Personalized recommendation system based on semantic analysis

Abstract

The invention discloses a personalized recommendation system based on semantic analysis, and relates to the technical field of personalized recommendation systems.The personalized recommendation system comprises a data acquisition module, a dynamic situational semantic modeling module and a recommendation engine module, the data acquisition module transmits acquired user behavior data and situational data to the dynamic situational semantic modeling module, and the dynamic situational semantic modeling module transmits the acquired user behavior data and situational data to the recommendation engine module; the dynamic situational semantic modeling module extracts sequence features of user behavior data through a behavior context analyzer, generates context embedding through a situational semantic embedding network, completes cross-modal semantic alignment of multi-modal data through a multi-modal fusion layer, and generates multi-level semantic embedding features. According to the method, the dynamic situational semantic modeling module is designed, the semantic alignment function of the situation and the behavior is achieved, the problem that a traditional recommendation system is insufficient in situation data utilization is solved, the adaptability of the recommendation result to the dynamic situation is enhanced, and the method is suitable for dynamically changing user demand scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of personalized recommendation systems, and in particular to a personalized recommendation system based on semantic analysis. Background Art

[0002] With the explosive growth of Internet information and the increasing diversity of user needs, traditional recommendation systems are unable to meet users' needs for personalized content and real-time recommendations. Existing recommendation technologies mainly rely on users' historical behaviors and basic portraits, ignoring the correlation between users' current contextual factors and complex multimodal data. This leads to a lack of dynamic adaptability in recommendation results, which cannot effectively meet users' personalized needs in different events, locations, and situations. At the same time, technical difficulties such as cold start problems, data sparsity, and missing modalities further limit the performance of recommendation systems.

[0003] Patent document CN117217865B discloses a personalized recommendation system based on big data analysis. The above patent realizes a wider range of collected data. The calculated product type attention index shows the user's attention to different types of products, and the product type demand index shows the user's actual demand for different types of products. The comprehensive purchase possibility index of the product calculated based on the two is more in line with the user's shopping preferences and the prediction of the user's product purchase is also more accurate.

[0004] The above patent recommends products that are more in line with the user's preferences through extensive data collection, but the above patent does not consider the effective combination of real-time context and behavioral semantics in the recommendation process;

[0005] To this end, the present application proposes a personalized recommendation system that can realize the semantic alignment function of context and behavior and dynamically adjust the weight of recommendation features. Summary of the invention

[0006] The purpose of the present invention is to provide a personalized recommendation system based on semantic analysis. In order to break through the limitations, by introducing dynamic contextual semantic modeling, multimodal data fusion and reinforcement learning feedback mechanism, it can fully explore the relationship between users' short-term behavior and long-term preferences, and generate highly personalized recommendation results in combination with real-time context information. This application has discovered a technical solution that can realize the semantic alignment function of context and behavior and dynamically adjust the weight of recommendation features, realize the semantic alignment function of context and behavior, the function of dynamically adjusting the weight of recommendation features, optimize the interaction relationship between users and items and adjust the recommendation weight through context perception technology, overcome the performance bottleneck problem of traditional recommendation systems in large-scale high-concurrency scenarios, so as to solve the technical problems of inaccurate recommendation results, poor user satisfaction, inability to dynamically optimize recommendation strategies and performance bottlenecks in high-concurrency scenarios proposed in the above background technology.

[0007] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a personalized recommendation system based on semantic analysis, comprising a data collection module, a dynamic contextualized semantic modeling module and a recommendation engine module, wherein the data collection module transmits the collected user behavior data and contextual data to the dynamic contextualized semantic modeling module, the dynamic contextualized semantic modeling module extracts the sequence features of the user behavior data through a behavior context parser, generates context embedding through a contextual semantic embedding network, and completes cross-modal semantic alignment of multimodal data through a multimodal fusion layer to generate multi-level semantic embedding features, and the recommendation engine module generates a recommendation list according to the multi-level semantic embedding features, user portraits and the interaction relationship between users and items.

[0008] Preferably, the behavior context parser extracts context features of user behavior data through the following steps:

[0009] Sort the user behavior data sequence by timestamp and split it based on the sliding window;

[0010] Use time series analysis models to extract behavior types, behavior time intervals, and historical path characteristics;

[0011] Graph neural networks are used to model the user's operation path and generate contextual feature representations.

[0012] Preferably, the context semantic embedding network adopts a Transformer structure, which specifically includes the following steps:

[0013] Encode time information, geographic location information and external context data to generate multi-dimensional context features;

[0014] Use multi-head self-attention mechanism to calculate the correlation between short-term behavior and long-term preference;

[0015] The embedding layer generates high-dimensional semantic vectors for different context data and dynamically updates the context weights.

[0016] Preferably, the multimodal fusion layer achieves cross-modal alignment by the following method:

[0017] Build embeddings of text, image, and video data;

[0018] Using positive and negative sample pairs, we optimize the consistency between modalities through contrastive learning methods;

[0019] In the scenario of missing modality, a generative adversarial network is used to generate missing modality features to ensure the integrity of the recommendation process.

[0020] Preferably, the positive and negative sample pairs of the contrastive learning are constructed as follows:

[0021] Positive sample pairs: generated from different modal data of the same recommended content;

[0022] Negative sample pairs: randomly select modal data from unrelated recommended content;

[0023] The loss function uses contrast loss to ensure that the positive sample pairs are closer and the negative sample pairs are farther away.

[0024] Preferably, the recommendation engine module includes a dynamic feature optimizer and a collaborative filtering enhancement module, and the dynamic feature optimizer dynamically adjusts the weight of the recommendation feature by the following method:

[0025] Dynamically reduce the weight of short-term behavior features based on the time decay function;

[0026] Use the policy gradient method of reinforcement learning to adjust the recommendation engine parameters in real time based on user feedback;

[0027] In the recommendation process, the graph-based collaborative filtering algorithm is combined to optimize the interaction model between users and items.

[0028] Preferably, the collaborative filtering enhancement module optimizes the recommendation results by the following steps:

[0029] Construct a user-item interaction matrix and calculate the similarity between users and items by combining semantic embedding features;

[0030] Use matrix decomposition algorithm to perform low-rank decomposition on the interaction matrix to generate recommended content based on sparse user behaviors;

[0031] The recommendation results are weighted and adjusted based on context-aware technology to solve the cold start problem.

[0032] Preferably, the personalized recommendation system further includes a feedback learning module, which updates the recommendation strategy in the recommendation engine module based on the user's interactive feedback information on the recommended content. The feedback learning module optimizes the recommendation engine module in the following ways:

[0033] Generate instant reward signals based on the user's click-through rate (CTR), dwell time (Time), and bounce rate (BounceRate) of the recommended content;

[0034] Use deep Q-learning algorithms to adjust recommendation strategies in real time;

[0035] In the immediate reward signal, the comprehensive score is calculated by the following formula:

[0036] Reward=α·CTR+β·Time-γ·BounceRate;

[0037] Among them, Reward represents the immediate reward signal, and α, β, and γ are hyperparameters for adjusting weights.

[0038] Preferably, the data acquisition module obtains the context data in the following manner:

[0039] Integrate third-party APIs to obtain external contextual information on weather, holidays, and hot events in real time;

[0040] Use ETL pipeline to clean, transform and load collected data;

[0041] In geographic location feature extraction, a spatial hierarchical model is used to map geographic location into spatial embedding features.

[0042] Preferably, the personalized recommendation system adopts the following distributed architecture design:

[0043] Data storage uses MongoDB or HBase as a distributed database, and combines it with Redis cache to optimize data access performance;

[0044] Data flow is distributed in real time through the Kafka message queue to ensure the efficiency and stability of data processing;

[0045] Model training is accelerated by GPU clusters, and low-latency reasoning is achieved by exporting models using ONNX and combining them with TensorRT.

[0046] Service management uses Kubernetes to orchestrate microservices, achieving dynamic expansion and high concurrent processing capabilities.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The present invention realizes the semantic alignment function of context and behavior by designing a dynamic contextual semantic modeling module, solves the problem of insufficient utilization of context data in traditional recommendation systems, enhances the adaptability of recommendation results to dynamic contexts, significantly improves the accuracy of recommendations, and is suitable for dynamically changing user demand scenarios;

[0049] 2. The present invention realizes the function of dynamically adjusting the weight of recommendation features by designing a dynamic feature optimizer, solves the problem of excessive influence of short-term behavior on recommendation results, and improves the responsiveness of recommendations to real-time needs of users, realizes dynamic optimization of recommendation features, and significantly improves the real-time performance of recommendation results and user satisfaction;

[0050] 3. The present invention has a collaborative filtering enhancement module designed to optimize the interaction between users and items and adjust the recommendation weights through context-aware technology, effectively solving the cold start problem and user behavior sparseness problem, and improving the accuracy and usability of recommendation results in new user and new item scenarios;

[0051] 4. The present invention overcomes the performance bottleneck problem of traditional recommendation systems in large-scale high-concurrency scenarios by designing a distributed architecture. It has the characteristics of high efficiency, scalability and low latency, and meets the needs of large-scale user recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a schematic diagram of the system framework of the present invention;

[0053] Figure 2 Schematic diagram of the personalized recommendation process of the present invention. DETAILED DESCRIPTION

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] See also Figure 1 and Figure 2 , an embodiment provided by the present invention: a personalized recommendation system based on semantic analysis, comprising a data collection module, a dynamic contextualized semantic modeling module and a recommendation engine module, wherein the data collection module transmits the collected user behavior data and contextual data to the dynamic contextualized semantic modeling module, the dynamic contextualized semantic modeling module extracts the sequence features of the user behavior data through a behavior context parser, generates contextual embedding through a contextual semantic embedding network, completes cross-modal semantic alignment of multimodal data through a multimodal fusion layer, and generates multi-level semantic embedding features, and the recommendation engine module generates a recommendation list according to the multi-level semantic embedding features, user portraits and the interactive relationship between users and items;

[0056] Furthermore, the system collects user behavior data in real time through tracking technology, browser plug-ins or mobile SDK, including clicks, browsing, searches, favorites and purchase operations. The data is transmitted to the data storage module through a unified interface and formatted. The system obtains geographic location information and timestamps from device sensors such as GPS and accelerometers, and obtains external contextual data in real time through third-party APIs such as weather services or news hotspot services, including weather conditions, hot events and holiday information. The ETL pipeline is used to clean and normalize the raw data, and the structured data is transmitted to the dynamic contextual semantic modeling module to provide support for subsequent analysis.

[0057] The collected user behavior data is sorted according to the order of events, and the user behavior sequence is segmented using a sliding window to generate short-term behavior fragments. The time interval characteristics of user behavior are extracted based on the event sequence analysis model, and the operation relationship in the user behavior sequence is extracted through path analysis. The user operation path is modeled using a graph neural network to generate contextual embedding features of the user behavior sequence. The context data is encoded using the Transformer architecture to generate a high-dimensional embedding vector of the time position sequence. The multi-head self-attention mechanism dynamically adjusts the weight of the context information to ensure that different contexts have appropriate importance in the recommendation process. The context embedding vector and the behavior context embedding vector are fused through the attention module to generate a semantic feature vector. Modal embedding is generated for multimodal data in user behavior and context, such as text, images, and videos, to construct a unified semantic representation. Contrastive learning is used to optimize the consistency between different modal features, and all modalities are mapped to a unified embedding space. For data with missing modalities, the generative adversarial network is used to complete the features to ensure the integrity of the semantic embedding.

[0058] The dynamic feature optimizer applies a time decay function to semantic features to reduce the impact of outdated short-term behaviors on recommendation results. It uses a reinforcement learning strategy to optimize the recommendation process. The policy gradient algorithm updates the recommendation weight in real time according to user behavior, dynamically generates user portraits, and adjusts the order and content of recommendation results in real time based on multi-layer semantic features. It builds a user-item interaction matrix based on context awareness, uses matrix decomposition methods to calculate the potential association between users and items, and optimizes the matrix decomposition results through semantic embedding features. In sparse interaction scenarios such as cold start users, the recommendation effect is improved. The recommendation list is sorted and adjusted according to the current user context, so that the recommendation results are more in line with the user's immediate needs.

[0059] The feedback learning module optimizes the recommendation strategy based on user feedback, realizes adaptive improvement of the recommendation engine, and continuously improves user satisfaction.

[0060] See also Figure 1 and Figure 2 , an embodiment provided by the present invention: a personalized recommendation system based on semantic analysis, the dynamic contextualized semantic modeling module extracts the sequence features of user behavior data through a behavior context parser, generates context embedding through a contextual semantic embedding network, and completes the cross-modal semantic alignment of multimodal data through a multimodal fusion layer to generate multi-level semantic embedding features, and the behavior context parser extracts the context features of user behavior data through the following steps:

[0061] Sort the user behavior data sequence by timestamp and split it based on the sliding window;

[0062] Use time series analysis models to extract behavior types, behavior time intervals, and historical path characteristics;

[0063] Use graph neural networks to model the user's operation path and generate context feature representations;

[0064] The context semantic embedding network adopts a Transformer structure and specifically includes the following steps:

[0065] Encode time information, geographic location information and external context data to generate multi-dimensional context features;

[0066] Use multi-head self-attention mechanism to calculate the correlation between short-term behavior and long-term preference;

[0067] Generate high-dimensional semantic vectors for different context data through the embedding layer and dynamically update the context weights;

[0068] The multimodal fusion layer achieves cross-modal alignment by the following method:

[0069] Build embeddings of text, image, and video data;

[0070] Using positive and negative sample pairs, we optimize the consistency between modalities through contrastive learning methods;

[0071] In the scenario of missing modalities, a generative adversarial network is used to generate missing modal features to ensure the integrity of the recommendation process;

[0072] The positive and negative sample pairs of the contrastive learning are constructed as follows:

[0073] Positive sample pairs: generated from different modal data of the same recommended content;

[0074] Negative sample pairs: randomly select modal data from unrelated recommended content;

[0075] The loss function uses contrast loss to ensure that the positive sample pairs are closer and the negative sample pairs are farther away;

[0076] Furthermore, the e-commerce platform records the user's behavior sequence, including browsing products, clicking details, adding to shopping carts, and completing purchases. Each record has a timestamp. After sorting the behavior data according to the timestamp, a sliding window such as a 5-minute interval is used to divide the behavior sequence into multiple time segments. Time series analysis modules such as ARIMA or LSTM are used to extract the following features: behavior type, time interval, and historical path; the user's behavior sequence is represented as a directed graph, where nodes represent behavior operations and edges represent behavior sequences. A graph neural network is used to model the graph, generate context feature representations, capture potential associations between user operations, and output context embedding features of the user's behavior sequence to describe the user's interests and operation habits in a specific time segment;

[0077] Time information is converted into vectors through periodic coding such as sine and cosine functions to represent day and night or holiday characteristics. Geographic location information such as the city or business district where the user is located is converted into vectors through geocoding to represent location correlation. External contexts such as weather and hot events are embedded through Transformer to generate multi-dimensional context features. The system uses a multi-head self-attention mechanism to dynamically analyze the relationship between short-term behaviors such as purchase behaviors in the last 5 minutes and long-term preferences such as historical purchase records. According to the behavior weight and context influence, the context weight is updated to adjust the influence of context data. Different types of context data are mapped to a unified high-dimensional semantic space through the embedding layer, and a dynamically updated context semantic vector is generated in combination with the user's current behavior to provide a basis for subsequent recommendations. A high-dimensional semantic vector representing the user's context characteristics is output to reflect the user's immediate needs.

[0078] For product text descriptions such as product names and detailed descriptions, pre-trained language models such as BERT are used to generate text embeddings. For product images, convolutional neural networks such as ResNET are used to extract visual embedding features. For product videos, time series convolutional neural networks are used to extract dynamic embedding features. Positive sample pairs: the text, image, and video embeddings of the same product are used as positive sample pairs. Negative sample pairs: modal embeddings of unrelated products are randomly selected as negative sample pairs. The contrast loss function is used to optimize the embedding space so that the positive sample pairs are closer and the negative sample pairs are farther away. In the case of missing modalities (such as some products without videos), a generative adversarial network is used to generate embedding features of the missing modality. The authenticity of the generated features is verified by the discriminator to ensure that the completed modal features are consistent with the actual modality. The unified semantic embedding of the products after cross-modal alignment is output for the recommendation engine to generate a content list.

[0079] See also Figure 1 and Figure 2 , an embodiment of the present invention provides: a personalized recommendation system based on semantic analysis, the recommendation engine module includes a dynamic feature optimizer and a collaborative filtering enhancement module, the dynamic feature optimizer dynamically adjusts the recommendation feature weights by the following method:

[0080] Dynamically reduce the weight of short-term behavior features based on the time decay function;

[0081] Use the policy gradient method of reinforcement learning to adjust the recommendation engine parameters in real time based on user feedback;

[0082] In the recommendation process, the graph-based collaborative filtering algorithm is combined to optimize the interaction model between users and items;

[0083] The collaborative filtering enhancement module optimizes the recommendation results by the following steps:

[0084] Construct a user-item interaction matrix and calculate the similarity between users and items by combining semantic embedding features;

[0085] Use matrix decomposition algorithm to perform low-rank decomposition on the interaction matrix to generate recommended content based on sparse user behaviors;

[0086] Based on context-aware technology, the recommendation results are weighted and adjusted to solve the cold start problem;

[0087] The personalized recommendation system further includes a feedback learning module, which updates the recommendation strategy in the recommendation engine module based on the user's interactive feedback information on the recommended content. The feedback learning module optimizes the recommendation engine module in the following ways:

[0088] Generate instant reward signals based on the user's click-through rate (CTR), dwell time (Time), and bounce rate (BounceRate) of the recommended content;

[0089] Use deep Q-learning algorithms to adjust recommendation strategies in real time;

[0090] In the immediate reward signal, the comprehensive score is calculated by the following formula:

[0091] Reward=α·CTR+β·Time-γ·BounceRate;

[0092] Among them, Reward represents the immediate reward signal, α, β, and γ are hyperparameters for adjusting weights;

[0093] The data acquisition module obtains context data in the following ways:

[0094] Integrate third-party APIs to obtain external contextual information on weather, holidays, and hot events in real time;

[0095] Use ETL pipeline to clean, transform and load collected data;

[0096] In the extraction of geographic location features, a spatial hierarchical model is used to map geographic locations into spatial embedding features;

[0097] Furthermore, the user's short-term behavior (such as clicks and browsing records in the past hour) has a higher weight for instant recommendation, but its importance will gradually decay over time. The calculation formula of the time decay function is: W t =W0·e -λ·Δt , where W t is the weight after time decay, W0 is the initial feature weight, Δt is the difference between the time when the behavior occurred and the current time, λ is the time decay coefficient (adjustable), t represents the time node, and 0 represents the initial time. The system adjusts the impact of short-term behavior on the recommendation results according to time decay, so that the recommended content is more in line with the user's immediate needs;

[0098] A reinforcement learning model with the recommendation strategy as the action space is constructed, and user behavior feedback is used as a reward signal to optimize the recommendation strategy. The reinforcement learning model uses the policy gradient method, and the optimization process is as follows:

[0099] Recommended content is the action, and user behavior is the reward;

[0100] Reward signal formula: Reward = α·CTR + β·Time - γ·BounceRate;

[0101] Among them, CTR reflects the attractiveness of the recommended content, Time measures the depth of user interest, BounceRate indicates the user's dissatisfaction with the recommended content, Reward represents the immediate reward signal, and α, β, and γ are hyperparameters for adjusting weights;

[0102] The policy update is achieved through the gradient ascent method:

[0103] θ t+1 =θ t +η·φ θ J(θ);

[0104] Among them, θ represents the strategy parameter; θ t is the strategy parameter at the current time t, indicating the parameter set currently used by the recommendation system; θ t+1 is the policy parameter at time t+1 after the update, which represents the new parameter set obtained through policy gradient optimization; J(θ) is the policy function, which represents the optimization goal of the recommendation system under the condition of policy parameter θ; φ is a gradient operator used to calculate the update direction of policy parameter θ; θ represents the gradient operator calculated for the policy parameter θ; φ θ J(θ) is the policy gradient, which represents the gradient of the policy function J(θ) to the policy parameter θ; η is the learning rate;

[0105] Reinforcement learning strategies dynamically update recommendation engine parameters, making recommendation results more relevant to user interests;

[0106] The collaborative filtering enhancement module optimizes user-item recommendation results through semantic embedding features and context-aware technology, especially in cold start problems:

[0107] Step 1: Construct a user-item interaction matrix

[0108] User behavior data is represented by a sparse matrix:

[0109]

[0110] The weight w in the interaction matrix can be determined by the behavior type and its importance, M ij represents the user-item interaction matrix;

[0111] For example:

[0112] i represents the user variable, j represents the item variable, click weight: 1; collection weight: 2; purchase weight: 5;

[0113] Step 2: Optimize similarity calculation by combining semantic embedding features

[0114] On the basis of traditional collaborative filtering, semantic embedding features are used to enhance similarity calculation:

[0115] Sim(i,j)=Cosine(E i ,E j )

[0116] Among them, i represents the user variable, j represents the item variable, and E i is the user semantic embedding vector, E j is the semantic embedding vector of the item, and Cosine represents the cosine similarity calculation;

[0117] Step 3: Matrix decomposition to generate sparse user behavior recommendations

[0118] Use matrix decomposition algorithms such as SVD and ALS to perform low-rank decomposition of the interaction matrix:

[0119] M≈P·Q T

[0120] Among them, P is the user potential factor matrix, Q is the item potential factor matrix, Q T represents the transposed matrix of the item latent factor matrix, and M is the approximate representation of the user-item interaction matrix;

[0121] After decomposition, a recommendation list can be generated for sparse scenarios to provide reasonable recommendations for cold start users.

[0122] Step 4: Weighted adjustment of recommendation results based on context-aware technology

[0123] Use the user's current context to make weighted adjustments to recommended content: Time weight: prioritize content with high relevance in the current time period (such as holiday promotional products); Geographic weight: recommend products or services near the user's location; External event weight: combine hot events to recommend products related to the event;

[0124] Adjustment formula:

[0125] R ij =S ij ·(1+λ·C j )

[0126] Among them, i represents the user variable, j represents the item variable, and R ij is the weighted adjusted recommendation score, S ij is the initial recommendation score, C j is the situational weight, and λ is the situational adjustment coefficient.

[0127] See also Figure 1 and Figure 2 , an embodiment of the present invention provides: a personalized recommendation system based on semantic analysis, the personalized recommendation system adopts the following distributed architecture design:

[0128] Data storage uses MongoDB or HBase as a distributed database, and combines it with Redis cache to optimize data access performance;

[0129] Data flow is distributed in real time through the Kafka message queue to ensure the efficiency and stability of data processing;

[0130] Model training is accelerated by GPU clusters, and low-latency reasoning is achieved by exporting models using ONNX and combining them with TensorRT.

[0131] Service management uses Kubernetes to orchestrate microservices, achieving dynamic expansion and high concurrent processing capabilities;

[0132] Furthermore, MongoDB or HBase, as distributed databases, provides high throughput and strong consistency. MongoDB is used to store user behavior data and contextual data, and organizes data according to the document structure. Each document contains user ID, behavior type, timestamp, contextual information, etc.; HBase is used for high-frequency writing and querying scenarios such as recommendation logs. The table structure includes row key (user ID or item ID), column family (behavior type) and timestamp, etc.; data partitioning strategy: horizontal sharding according to user ID or item ID to achieve load balancing; data copy strategy: set up a multi-copy synchronization mechanism to ensure high availability and disaster recovery capabilities, Redis caches hot user behavior data and hot recommendation results, uses the LRU strategy to dynamically manage cache capacity, sets the cache expiration time to 10 minutes, and refreshes recommended content regularly;

[0133] Use Spark clusters to preprocess historical data, including deduplication, normalization, and feature engineering. Use deep learning frameworks such as TensorFlow for model training. GPU clusters use the Horovod framework to implement multi-node parallel training to accelerate large-scale semantic modeling and iterative optimization of recommendation algorithms. After training, use the validation set to calculate metrics and convert the optimal model into ONNX format to make it compatible with multiple inference engines.

[0134] Split the recommendation system into the following microservices:

[0135] Data collection service: processing user behavior data and context data upload;

[0136] Semantic modeling service: Generate user semantic embedding features;

[0137] Recommendation engine service: generates recommendation lists and responds to user requests;

[0138] Feedback learning service: optimize recommendation strategies based on user feedback;

[0139] Use Docker to containerize each microservice, deploy it to the cluster through Kubernetes, set HPA to dynamically adjust the number of Pods based on CPU and memory usage, use Ingress Controller and load balancer to distribute user requests to ensure service stability; use Prometheus and Grafana to monitor the performance indicators of each microservice, such as QPS, latency, and error rate, Kubernetes automatically restarts failed Pods, and seamlessly deploys new versions through a rolling update mechanism.

[0140] Working principle: The system collects user behavior data and context data in real time through the data acquisition module, and transmits the user behavior data and context data to the dynamic contextualized semantic modeling module. The contextualized semantic modeling module converts multi-dimensional data into multi-level semantic embedding features through the behavior context parser, contextual semantic embedding network and multimodal fusion layer;

[0141] The recommendation engine module generates a personalized recommendation list based on semantic embedding features, user portraits, and the interaction between users and items. The collaborative filtering enhancement module further combines context-aware technology to optimize recommendation results and solve the cold start problem. At the same time, the dynamic feature optimizer uses reinforcement learning and time decay mechanisms to adjust the recommendation weights to improve the relevance and accuracy of recommendations.

[0142] The feedback learning module monitors the user's interaction with the recommended content, updates the recommendation strategy through reinforcement learning, and dynamically adjusts the recommendation engine parameters of the recommendation engine module to improve the adaptability of the recommendation results and user satisfaction.

[0143] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. A personalized recommendation system based on semantic analysis, comprising a data collection module, a dynamic contextualized semantic modeling module and a recommendation engine module, characterized in that: The data collection module transmits the collected user behavior data and context data to the dynamic contextualized semantic modeling module. The dynamic contextualized semantic modeling module extracts the sequence features of the user behavior data through the behavior context parser, generates context embedding through the contextual semantic embedding network, and completes the cross-modal semantic alignment of multimodal data through the multimodal fusion layer to generate multi-level semantic embedding features. The recommendation engine module generates a recommendation list based on the multi-level semantic embedding features, user portraits, and the interactive relationship between users and items.

2. The personalized recommendation system based on semantic analysis according to claim 1, characterized in that: The behavior context parser extracts context features of user behavior data through the following steps: Sort the user behavior data sequence by timestamp and split it based on the sliding window; Use time series analysis models to extract behavior types, behavior time intervals, and historical path characteristics; Graph neural networks are used to model the user's operation path and generate contextual feature representations.

3. The personalized recommendation system based on semantic analysis according to claim 1, characterized in that: The context semantic embedding network adopts a Transformer structure, which specifically includes the following steps: Encode time information, geographic location information and external context data to generate multi-dimensional context features; Use multi-head self-attention mechanism to calculate the correlation between short-term behavior and long-term preference; The embedding layer generates high-dimensional semantic vectors for different context data and dynamically updates the context weights.

4. The personalized recommendation system based on semantic analysis according to claim 1, characterized in that: The multimodal fusion layer achieves cross-modal alignment by the following method: Build embeddings of text, image, and video data; Using positive and negative sample pairs, we optimize the consistency between modalities through contrastive learning methods; In the scenario of missing modality, a generative adversarial network is used to generate missing modality features to ensure the integrity of the recommendation process.

5. The personalized recommendation system based on semantic analysis according to claim 4, characterized in that: The positive and negative sample pairs of the contrastive learning are constructed as follows: Positive sample pairs: generated from different modal data of the same recommended content; Negative sample pairs: randomly select modal data from unrelated recommended content; The loss function uses contrast loss to ensure that the positive sample pairs are closer and the negative sample pairs are farther away.

6. The personalized recommendation system based on semantic analysis according to claim 1, characterized in that: The recommendation engine module includes a dynamic feature optimizer and a collaborative filtering enhancement module. The dynamic feature optimizer dynamically adjusts the weight of the recommended features by the following method: Dynamically reduce the weight of short-term behavior features based on the time decay function; Use the policy gradient method of reinforcement learning to adjust the recommendation engine parameters in real time based on user feedback; In the recommendation process, the graph-based collaborative filtering algorithm is combined to optimize the interaction model between users and items.

7. The personalized recommendation system based on semantic analysis according to claim 6, characterized in that: The collaborative filtering enhancement module optimizes the recommendation results by the following steps: Construct a user-item interaction matrix and calculate the similarity between users and items by combining semantic embedding features; Use matrix decomposition algorithm to perform low-rank decomposition on the interaction matrix to generate recommended content based on sparse user behaviors; The recommendation results are weighted and adjusted based on context-aware technology to solve the cold start problem.

8. The personalized recommendation system based on semantic analysis according to claim 1, characterized in that: The personalized recommendation system further includes a feedback learning module, which updates the recommendation strategy in the recommendation engine module based on the user's interactive feedback information on the recommended content. The feedback learning module optimizes the recommendation engine module in the following ways: Generate instant reward signals based on the user's click-through rate (CTR), dwell time (Time), and bounce rate (BounceRate) of the recommended content; Use deep Q-learning algorithms to adjust recommendation strategies in real time; In the immediate reward signal, the comprehensive score is calculated by the following formula: Reward=α·CTR+β·Time-γ·BounceRate; Among them, Reward represents the immediate reward signal, and α, β, and γ are hyperparameters for adjusting weights.

9. The personalized recommendation system based on semantic analysis according to claim 1, characterized in that: The data acquisition module obtains context data in the following ways: Integrate third-party APIs to obtain external contextual information on weather, holidays, and hot events in real time; Use ETL pipeline to clean, transform and load collected data; In geographic location feature extraction, a spatial hierarchical model is used to map geographic location into spatial embedding features.

10. The personalized recommendation system based on semantic analysis according to claim 1, characterized in that: The personalized recommendation system adopts the following distributed architecture design: Data storage uses MongoDB or HBase as a distributed database, and combines it with Redis cache to optimize data access performance; Data flow is distributed in real time through the Kafka message queue to ensure the efficiency and stability of data processing; Model training is accelerated by GPU clusters, and low-latency reasoning is achieved by exporting models using ONNX and combining them with TensorRT. Service management uses Kubernetes to orchestrate microservices, achieving dynamic expansion and high concurrent processing capabilities.

Images