Precise targeted AI recommendation system

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a precise targeted AI recommendation system. The method comprises the following steps: firstly, acquiring multi-source heterogeneous data, constructing a global space-time database through space-time alignment, complementing and repairing the data by using a generative adversarial network, and cleaning and verifying; and constructing a three-dimensional topological model based on the improved graph neural network, and fusing multiple features to generate a multi-resolution model. A scene simulation platform is established by adopting a multi-objective reinforcement learning algorithm, multi-objective optimization is set, and a strategy library is constructed. And screening key factors by using a sparse Bayesian theory to construct a lightweight model. And designing a layered dynamic optimization framework, solving an optimal solution of each sub-layer by using a hybrid algorithm, realizing cross-layer constraint synchronous decoupling through a block chain smart contract, and outputting a global optimal three-dimensional design scheme. The method improves the application level of the digital achievement of the power grid design, gives consideration to stability, economy and environmental protection, and has important application value.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and specifically to a precise targeting AI recommendation system. Background Art

[0002] In today's digital age, the Internet information has exploded, and users are faced with a vast amount of data. How to quickly and accurately obtain the content they are interested in from this vast amount of information has become an urgent problem to be solved. Recommendation systems have emerged accordingly. They analyze the behavior data of users and provide personalized recommendations for users, helping users save time and effort in information screening and improving the user experience. However, there are still many deficiencies in the accuracy and adaptability of existing recommendation systems.

[0003] Most traditional recommendation systems are based on simple collaborative filtering algorithms or content-based recommendation algorithms. The collaborative filtering algorithm makes recommendations by analyzing user groups with similar behaviors, but it has the problem of data sparsity. When the user behavior data is scarce, it is difficult to find similar users, resulting in poor recommendation effects. For example, in some niche interest fields, due to the limited number of users, it is difficult for the collaborative filtering algorithm to accurately recommend content that meets the user's needs. The content-based recommendation algorithm mainly relies on the extraction and matching of content features, and this method often ignores the potential interests of users and the dynamic changes of behaviors. For example, a user's interests may change with factors such as time and scenario, but it is difficult for the content-based recommendation algorithm to quickly adapt to these changes.

[0004] In addition, there are also defects in the data collection and processing of existing recommendation systems. On the one hand, the data collection method is relatively single, and it can often only obtain the basic behavior data of users, such as clicks and browsing, and cannot comprehensively capture the multi-dimensional behaviors of users. For example, when recommending travel products, it is impossible to obtain information such as the user's real-time feelings during the travel and the impact of local weather on the user's decision-making. On the other hand, for the collected data, there is a lack of effective processing and analysis methods, and the value behind the data cannot be fully explored. For example, when processing time-series behavior data, it is impossible to accurately extract the periodic pattern features of user interests, resulting in insufficient timeliness and pertinence of recommendations.

[0005] In terms of interest modeling, it is difficult for existing recommendation systems to build a comprehensive and accurate user interest model. Most systems simply model based on the user's historical behaviors, without fully considering the influence of factors such as content and scenario on user interests. For example, when recommending movies, the preference differences of users for movie types at different times (such as weekends and weekday evenings) and different moods (such as happy and tired) are not considered, resulting in a deviation between the recommendation results and the actual needs of users.

[0006] In terms of recommendation generation and feedback optimization, there is still much room for improvement in existing recommendation systems. The recommendation generation process lacks effective evaluation of long-term benefits, resulting in recommendation results that may only meet users' short-term needs and cannot provide users with continuously valuable recommendations. At the same time, when faced with users' feedback, the system has insufficient optimization capabilities and cannot adjust the recommendation strategy in a timely manner to adapt to the dynamic changes of users. For example, when users show negative feedback on recommended content, the system cannot quickly analyze the reasons and improve the recommendation algorithm, thus affecting users' trust in the recommendation system and their usage experience. Summary of the Invention

[0007] The purpose of the present invention is to provide a precise targeting AI recommendation system to solve the problems raised in the above background technology.

[0008] To achieve the above purpose, the present invention provides the following technical solution: A precise targeting AI recommendation system, the system includes:

[0009] A data collection module, which acquires multi-dimensional behavioral data streams of users, dynamically adjusts the data collection frequency and weight, and assigns different weights to behavioral data at different time nodes based on a time decay function;

[0010] A feature extraction module, which uses a multi-modal fusion algorithm to hierarchically analyze the behavioral data stream, extracts users' explicit preference features and implicit association features, and constructs a feature vector set;

[0011] An interest modeling module, which constructs a three-dimensional relationship topology of user-content-scene through a heterogeneous graph neural network, and quantifies the association strength between nodes by combining a dynamic attention mechanism;

[0012] A recommendation generation module, which generates candidate recommendation sequences based on a reinforcement learning framework, and evaluates the long-term benefits of the sequences through the Monte Carlo tree search algorithm and generates an optimal recommendation path;

[0013] A feedback optimization module, which monitors users' interaction behaviors in real time, uses a generative adversarial network to simulate negative feedback scenarios, and dynamically corrects the deviation parameters of the recommendation strategy.

[0014] Preferably, the dynamic adjustment of the data collection frequency and weight includes the following steps:

[0015] Divide time windows according to users' active periods, and calculate the real-time density distribution of behavioral data based on a sliding window mechanism;

[0016] When it is detected that the variance of the density distribution within consecutive time windows exceeds a preset threshold, trigger a high-frequency collection mode and increase the weight of the current window;

[0017] For inactive window data that exceeds the set time threshold, an exponential decay function is used to reduce its weight until it reaches zero.

[0018] Preferably, the multi-modal fusion algorithm includes the following steps:

[0019] Input the text data into a bidirectional Transformer model to extract semantic features, and use convolutional sparse coding for the image data to extract visual features;

[0020] Use a long short-term memory network for the time-series behavior data to extract periodic pattern features;

[0021] Calculate the correlation matrix between different features through a cross-modal attention layer, and fuse them to generate a unified feature vector.

[0022] Preferably, the construction of the three-dimensional relationship topology of user-content-scene includes the following steps:

[0023] Map the user node, content node, and scene node to three independent subgraphs of a heterogeneous graph respectively;

[0024] Establish bidirectional hyperedges between the subgraphs, and the hyperedge weights are jointly calculated from the user's historical behavior and scene context;

[0025] Iteratively update the node embedding representation through the graph diffusion algorithm until the topological structure converges.

[0026] Preferably, the Monte Carlo tree search algorithm for evaluating long-term rewards includes the following steps:

[0027] Decompose the recommended path into a sequence of state-action pairs and construct a virtual decision tree;

[0028] Integrate a Bayesian probability model in the tree nodes to predict the potential behavior branches of the user;

[0029] Accumulate the expected reward values of each path through backpropagation, and select the path with the smallest reward variance as the output.

[0030] Preferably, the adversarial generative network for simulating negative feedback scenarios includes the following steps:

[0031] The generator network synthesizes virtual negative feedback samples based on the user's historical positive feedback data;

[0032] The discriminator network outputs an adversarial loss value by comparing the feature distribution differences between the real negative feedback and the synthesized samples;

[0033] Inject the adversarial loss value back into the gradient update process of the recommendation strategy to optimize the robustness of the strategy.

[0034] Preferably, the dynamic attention mechanism for quantifying the association strength includes the following steps:

[0035] Define a multi-head attention calculation unit between nodes in the heterogeneous graph;

[0036] Adaptively select the number of attention heads according to the node type, where four-head attention is used for user-content association and two-head attention is used for content-scenario association;

[0037] Fuse the outputs of multi-head attention through a gating mechanism to generate the final association strength coefficient.

[0038] Preferably, the hierarchical parsing of the behavioral data stream includes the following steps:

[0039] Divide the original data stream into a basic operation layer, a semantic understanding layer, and an intention reasoning layer;

[0040] Identify click, stay, and favorite behavior patterns in the basic operation layer;

[0041] Parse the co-occurrence relationship between search keywords and content tags in the semantic understanding layer; ]>

[0042] Construct a probability transition matrix of the behavior chain in the intention reasoning layer to identify potential intention evolution paths.

[0043] Preferably, the steps for correcting the deviation parameters of the recommendation strategy include the following:

[0044] Establish a three-level evaluation index for policy deviation, including short-term deviation, medium-term stability, and long-term trend;

[0045] Use a temporal convolutional network to extract the time series features of each level index respectively;

[0046] Fuse multi-level features through residual connections to generate a deviation correction amount and update the policy parameters.

[0047] Preferably, the steps for constructing the feature vector set include the following:

[0048] Perform principal component analysis dimensionality reduction on the explicit preference features, and retain the principal components with a variance contribution rate exceeding 85%;

[0049] Use an autoencoder to perform non-linear compression on the implicit association features to generate low-dimensional dense vectors;

[0050] Calculate the interaction weights between the explicit features and the implicit features through the feature cross layer, and combine them to generate the final feature vector.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] In the data collection stage, by dynamically adjusting the data collection frequency and weight, it is possible to obtain user behavior data more efficiently and accurately. Time windows are divided according to the user's active periods, and the sliding window mechanism is used to calculate the real-time density distribution of behavior data. Once it is found that the variance of the density distribution in consecutive time windows exceeds the preset threshold, the high-frequency collection mode is triggered and the weight of the current window is increased. This can timely capture the behavior data during the user's active interest period and ensure that important information is not missed. For the data in inactive windows that exceed the set time threshold, the exponential decay function is used to reduce its weight until it reaches zero, effectively avoiding the interference of stale data on the recommendation results and enabling the data to always reflect the user's latest interest status. Compared with the traditional fixed-frequency collection, this data collection method can obtain more timely and targeted data, providing a more solid data foundation for subsequent recommendations.

[0053] The feature extraction module uses a multi-modal fusion algorithm to extract rich features from multi-dimensional data such as text, images, and temporal behaviors. The bidirectional Transformer model extracts semantic features from text data and can deeply understand the semantics of text information such as the user's search keywords and browsing content; convolutional sparse coding extracts visual features from image data and accurately grasps the key elements in the images; the long short-term memory network effectively extracts the periodic pattern features of temporal behavior data. By calculating the correlation matrix between different features through the cross-modal attention layer and fusing them to generate a unified feature vector, it comprehensively covers the user's explicit preference features and implicit association features, providing more comprehensive and accurate feature information for interest modeling and recommendation generation, and greatly improving the recommendation system's ability to understand user interests.

[0054] The interest modeling module constructs a three-dimensional relationship topology of user-content-scene with the help of a heterogeneous graph neural network, comprehensively considering the complex associations between users, content, and scenes. The user nodes, content nodes, and scene nodes are respectively mapped into independent subgraphs, and the hyper-edge weights are calculated by combining the user's historical behaviors and scene contexts to establish bidirectional hyper-edges between the subgraphs. This method can more accurately depict the interactions between different factors. Then, combined with the graph diffusion algorithm, the node embedding representations are iteratively updated until the topological structure converges, enabling the interest model to be continuously optimized as the data changes. At the same time, the dynamic attention mechanism adaptively selects the number of attention heads according to the node type and fuses the outputs of multi-head attention through a gating mechanism to accurately quantify the association strength between nodes, thus more accurately capturing the associations between the user's interests and various contents in different scenarios and providing a more accurate basis for personalized recommendations.

[0055] The recommendation generation module generates candidate recommendation sequences based on a reinforcement learning framework, and uses the Monte Carlo tree search algorithm to evaluate the long-term benefits of the sequences, generating an optimal recommendation path. This approach overcomes the drawback of traditional recommendation systems that only focus on short-term benefits, and considers the impact of recommendations on users from a long-term perspective. By disassembling the recommendation path into a sequence of state-action pairs to construct a virtual decision tree, integrating a Bayesian probability model in the tree nodes to predict the potential behavior branches of users, and then accumulating the expected benefit values of each path through backpropagation, and selecting the path with the smallest benefit variance as the output, it can provide users with more stable and long-term valuable recommended content, improving users' satisfaction with the recommendation results and their willingness to use the recommendation system in the long term.

[0056] The feedback optimization module monitors users' interaction behaviors in real time, and uses a generative adversarial network to simulate negative feedback scenarios, effectively making up for the deficiencies of traditional recommendation systems in feedback processing. The generator network synthesizes virtual negative feedback samples based on users' historical positive feedback data, and the discriminator network outputs an adversarial loss value by comparing the feature distribution differences between real negative feedback and synthetic samples, and injects this loss value reversely into the gradient update process of the recommendation strategy, which can dynamically correct the deviation parameters of the recommendation strategy. At the same time, by establishing a three-level evaluation index for strategy deviation, using a temporal convolutional network to extract the time series features of each level of index, and fusing multi-level features through residual connections to generate a deviation correction amount to update the strategy parameters, the recommendation system can quickly and accurately adjust the recommendation strategy according to users' feedback, continuously optimize the recommendation effect, improve the adaptability and robustness of the recommendation system, and continuously provide users with accurate recommendation services. Brief Description of the Drawings

[0057] Figure 1 is the working principle diagram of the precise targeting AI recommendation system described in the present invention;

[0058] Figure 2 is the flow chart of the Monte Carlo tree search algorithm for evaluating recommendation benefits;

[0059] Figure 3 is the flow chart of the generative adversarial network for simulating negative feedback scenarios and strategy optimization;

[0060] Figure 4 is the flow chart of the dynamic attention mechanism for quantifying the association strength. Detailed Embodiments

[0061] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

[0062] Please refer to Figures 1-4 , the present invention provides a precise targeted AI recommendation system, which mainly includes a data collection module, a feature extraction module, an interest modeling module, a recommendation generation module, and a feedback optimization module. Each module works collaboratively to achieve precise personalized recommendations. The overall implementation solution is as follows:

[0063] The data collection module is responsible for obtaining multi-dimensional behavioral data streams of users. During the data collection process, the data collection frequency and weight are dynamically adjusted. The system assigns differential weights to the behavioral data at different time nodes based on the time decay function, making the recent behavioral data have a greater impact on the recommendation and better reflecting the user's current interests.

[0064] After receiving the behavioral data stream, the feature extraction module uses a multi-modal fusion algorithm to perform hierarchical parsing on it. In this way, the explicit preference features and implicit association features of the user can be extracted, and a feature vector set can be constructed to provide comprehensive and valuable feature information for subsequent recommendations.

[0065] The interest modeling module constructs a three-dimensional relationship topology of user-content-scene by means of a heterogeneous graph neural network. At the same time, combined with the dynamic attention mechanism, the association strength between nodes is quantified, so as to more accurately depict the relationship between the user's interests and different contents and scenes.

[0066] The recommendation generation module generates a candidate recommendation sequence based on the reinforcement learning framework. Then, the Monte Carlo tree search algorithm is used to evaluate the long-term benefits of the sequence, and finally the optimal recommendation path is generated to achieve precise recommendations for users.

[0067] The feedback optimization module monitors the user interaction behavior in real time. Using the generative adversarial network to simulate the negative feedback scenario, and then dynamically correct the deviation parameters of the recommendation strategy, so that the recommendation system can be continuously optimized to improve the accuracy and adaptability of the recommendation.

[0068] The following further illustrates the implementation of the present invention in combination with Embodiments 1 to 6.

[0069] Embodiment 1:

[0070] In this embodiment, the specific process of dynamically adjusting the data collection frequency and weight in the data collection module is mainly described in detail.

[0071] The system first divides the time window according to the user's active period. For example, by analyzing the historical behavior data of a large number of users, it is found that the user activity is relatively high from 19:00 to 22:00 every day, and relatively low in other time periods. Therefore, a day is divided into multiple time windows, divided by the hour, and each hour is a time window.

[0072] Next, calculate the real-time density distribution of behavioral data based on the sliding window mechanism. The sliding window mechanism is like a window sliding on the time series of behavioral data, sliding by a fixed time interval (such as 1 minute) each time. Within each window, count the number of user behavioral data, such as click behavior, browsing behavior, etc. Suppose within a 1-hour time window, there are 50 click behaviors in the first 30 minutes and 30 click behaviors in the last 30 minutes. By calculation, the density distribution of behavioral data within this window can be obtained.

[0073] When the variance of the density distribution within consecutive time windows is detected to exceed the preset threshold, trigger the high-frequency acquisition mode and increase the weight of the current window. The preset threshold is a value preset according to historical data and business requirements. For example, after a large number of experiments and data analyses, it is determined that when the variance of the behavioral data density distribution within 3 consecutive time windows is greater than 10, it is considered that there are significant fluctuations in user behavior and the user may be in an active interest period. At this time, the system will increase the data acquisition frequency. Originally, data was acquired once every 1 minute, and now it may be acquired once every 30 seconds. At the same time, increase the weight of the current window, for example, from the default weight of 1 to 1.5, so that the data within this window has higher importance in subsequent analyses.

[0074] For the data in inactive windows that exceed the set time threshold, use an exponential decay function to reduce its weight until it reaches zero. The set time threshold is also set according to the actual situation. Suppose it is set to 6 hours. If the data within a certain time window is more than 6 hours away from the current time, it is considered an inactive window. The exponential decay function can be expressed as: W(t) = W0 × e -λt where W(t) represents the weight of the data at time t, W0 is the initial weight, λ is the decay coefficient, and t is the duration from the current time. As t increases, W(t) will gradually decrease. When t is large enough, W(t) approaches 0, that is, the influence of these inactive window data on recommendations gradually disappears.

[0075] By such a way of dynamically adjusting the data acquisition frequency and weight, it is possible to collect and utilize user behavioral data more efficiently, providing a more accurate data basis for subsequent recommendations.

[0076] Example 2:

[0077] When processing text data, it is input into a bidirectional Transformer model to extract semantic features. The bidirectional Transformer model has powerful language understanding capabilities. It can encode text both forward and backward simultaneously, capturing the context information in the text. For example, for the keyword "travel guide" searched by the user, the bidirectional Transformer model will encode each word, taking into account the semantic association between "travel" and "guide", as well as their position information in the entire search intention, so as to extract a feature vector containing rich semantic information.

[0078] For image data, convolutional sparse coding is used to extract visual features. Convolutional sparse coding extracts features from images through convolutional operations, and can extract key visual information such as edges and textures in the images. Taking a picture of a tourist attraction as an example, convolutional sparse coding can extract features such as the outline and color distribution of the scenic spot in the picture, and convert these features into a vector representation in numerical form.

[0079] For time-series behavior data, a long short-term memory network (LSTM) is used to extract periodic pattern features. LSTM can effectively process time-series data and remember long-term dependencies. For example, if a user browses travel-related content multiple times within a period of time, and the browsing time intervals show a certain pattern, LSTM can capture this periodic pattern and judge whether the user's interest in travel is a short-term concentrated focus or a long-term continuous focus.

[0080] After the features of text, image, and time-series behavior data are extracted respectively, the cross-modal attention layer calculates the correlation matrix between different features. The cross-modal attention layer will calculate the correlation degree between text features and image features, text features and time-series behavior features, and image features and time-series behavior features. For example, it is found that there is a certain correlation between the travel keywords searched by the user and the travel pictures browsed, and the correlation matrix is obtained by calculating this correlation degree. Finally, a unified feature vector is generated according to the correlation matrix. The features of different modalities are weighted and fused according to the correlation degree, so that the finally generated feature vector contains both the semantic information of the text, the visual information of the image, and the periodic pattern information of the time-series behavior, providing more comprehensive and representative features for subsequent recommendations.

[0081] Example 3:

[0082] This example mainly describes the specific implementation method of constructing a three-dimensional relationship topology of user-content-scene and quantifying the association strength by a dynamic attention mechanism in the interest modeling module.

[0083] When constructing the three-dimensional relationship topology of user-content-scene, first map the user nodes, content nodes, and scene nodes into three independent subgraphs of the heterogeneous graph respectively. For example, each user corresponds to a user node, each piece of content (such as a travel article, a travel video) corresponds to a content node, and each scene (such as travel season, travel destination) corresponds to a scene node. These nodes in their respective subgraphs are represented by their attributes and the connection relationships between them.

[0084] Then establish bidirectional hyperedges between the subgraphs. The hyperedge weights are calculated jointly by the user's historical behavior and the scene context. Suppose a user has browsed a large amount of content about seaside travel in summer. Then a hyperedge will be established between the user node and the scene node representing summer and the content node of seaside travel. The calculation of the hyperedge weight will comprehensively consider historical behaviors such as the frequency and stay time of the user browsing this content, as well as relevant information about the summer scene, such as that summer is the peak tourist season and seaside travel is more popular in summer. Through such a calculation method, the hyperedge weight can more accurately reflect the degree of association between users, content, and scenes.

[0085] Then iteratively update the node embedding representations through the graph diffusion algorithm until the topological structure converges. The graph diffusion algorithm will spread information in the heterogeneous graph and continuously update the embedding representations of the nodes. In each iteration, the node will update its own embedding vector according to the information of the nodes connected to it, so that the representation of the node can better reflect its position in the entire graph and its relationship with other nodes. After multiple iterations, when the change in the node embedding representation is very small and meets certain convergence conditions, the topological structure converges.

[0086] In terms of quantifying the association strength by the dynamic attention mechanism, first define the multi-head attention calculation unit between nodes in the heterogeneous graph. The multi-head attention mechanism can capture the relationships between nodes from different perspectives. The number of attention heads is adaptively selected according to the node type. Among them, four heads of attention are used for user-content association, and two heads of attention are used for content-scene association. Taking the user-content association as an example, the four heads of attention respectively focus on the associations between different behavioral characteristics of the user (such as browsing behavior, collection behavior, comment behavior) and the content, and calculate the degree of attention of the user to the content from multiple dimensions. While for the content-scene association, two heads of attention are used, which respectively focus on the matching degree between the theme and style of the content and the scene.

[0087] Finally, the gating mechanism fuses the outputs of multiple attention heads to generate the final correlation strength coefficient. The gating mechanism dynamically adjusts the weights of the attention heads based on their outputs, fusing the results of multiple heads. For example, when a user expresses strong interest in a certain type of content in a specific scenario, the gating mechanism increases the weight of the corresponding attention head. This ensures that the resulting correlation strength coefficient more accurately reflects the true strength of the correlation between the user, content, and scenario, providing a more precise basis for interest modeling in subsequent recommendations.

[0088] Example 4:

[0089] This embodiment details the specific steps of evaluating long-term benefits using the Monte Carlo tree search algorithm in the recommendation generation module.

[0090] First, the recommendation path is broken down into a sequence of state-action pairs to construct a virtual decision tree. The recommendation path can be understood as the series of decision-making processes by which the system recommends content to users. For example, the system recommends a travel guide article, and the user clicks on it. This constitutes a state-action pair (the state represents the user's current interest, and the action is recommending the article). Arranging these state-action pairs throughout the recommendation process sequentially forms the recommendation path. A virtual decision tree is then constructed based on these pairs. Each node in the tree represents a state, and each edge from one node to another represents an action.

[0091] A Bayesian probability model is integrated into tree nodes to predict potential user behavior branches. Based on historical data and current status, the Bayesian probability model can calculate the probability of different possible user behaviors in response to a recommendation. For example, when a travel guide is recommended, the Bayesian probability model can be used to predict the probability of the user continuing to browse related articles, adding the current article to favorites, or leaving the current page based on the user's previous browsing and saving of similar guides, as well as the behavioral data of other users with similar interests.

[0092] Cumulatively calculate the expected revenue values of each path through backpropagation, and select the path with the smallest revenue variance as the output. After the virtual decision tree is constructed and the potential behavior branches of the user are predicted, starting from the leaf nodes of the tree, calculate the expected revenue values of each path through backpropagation according to the expected revenue and probability of each action. The expected revenue value can be defined according to the business objective. For example, the number of user clicks, the stay time, the purchase conversion rate, etc. of the user can all be used as the measurement indicators of the revenue. Suppose a recommended path finally brings a large number of user clicks and a long stay time, then the expected revenue value of this path is relatively high. After calculating the expected revenue values of all paths, select the path with the smallest revenue variance as the final recommended path. The revenue variance reflects the stability of the path revenue. If the variance is small, it means that the revenue of this path is relatively stable and is more suitable as a recommended path, which can provide more reliable recommendation results for users.

[0093] Example 5:

[0094] In terms of the adversarial generative network simulating the negative feedback scenario, the generator network synthesizes virtual negative feedback samples based on the user's historical positive feedback data. For example, when the user browses travel-related content and gives positive feedback behaviors such as liking and collecting some recommended articles about tourist attractions. The generator network will synthesize some virtual negative feedback samples according to this positive feedback data by learning the characteristics and distribution laws of the data. These samples may indicate that the user is not interested in certain travel content, such as generating virtual behavior data that the user quickly leaves the page after browsing or gives negative evaluations to the recommended content.

[0095] The discriminator network outputs the adversarial loss value by comparing the feature distribution differences between the real negative feedback and the synthesized samples. The real negative feedback data is the user's uninterested behavior data actually collected by the system. The discriminator network will extract and analyze the features of the real negative feedback data and the virtual negative feedback samples synthesized by the generator, and compare their feature distribution situations. If the feature distribution of the synthesized samples is quite different from the real negative feedback data, it means that the samples generated by the generator are not real enough, and the discriminator will output a relatively large adversarial loss value.

[0096] Inject the adversarial loss value reversely into the gradient update process of the recommendation strategy to optimize the robustness of the strategy. The adversarial loss value will be used to adjust the parameters of the recommendation strategy. During the training process of the recommendation strategy, through the backpropagation algorithm, the adversarial loss value is passed to the model parameters of the recommendation strategy, so that the model can adjust the parameters according to this loss value to generate more recommendation results that conform to the user's real feedback, thereby improving the robustness of the recommendation strategy and enabling it to better handle various user feedback situations.

[0097] When correcting the deviation parameters of the recommendation strategy, first establish a three-level evaluation index for strategy deviation, including short-term deviation, medium-term stability, and long-term trend. The short-term deviation is used to measure the deviation degree between the recommendation result and the user's actual interest in the short term. For example, the matching degree between the content recommended within a day and the user's click and browsing behaviors. The medium-term stability focuses on the stability degree of the recommendation strategy within a certain period (such as a week), whether there will be frequent fluctuations. The long-term trend is to examine whether the recommendation strategy can keep up with the changing trend of the user's interest in the long term (such as a month).

[0098] Then, use a temporal convolutional network to extract the time series features of each level of indicators respectively. The temporal convolutional network can effectively process time series data and extract the time features in the data. For the time series data of short-term deviation, the temporal convolutional network can capture the changing trend of the recommendation deviation in the short term; for the time series data of medium-term stability and long-term trend, the corresponding features can also be extracted.

[0099] Finally, fuse multi-level features through residual connections to generate a deviation correction amount and update the strategy parameters. Residual connections can avoid the loss of information during the feature fusion process and effectively fuse features at different levels. The fused features are calculated to generate a deviation correction amount, which is used to update the parameters of the recommendation strategy, enabling the recommendation strategy to be continuously optimized, reduce deviations, and improve the accuracy of recommendations.

[0100] Example 6:

[0101] This example details the specific implementation method of hierarchically parsing the behavior data stream and constructing a feature vector set in the feature extraction module.

[0102] When hierarchically parsing the behavior data stream, first divide the original data stream into a basic operation layer, a semantic understanding layer, and an intention reasoning layer. The original data stream contains various behavior data of users, such as click, slide, search and other operations.

[0103] Identify click, stay, and favorite behavior patterns in the basic operation layer. For example, the system can record which travel attraction introduction pages the user clicks on when browsing a travel website, the time the user stays on each page, and whether the user favorites certain pages. By statistically analyzing these behaviors, the user's interest tendency can be initially understood. If the user frequently clicks on the pages of seaside travel attractions, stays for a long time, and also performs a favorite operation, it indicates that the user may have a high interest in seaside travel.

[0104] Parse the co-occurrence relationship between search keywords and content tags at the semantic understanding layer. When the user searches for "Sanya travel guide", the system analyzes the co-occurrence of this keyword with travel content tags on the website (such as "Sanya attractions", "beach vacation", "food recommendation", etc.). If the search keyword frequently co-occurs with the tags "Sanya attractions" and "beach vacation", it indicates that the user may be more interested in Sanya's beach tourism-related content.

[0105] Construct a probability transition matrix of the behavior chain at the intention inference layer to identify potential intention evolution paths. The behavior chain is a sequence of a user's series of behaviors. By analyzing the user's behavior chain, a probability transition matrix is constructed. For example, the user first searches for "Sanya travel guide", then clicks on the "Sanya hotel recommendation" page, and then browses the "Sanya food" page. Through a large amount of such behavior chain data, a probability transition matrix is constructed to calculate the transition probability from one behavior to another. Based on this matrix, the user's possible next behavior can be predicted, potential intention evolution paths can be identified, and the user's interests can be further explored.

[0106] In terms of constructing a set of feature vectors, perform principal component analysis dimensionality reduction on the explicit preference features, and retain the principal components with a variance contribution rate exceeding 85%. The explicit preference features can be the interest features explicitly shown by the user, such as the content categories collected and the types of pages frequently browsed. Principal component analysis can perform dimensionality reduction on these high-dimensional feature data, remove redundant information, and retain most of the key information at the same time. For example, if there are originally 10 explicit preference features, through principal component analysis, only 3-4 principal components may be retained, and these principal components can explain more than 85% of the variance of the original data, reducing the data dimension without losing too much important information.

[0107] Use an autoencoder to perform non-linear compression on the implicit association features to generate low-dimensional dense vectors. The implicit association features are potential association features mined by analyzing the user's behavior data. The autoencoder can learn the internal structure of these features and compress the high-dimensional implicit association features into low-dimensional dense vectors. For example, through the autoencoder, the complex association relationships between the user's browsing behaviors are learned and compressed to generate more concise and representative low-dimensional vectors.

[0108] Finally, the interaction weights between the explicit features and the implicit features are calculated through the feature crossing layer, and the final feature vector is generated by combination. The feature crossing layer will consider the interaction between the explicit features and the implicit features and calculate their interaction weights. For example, if the explicit feature represents the user's preference for seaside tourism and the implicit feature represents the user's attention pattern to tourism content during a specific period, the interaction weights between the two are calculated through the feature crossing layer, and the explicit feature and the implicit feature are combined according to this weight to generate the final feature vector. This feature vector synthesizes the explicit and implicit interest features of the user, can more comprehensively reflect the user's interests, and provides a more accurate feature basis for subsequent recommendations.

[0109] It should be noted that in this document, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0110] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A precise targeted AI recommendation system, characterized in that, Including: A data acquisition module, which acquires the multi-dimensional behavior data stream of users, dynamically adjusts the data acquisition frequency and weight, and assigns differential weights to the behavior data at different time nodes based on a time decay function; A feature extraction module, which uses a multi-modal fusion algorithm to hierarchically analyze the behavior data stream, extracts the explicit preference features and implicit association features of users, and constructs a feature vector set; An interest modeling module, which constructs a three-dimensional relationship topology of user-content-scene through a heterogeneous graph neural network, and combines a dynamic attention mechanism to quantify the association strength between nodes; A recommendation generation module, which generates a candidate recommendation sequence based on a reinforcement learning framework, and evaluates the long-term benefits of the sequence through a Monte Carlo tree search algorithm and generates an optimal recommendation path; A feedback optimization module, which monitors the user interaction behavior in real time, uses a generative adversarial network to simulate a negative feedback scenario, and dynamically corrects the deviation parameters of the recommendation strategy.

2. The precise targeting AI recommendation system according to claim 1, characterized in that, The dynamic adjustment of the data acquisition frequency and weight includes the following steps: Dividing time windows according to the user's active period, and calculating the real-time density distribution of behavior data based on a sliding window mechanism; When the variance of the density distribution within consecutive time windows is detected to exceed a preset threshold, triggering a high-frequency acquisition mode and increasing the weight of the current window; For the data of inactive windows that exceed the set time threshold, using an exponential decay function to reduce its weight until it reaches zero.

3. The precise targeting type AI recommendation system according to claim 2, wherein The multi-modal fusion algorithm includes the following steps: Inputting text data into a bidirectional Transformer model to extract semantic features, and using convolutional sparse coding for image data to extract visual features; Using a long short-term memory network for time-series behavior data to extract periodic pattern features; Calculating the association matrix between different features through a cross-modal attention layer, and fusing to generate a unified feature vector.

4. The precise targeting AI recommendation system according to claim 3, characterized in that, The construction of the three-dimensional relationship topology of user-content-scene includes the following steps: Mapping user nodes, content nodes, and scene nodes into three independent subgraphs of a heterogeneous graph respectively; Establishing bidirectional hyperedges between subgraphs, and the hyperedge weights are jointly calculated by the user's historical behavior and scene context; Iteratively updating the node embedding representation through a graph diffusion algorithm until the topological structure converges.

5. The precise targeting AI recommendation system according to claim 4, wherein The Monte Carlo tree search algorithm for evaluating long-term benefits includes the following steps: Decomposing the recommendation path into a sequence of state-action pairs and constructing a virtual decision tree; Integrating a Bayesian probability model in the tree nodes to predict the potential behavior branches of users; Accumulating the expected benefit values of each path through backpropagation, and selecting the path with the smallest benefit variance as the output.

6. The precise targeting type AI recommendation system according to claim 5, wherein, The generative adversarial network for simulating a negative feedback scenario includes the following steps: The generator network synthesizes virtual negative feedback samples based on the user's historical positive feedback data; The discriminator network outputs an adversarial loss value by comparing the feature distribution differences between real negative feedback and synthetic samples; Injecting the adversarial loss value backward into the gradient update process of the recommendation strategy to optimize the robustness of the strategy.

7. The precise targeting AI recommendation system according to claim 6, characterized in that, The dynamic attention mechanism for quantifying the association strength includes the following steps: Defining a multi-head attention calculation unit between nodes in a heterogeneous graph; Adaptive selection of the number of attention heads according to the node type, where four attention heads are used for user-content association and two attention heads are used for content-scenario association; Fuse the outputs of multi-head attention through a gating mechanism to generate the final association strength coefficient.

8. The precise targeting AI recommendation system according to claim 7, wherein The hierarchical parsing of the behavioral data stream includes the following steps: Divide the original data stream into a basic operation layer, a semantic understanding layer, and an intention reasoning layer; Identify click, stay, and favorite behavior patterns in the basic operation layer; Parse the co-occurrence relationship between search keywords and content tags in the semantic understanding layer; Construct a probability transition matrix of the behavior chain in the intention reasoning layer to identify potential intention evolution paths.

9. The precise targeting type AI recommendation system according to claim 8, characterized in that, The steps for correcting the deviation parameters of the recommendation strategy include the following: Establish a three-level evaluation index for policy deviation, including short-term deviation, medium-term stability, and long-term trend; Use a temporal convolutional network to extract the time series features of each level of index respectively; Fuse multi-level features through residual connections to generate a deviation correction amount and update the policy parameters.

10. The precise targeting AI recommendation system according to claim 9, wherein The steps for constructing the feature vector set include the following: Perform principal component analysis dimensionality reduction on the explicit preference features, and retain the principal components with a variance contribution rate exceeding 85%; Use an autoencoder to perform non-linear compression on the implicit association features to generate low-dimensional dense vectors; Calculate the interaction weights of the explicit features and the implicit features through the feature crossing layer, and combine them to generate the final feature vector.

Images