An artificial intelligence-based digital economy platform intelligent recommendation optimization method

Abstract

The application discloses an intelligent recommendation optimization method of a digital economic platform based on artificial intelligence, and comprises the following steps: S1, acquiring user behavior and content description data, pre-processing the user behavior data, and constructing a user behavior sequence; S2, constructing a user feature set based on the user behavior sequence through a bidirectional long short-term memory network, and constructing a content feature set based on the content description data; S3, constructing a recommendation state set according to the user and the content feature sets through a twin neural network; S4, constructing a DreamerV3 world model, and generating a recommendation ranking sequence according to the recommendation state set; S5, applying the recommendation ranking sequence, and generating a feedback result; and S6, updating the DreamerV3 world model and the bidirectional long short-term memory network parameters based on the feedback result. The application proposes an intelligent recommendation optimization method fusing deep learning and reinforcement learning, and realizes high-precision modeling and strategy closed-loop optimization based on user feedback.

Description

Technical Field

[0001] This invention relates to the field of digital economy platform technology, and in particular to an intelligent recommendation optimization method for digital economy platforms based on artificial intelligence. Background Technology

[0002] With the rapid development of the digital economy, online content consumption and personalized recommendations have become key means to improve user engagement and platform operational efficiency. Existing recommendation systems widely employ collaborative filtering, content matching, or shallow neural network models to model user historical behavior and content features, thereby generating recommendation results. However, these methods generally suffer from insufficient dimensions in user behavior modeling, limited state abstraction capabilities, and inadequate feedback mechanism updates, making it difficult to support the real-time recommendation accuracy requirements of highly dynamic and multimodal digital economy scenarios.

[0003] Current research attempts to incorporate recurrent neural networks, convolutional neural networks, or Transformer structures to enhance the model's ability to represent temporal behavior and content semantics. For example, Bidirectional Long Short-Term Memory (BiLSTM) networks are used to extract contextual dependencies in user behavior sequences, and convolutional neural networks (CNNs) are used to obtain local features of content text. This has improved the recommendation model's ability to perceive the evolution of user interests to some extent. However, in modeling the feature association between users and content, there is still a lack of structured, multi-scale measurement mechanisms, resulting in inaccurate representation of the recommendation state space and affecting the effectiveness of policy decisions.

[0004] Furthermore, most existing recommender systems rely solely on static ranking scores for recommendation decisions, lacking effective policy optimization processes and value assessment mechanisms, making it difficult to form a closed-loop optimization structure between behavioral feedback and model adjustment. While some reinforcement learning methods have been used for recommender policy learning, they often neglect the accuracy of state modeling, the hierarchical differences in behavioral feedback, and the simultaneous optimization design of the policy network and value network, resulting in poor model stability and low feedback utilization.

[0005] Especially in the feedback utilization phase of the recommendation system, there are significant differences and delays between user click behavior, browsing path, dwell time and conversion behavior. Using only a single dimension of click signal as the optimization target cannot fully reflect the true value of the recommendation effect and has limitations in improving long-term user interest and platform revenue.

[0006] Therefore, how to provide an intelligent recommendation optimization method for digital economy platforms based on artificial intelligence is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose an intelligent recommendation optimization method for digital economy platforms based on artificial intelligence. This invention comprehensively utilizes bidirectional long short-term memory networks, Siamese neural networks, and the DreamerV3 world model, integrating user behavior modeling, content feature extraction, and reinforcement learning strategy optimization mechanisms. It constructs a detailed algorithm flow that enables dynamic recommendation state generation, recommendation strategy iteration, and closed-loop training of behavioral feedback, possessing the advantages of high modeling accuracy, high recommendation accuracy, and significant user response optimization effects.

[0008] An intelligent recommendation optimization method for a digital economy platform based on artificial intelligence, according to an embodiment of the present invention, includes the following steps: S1. Obtain user behavior data and content description data recorded in chronological order, preprocess the user behavior data, and construct a user behavior sequence; S2. Perform bidirectional state extraction and fusion operations on user behavior sequences through a bidirectional long short-term memory network to construct a user feature set. At the same time, perform word vector embedding, one-dimensional convolution and pooling operations on content description data to construct a content feature set. S3. The user feature set and content feature set are shared and standardized through a Siamese neural network. Euclidean distance and Mahalanobis distance are calculated based on the standardization results. Adjacency features of the two sets are extracted according to time steps. Recommendation state set is constructed based on the calculation results and extraction results. S4. Construct the DreamerV3 world model. Through the policy network and value network of this model, perform forward inference operations and state value estimation on the set of recommended states respectively to generate a recommendation ranking sequence. At the same time, update the policy network parameters based on the estimation results. S5. Use the recommended ranking sequence in the actual recommendation process, record the user's click behavior, browsing path, dwell time and conversion identifier, and generate feedback results in chronological order; S6. Based on the feedback results, the parameters of the DreamerV3 world model and the bidirectional long short-term memory network are updated using the policy gradient algorithm.

[0009] Optionally, the user behavior data includes click records, browsing paths, dwell time, and conversion identifiers. The conversion identifiers indicate whether the user has achieved their target behavior after receiving recommended content, including whether or not they have placed an order, submitted a form, shared content, or saved the content. The content description data includes structured or unstructured information corresponding to the recommendable content in the digital economy platform. The preprocessing includes missing information completion, format unification, and time alignment operations. The adjacency feature represents the local correspondence between the user feature set and the content feature set under time alignment.

[0010] Optionally, S2 specifically includes: S21. Arrange the user behavior segments at each time step in the user behavior sequence into behavior vectors according to a preset index order, and combine all behavior vectors into a time series matrix. S22. Perform bidirectional state extraction and fusion operations through a bidirectional long short-term memory network. Based on the time series matrix, update the forward hidden state sequentially along the forward time direction and update the backward hidden state sequentially along the reverse time direction to construct the forward state sequence and the backward state sequence. S23. Connect the two sequences by time step, and perform dimension normalization, scale correction and ReLU function activation on the connection result to construct a user feature set; S24. Perform a word vector lookup operation on each term in the content description data according to the preset embedding rule table, arrange the lookup results in the order of the content description data to generate an embedding matrix, and perform a one-dimensional convolution operation on the embedding matrix using a fixed-length sliding window to calculate the convolution response of the local region and form a convolution response matrix. S25. Set a pooling window of fixed size on the convolution response matrix, perform a maximum value selection operation on each window region, and combine all selection results into a content feature set.

[0011] Optionally, S22 specifically includes: S221. The forward hidden state is updated sequentially along the forward direction of time through a bidirectional long short-term memory network. The behavior vector of each time step in the time series matrix is ​​concatenated with the forward hidden state of the previous time step to construct a forward joint vector. In the first time step, the forward hidden state of the previous time step is initialized to a zero vector. S222. Perform forget gate transformation, update gate transformation and output gate transformation operations on the forward joint vector respectively, and use the Sigmoid function to activate the transformation results to generate forget gate vector, update gate vector and output gate vector. S223. Perform a linear transformation operation on the forward joint vector, and use the Tanh function to perform element-wise activation processing on the transformation result to generate candidate memory vectors. S224. Perform element-wise multiplication on the candidate memory vector and the update gate vector to generate the memory update vector. Perform element-wise multiplication on the forward memory vector of the previous time step and the forget gate vector to generate the memory retention vector. Perform element-wise addition on the memory update vector to generate the forward memory vector of the current time step. In the first time step, use the preset initialization vector as the forward memory vector of the previous time step. S225. Perform element-wise multiplication between the forward memory vector and the output gate vector to form the forward hidden state at the current time step; S226. The backward hidden state is updated sequentially in reverse time through a bidirectional long short-term memory network. At each time step, the current action vector is concatenated with the backward hidden state of the next time step, and the forget gate transformation, update gate transformation, output gate transformation, Tanh function activation, element-wise multiplication and addition operations are performed to generate the backward hidden state corresponding to each time step. At the last time step, the backward hidden state of the next time step is initialized to a zero vector. S227. Combine all forward hidden states and backward hidden states to form forward state sequences and backward state sequences respectively.

[0012] Optionally, S3 specifically includes: S31. Input the user feature set and the content feature set into two network branches with shared weights in the Siamese neural network, respectively, and perform feature mapping and embedding transformation operations to obtain the shared encoding result of the two sets; S32. Adjust the mean and scale of each encoding vector in the two shared encoding results according to the element dimension to obtain the user encoding set and the content encoding set. S33. Based on the user coding set and the content coding set, calculate the Euclidean distance and Mahalanobis distance between the user coding vector and the content coding vector at the corresponding time step, and generate the Euclidean distance sequence and the Mahalanobis distance sequence. S34. Align the user feature set and the content feature set in time, and extract the user feature vector and the content feature vector at each time step to form an adjacent feature sequence. S35. Concatenate the Euclidean distance sequence, the Mahalanobis distance sequence, and the adjacency feature sequence to construct a set of recommended states.

[0013] Optionally, the Siamese neural network includes two network branches that share weight matrices and bias parameters. Each network branch includes an embedding layer and two fully connected layers. The processing procedure of this network specifically includes: The user feature set and the content feature set are respectively input into two network branches and processed sequentially according to time steps; In the first fully connected layer, matrix multiplication and bias addition operations are performed on the input user feature vector or content feature vector, and ReLU function is used for activation. The activation result is fed into the second fully connected layer, where matrix multiplication, bias addition, activation and batch normalization operations are performed to generate standardized features. In the embedding layer, a dimension mapping operation is performed on the standardized features. The standardized features are projected using a preset linear weight matrix, and the projection result is activated by the GELU function to generate an embedding vector. The user encoding set and the content encoding set are formed based on the embedding vectors output by the two network branches at each time step.

[0014] Optionally, S33 specifically includes: S331. Perform time pairing on the user coding set and the content coding set, and extract the user coding vector and content coding vector corresponding to each time step as a coding pair. S332. Perform Euclidean distance calculation operation on each encoding pair, calculate the Euclidean distance between the two vectors by summing the square differences and then taking the square root, and generate an Euclidean distance sequence in time order; S333. Perform Mahalanobis distance calculation for each encoding pair, construct a covariance matrix based on all encoding pairs, calculate the Mahalanobis distance of the encoding pairs based on the covariance matrix, and generate a Mahalanobis distance sequence in chronological order.

[0015] Optionally, the Mahalanobis distance calculation process specifically includes: Extract the vector differences of all encoded pairs and generate a set of difference vectors according to time steps; The mean and deviation of the difference vector set are calculated in each dimension to construct the covariance matrix, and the inverse matrix is ​​obtained by performing the inverse matrix operation on the covariance matrix. The difference vector for this time step is constructed by subtracting the corresponding elements from the user encoding vector and the content encoding vector in the encoding pair. The difference vector is taken as a row matrix input. Each element of the matrix input in the column direction is rearranged according to its index to the row direction position of the new matrix to obtain the transposed difference vector. Then, matrix multiplication is performed with the covariance inverse matrix to obtain the intermediate matrix result. Perform a matrix multiplication operation between the intermediate matrix result and the difference vector to generate the Mahalanobis distance for that time step; Arrange all Mahalanobis distance values ​​in chronological order to form a Mahalanobis distance sequence.

[0016] Optionally, S4 specifically includes: S41. Construct the DreamerV3 world model, perform forward inference operations on the set of recommended states in the policy network of the world model, generate action feature vectors corresponding to each time step, and arrange them into a recommendation ranking sequence in chronological order. S42. Through the value network of the world model, perform state encoding and score prediction operations on the recommended state set at time steps to generate state score vectors at corresponding time steps, and combine all state score vectors in time order to form a state value sequence. S43. Align the recommended ranking sequence with the state value sequence in time. Calculate the policy loss based on the difference between the state score vector and the action feature vector at each time step in the alignment result. Update the policy network parameters using the gradient backpropagation algorithm.

[0017] Optionally, the execution process within the DreamerV3 world model specifically includes: In the policy network, a vector dimension mapping operation is performed on the recommended state vector at each time step to transform the recommended state vector into a policy representation vector with fixed dimensions, and batch normalization and ReLU function activation are performed on the policy representation vector. The activated policy representation vector is subjected to two rounds of Dropout processing and linear transformation operations. After the first round of transformation, layer normalization and Leaky ReLU function activation are performed. After the second round of transformation, a residual connection structure is introduced to perform dimensionality compression processing, generating the action feature vector for the corresponding time step. Arrange all action feature vectors in chronological order to construct a recommended ranking sequence; In the value network, embedding mapping and Tanh function activation are performed on the recommended state vector at each time step, and linear mapping, Z-score standardization and numerical compression are performed on the activation results in sequence to generate the state score vector for the corresponding time step. Combine all state score vectors in chronological order to construct a state value sequence; Align the recommended ranking sequence with the state value sequence along the time dimension to construct a paired sequence; Perform time-step error calculation on the paired sequences, calculate the policy difference value between action features and state scores at each time step, perform loss aggregation on all policy difference values ​​by time step, and perform backpropagation based on the aggregation result to update the policy network parameters.

[0018] The beneficial effects of this invention are: First, by introducing a bidirectional long short-term memory network to perform forward and reverse temporal modeling operations on user behavior sequences, this invention can effectively capture the historical behavioral characteristics and evolution trends of users at different time scales, thereby improving the completeness and continuity of user interest expression, and possessing stronger time-dependent modeling capabilities and context-related capture capabilities, laying a high-quality user feature foundation for the subsequent construction of recommendation states.

[0019] Secondly, this invention uses a Siamese neural network to perform shared encoding and similarity measurement operations on user feature sets and content feature sets, and introduces Euclidean distance and Mahalanobis distance to construct a multi-perspective representation of recommendation states. This enables the model to not only mine the shallow matching relationship between users and content, but also to analyze the deep distribution differences based on the covariance structure, effectively improving the discriminativeness and generalization ability of recommendation states. At the same time, the temporal alignment mechanism of adjacent features enhances the expression of interaction features between behavior and content, and improves the information distortion problem in recommendation state modeling.

[0020] Finally, this invention introduces the DreamerV3 world model to generate policies and evaluate values ​​for the recommendation state set, constructs a recommendation ranking sequence and combines it with user behavior feedback results, and uses a policy gradient algorithm to achieve joint training and dynamic updating of the policy network and the temporal modeling network, forming a closed-loop mechanism of recommendation action—user response—policy optimization. While improving recommendation accuracy, it effectively enhances the model's focus on long-term user value and policy stability, demonstrating higher recommendation response capability, user conversion capability and system adaptability, and has significant application value and promotion prospects. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Fig. 1 This is a flowchart of an intelligent recommendation optimization method for a digital economy platform based on artificial intelligence, as proposed in this invention. Fig. 2 This is a schematic diagram illustrating the user and content feature modeling process of an AI-based intelligent recommendation optimization method for digital economy platforms proposed in this invention. Fig. 3 This is a flowchart of the DreamerV3 strategy optimization and feedback learning structure for an intelligent recommendation optimization method for a digital economy platform based on artificial intelligence, as proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figs. 1-3 An intelligent recommendation optimization method for digital economy platforms based on artificial intelligence includes the following steps: S1. Obtain user behavior data and content description data recorded in chronological order, preprocess the user behavior data, and construct a user behavior sequence; S2. Perform bidirectional state extraction and fusion operations on user behavior sequences through a bidirectional long short-term memory network to construct a user feature set. At the same time, perform word vector embedding, one-dimensional convolution and pooling operations on content description data to construct a content feature set. S3. The user feature set and content feature set are shared and standardized through a Siamese neural network. Euclidean distance and Mahalanobis distance are calculated based on the standardization results. Adjacency features of the two sets are extracted according to time steps. Recommendation state set is constructed based on the calculation results and extraction results. S4. Construct the DreamerV3 world model. Through the policy network and value network of this model, perform forward inference operations and state value estimation on the set of recommended states respectively to generate a recommendation ranking sequence. At the same time, update the policy network parameters based on the estimation results. S5. Use the recommended ranking sequence in the actual recommendation process, record the user's click behavior, browsing path, dwell time and conversion identifier, and generate feedback results in chronological order; S6. Based on the feedback results, the parameters of the DreamerV3 world model and the bidirectional long short-term memory network are updated using the policy gradient algorithm.

[0024] In this embodiment, the user behavior data includes click records, browsing paths, dwell time, and conversion identifiers. The conversion identifiers indicate the achievement of the user's target behavior after receiving recommended content, including whether or not order placement, form submission, content sharing, or collection behavior has occurred. The content description data includes structured or unstructured information corresponding to the recommendable content in the digital economy platform. The preprocessing includes missing information completion, format unification, and time alignment operations. The adjacency feature represents the local correspondence between the user feature set and the content feature set under time alignment.

[0025] In this embodiment, S2 specifically includes: S21. Arrange the user behavior segments at each time step in the user behavior sequence into behavior vectors according to a preset index order, and combine all behavior vectors into a time series matrix. S22. Perform bidirectional state extraction and fusion operations through a bidirectional long short-term memory network. Based on the time series matrix, update the forward hidden state sequentially along the forward time direction and update the backward hidden state sequentially along the reverse time direction to construct the forward state sequence and the backward state sequence. S23. Connect the two sequences by time step, and perform dimension normalization, scale correction and ReLU function activation on the connection result to construct a user feature set; S24. Perform a word vector lookup operation on each term in the content description data according to the preset embedding rule table, arrange the lookup results in the order of the content description data to generate an embedding matrix, and perform a one-dimensional convolution operation on the embedding matrix using a fixed-length sliding window to calculate the convolution response of the local region and form a convolution response matrix. S25. Set a pooling window of fixed size on the convolution response matrix, perform a maximum value selection operation on each window region, and combine all selection results into a content feature set.

[0026] In this embodiment, S22 specifically includes: S221. The forward hidden state is updated sequentially along the forward direction of time through a bidirectional long short-term memory network. The behavior vector of each time step in the time series matrix is ​​concatenated with the forward hidden state of the previous time step to construct a forward joint vector. In the first time step, the forward hidden state of the previous time step is initialized to a zero vector. S222. Perform forget gate transformation, update gate transformation and output gate transformation operations on the forward joint vector respectively, and use the Sigmoid function to activate the transformation results to generate forget gate vector, update gate vector and output gate vector. S223. Perform a linear transformation operation on the forward joint vector, and use the Tanh function to perform element-wise activation processing on the transformation result to generate candidate memory vectors. S224. Perform element-wise multiplication on the candidate memory vector and the update gate vector to generate the memory update vector. Perform element-wise multiplication on the forward memory vector of the previous time step and the forget gate vector to generate the memory retention vector. Perform element-wise addition on the memory update vector to generate the forward memory vector of the current time step. In the first time step, use the preset initialization vector as the forward memory vector of the previous time step. S225. Perform element-wise multiplication between the forward memory vector and the output gate vector to form the forward hidden state at the current time step; S226. The backward hidden state is updated sequentially in reverse time through a bidirectional long short-term memory network. At each time step, the current action vector is concatenated with the backward hidden state of the next time step, and the forget gate transformation, update gate transformation, output gate transformation, Tanh function activation, element-wise multiplication and addition operations are performed to generate the backward hidden state corresponding to each time step. At the last time step, the backward hidden state of the next time step is initialized to a zero vector. S227. Combine all forward hidden states and backward hidden states to form forward state sequences and backward state sequences respectively.

[0027] In this embodiment, S3 specifically includes: S31. Input the user feature set and the content feature set into two network branches with shared weights in the Siamese neural network, respectively, and perform feature mapping and embedding transformation operations to obtain the shared encoding result of the two sets; S32. Adjust the mean and scale of each encoding vector in the two shared encoding results according to the element dimension to obtain the user encoding set and the content encoding set. S33. Based on the user coding set and the content coding set, calculate the Euclidean distance and Mahalanobis distance between the user coding vector and the content coding vector at the corresponding time step, and generate the Euclidean distance sequence and the Mahalanobis distance sequence. S34. Align the user feature set and the content feature set in time, and extract the user feature vector and the content feature vector at each time step to form an adjacent feature sequence. S35. Concatenate the Euclidean distance sequence, the Mahalanobis distance sequence, and the adjacency feature sequence to construct a set of recommended states.

[0028] In this embodiment, the Siamese neural network includes two network branches that share weight matrices and bias parameters. Each network branch includes an embedding layer and two fully connected layers. The processing of this network specifically includes: The user feature set and the content feature set are respectively input into two network branches and processed sequentially according to time steps; In the first fully connected layer, matrix multiplication and bias addition operations are performed on the input user feature vector or content feature vector, and ReLU function is used for activation. The activation result is fed into the second fully connected layer, where matrix multiplication, bias addition, activation and batch normalization operations are performed to generate standardized features. In the embedding layer, a dimension mapping operation is performed on the standardized features. The standardized features are projected using a preset linear weight matrix, and the projection result is activated by the GELU function to generate an embedding vector. The user encoding set and the content encoding set are formed based on the embedding vectors output by the two network branches at each time step.

[0029] In this embodiment, S33 specifically includes: S331. Perform time pairing on the user coding set and the content coding set, and extract the user coding vector and content coding vector corresponding to each time step as a coding pair. S332. Perform Euclidean distance calculation operation on each encoding pair, calculate the Euclidean distance between the two vectors by summing the square differences and then taking the square root, and generate an Euclidean distance sequence in time order; S333. Perform Mahalanobis distance calculation for each encoding pair, construct a covariance matrix based on all encoding pairs, calculate the Mahalanobis distance of the encoding pairs based on the covariance matrix, and generate a Mahalanobis distance sequence in chronological order.

[0030] In this embodiment, the Mahalanobis distance calculation process specifically includes: Extract the vector differences of all encoded pairs and generate a set of difference vectors according to time steps; The mean and deviation of the difference vector set are calculated in each dimension to construct the covariance matrix, and the inverse matrix is ​​obtained by performing the inverse matrix operation on the covariance matrix. The difference vector for this time step is constructed by subtracting the corresponding elements from the user encoding vector and the content encoding vector in the encoding pair. The difference vector is taken as a row matrix input. Each element of the matrix input in the column direction is rearranged according to its index to the row direction position of the new matrix to obtain the transposed difference vector. Then, matrix multiplication is performed with the covariance inverse matrix to obtain the intermediate matrix result. Perform a matrix multiplication operation between the intermediate matrix result and the difference vector to generate the Mahalanobis distance for that time step; Arrange all Mahalanobis distance values ​​in chronological order to form a Mahalanobis distance sequence.

[0031] In this embodiment, S34 specifically includes: S341. Align the user feature set and the content feature set in the time dimension, and extract the user feature vector and the corresponding content feature vector at each time step to form a feature pairing sequence. S342. In each feature pairing, perform vector concatenation operation in the order of user feature vector first and content feature vector last to form a single adjacency vector. S343. Perform Min-Max normalization on each adjacent vector, set a fixed-length sliding window on the normalized adjacent vector, and extract the maximum, minimum and mean values ​​within the sliding window, and concatenate them into a local statistical vector. S344. Arrange all the local statistical vectors generated by the sliding windows in sequence to form the adjacency feature terms of the corresponding adjacency vectors, and arrange all the adjacency feature terms in chronological order to construct the adjacency feature sequence.

[0032] In this embodiment, S4 specifically includes: S41. Construct the DreamerV3 world model, perform forward inference operations on the set of recommended states in the policy network of the world model, generate action feature vectors corresponding to each time step, and arrange them into a recommendation ranking sequence in chronological order. S42. Through the value network of the world model, perform state encoding and score prediction operations on the recommended state set at time steps to generate state score vectors at corresponding time steps, and combine all state score vectors in time order to form a state value sequence. S43. Align the recommended ranking sequence with the state value sequence in time. Calculate the policy loss based on the difference between the state score vector and the action feature vector at each time step in the alignment result. Update the policy network parameters using the gradient backpropagation algorithm.

[0033] In this embodiment, the execution process within the DreamerV3 world model specifically includes: In the policy network, a vector dimension mapping operation is performed on the recommended state vector at each time step to transform the recommended state vector into a policy representation vector with fixed dimensions, and batch normalization and ReLU function activation are performed on the policy representation vector. The activated policy representation vector is subjected to two rounds of Dropout processing and linear transformation operations. After the first round of transformation, layer normalization and Leaky ReLU activation are performed. After the second round of transformation, a residual connection structure is introduced to perform dimensionality compression processing, generating the action feature vector for the corresponding time step. The dimensionality compression processing specifically includes: The activation results before the second round of transformation are extracted as residual vectors and fused with the results of the second round of transformation by element-wise weighting to form a fusion vector. The weight set of the weighted fusion is automatically generated by the policy network based on the recommended state vector. The fusion vector is compressed by performing a dimension compression operation, which maps the fusion vector to the target dimension using a preset linear projection matrix, and the compression result is normalized by layer to generate an action feature vector. Arrange all action feature vectors in chronological order to construct a recommended ranking sequence; In the value network, embedding mapping and Tanh function activation are performed on the recommended state vector at each time step. The activation results are then subjected to linear mapping, Z-score normalization, and numerical compression to generate the state score vector for the corresponding time step. The numerical compression operation specifically includes: For each time step, the state score vector is compressed using the Sigmoid function, which maps the values ​​of each element in the state score vector to between 0 and 1, and performs maximum value truncation, replacing elements that exceed the set threshold with the set threshold. Combine all state score vectors in chronological order to construct a state value sequence; Align the recommended ranking sequence with the state value sequence along the time dimension to construct a paired sequence; Perform time-step error calculation on the paired sequences, calculate the policy difference value between action features and state scores at each time step, perform loss aggregation on all policy difference values ​​by time step, and perform backpropagation based on the aggregation result to update the policy network parameters.

[0034] In this embodiment, S6 specifically includes: S61. Pair the feedback results with the recommended ranking sequence by time step to calculate the action reward value sequence; S62. Based on the action reward value sequence and the recommendation ranking sequence, a policy loss function is constructed using the policy gradient algorithm. The gradient backpropagation operation is performed on the policy loss function to update the policy network parameters in the DreamerV3 world model. S63. Based on the feedback results, the user behavior sequence is re-labeled, and the behavior segments corresponding to effective click behavior and high-value conversion behavior are marked to construct a user behavior sample set. S64. Input the user behavior sample set into the bidirectional long short-term memory network, perform error propagation and parameter update operations, and update the parameters of the bidirectional long short-term memory network.

[0035] Example 1: To verify the feasibility of this invention in practice, it was applied to a recommendation scenario of a digital economy platform. The massive user behavior data and content description data within the platform were processed, and compared with existing mainstream recommendation algorithms. The recommendation effect, user behavior response indicators, and system feedback optimization capabilities were comprehensively evaluated to verify the performance advantages and stability of the proposed method in actual business scenarios.

[0036] The recommendation scenario applied in this invention is a typical content-driven digital platform with over one million daily active users, high content update frequency, and complex user behavior exhibiting strong temporality, diversity, and sparsity. In this scenario, user page browsing, content clicks, navigation paths, dwell time, and final conversion behavior are all recorded in real time. Simultaneously, the platform provides recommendable content in the form of videos, articles, and product cards, all accompanied by structured and unstructured descriptive information, such as tags, titles, summaries, and image descriptions. Traditional collaborative filtering and content-based recommendation methods struggle to handle high-frequency dynamic behaviors and complex feature interactions, often resulting in insufficient recommendation accuracy and decreased user response efficiency.

[0037] In the specific implementation process, user behavior data is first collected from the platform over a certain period, including click records, browsing paths, dwell time, and conversion indicators. This data is then combined with descriptive data of all recommendable items in the platform's content library to construct a raw dataset. Missing data is then filled in and time-aligned on the user behavior data to generate user behavior sequences. Simultaneously, word vector embedding and convolutional pooling are performed on the content description data to construct a content feature set.

[0038] Next, a bidirectional long short-term memory network is used to extract forward and backward states from the user behavior sequence and then fuse them to generate a high-quality user feature set. Then, a Siamese neural network is used to share the encoding between the user feature set and the content feature set, calculating the Euclidean and Mahalanobis distances between each user and content combination, and combining this with time-aligned adjacency features to construct a recommendation state set. This recommendation state set is then input into the policy network and value network of the DreamerV3 world model to generate recommendation ranking sequences and state-value sequences, respectively. Finally, a policy gradient algorithm combined with real user feedback is used to optimize the policy and update network parameters, achieving closed-loop improvement of the recommendation policy.

[0039] During the effectiveness verification phase, comparative tests were conducted with three common recommendation methods: collaborative filtering (CF), gradient boosting tree-based ranking model (GBDT-Rank), and the classic deep neural network recommendation method (DNN-Recommender). A unified evaluation period and test user group were set, and metrics such as precision, recall, click-through rate (CTR), and conversion rate (CVR) were used for evaluation.

[0040] The test results are shown in the table below: Table 1. Comparative experimental data statistics of the present invention and existing recommended methods.

[0041] First, judging from the precision and recall in Table 1, the precision of the method of this invention reaches 0.468 and the recall is 0.583, which are 49.9% and 36.5% higher than the collaborative filtering method, respectively. This indicates that it has a higher recognition ability in terms of content and user feature matching, and the recommendation results are more in line with user interests.

[0042] Secondly, in terms of click-through rate and conversion rate, this invention achieves a click-through rate of 0.317 and a conversion rate of 0.128, which are 14.8% and 26.7% higher than the classic DNN method, respectively. This shows that by introducing a reinforcement learning mechanism, the recommended content is more attractive and can effectively guide users to achieve their target behaviors, such as purchasing, submitting forms, or adding to favorites.

[0043] In terms of user behavior response, this invention achieves an average dwell time of 112.5 seconds, which is much higher than other methods, and the user bounce rate is only 24.2%, which is significantly lower than traditional methods. This indicates that this invention can not only attract users to click, but also maintain high content stickiness and retention.

[0044] Finally, in terms of recommendation response time, this invention controls the response latency to 136ms while achieving complex strategy learning, which is similar to traditional models, maintaining good online availability and making it suitable for deployment in real-time recommendation scenarios.

[0045] In summary, this invention outperforms existing mainstream recommendation methods across multiple key metrics, not only improving recommendation accuracy and user satisfaction but also demonstrating excellent response performance and continuous optimization capabilities, fully validating its practicality and advanced nature in recommendation scenarios for digital economy platforms.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent recommendation optimization of a digital economy platform based on artificial intelligence, characterized in that, Includes the following steps: S1. Obtain user behavior data and content description data recorded in chronological order, preprocess the user behavior data, and construct a user behavior sequence; S2. Perform bidirectional state extraction and fusion operations on user behavior sequences through a bidirectional long short-term memory network to construct a user feature set. At the same time, perform word vector embedding, one-dimensional convolution and pooling operations on content description data to construct a content feature set. S3. The user feature set and content feature set are shared and standardized through a Siamese neural network. Euclidean distance and Mahalanobis distance are calculated based on the standardization results. Adjacency features of the two sets are extracted according to time steps. Recommendation state set is constructed based on the calculation results and extraction results. S4. Construct the DreamerV3 world model. Through the policy network and value network of this model, perform forward inference operations and state value estimation on the set of recommended states respectively to generate a recommendation ranking sequence. At the same time, update the policy network parameters based on the estimation results. S5. Use the recommended ranking sequence in the actual recommendation process, record the user's click behavior, browsing path, dwell time and conversion identifier, and generate feedback results in chronological order; S6. Based on the feedback results, the parameters of the DreamerV3 world model and the bidirectional long short-term memory network are updated using the policy gradient algorithm.

2. The intelligent recommendation optimization method for a digital economy platform based on artificial intelligence according to claim 1, characterized in that, The user behavior data includes click records, browsing paths, dwell time, and conversion identifiers. The conversion identifiers indicate whether the user has achieved their target behavior after receiving recommended content, including whether or not they have placed an order, submitted a form, shared content, or saved the content. The content description data includes structured or unstructured information corresponding to the recommendable content in the digital economy platform. The preprocessing includes missing information completion, format standardization, and time alignment operations. The adjacency feature represents the local correspondence between the user feature set and the content feature set under time alignment.

3. The intelligent recommendation optimization method for a digital economy platform based on artificial intelligence according to claim 1, characterized in that, S2 specifically includes: S21. Arrange the user behavior segments at each time step in the user behavior sequence into behavior vectors according to a preset index order, and combine all behavior vectors into a time series matrix. S22. Perform bidirectional state extraction and fusion operations through a bidirectional long short-term memory network. Based on the time series matrix, update the forward hidden state sequentially along the forward time direction and update the backward hidden state sequentially along the reverse time direction to construct the forward state sequence and the backward state sequence. S23. Connect the two sequences by time step, and perform dimension normalization, scale correction and ReLU function activation on the connection result to construct a user feature set; S24. Perform a word vector lookup operation on each term in the content description data according to the preset embedding rule table, arrange the lookup results in the order of the content description data to generate an embedding matrix, and perform a one-dimensional convolution operation on the embedding matrix using a fixed-length sliding window to calculate the convolution response of the local region and form a convolution response matrix. S25. Set a pooling window of fixed size on the convolution response matrix, perform a maximum value selection operation on each window region, and combine all selection results into a content feature set.

4. The intelligent recommendation optimization method for a digital economy platform based on artificial intelligence according to claim 3, characterized in that, S22 specifically includes: S221. The forward hidden state is updated sequentially along the forward direction of time through a bidirectional long short-term memory network. The behavior vector of each time step in the time series matrix is ​​concatenated with the forward hidden state of the previous time step to construct a forward joint vector. In the first time step, the forward hidden state of the previous time step is initialized to a zero vector. S222. Perform forget gate transformation, update gate transformation and output gate transformation operations on the forward joint vector respectively, and use the Sigmoid function to activate the transformation results to generate forget gate vector, update gate vector and output gate vector. S223. Perform a linear transformation operation on the forward joint vector, and use the Tanh function to perform element-wise activation processing on the transformation result to generate candidate memory vectors. S224. Perform element-wise multiplication on the candidate memory vector and the update gate vector to generate the memory update vector. Perform element-wise multiplication on the forward memory vector of the previous time step and the forget gate vector to generate the memory retention vector. Perform element-wise addition on the memory update vector to generate the forward memory vector of the current time step. In the first time step, use the preset initialization vector as the forward memory vector of the previous time step. S225. Perform element-wise multiplication between the forward memory vector and the output gate vector to form the forward hidden state at the current time step; S226. The backward hidden state is updated sequentially in reverse time through a bidirectional long short-term memory network. At each time step, the current action vector is concatenated with the backward hidden state of the next time step, and the forget gate transformation, update gate transformation, output gate transformation, Tanh function activation, element-wise multiplication and addition operations are performed to generate the backward hidden state corresponding to each time step. At the last time step, the backward hidden state of the next time step is initialized to a zero vector. S227. Combine all forward hidden states and backward hidden states to form forward state sequences and backward state sequences respectively.

5. The intelligent recommendation optimization method for a digital economy platform based on artificial intelligence according to claim 1, characterized in that, S3 specifically includes: S31. Input the user feature set and the content feature set into two network branches with shared weights in the Siamese neural network, respectively, and perform feature mapping and embedding transformation operations to obtain the shared encoding result of the two sets; S32. Adjust the mean and scale of each encoding vector in the two shared encoding results according to the element dimension to obtain the user encoding set and the content encoding set. S33. Based on the user coding set and the content coding set, calculate the Euclidean distance and Mahalanobis distance between the user coding vector and the content coding vector at the corresponding time step, and generate the Euclidean distance sequence and the Mahalanobis distance sequence. S34. Align the user feature set and the content feature set in time, and extract the user feature vector and the content feature vector at each time step to form an adjacent feature sequence. S35. Concatenate the Euclidean distance sequence, the Mahalanobis distance sequence, and the adjacency feature sequence to construct a set of recommended states.

6. The intelligent recommendation optimization method for a digital economy platform based on artificial intelligence according to claim 5, characterized in that, The Siamese neural network comprises two branches that share weight matrices and bias parameters. Each branch includes an embedding layer and two fully connected layers. The processing of this network specifically includes: The user feature set and the content feature set are respectively input into two network branches and processed sequentially according to time steps; In the first fully connected layer, matrix multiplication and bias addition operations are performed on the input user feature vector or content feature vector, and ReLU function is used for activation. The activation result is fed into the second fully connected layer, where matrix multiplication, bias addition, activation and batch normalization operations are performed to generate standardized features. In the embedding layer, a dimension mapping operation is performed on the standardized features. The standardized features are projected using a preset linear weight matrix, and the projection result is activated by the GELU function to generate an embedding vector. The user encoding set and the content encoding set are formed based on the embedding vectors output by the two network branches at each time step.

7. The intelligent recommendation optimization method for a digital economy platform based on artificial intelligence according to claim 5, characterized in that, Specifically, S33 includes: S331. Perform time pairing on the user coding set and the content coding set, and extract the user coding vector and content coding vector corresponding to each time step as a coding pair. S332. Perform Euclidean distance calculation operation on each encoding pair, calculate the Euclidean distance between the two vectors by summing the square differences and then taking the square root, and generate an Euclidean distance sequence in time order; S333. Perform Mahalanobis distance calculation for each encoding pair, construct a covariance matrix based on all encoding pairs, calculate the Mahalanobis distance of the encoding pairs based on the covariance matrix, and generate a Mahalanobis distance sequence in chronological order.

8. The intelligent recommendation optimization method for a digital economy platform based on artificial intelligence according to claim 7, characterized in that, The Mahalanobis distance calculation process specifically includes: Extract the vector differences of all encoded pairs and generate a set of difference vectors according to time steps; The mean and deviation of the difference vector set are calculated in each dimension to construct the covariance matrix, and the inverse matrix is ​​obtained by performing the inverse matrix operation on the covariance matrix. The difference vector for this time step is constructed by subtracting the corresponding elements from the user encoding vector and the content encoding vector in the encoding pair. The difference vector is taken as a row matrix input. Each element of the matrix input in the column direction is rearranged according to its index to the row direction position of the new matrix to obtain the transposed difference vector. Then, matrix multiplication is performed with the covariance inverse matrix to obtain the intermediate matrix result. Perform a matrix multiplication operation between the intermediate matrix result and the difference vector to generate the Mahalanobis distance for that time step; Arrange all Mahalanobis distance values ​​in chronological order to form a Mahalanobis distance sequence.

9. The intelligent recommendation optimization method for a digital economy platform based on artificial intelligence according to claim 1, characterized in that, S4 specifically includes: S41. Construct the DreamerV3 world model, perform forward inference operations on the set of recommended states in the policy network of the world model, generate action feature vectors corresponding to each time step, and arrange them into a recommendation ranking sequence in chronological order. S42. Through the value network of the world model, perform state encoding and score prediction operations on the recommended state set at time steps to generate state score vectors at corresponding time steps, and combine all state score vectors in time order to form a state value sequence. S43. Align the recommended ranking sequence with the state value sequence in time. Calculate the policy loss based on the difference between the state score vector and the action feature vector at each time step in the alignment result. Update the policy network parameters using the gradient backpropagation algorithm.

10. The intelligent recommendation optimization method for a digital economy platform based on artificial intelligence according to claim 9, characterized in that, The execution process within the DreamerV3 world model specifically includes: In the policy network, a vector dimension mapping operation is performed on the recommended state vector at each time step to transform the recommended state vector into a policy representation vector with fixed dimensions, and batch normalization and ReLU function activation are performed on the policy representation vector. The activated policy representation vector is subjected to two rounds of Dropout processing and linear transformation operations. After the first round of transformation, layer normalization and Leaky ReLU function activation are performed. After the second round of transformation, a residual connection structure is introduced to perform dimensionality compression processing, generating the action feature vector for the corresponding time step. Arrange all action feature vectors in chronological order to construct a recommended ranking sequence; In the value network, embedding mapping and Tanh function activation are performed on the recommended state vector at each time step, and linear mapping, Z-score standardization and numerical compression are performed on the activation results in sequence to generate the state score vector for the corresponding time step. Combine all state score vectors in chronological order to construct a state value sequence; Align the recommended ranking sequence with the state value sequence along the time dimension to construct a paired sequence; Perform time-step error calculation on the paired sequences, calculate the policy difference value between action features and state scores at each time step, perform loss aggregation on all policy difference values ​​by time step, and perform backpropagation based on the aggregation result to update the policy network parameters.

Images