Interactive digital media content intelligent recommendation method based on big data

Abstract

The application relates to the technical field of intelligent recommendation, in particular to an interactive digital media content intelligent recommendation method based on big data, which comprises the following steps: collecting heterogeneous interaction behavior data of users on multiple platforms, generating a unified user full-cycle behavior profile through cross-platform identity mapping and behavior track fusion, and constructing a dynamic evolving user interest graph based on the profile. Newly added content is processed in real time, deep semantic deconstruction and feature extraction are carried out, and a content feature vector is obtained. The graph and the vector are input into an interactive matching engine, a dynamic matching degree is calculated through a multi-round bidirectional attention mechanism, an initial recommendation set is generated according to the matching degree, a simulated feedback environment is injected to simulate expected interaction, and the final recommendation list is generated through sorting and diversity optimization according to the simulation result. The method realizes deep understanding of the panoramic interest of users across platforms, and improves the accuracy and diversity of the recommendation through dynamic interactive matching and simulation optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology, and in particular to an intelligent recommendation method for interactive digital media content based on big data. Background Technology

[0002] Current digital media content recommendation technologies primarily rely on analyzing user behavior data within a single platform or application. User models built using these methods can only reflect localized interests and preferences within specific service scenarios, exhibiting significant data limitations. Due to the varying user identification and behavioral data formats across different platforms, cross-platform heterogeneous interactions cannot be effectively correlated and integrated. This results in a fragmented understanding of user interests, making it difficult to depict the complete lifecycle and dynamic evolution of interests across multiple platforms. At the recommendation matching computation level, existing technologies generally employ methods that embed user interests and content features into fixed-dimensional vectors, completing recommendations through a one-time similarity calculation. This static matching model simplifies the user interest structure to a single vector representation, failing to characterize the complex network of relationships between entities and tags within an interest. Furthermore, unidirectional similarity calculations struggle to dynamically assess the non-linear interactions between multi-dimensional content features and different nodes in the user interest graph, and the matching process lacks an interpretable deep reasoning mechanism.

[0003] The core problem facing existing technologies lies in the fragmentation of user behavior data and the static nature of recommendation matching models. This requires new technical solutions to achieve a unified representation of user behavior across platforms and to establish a matching mechanism capable of deep, dynamic interactive computation, in order to improve the understanding of users' overall interests and the accuracy of recommendation results. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent recommendation method for interactive digital media content based on big data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent recommendation method for interactive digital media content based on big data, comprising: Collect raw data of heterogeneous user interaction behavior on multiple platforms, perform cross-platform identity mapping and behavior trajectory fusion processing on the raw data of heterogeneous interaction behavior, and generate a unified user full-lifecycle behavior profile. Based on the user's full-lifecycle behavior profile, a dynamically evolving user interest graph is constructed, which includes content entity nodes, interest tag nodes, and the multidimensional relationships between them. Real-time access to new content streams from the digital media content library; deep semantic deconstruction and multi-dimensional feature extraction for each new piece of content; generating content feature vectors. The user interest graph and the content feature vector are input into the interactive matching engine, and the dynamic matching degree between the nodes in the user interest graph and the content feature vector is calculated through a multi-round bidirectional attention mechanism. An initial set of recommended content is generated based on the dynamic matching degree, and the initial set of recommended content is injected into a simulated feedback environment to simulate the expected interaction. Based on the results of the expected interaction simulation, the initial set of recommended content is sorted, adjusted, and its diversity is controlled to generate an optimized recommendation list.

[0006] As a further aspect of the present invention, the step of performing cross-platform identity mapping and behavior trajectory fusion processing on the heterogeneous interaction behavior raw data to generate a unified user full-lifecycle behavior profile includes: The raw data of the heterogeneous interaction behavior includes explicit feedback records, implicit behavior sequences, and social association data; Device fingerprints, login accounts, and behavioral pattern features are extracted from the raw data of heterogeneous interaction behaviors as identity recognition anchors; The similarity between the identity recognition anchors on different platforms is calculated by using a pre-trained cross-platform identity recognition model. Anchors with similarity exceeding a threshold are clustered and associated to establish a unified identity identifier. Using the unified identity identifier as an index, retrieve all user behavior records across all platforms and sort and align them according to a unified timestamp standard; Standardize the behavior type coding of time-aligned behavior records to eliminate differences in the definition of the same type of behavior across different platforms; The standardized coded behavior records are bound to the corresponding original context information and organized into a structured behavior trajectory chain according to the time sequence, which serves as the user's full-cycle behavior profile.

[0007] As a further aspect of the present invention, the step of constructing a dynamically evolving user interest graph based on the user's full-lifecycle behavior profile includes: Identify interaction events related to digital media content from the behavioral trajectory chain of the user's full-lifecycle behavior profile; Analyze each interaction event to extract the content object it affects, the type of behavior it applies, and the context in which it occurs; The extracted content objects are used as content entity nodes, and the semantic themes implied in the behavior type and context are used as interest tag nodes, which are then added to the graph. Calculate the association strength between the content entity node and the interest tag node, whereby the association strength is determined based on behavior frequency, behavior depth, and time decay factor. Based on the relationships between all nodes, a weighted directed graph is constructed as the initial user interest graph; Set a graph update cycle. Whenever a new full-cycle user behavior profile is generated, the initial user interest graph is iteratively evolved by adding nodes and adjusting existing edge weights.

[0008] As a further aspect of the present invention, the real-time access to the new content stream of the digital media content library involves performing deep semantic deconstruction and multi-dimensional feature extraction on each new piece of content to generate a content feature vector, including: Monitor the update interface of the digital media content library and capture the original data of newly added media content in real time; Multimodal parsing is performed on the raw media content data to separate text description information, visual thumbnails, audio samples, and metadata tags; The text description information is analyzed for topic analysis, sentiment identification, and key entity extraction using a natural language processing model. The visual thumbnails are used to perform scene recognition, subject detection, and style classification using computer vision models. The topic analysis results, sentiment identification results, key entity extraction results, scene identification results, subject object detection results, style classification results, and metadata tags are all input into a feature fusion network. The feature fusion network integrates multi-dimensional information into a high-dimensional dense vector representation, namely the content feature vector, through nonlinear transformation.

[0009] As a further aspect of the present invention, the step of inputting the user interest graph and the content feature vector into an interactive matching engine, and calculating the dynamic matching degree between nodes in the user interest graph and the content feature vector through a multi-round bidirectional attention mechanism, includes: The node set of the user interest graph and the content feature vector are simultaneously input into the interactive matching engine; In the first round of attention calculation, the content feature vector is used as the query vector, and all node features in the user interest graph are used as keys and values. The attention distribution of content to user interest is calculated to obtain the content-aware user interest representation. In the second round of attention calculation, the content-aware user interest representation obtained in the previous round is used as a new query vector, and the content feature vector is used as a key and value. The attention distribution of user interest on the content is calculated to obtain the interest-aware content representation. The content-aware user interest representation and the interest-aware content representation are concatenated as vectors. The concatenated vectors are subjected to dimensionality reduction and comprehensive scoring using a multilayer perceptron, and the output scalar score is the dynamic matching degree.

[0010] As a further aspect of the present invention, the step of generating an initial recommended content set based on the dynamic matching degree and injecting the initial recommended content set into a simulated feedback environment for expected interaction simulation includes: The dynamic matching degree of all candidate content in the digital media content library is sorted, and a preset number of top-ranked content are selected to form the initial recommended content set. Construct a simulated feedback environment, which includes a simulated user model, and the simulated user model is parameterized based on the user interest graph and historical behavior patterns; In the simulated feedback environment, each piece of content in the initial recommended content set is presented to the simulated user model in sequence; Based on the built-in strategy of the simulated user model, predict the type and intensity of its interactive behavior for each displayed content, and generate a simulated feedback sequence. Record the simulated feedback sequence triggered by each piece of content and summarize it into the result of the expected interaction simulation.

[0011] As a further aspect of the present invention, the step of sorting and adjusting the initial recommended content set and controlling its diversity based on the results of the expected interaction simulation to generate an optimized recommended list includes: Analyze the results of the expected interaction simulation and extract the predicted interaction intensity score for each piece of content; Based on the predicted interaction intensity score, the content in the initial recommended content set is reordered; Based on the reordering, calculate the feature similarity matrix between the contents; The maximum boundary correlation algorithm is used to iteratively adjust the sorted list: each time the content of the next position in the list is selected, the content with low feature similarity to the already selected content is selected first, provided that the predicted interaction intensity does not decrease significantly. After multiple rounds of iterative adjustments, a final sequence that strikes a balance between predicted interaction intensity and content diversity is formed, serving as the optimized recommendation list.

[0012] As a further aspect of the present invention, the method further includes: While displaying the optimized recommendation list to target users, implicit feedback tracking is deployed to capture users' fine-grained interaction behaviors with the recommended content in real time, specifically including: On the user interface that renders and displays the optimized recommendation list, invisible monitoring code is embedded for each recommended content item; The monitoring code records the duration of the user's cursor hovering over content items, the speed at which the cursor scrolls, whether the user clicks to enter the details page, the duration of the user's stay on the details page, whether the user performs actions such as liking, collecting, or sharing, and the percentage of time the user completes viewing or reading the content. The recorded raw interaction events are packaged in chronological order of occurrence and contextual information is added to form the fine-grained interaction behavior data stream. The fine-grained interactive behavior data stream is sent to the background analysis server in real time.

[0013] As a further aspect of the present invention, the method further includes: The user interest graph is updated incrementally in real time using the fine-grained interaction behaviors captured in real time, and the next round of recommendation calculation is triggered. Specifically, this includes: After receiving the fine-grained interactive behavior data stream, the background analysis server immediately parses it and identifies behavioral patterns that represent positive interest, negative interest, or neutral exploration. Based on the identified behavioral patterns, the type of update operation that needs to be performed on the user interest graph is determined. The update operation type includes strengthening existing node associations, weakening existing node associations, and adding new nodes or new associations. Based on the intensity information and content identifiers contained in the fine-grained interactive behaviors, the weight values ​​that need to be adjusted for the update operation type are quantitatively calculated. Without interrupting the current graph service, the weight values ​​are updated and applied to the corresponding nodes and edges of the user interest graph in an incremental manner. After the graph update is completed, the updated user interest graph and the latest content feature vector are immediately used as input to automatically trigger a new round of the complete process from matching calculation to generating the recommendation list.

[0014] As a further aspect of the present invention, the construction steps of the interactive matching engine include: Define the architecture of the matching engine, which includes an input interface, at least two attention calculation layers, a feature fusion layer, and a scoring output layer; Configure the input interface to simultaneously receive the node feature matrix from the user interest graph and the content feature vector from the content processing flow; A first attention calculation layer is constructed, in which the content feature vector is used as the query input and the node feature matrix is ​​used as the key and value input to realize the attention calculation from content to user interest and output the content-aware interest representation vector. A second attention calculation layer is constructed, which takes the content-aware interest representation vector as the query input and the content feature vector as the key and value input to realize the attention calculation from user interest to content and output the interest-aware content representation vector. In the feature fusion layer, the content-aware interest representation vector and the interest-aware content representation vector are concatenated to form a joint representation vector; In the scoring output layer, a fully connected neural network is set up to perform nonlinear transformation and dimensionality reduction on the joint representation vector, and finally output a scalar score as the dynamic matching degree. The completed interactive matching engine is trained offline, using historical user interaction data as a supervision signal to optimize the model parameters in the attention calculation layer and the scoring output layer.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: By implementing cross-platform identity mapping and behavioral trajectory fusion, raw user interaction data from different applications and with heterogeneous formats can be uniformly anchored to a unique virtual identity. This process connects scattered behavioral points in chronological order, forming a continuous behavioral sequence covering multiple scenarios and spanning long time periods. The resulting unified user lifecycle behavior profile breaks down data barriers between platforms, enabling subsequent analysis to be based on a complete and coherent data foundation. This profile can comprehensively reflect the cross-domain migration and long-term evolution of user interests, providing real and comprehensive data support for building a more stable and deeper user interest model. An interactive matching engine based on a multi-round bidirectional attention mechanism changes the computational paradigm of traditional static vector matching. In each round of calculation, this engine allows bidirectional attention weight allocation and information interaction between content entity nodes, interest tag nodes in the user interest graph, and feature vectors of the content to be recommended. This multi-round, bidirectional cross-attention calculation can dynamically evaluate the correlation between different semantic dimensions of content and different nodes and relationships between nodes in the user interest graph. It makes the matching process an iterative and adjustable interactive reasoning process, rather than a one-time judgment, thereby capturing and quantifying the subtle and dynamic relationship between the user's complex interest structure and the deep features of the content, enhancing the rationality of the recommendation reasons and the accuracy of the matching results. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intelligent recommendation method for interactive digital media content based on big data according to the present invention; Figure 2 A flowchart for cross-platform identity mapping and behavior trajectory fusion processing; Figure 3Heatmap of bidirectional attention weight distribution for the interactive matching engine; Figure 4 A correlation analysis diagram to balance the two-dimensional impact of hyperparameter λ on recommendation performance; Figure 5 A comparative analysis chart of incremental updates to the weights of nodes in the user interest graph. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] Please see Figure 1 The present invention provides an intelligent recommendation method for interactive digital media content based on big data. The overall implementation scheme is as follows: Raw data of heterogeneous user interaction behaviors across multiple platforms is collected. This data undergoes cross-platform identity mapping and behavior trajectory fusion processing to generate a unified user lifecycle behavior profile. Based on this user lifecycle behavior profile, a dynamically evolving user interest graph is constructed. This user interest graph includes content entity nodes, interest tag nodes, and their multi-dimensional relationships. New content streams from the digital media content library are accessed in real time. Each new piece of content undergoes deep semantic deconstruction and multi-dimensional feature extraction to generate a content feature vector. The user interest graph and the content feature vector are input into an interactive matching engine. A multi-round bidirectional attention mechanism is used to calculate the dynamic matching degree between nodes in the user interest graph and the content feature vector. An initial recommended content set is generated based on this dynamic matching degree and injected into a simulated feedback environment for expected interaction simulation. Based on the results of the expected interaction simulation, the initial recommended content set is sorted, adjusted, and its diversity controlled to generate an optimized recommendation list.

[0020] In one embodiment of the present invention, the raw data of heterogeneous interaction behavior includes explicit feedback records, implicit behavior sequences, and social association data. Device fingerprints, login accounts, and behavioral pattern features are extracted from the raw data of heterogeneous interaction behavior as identity recognition anchors. The similarity between identity recognition anchors on different platforms is calculated using a pre-trained cross-platform identity recognition model. Anchors with similarity exceeding a threshold are clustered and associated to establish a unified identity identifier. Using the unified identity identifier as an index, all user behavior records on all platforms are retrieved and sorted and aligned according to a unified timestamp standard. The time-aligned behavior records are standardized and encoded to eliminate differences in the definition of similar behaviors on different platforms. The standardized and encoded behavior records are bound to the corresponding original context information and organized into a structured behavior trajectory chain according to the time sequence, serving as a full-cycle user behavior profile. The construction steps of the interactive matching engine include defining the architecture of the matching engine, which includes an input interface, at least two attention calculation layers, a feature fusion layer, and a scoring output layer. The input interface is configured to simultaneously receive node feature matrices from the user's interest graph and content feature vectors from the content processing flow. A first attention computation layer is constructed, using content feature vectors as query input and node feature matrices as key and value inputs to perform attention computation from content to user interests, and outputting a content-aware interest representation vector. A second attention computation layer is constructed, using the content-aware interest representation vector as query input and content feature vectors as key and value inputs to perform attention computation from user interests to content, and outputting an interest-aware content representation vector. In the feature fusion layer, the content-aware interest representation vector and the interest-aware content representation vector are concatenated to form a joint representation vector. In the scoring output layer, a fully connected neural network is used to perform nonlinear transformation and dimensionality reduction on the joint representation vector, ultimately outputting a scalar score as the dynamic matching degree. The constructed interactive matching engine is trained offline using historical user interaction data as supervision signals to optimize the model parameters in the attention computation layer and the scoring output layer.

[0021] In practical implementation, the input for cross-platform identity mapping and behavior trajectory fusion processing is heterogeneous interaction behavior raw data. This data includes explicit feedback records, implicit behavior sequences, and social association data collected from multiple platforms. The data processing system extracts device fingerprints, login accounts, and behavioral pattern features from the heterogeneous interaction behavior raw data as identity recognition anchors. Device fingerprints include a combined hash value of the device hardware serial number, operating system version, screen resolution, and installed font list. Login accounts include email address, mobile phone number, and a unique user identifier from a third-party open platform. Behavioral pattern features include typical active time periods, commonly used interaction gesture sequences, and average scrolling speed of content browsing.

[0022] In some embodiments, the pre-trained cross-platform identity recognition model calculates the similarity between identity anchors on different platforms. The pre-trained cross-platform identity recognition model is a deep metric learning-based neural network that receives two sets of identity anchor vectors from two different platforms as input and outputs a similarity score between 0 and 1. The calculation of the similarity score follows the following relationship:

[0023] Where: symbol This represents the similarity score between two sets of identity recognition anchors on platform A and platform B; (symbol) Indicates parameters Defined multilayer perceptron network; symbol and These represent independent feature encoders for the feature distributions of platform A and platform B, respectively. When the similarity score exceeds a preset threshold (e.g., 0.85), the system clusters these two sets of identity recognition anchors and assigns a globally unique unified identity identifier to each cluster. Using the unified identity identifier as an index, the system retrieves all user behavior records across all associated platforms from the databases of all associated platforms, including clicks, plays, favorites, comments, shares, and dwell events.

[0024] Understandably, the system converts all timestamps in behavior records to Coordinated Universal Time (UTC) and performs global sorting and alignment based on the converted timestamps. For the time-aligned behavior records, the system performs standardized behavior type encoding. For example, behaviors such as "like," "share," and "upvote" defined on different platforms are uniformly mapped to the code "LIKE," and "play," "start watching," and "open audio" are uniformly mapped to the code "PLAY." This standardized encoding process eliminates semantic differences in the definitions of similar interactive behaviors across different platforms. After encoding, the system binds the standardized behavior records to the original context information at the time the behavior occurred. The original context information includes content identifiers, platform identifiers, network environment, and device type. The bound data is organized into a structured behavior trajectory chain according to time sequence. This behavior trajectory chain serves as the user's full-lifecycle behavior profile used in subsequent processes.

[0025] The construction steps of the interactive matching engine include defining the engine's architecture, which comprises an input interface, a first attention calculation layer, a second attention calculation layer, a feature fusion layer, and a scoring output layer. The input interface is configured to simultaneously receive node feature matrices from the user's interest graph and content feature vectors from the content processing flow. The first attention calculation layer uses the latter as the query input and the former as the key and value input to perform attention calculation from content to user interests. The output of the first attention calculation layer is a content-aware interest representation vector. The second attention calculation layer uses the content-aware interest representation vector output from the previous layer as the query input and the content feature vector as the key and value input to perform attention calculation from user interests to content. The output of the second attention calculation layer is an interest-aware content representation vector.

[0026] In its implementation, the feature fusion layer performs vector concatenation, connecting the content-aware interest representation vector and the interest-aware content representation vector along the feature dimension to form a joint representation vector. The scoring output layer consists of a fully connected neural network, which performs non-linear transformation and dimensionality reduction on the joint representation vector, ultimately outputting a scalar score as the dynamic matching degree. The completed interactive matching engine requires offline training, using historical user interaction data as supervision. This historical user interaction data includes positive and negative user interaction labels. The training process aims to minimize the gap between the predicted dynamic matching degree and the actual interaction labels, optimizing the trainable model parameters of the fully connected neural networks in the first attention calculation layer, the second attention calculation layer, and the scoring output layer through backpropagation.

[0027] See Figure 2In one embodiment of the present invention, interactive events related to digital media content are identified from the behavioral trajectory chain of a user's full-lifecycle behavior profile. Each interactive event is parsed to extract the content object it affects, the type of behavior performed, and the context in which it occurs. The extracted content object is used as a content entity node, and the semantic themes implied in the behavior type and context are used as interest tag nodes, which are then added to the graph. The association strength between content entity nodes and interest tag nodes is calculated, based on behavior frequency, behavior depth, and a time decay factor. A weighted directed graph is constructed as the initial user interest graph based on the association relationships of all nodes. A graph update cycle is set, and whenever a new user full-lifecycle behavior profile is generated, the initial user interest graph is iteratively evolved by adding nodes and adjusting existing edge weights. The update interface of the digital media content library is monitored to capture the raw data of newly added media content in real time. Multimodal parsing is performed on the raw media content data to separate text description information, visual thumbnails, audio samples, and metadata tags. A natural language processing model is used to perform topic analysis, sentiment identification, and key entity extraction on the text description information. Computer vision models are used to perform scene recognition, subject detection, and style classification on visual thumbnails. The results of topic analysis, sentiment analysis, key entity extraction, scene recognition, subject detection, style classification, and metadata tags are input into a feature fusion network. This feature fusion network integrates multi-dimensional information into a high-dimensional, dense vector representation, namely the content feature vector, through nonlinear transformation.

[0028] In practical implementation, the initial input for constructing a dynamically evolving user interest graph is the user's full-lifecycle behavior profile. From the structured behavioral trajectory chain of the user's full-lifecycle behavior profile, the data processing module scans and identifies interactive events related to digital media content. The identifiers of interactive events include content playback, content collection, content rating, and content sharing. Parsing each identified interactive event involves extracting the content object it affects, the type of behavior performed, and the context in which it occurred. The content object is represented by a globally unique content identifier, the behavior type is a standardized type label, and the context includes the timestamp of the event, geographical location information, the type of terminal device used, and the current network connection status.

[0029] In some embodiments, the system adds extracted content objects as content entity nodes to the user interest graph being constructed, and adds semantic topics implied in behavior types and contextual situations as interest tag nodes. Behavior types are directly used as a type of interest tag node, while the semantic topics of the contextual situation need to be extracted from the contextual description fields using natural language processing techniques. For example, from the context of "watching on Wi-Fi at home on a weekend night," "family evening entertainment" might be extracted as an interest tag node. The association strength between content entity nodes and interest tag nodes is calculated. The value of the association strength is jointly determined by behavior frequency, behavior depth, and time decay factor, and its calculation follows the following relationship:

[0030] Where: symbol Represents content entity nodes With interest tag nodes The strength of the association between them; symbols This indicates the node pairs associated in the behavior trajectory chain. The total number of interaction events; symbol Indicates the first The behavioral depth quantification value of each interaction event (e.g., 1.0 for complete viewing, and 0.5 for exiting midway); symbol It is the preset time decay coefficient; symbol Indicates the current calculation time; symbol Indicates the first The time of occurrence of the next interaction event; symbol and Representing content entity nodes respectively and interest tag nodes The total frequency of occurrence across all user behaviors. Based on all identified content entity nodes, interest tag nodes, and the calculated association strength between node pairs, the system constructs a weighted directed graph as the initial user interest graph. In the graph, nodes represent entities or interest tags, and the weights of the directed edges represent the association strength.

[0031] It's understandable that the user interest graph is set to evolve dynamically. The system has a preset graph update cycle, which can be a fixed time interval or triggered by the accumulation of new behavioral data. Whenever a new full-lifecycle user behavior profile is generated, the graph update engine starts, iteratively evolving the initial user interest graph by adding nodes and adjusting existing edge weights. New nodes come from content objects or interest tags parsed from the new behavior profile that have not appeared in the graph before. Adjusting existing edge weights is based on interaction events in the new behavior profile, reusing the association strength calculation formula to update the edge weight values ​​between corresponding node pairs.

[0032] Real-time access to new content streams to the digital media content library is achieved by monitoring the library's update interface, which pushes a notification message each time new content is added. The system captures the raw data of newly added media content, which typically includes structured metadata fields, text descriptions, cover images, and audio / video files. Multimodal parsing is performed on the raw media content data, separating it into text description information, visual thumbnails, audio samples, and metadata tags. The text description information includes the title, summary, and subtitle text; the visual thumbnail is the main cover or keyframe of the content; the audio sample is a representative audio clip extracted from the audio / video content; and the metadata tags include content category, publisher, duration, and resolution.

[0033] In practice, natural language processing models are used to perform topic analysis, sentiment analysis, and key entity extraction on the separated text description information. Topic analysis maps the text to a predefined topic distribution vector, sentiment analysis outputs the overall sentiment polarity score of the text, and key entity extraction identifies named entities such as people, places, organizations, and technical terms mentioned in the text. In parallel, computer vision models are used to perform scene recognition, subject detection, and style classification on the separated visual thumbnails. Scene recognition determines the environment category of the image, subject detection locates and identifies the main objects or people in the image, and style classification determines the visual art style of the image.

[0034] Optionally, auditory features can also be extracted from audio samples, such as identifying music genres, rhythms, or emotional states of speech through audio analysis models. The results of topic analysis, sentiment identification, key entity extraction, scene identification, subject detection, style classification, auditory features, and metadata tags are input into a feature fusion network. This feature fusion network is a deep neural network that integrates and compresses the aforementioned multi-dimensional and heterogeneous information into a high-dimensional, dense vector representation through multiple layers of nonlinear transformations. This high-dimensional, dense vector representation serves as the content feature vector used in subsequent matching calculations.

[0035] In one embodiment of the present invention, the node set of the user interest graph and the content feature vector are simultaneously input into the interactive matching engine. In the first round of attention calculation, the content feature vector is used as the query vector, and all node features in the user interest graph are used as keys and values ​​to calculate the attention distribution of content on user interests, thus obtaining a content-aware user interest representation. In the second round of attention calculation, the content-aware user interest representation obtained in the previous round is used as the new query vector, and the content feature vector is used as the key and value to calculate the attention distribution of user interests on content, thus obtaining an interest-aware content representation. The content-aware user interest representation and the interest-aware content representation are concatenated. A multilayer perceptron is used to perform dimensionality reduction and comprehensive scoring on the concatenated vector, and the output scalar score is the dynamic matching degree. The dynamic matching degrees of all candidate content in the digital media content library are sorted, and a predetermined number of top-ranked content are selected to form an initial recommended content set. A simulated feedback environment is constructed, which includes a simulated user model, which is parameterized based on the user interest graph and historical behavior patterns. In this simulated feedback environment, each piece of content in the initial recommended content set is sequentially displayed to the simulated user model. Based on the built-in strategies of the simulated user model, predict the type and intensity of user interaction behavior for each displayed content, and generate a simulated feedback sequence. Record the simulated feedback sequence triggered by each piece of content, and summarize them into the result of the expected interaction simulation.

[0036] In practice, the node set of the user interest graph and the content feature vector are simultaneously input into the input interface of the interactive matching engine. The interactive matching engine then initiates a multi-round bidirectional attention calculation process. In the first round of attention calculation, the content feature vector is used as the query vector, and all node features in the user interest graph are used as key and value vectors. The attention distribution of content to user interests is calculated using a scaled dot product attention mechanism. The calculation of the attention distribution involves similarity scores between the query vector and each key vector, which are then normalized using Softmax. The value vectors are weighted and summed based on attention weights to output a content-aware user interest representation vector. In the second round of attention calculation, the content-aware user interest representation vector obtained in the first round is used as the new query vector, and the content feature vector is used as the key and value vectors. The same attention mechanism is used again to calculate the attention distribution of user interests to content, outputting an interest-aware content representation vector.

[0037] In some embodiments, the content-aware user interest representation vector and the interest-aware content representation vector are passed to a feature fusion layer for vector concatenation. The concatenation operation joins the two vectors along the feature dimension into a joint representation vector. The joint representation vector is then fed into a scoring output layer, which contains a multilayer perceptron. The multilayer perceptron performs a nonlinear transformation and dimensionality reduction on the joint representation vector to calculate the dynamic matching degree. The calculation of the dynamic matching degree follows the following relationship:

[0038] Where: symbol Represents the final output dynamic matching degree scalar score; symbol Represents the joint representation vector of the inputs; symbol and This represents the weight matrix and bias vector of the first layer of a multilayer perceptron; (symbol) Represents the ReLU activation function; symbol and This represents the weight matrix and bias vector of the second layer of a multilayer perceptron; (symbol) This indicates that the Sigmoid activation function ensures the output range is between 0 and 1. The interactive matching engine independently performs the above calculation process for each candidate content in the digital media content library, generating a dynamic matching degree for each candidate content.

[0039] Understandably, the system sorts all candidate content by dynamic matching degree in descending order, and selects the top K candidate content with the highest dynamic matching degree based on a preset threshold number K (e.g., K=100). These selected candidate content pieces together form the initial recommended content set. Constructing the simulated feedback environment involves initializing a simulated user model. The simulated user model is a parameterized computational model whose parameters are derived from the structured information of the current user interest graph and the historical behavioral patterns statistically analyzed in the user's full-cycle behavior profile. These historical behavioral patterns include the user's average interaction probability with various types of content, typical session duration, and frequency of exploratory behaviors. In the simulated feedback environment, the system sequentially displays each piece of content in the initial recommended content set to the simulated user model, presented in the form of a content feature vector and associated metadata.

[0040] In practice, the simulated user model processes each displayed content according to a built-in strategy. This strategy predicts interactive behavior based on the matching degree between the content feature vector and the simulated user model's internal interest state. Interactive behavior types include clicking, playing, saving, sharing, and skipping, while the interaction intensity is quantified as a value between 0 and 1. For each displayed content, the simulated user model outputs a predicted interactive behavior type and its corresponding interaction intensity value. The predicted outputs of multiple contents are arranged in the order of display to generate a simulated feedback sequence. The system records the simulated feedback sequence triggered by each piece of content in the initial recommended content set. This sequence includes the content identifier, the predicted interactive behavior type, and the interaction intensity value. All simulated feedback sequences are aggregated into the expected interaction simulation result data.

[0041] Optionally, the built-in strategies of the simulated user model can be trained based on a reinforcement learning framework to better simulate the user decision-making process. The operation of the simulated feedback environment can be parallelized to improve efficiency, i.e., different content can be presented to multiple simulated user model instances simultaneously and feedback can be collected. The results of the expected interaction simulation are stored in a structured format, including a list of content, predicted sequences of interactive behaviors, and timestamp information. This results are directly used in subsequent ranking adjustments and diversity control processes.

[0042] See Figure 3 In the bidirectional attention calculation process of the interactive matching engine, the heatmap visually presents the distribution characteristics of the attention weights in the two rounds under different content display orders. Specifically, the first round of attention weights uses content features as the query and weights the user interest graph nodes, with values ​​fluctuating between 0.563 and 0.950. Among them, the weights of items with content display orders of 2 and 13 reach a peak of 0.950, indicating that these contents have a strong correlation with the core nodes of the user interest graph. The second round of attention weights uses the content-aware interest representation output in the first round as the query and weights the content features, with values ​​ranging from 0.457 to 0.950. Among them, the weight of items with content display order 7 is the highest (0.950), while the weight of items with order 9 is the lowest (0.457), reflecting that there are significant differences in the user's attention to different contents. This difference in bidirectional weight distribution reveals an asymmetric matching relationship between content and interests: some content can strongly activate user interest nodes, but user interest has low reverse attention to them (such as items in order 2 and 13), while some content, although initially only moderately activated, can trigger high reverse attention to interests (such as item in order 7).

[0043] In one embodiment of the present invention, the results of the expected interaction simulation are analyzed to extract the predicted interaction intensity score for each content. Based on the predicted interaction intensity score, the content in the initial recommended content set is reordered. Based on the reordering, a feature similarity matrix between the content is calculated. A maximum boundary correlation algorithm is used to iteratively adjust the sorted list: each time a content at the next position in the list is selected, content with low feature similarity to already selected content is prioritized, provided that the predicted interaction intensity does not decrease significantly. After multiple rounds of iterative adjustments, a final sequence that balances predicted interaction intensity and content diversity is formed, serving as the optimized recommended list. While displaying the optimized recommended list to the target user, implicit feedback tracking is deployed to capture fine-grained user interaction behaviors with the recommended content in real time. On the user interface that renders and displays the optimized recommended list, invisible monitoring code is embedded for each recommended content item. This monitoring code records the duration of the user's cursor hovering over the content item, the scrolling speed, whether the user clicks to enter the details page, the duration of stay on the details page, whether the user performs like, favorite, or share actions, and the completion rate of viewing or reading. The recorded raw interaction events are packaged in chronological order and contextual information is added to form a fine-grained interaction behavior data stream. This fine-grained interaction behavior data stream is then sent to the backend analysis server in real time.

[0044] In practical implementation, analyzing the results of the anticipated interaction simulation involves processing the simulation feedback sequence data. The system extracts the predicted interaction intensity score for each piece of content from the simulation results. This score is a numerical value between 0 and 1, output by the simulated user model for that content. Based on the predicted interaction intensity score, the system reorders the content in the initial recommended content set in descending order, forming an intermediate list with predicted interaction intensity as the sole ranking criterion. Following this reordering, the system calculates a feature similarity matrix between the content. This matrix is ​​obtained by calculating the cosine similarity between the feature vectors of every two pieces of content. The feature similarity matrix is ​​a symmetric matrix, with element values ​​ranging from -1 to 1.

[0045] In some embodiments, the maximum boundary correlation algorithm is used to iteratively adjust the intermediate list. The goal of the maximum boundary correlation algorithm is to balance the predicted interaction strength and content diversity. The iterative adjustment process starts from the first position of the intermediate list, directly placing the top-ranked content into the final sequence. For each subsequent candidate position, the algorithm selects the content with the lowest feature similarity to the already selected content from the remaining candidates in the intermediate list, ensuring that the predicted interaction strength does not decrease significantly. The selection criteria follow the following relationship:

[0046] Where: symbol Indicates the content identifier selected in this iteration; symbol This represents the set of remaining items in the current intermediate list that have not yet been selected; symbol Content Predicted interaction intensity score; symbol Represents the set of contents that have been selected for the final sequence; symbol Content With content eigenvectors and Cosine similarity between them; symbols This is a preset hyperparameter used to balance the weights of predicted interaction strength and diversity; its value ranges between 0 and 1. After selecting one piece of content in each iteration, that content is removed from the remaining set. Move to the selected set The iterative process continues until the length of the final sequence reaches a preset value. After multiple rounds of iterative adjustments, the system forms a final sequence that strikes a balance between predicted interaction intensity and content diversity; this final sequence serves as the optimized recommendation list. Table 1 shows key data from the adjustment process for an initial set of recommended content containing five example items.

[0047] Table 1: Initial Recommended Content Set Data Table

[0048] It's understandable that while rendering and displaying the optimized recommendation list to the target user's terminal device, the system deploys implicit feedback tracking points to capture fine-grained user interactions with the recommended content in real time. Deploying implicit feedback tracking points means embedding invisible monitoring code into the user interface of the rendered and optimized recommendation list for each recommended content item. This monitoring code typically exists in the form of JavaScript scripts or mobile SDK functions. The monitoring code running on the user's device records the duration the user's cursor hovers over content items, the speed at which the mouse or finger scrolls through the content item's display area, whether the user clicks on a content item to enter the details page, the duration the user stays on the details page, whether the user performs actions such as liking, saving, or sharing, and the percentage of time the user completes watching or reading audio, video, or text / image content.

[0049] In practice, the recorded raw interaction events are packaged in chronological order and supplemented with contextual information to form a fine-grained interactive behavior data stream. This contextual information includes the event type, timestamp, corresponding content identifier, content position in the recommendation list, current session identifier, and device identifier. This fine-grained interactive behavior data stream is then sent in real-time to the data receiving endpoint of the backend analytics server via a network connection. The deployment of implicit feedback tracking ensures no user disruption while capturing a series of continuous behaviors from content exposure to deep interaction.

[0050] Optionally, the monitoring code can be configured to send data packets in batches at a specific frequency to reduce the number of network requests, or to send them uniformly when a user ends a session. The background analytics server buffers and parses the received fine-grained interaction behavior data stream, and the parsed data will be used to update the user interest graph. The entire capture and transmission process must comply with data privacy and security protocols, and sensitive information must be anonymized.

[0051] See Figure 4 In recommendation sequence optimization based on the maximum boundary correlation algorithm, the value of the balancing hyperparameter λ directly determines the trade-off strategy between "predicted interaction intensity" and "content diversity" in the recommendation system. Quantitatively, when λ=0 (focusing only on interaction intensity), the system's average predicted interaction intensity score reaches its peak (approximately 0.922), but the content diversity index (the inverse of the average similarity) is at its lowest level (approximately 1.2), reflecting the high homogeneity of the recommended content. As λ gradually increases to 1.0 (focusing only on diversity), the content diversity index shows an almost linear growth trend (reaching a maximum of approximately 6.8), but the average predicted interaction intensity score continuously decreases to approximately 0.865, verifying the inverse relationship between the two. In terms of engineering implementation, the dual-axis visualization design of the graph intuitively presents the dynamic adjustment effect of the hyperparameter λ in the range of 0 (focusing on intensity) to 1 (focusing on diversity): the decaying trend of the bar chart contrasts sharply with the upward trend of the line chart, clearly revealing the direct impact of weight allocation on the recommendation effect when the maximum boundary correlation algorithm iteratively selects content. This result provides a clear quantitative basis for parameter tuning of recommendation systems. For example, in scenarios with highly active users, λ=0.2 can be selected to maintain a high level of interaction; while in scenarios involving the exploration of user interests, λ=0.8 can be selected to improve content diversity.

[0052] In one embodiment of the present invention, the user interest graph is updated incrementally in real time using fine-grained interactive behaviors captured in real time, triggering the next round of recommendation calculation. Upon receiving the fine-grained interactive behavior data stream, the background analysis server immediately parses it to identify behavioral patterns representing positive, negative, or neutral exploration interests. Based on the identified behavioral patterns, the type of update operation required for the user interest graph is determined, including strengthening existing node associations, weakening existing node associations, and adding new nodes or associations. Based on the intensity information and content identifiers contained in the fine-grained interactive behaviors, the weight values ​​that need to be adjusted for each update operation type are quantitatively calculated. Without interrupting the current graph service, the weight values ​​are incrementally updated and applied to the corresponding nodes and edges of the user interest graph. After the graph update is completed, the updated user interest graph and the latest content feature vector are immediately used as input to automatically trigger a new round of the complete process from matching calculation to generating the recommendation list.

[0053] In practice, the backend analysis server receives and parses fine-grained interactive behavior data streams from user terminal devices in real time. These data streams contain raw interactive events packaged in chronological order, along with additional contextual information. Parsing these data streams involves decoding the data packet format, extracting event types, intensity information, and associated content identifiers, and identifying behavioral patterns. Behavioral pattern identification involves categorizing the raw event sequences into patterns representing positive interest, negative interest, or neutral exploration based on predefined rule mappings. For example, the event sequence "click to enter details page - dwell time exceeds threshold - completion rate exceeds threshold - perform like operation" is identified as a strong positive interest pattern; the event sequence "extremely short cursor hover time - rapid scrolling" is identified as a negative interest pattern; and the event sequence "click to enter details page - medium dwell time - exit without performing any operation" is identified as a neutral exploration pattern.

[0054] In some embodiments, based on the identified behavioral patterns, the system determines the type of update operation required for the user interest graph. The update operation types include strengthening existing node associations, weakening existing node associations, and adding new nodes or new associations. For strong positive interest patterns, the determined update operation type typically strengthens existing node associations and may add new associations; for negative interest patterns, the determined update operation type weakens existing node associations; for neutral exploration patterns, the determined update operation type may involve adding new nodes but establishing associations with lower initial weights. Based on the intensity information and content identifiers contained in the fine-grained interaction behaviors, the system quantitatively calculates the weight values ​​that need to be adjusted for each update operation type, following the following relationship:

[0055] Where: symbol This indicates the weight adjustment values ​​that need to be applied to specific nodes and edges in the user interest graph; symbol Represents the global learning rate coefficient; symbol This represents the total number of events related to this content identifier in the current fine-grained interactive behavior data stream; symbol Indicates the first The type coefficient of each event (e.g., a click event coefficient of 0.3, a full view event coefficient of 1.0, and a like event coefficient of 0.5); symbol Indicates the first The intensity normalized value of an event (e.g., the proportion of dwell time to total time, or the log-normalized value of hover time). If If so, the operation is to strengthen the association by increasing the weight of the corresponding edge; if If the problem involves adding a new node or a new association, the operation is to weaken the association by reducing the weight of the corresponding edge; if it involves adding a new node or a new association, then the operation is to reduce the weight of the corresponding edge. The absolute value is used as the initial weight.

[0056] Understandably, the system incrementally updates and applies the calculated weight values ​​to the corresponding nodes and edges of the user interest graph without interrupting the current graph service. This incremental update is achieved by loading the current state of the user interest graph into memory, directly modifying the weight values ​​of the target nodes and edges in memory, and then asynchronously persisting the modified subgraph structure to the storage database. After completing the graph update, the system immediately uses the updated user interest graph and the latest content feature vectors as input to automatically trigger a new round of the complete process from matching calculation to generating the recommendation list. This means that the matching calculation engine, the simulated feedback environment, and the ranking adjustment module will be called sequentially, ultimately generating an optimized recommendation list based on the latest user interest state. This list can be cached or directly pushed to the user interface for the next round of display.

[0057] Optionally, the timing for triggering a new round of recommendation calculations can be configured to be executed immediately after each graph update, or it can be set to be executed in batches after accumulating multiple minor updates within a short time window, in order to balance system real-time performance and computational resource consumption. The real-time incremental update mechanism of the user interest graph enables the recommendation system to respond to users' immediate feedback and dynamically adjust the recommendation results. The entire process forms a closed loop from user interaction to system updates and then to a new round of recommendations.

[0058] See Figure 5This study presents the weight change characteristics of different interest graph node types. Specifically, it simultaneously displays three core indicators using a dual-axis structure: the left vertical axis represents node weight (0-1), including the weight before the update (blue bars) and the weight after the update (orange bars); the right vertical axis represents the weight update magnitude (red dotted line), reflecting the direction and intensity of weight increases and decreases. Looking at the changes in node weights, the updated weights of the "Education-Programming" and "Technology-Artificial Intelligence" nodes are at relatively high levels, close to 0.95 and 0.95 respectively, and their pre-update weights also remained above 0.8, indicating that these two interest node types had a good initial weight base and were further strengthened after incremental updates. While the updated weight of the "Entertainment-Short Video" node increased, the weight update magnitude was negative, indicating that user interest in this node showed a declining trend during this stage, and only maintained a certain weight level through incremental updates. The line graph showing the change in weight update magnitude reveals the differences in interest dynamics: "Technology-Artificial Intelligence" has the highest weight update magnitude (approximately 0.125), corresponding to the weight strengthening driven by a strong positive interest mode; "Finance-Wealth Management" has a negative weight update magnitude (approximately -0.025), reflecting the weight weakening caused by a negative interest mode; "Lifestyle-Food" and "Health-Exercise" have weight update magnitudes in the range of 0.025-0.03, reflecting a mild weight adjustment under a neutral exploration mode.

[0059] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent recommendation of interactive digital media content based on big data, characterized in that, The method includes: Collect raw data of heterogeneous user interaction behaviors across multiple platforms, perform cross-platform identity mapping and behavior trajectory fusion processing on the raw data of heterogeneous interaction behaviors, and generate a unified user full-lifecycle behavior profile. Based on the user's full-lifecycle behavior profile, a dynamically evolving user interest graph is constructed, which includes content entity nodes, interest tag nodes, and the multidimensional relationships between them. Real-time access to new content streams from the digital media content library; deep semantic deconstruction and multi-dimensional feature extraction for each new piece of content; generating content feature vectors. The user interest graph and the content feature vector are input into the interactive matching engine, and the dynamic matching degree between the nodes in the user interest graph and the content feature vector is calculated through a multi-round bidirectional attention mechanism. An initial set of recommended content is generated based on the dynamic matching degree, and the initial set of recommended content is injected into a simulated feedback environment to simulate the expected interaction. Based on the results of the expected interaction simulation, the initial set of recommended content is sorted, adjusted, and its diversity is controlled to generate an optimized recommendation list.

2. The intelligent recommendation method for interactive digital media content based on big data according to claim 1, characterized in that, The process of performing cross-platform identity mapping and behavior trajectory fusion on the original data of heterogeneous interaction behaviors to generate a unified user lifecycle behavior profile includes: The raw data of the heterogeneous interaction behavior includes explicit feedback records, implicit behavior sequences, and social association data; Device fingerprints, login accounts, and behavioral pattern features are extracted from the raw data of heterogeneous interaction behaviors as identity recognition anchors; The similarity between the identity recognition anchors on different platforms is calculated by using a pre-trained cross-platform identity recognition model. Anchors with similarity exceeding a threshold are clustered and associated to establish a unified identity identifier. Using the unified identity identifier as an index, retrieve all user behavior records across all platforms and sort and align them according to a unified timestamp standard; Standardize the behavior type coding of time-aligned behavior records to eliminate differences in the definition of the same type of behavior across different platforms; The standardized coded behavior records are bound to the corresponding original context information and organized into a structured behavior trajectory chain according to the time sequence, which serves as the user's full-cycle behavior profile.

3. The intelligent recommendation method for interactive digital media content based on big data according to claim 2, characterized in that, The construction of a dynamically evolving user interest graph based on the user's full-lifecycle behavior profile includes: Identify interaction events related to digital media content from the behavioral trajectory chain of the user's full-lifecycle behavior profile; Analyze each interaction event to extract the content object it affects, the type of behavior it applies, and the context in which it occurs; The extracted content objects are used as content entity nodes, and the semantic themes implied in the behavior type and context are used as interest tag nodes, which are then added to the graph. Calculate the association strength between the content entity node and the interest tag node, whereby the association strength is determined based on behavior frequency, behavior depth, and time decay factor. Based on the relationships between all nodes, a weighted directed graph is constructed as the initial user interest graph; Set a graph update cycle. Whenever a new full-cycle user behavior profile is generated, the initial user interest graph is iteratively evolved by adding nodes and adjusting existing edge weights.

4. The intelligent recommendation method for interactive digital media content based on big data according to claim 3, characterized in that, The newly added content stream that is accessed in real time from the digital media content library performs deep semantic deconstruction and multi-dimensional feature extraction on each new piece of content to generate a content feature vector, including: Monitor the update interface of the digital media content library and capture the original data of newly added media content in real time; Multimodal parsing is performed on the raw media content data to separate text description information, visual thumbnails, audio samples, and metadata tags; The text description information is analyzed for topic analysis, sentiment identification, and key entity extraction using a natural language processing model. The visual thumbnails are used to perform scene recognition, subject detection, and style classification using computer vision models. The topic analysis results, sentiment identification results, key entity extraction results, scene identification results, subject object detection results, style classification results, and metadata tags are all input into a feature fusion network. The feature fusion network integrates multi-dimensional information into a high-dimensional dense vector representation, namely the content feature vector, through nonlinear transformation.

5. The intelligent recommendation method for interactive digital media content based on big data according to claim 4, characterized in that, The step of inputting the user interest graph and the content feature vector into the interactive matching engine, and calculating the dynamic matching degree between nodes in the user interest graph and the content feature vector through a multi-round bidirectional attention mechanism, includes: The node set of the user interest graph and the content feature vector are simultaneously input into the interactive matching engine; In the first round of attention calculation, the content feature vector is used as the query vector, and all node features in the user interest graph are used as keys and values. The attention distribution of content to user interest is calculated to obtain the content-aware user interest representation. In the second round of attention calculation, the content-aware user interest representation obtained in the previous round is used as a new query vector, and the content feature vector is used as a key and value. The attention distribution of user interest on the content is calculated to obtain the interest-aware content representation. The content-aware user interest representation and the interest-aware content representation are concatenated as vectors. The concatenated vectors are subjected to dimensionality reduction and comprehensive scoring using a multilayer perceptron, and the output scalar score is the dynamic matching degree.

6. The intelligent recommendation method for interactive digital media content based on big data according to claim 5, characterized in that, The step of generating an initial recommended content set based on the dynamic matching degree and injecting the initial recommended content set into the simulated feedback environment for expected interaction simulation includes: The dynamic matching degree of all candidate content in the digital media content library is sorted, and a preset number of top-ranked content are selected to form the initial recommended content set. Construct a simulated feedback environment, which includes a simulated user model, and the simulated user model is parameterized based on the user interest graph and historical behavior patterns; In the simulated feedback environment, each piece of content in the initial recommended content set is presented to the simulated user model in sequence; Based on the built-in strategy of the simulated user model, predict the type and intensity of its interactive behavior for each displayed content, and generate a simulated feedback sequence. Record the simulated feedback sequence triggered by each piece of content and summarize it into the result of the expected interaction simulation.

7. The intelligent recommendation method for interactive digital media content based on big data according to claim 6, characterized in that, The process of sorting and adjusting the initial recommended content set and controlling its diversity based on the results of the expected interaction simulation to generate an optimized recommended list includes: Analyze the results of the expected interaction simulation and extract the predicted interaction intensity score for each piece of content; Based on the predicted interaction intensity score, the content in the initial recommended content set is reordered; Based on the reordering, calculate the feature similarity matrix between the contents; The maximum boundary correlation algorithm is used to iteratively adjust the sorted list: each time the content of the next position in the list is selected, the content with low feature similarity to the already selected content is selected first, provided that the predicted interaction intensity does not decrease significantly. After multiple rounds of iterative adjustments, a final sequence that strikes a balance between predicted interaction intensity and content diversity is formed, serving as the optimized recommendation list.

8. The intelligent recommendation method for interactive digital media content based on big data according to claim 7, characterized in that, The method further includes: While displaying the optimized recommendation list to target users, implicit feedback tracking is deployed to capture users' fine-grained interaction behaviors with the recommended content in real time, specifically including: On the user interface that renders and displays the optimized recommendation list, invisible monitoring code is embedded for each recommended content item; The monitoring code records the duration of the user's cursor hovering over content items, the speed at which the user scrolls, whether the user clicks to enter the details page, the duration of the user stays on the details page, whether the user performs actions such as liking, collecting, or sharing, and the percentage of users who complete viewing or reading the content. The recorded raw interaction events are packaged in chronological order of occurrence and contextual information is added to form the fine-grained interaction behavior data stream. The fine-grained interactive behavior data stream is sent to the background analysis server in real time.

9. The intelligent recommendation method for interactive digital media content based on big data according to claim 8, characterized in that, The method further includes: The user interest graph is updated incrementally in real time using the fine-grained interaction behaviors captured in real time, and the next round of recommendation calculation is triggered. Specifically, this includes: After receiving the fine-grained interactive behavior data stream, the background analysis server immediately parses it and identifies behavioral patterns that represent positive interest, negative interest, or neutral exploration. Based on the identified behavioral patterns, the type of update operation that needs to be performed on the user interest graph is determined. The update operation type includes strengthening existing node associations, weakening existing node associations, and adding new nodes or new associations. Based on the intensity information and content identifiers contained in the fine-grained interactive behaviors, the weight values ​​that need to be adjusted for the update operation type are quantitatively calculated. Without interrupting the current graph service, the weight values ​​are updated and applied to the corresponding nodes and edges of the user interest graph in an incremental manner. After the graph update is completed, the updated user interest graph and the latest content feature vector are immediately used as input to automatically trigger a new round of the complete process from matching calculation to generating the recommendation list.

10. The intelligent recommendation method for interactive digital media content based on big data according to claim 5, characterized in that, The construction steps of the interactive matching engine include: Define the architecture of the matching engine, which includes an input interface, at least two attention calculation layers, a feature fusion layer, and a scoring output layer; Configure the input interface to simultaneously receive the node feature matrix from the user interest graph and the content feature vector from the content processing flow; A first attention calculation layer is constructed, in which the content feature vector is used as the query input and the node feature matrix is ​​used as the key and value input to realize the attention calculation from content to user interest and output the content-aware interest representation vector. A second attention calculation layer is constructed, which takes the content-aware interest representation vector as the query input and the content feature vector as the key and value input to realize the attention calculation from user interest to content and output the interest-aware content representation vector. In the feature fusion layer, the content-aware interest representation vector and the interest-aware content representation vector are concatenated to form a joint representation vector; In the scoring output layer, a fully connected neural network is set up to perform nonlinear transformation and dimensionality reduction on the joint representation vector, and finally output a scalar score as the dynamic matching degree. The completed interactive matching engine is trained offline, using historical user interaction data as a supervision signal to optimize the model parameters in the attention calculation layer and the scoring output layer.

Images