantitative cloud recommendation model is used. Local caching model The degree of difference between them is determined by the model out-of-sync metric function, which calculates the model synchronization offset as follows: ,in This is the model synchronization offset, used to reflect the "degree of model consistency" between the cloud and the local machine. This represents the parameter matrix of the cloud-based recommendation model at time t. For the local caching model at time The parameter matrix, Synchronization delay time represents the time difference between the local model and the cloud version. It is the L2 norm, used to characterize the overall drift magnitude in the model parameter space; It should be noted that the larger the model synchronization offset, the more significant the divergence in the parameters of the two models, the greater the difference in inference output, and the higher the risk of state drift the system faces. Model drift depends not only on the current offset but also on its trend of change. The rate of change of the recorded model synchronization offset is as follows: ,in This represents the rate of change of the model synchronization offset.
[0024] It should be noted that the above formulas are all dimensionless calculations. Commonly used methods for removing dimensions include Min-Max normalization and Z-Score standardization, which will not be elaborated here.
[0025] Real-time monitoring of cloud CPU, GPU, memory, and I / O queue utilization; constructing a resource balance function to calculate resource balance. In this embodiment of the invention, the global average resource utilization level is obtained by averaging the resource utilization of all nodes by real-time monitoring of the CPU, GPU, memory, and I / O queue utilization in the cloud. ,in This represents the average occupancy rate. , , , These represent the CPU, GPU, memory, and I / O queue utilization rates of computing node i at time t, respectively. This refers to the number of computing nodes within the cloud cluster. , These represent the preset weighting coefficients for CPU, GPU, memory, and I / O queue utilization, respectively. It should be noted that, , The settings should be tailored to the specific circumstances. For example, an expert weighting method can be used, which involves inviting experts in relevant fields to determine the pre-defined weighting coefficients for each indicator through professional opinion surveys and comprehensive evaluations. , The initial value can be 0.25, 0.25, 0.25, 0.25.
[0026] The standard deviation of resource usage is used to measure the degree of resource distribution balance among nodes. ,in For the standard deviation of resource usage, The overall resource utilization rate of computing node i is calculated by averaging the CPU, GPU, memory, and I / O queue utilization rates of computing node i. Based on the standard deviation of resource utilization and the average utilization rate, a resource balance function is constructed to calculate the resource balance: ,in For resource balance, This is to prevent division by zero by a very small constant (generally taken as...). ).
[0027] It should be noted that the above formulas are all dimensionless calculations. Commonly used methods for removing dimensions include Min-Max normalization and Z-Score standardization, which will not be elaborated here.
[0028] The cross-layer state drift risk index is calculated based on the model synchronization offset and resource balance, and the recommended feedback instability index is calculated based on the rate of change of the model synchronization offset. The cross-layer state drift risk index is calculated based on the model synchronization offset and resource balance, which reflects the combined degree of model asynchronous out-of-step and resource imbalance, as detailed below: ,in This is a cross-layer state drift risk index. These represent the preset weighting coefficients for model synchronization offset and resource balance, respectively. It should be noted that, The settings should be tailored to the specific circumstances. For example, an expert weighting method can be used, which involves inviting experts in relevant fields to determine the pre-defined weighting coefficients for each indicator through professional opinion surveys and comprehensive evaluations. The initial value can be 0.5 or 0.5.
[0029] The recommended feedback instability index is calculated based on the rate of change of the model synchronization offset. This index is used to measure the coupling effect between the feedback fluctuation of the system's recommended output and the model drift rate within a short period, as detailed below: ,in To provide feedback on the instability index, Let V be the variance of the confidence scores of the past k recommendation results. These are the preset weighting coefficients representing the rate of change of the model's synchronous offset and the variance of the confidence level of the recommendation results, respectively. It should be noted that, The settings should be tailored to the specific circumstances. For example, an expert weighting method can be used, which involves inviting experts in relevant fields to determine the pre-defined weighting coefficients for each indicator through professional opinion surveys and comprehensive evaluations. The initial value can be 0.5 or 0.5.
[0030] Based on the cross-level state drift risk index and the recommendation feedback instability index, we can identify whether the housing recommendation system has a combined risk of cross-level state drift and feedback instability. The identification logic for whether the housing recommendation system has a combined risk of cross-level state drift and feedback instability is as follows: If the cross-level state drift risk index is greater than the preset cross-level state drift risk index threshold and the recommendation feedback instability index is greater than the preset recommendation feedback instability index threshold, then the housing recommendation system will be marked as having a combined risk of cross-level state drift and feedback instability.
[0031] To address the combined risks of cross-level state drift and feedback instability, a two-way control strategy is initiated, and a drift suppression control equation is introduced. After risk control, a multi-objective comprehensive scoring is performed on the candidate housing set, and the recommendation results are finally output in descending order of the multi-objective comprehensive scores.
[0032] The bidirectional control strategy includes uplink synchronous calibration (Cloud → Edge): redistributing the latest model parameters in the cloud and locking version consistency; downlink adaptive feedback (Edge → Cloud): uploading the drift feedback weights of the edge inference results to the cloud; and introducing a drift suppression control equation. ,in Here is the model parameter matrix after drift suppression, representing the time step (i). The set of local model parameters, corrected by the drift suppression control equation, represents the optimal approximate weights after the system's self-stabilizing adjustment, and is used in the next round of recommendation inference or training iteration. This is the local model parameter matrix, referring to the edge nodes or local recommendation modules at time [time value missing]. The original model parameters, without drift suppression or uplink calibration correction, reflect the local state of the nodes during the independent learning phase. This refers to the dynamic learning rate. For gradient operators; After completing risk control and model synchronization correction, the system enters the dynamic comprehensive evaluation stage of the housing candidate set to achieve synergistic optimization of user preferences, housing quality, and system stability. The multi-objective comprehensive scoring of the housing candidate set is performed as follows: ,in This is a comprehensive score based on multiple objectives. The multimodal behavior vector of the current user With candidate housing The semantic similarity between feature vectors These represent the preset weight coefficients for semantic similarity and recommendation feedback instability index, respectively. It should be noted that the above formulas are all dimensionless calculations. Commonly used methods for removing dimensions include Min-Max normalization and Z-Score standardization, which will not be elaborated here. The settings should be tailored to the specific circumstances. For example, an expert weighting method can be used, which involves inviting experts in relevant fields to determine the pre-defined weighting coefficients for each indicator through professional opinion surveys and comprehensive evaluations. The initial value can be 0.5 or 0.5.
[0033] This invention constructs a multimodal housing recommendation and matching method based on cloud computing and user behavior. It achieves comprehensive perception and dynamic control of cross-layer model state drift and recommendation feedback instability under a centralized cloud training and edge real-time inference architecture. By introducing real-time measurement of model synchronization offset and rate of change, resource balance monitoring, and cross-layer risk index calculation, this invention can accurately identify the combined risks caused by asynchronous cloud updates, local cache out-of-sync issues, and sudden congestion of computing resources. Simultaneously, through a bidirectional collaborative control strategy and drift suppression control equation, it achieves self-stabilizing synchronization and feedback adjustment between cloud and edge nodes, significantly reducing the accumulation of model out-of-sync and inference bias. Combined with a multi-objective comprehensive scoring mechanism, this invention can dynamically balance user interest matching, overall housing quality, and system stability in a steady state, thereby improving the accuracy and continuity of recommendation results and reducing the risk of recommendation oscillations and user experience fragmentation. Overall, this invention significantly enhances the robustness and adaptability of housing recommendation systems in high-concurrency, multi-tenant environments, achieving efficient, stable, and intelligent recommendation processes, and effectively improving user satisfaction and housing matching efficiency.
[0034] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.
[0035] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0036] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for intelligent housing recommendation and matching based on cloud computing and user behavior, characterized by: Includes the following steps: Collect user operation logs, text search content, geographical location preferences, and page dwell time information on the housing platform to form a multimodal behavior vector set; The multimodal behavior vector set is used as the input benchmark for the cloud recommendation model to train the cloud recommendation model. After training, the model parameters are distributed to the edge nodes, and the current model version number and timestamp are recorded to form a model status registration table. By comparing the timestamps and parameter differences between the cloud and local models, a model synchronization metric function is constructed to calculate the model synchronization offset, and the rate of change of the model synchronization offset is recorded. Real-time monitoring of cloud CPU, GPU, memory, and I / O queue utilization; constructing a resource balance function to calculate resource balance. The cross-layer state drift risk index is calculated based on the model synchronization offset and resource balance, and the recommended feedback instability index is calculated based on the rate of change of the model synchronization offset. Based on the cross-level state drift risk index and the recommendation feedback instability index, we can identify whether the housing recommendation system has a combined risk of cross-level state drift and feedback instability. To address the combined risks of cross-level state drift and feedback instability, a two-way control strategy is initiated, and a drift suppression control equation is introduced. After risk control, a multi-objective comprehensive scoring is performed on the candidate housing set, and the recommendation results are finally output in descending order of the multi-objective comprehensive scores.
2. The intelligent housing recommendation and matching method based on cloud computing and user behavior according to claim 1, characterized in that: The cloud-based recommendation model adopts a two-layer time-series fusion structure, including a long-term preference sub-model and a short-term interest sub-model.
3. The intelligent housing recommendation and matching method based on cloud computing and user behavior according to claim 2, characterized in that: To quantify cloud-based recommendation models Local caching model The degree of difference between them is determined by the model out-of-sync metric function, which calculates the model synchronization offset as follows: ,in This is the model synchronization offset. This represents the parameter matrix of the cloud-based recommendation model at time t. For the local caching model at time The parameter matrix, Synchronization delay time represents the time difference between the local model and the cloud version. It is an L2 norm.
4. The intelligent housing recommendation and matching method based on cloud computing and user behavior according to claim 3, characterized in that: The rate of change of the synchronization offset of the recording model is as follows: ,in This represents the rate of change of the model synchronization offset.
5. The intelligent housing recommendation and matching method based on cloud computing and user behavior according to claim 4, characterized in that: The process involves real-time monitoring of cloud-based CPU, GPU, memory, and I / O queue utilization, and constructing a resource balance function based on the standard deviation and average utilization of resource utilization to calculate resource balance. ,in For resource balance, This is to prevent division by zero by a very small constant. For the standard deviation of resource usage, This represents the average occupancy rate.
6. The intelligent housing recommendation and matching method based on cloud computing and user behavior according to claim 5, characterized in that: The cross-layer state drift risk index is calculated based on the model synchronization offset and resource balance as follows: ,in This is a cross-layer state drift risk index. These represent the preset weighting coefficients for model synchronization offset and resource balance, respectively.
7. The intelligent housing recommendation and matching method based on cloud computing and user behavior according to claim 6, characterized in that: The recommended feedback instability index is calculated based on the rate of change of the model synchronization offset, as detailed below: ,in To provide feedback on the instability index, Let V be the variance of the confidence scores of the past k recommendation results. These are the preset weighting coefficients representing the rate of change of the model's synchronous offset and the variance of the confidence level of the recommendation results, respectively.
8. The intelligent housing recommendation and matching method based on cloud computing and user behavior according to claim 7, characterized in that: The identification logic for whether the housing recommendation system has a combined risk of cross-level state drift and feedback instability is as follows: If the cross-level state drift risk index is greater than the preset cross-level state drift risk index threshold and the recommendation feedback instability index is greater than the preset recommendation feedback instability index threshold, then the housing recommendation system will be marked as having a combined risk of cross-level state drift and feedback instability.
9. The intelligent housing recommendation and matching method based on cloud computing and user behavior according to claim 8, characterized in that: The bidirectional control strategy includes uplink synchronous calibration and downlink adaptive feedback; a drift suppression control equation is introduced: ,in This is the model parameter matrix after drift suppression. This is the local model parameter matrix. For dynamic learning rate, This is the gradient operator.
10. The intelligent housing recommendation and matching method based on cloud computing and user behavior according to claim 9, characterized in that: The multi-objective comprehensive scoring of the candidate housing set is performed as follows: ,in This is a comprehensive score based on multiple objectives. The multimodal behavior vector of the current user With candidate housing The semantic similarity between feature vectors These represent the preset weight coefficients for semantic similarity and recommendation feedback instability index, respectively.