Short drama personalized content pushing method and system based on big data mining

Abstract

This application relates to the field of data processing technology and discloses a method and system for personalized content recommendation of short dramas based on big data mining. The method includes: analyzing short drama features to generate content vectors; analyzing user behavior based on the content vectors to obtain emotional patterns; constructing an interest graph and training an emotional resonance network to form a recommendation model; applying an emotional rhythm algorithm to sort and generate a recommendation list; generating multimodal preview resources; analyzing user interaction behavior to adjust emotional weights and optimize the recommendation effect. This application addresses the problems of existing short drama recommendation systems that overemphasize content type matching while neglecting the emotional experience dimension, lack overall planning of emotional sequences, and have a single content preview method. By constructing a recommendation mechanism based on emotional resonance and a multimodal preview system, it improves the emotional matching accuracy of recommended content and user satisfaction.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for personalized content push of short dramas based on big data mining. Background Technology

[0002] The rapid development of short video platforms has spawned a massive amount of short drama content. Faced with this vast sea of ​​content, users often struggle to efficiently find content that matches their interests and emotional needs. Existing short drama recommendation technologies primarily rely on collaborative filtering, content matching, and deep learning. These technologies analyze users' historical behavioral data and content characteristics to calculate the match between users and content, thereby generating personalized recommendation lists. These technologies have achieved some success in improving content distribution efficiency and user experience, enabling users to more easily access short drama content that interests them, and allowing platforms to more accurately push content to potential audiences.

[0003] However, traditional recommendation algorithms overemphasize content type and theme matching, neglecting users' emotional experience needs. This results in recommendations with a singular emotional dimension, failing to meet users' diverse emotional expectations. Secondly, existing recommendation systems generally lack the ability to finely capture and analyze the emotional characteristics of content and users' emotional responses, making it difficult to accurately identify which content elements truly resonate with users. Furthermore, traditional recommendation systems lack overall planning in organizing recommended content, simply pursuing the matching degree of individual content with users, ignoring the impact of the content sequence as a whole on users' emotional experience. In addition, existing content preview methods are too simplistic, failing to effectively convey the emotional value and core appeal of short dramas, affecting the efficiency and accuracy of users' content selection. These problems severely restrict the user experience and content value delivery efficiency of short drama platforms. Summary of the Invention

[0004] This application provides a method and system for personalized content recommendation of short dramas based on big data mining, which addresses the problems of existing short drama recommendation systems that overemphasize content type matching while neglecting the emotional experience dimension, lack overall planning of emotional sequences, and have a single content preview method. By constructing a recommendation mechanism based on emotional resonance and a multimodal preview system, the method improves the emotional matching accuracy of recommended content and user satisfaction.

[0005] Firstly, this application provides a method for personalized short drama content recommendation based on big data mining. The method includes: extracting features and constructing tags for short drama content to generate a short drama content feature vector; collecting and performing sentiment analysis on user behavior data based on the short drama content feature vector to obtain user sentiment pattern data; constructing a user interest graph based on the short drama content feature vector and the user sentiment pattern data, and training the user interest graph using an emotional resonance network to obtain a content recommendation model; applying the content recommendation model to a candidate content pool, and sorting the candidate short dramas using an emotional rhythm planning algorithm to obtain a short drama recommendation list; generating multimodal preview resources for the short drama content in the short drama recommendation list to obtain content preview data; collecting user interaction behavior data with the content preview data, and calculating an emotional matching score based on user completion rate, repeated viewing times, and social sharing behavior; adjusting the emotional weight coefficients in the content recommendation model based on the emotional matching score to generate a target short drama recommendation list.

[0006] Secondly, this application provides a personalized short drama content push system based on big data mining. The personalized short drama content push system based on big data mining includes: a construction module, used to perform feature extraction and tag construction processing on short drama content to generate short drama content feature vectors;

[0007] The analysis module is used to collect and process user behavior data and sentiment analysis based on the feature vector of the short drama content to obtain user sentiment pattern data.

[0008] The training module is used to construct a user interest graph based on the short drama content feature vector and the user emotional pattern data, and to train the emotional resonance network on the user interest graph to obtain a content recommendation model.

[0009] The sorting module is used to apply the content recommendation model to the candidate content pool, and sort the candidate short dramas through the emotional rhythm planning algorithm to obtain a short drama recommendation list;

[0010] The generation module is used to perform multimodal preview resource generation processing on the short drama content in the short drama recommendation list to obtain content preview data;

[0011] The recommendation module is used to collect user interaction data with the content preview data, and calculate the sentiment matching score by combining user viewing completion rate, number of repeated viewings and social sharing behavior. Based on the sentiment matching score, the sentiment weight coefficient in the content recommendation model is adjusted to generate a target short drama recommendation list.

[0012] Thirdly, a short drama personalized content push device based on big data mining is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory to cause the short drama personalized content push device based on big data mining to execute the aforementioned short drama personalized content push method based on big data mining.

[0013] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, cause the computer to execute the above-described method for personalized short drama content push based on big data mining.

[0014] The technical solution provided in this application generates content feature vectors by extracting features and constructing tags from short drama content, thus achieving accurate expression of multi-dimensional information about the short drama content. Based on the short drama content feature vectors, user behavior data is collected and processed through sentiment analysis to obtain user emotional pattern data. This breaks through the limitation of traditional recommendation systems that only focus on explicit behavior. Innovative technologies such as micro-expression recognition capture users' true emotional preferences, significantly improving the accuracy and depth of user profiles. A user interest graph is constructed based on the short drama content feature vectors and user emotional pattern data, and an emotional resonance network is trained to obtain a content recommendation model. This innovative combination of graph neural networks and emotional resonance enhancement networks not only identifies users' content type preferences but also accurately captures content features that evoke emotional resonance, achieving a recommendation dimension upgrade from "content matching" to "emotional resonance." The recommendation model is applied to the candidate content pool and sorted using an emotional rhythm planning algorithm to obtain a short drama recommendation list. It creatively introduces the rhythm construction principle from music theory, transforming the recommendation results from isolated content sets into a sequence of content with artistic emotional rhythm, significantly improving the user's continuous viewing experience. Multimodal preview resource generation is performed on the short drama content in the recommendation list to obtain content preview data. This breaks through the traditional single preview mode of static covers and text descriptions, greatly improving the value delivery efficiency of short drama content through multimodal preview methods such as dynamic covers with emotional climaxes and emotional concentration heatmaps. User interaction with preview data is collected, and emotional matching scores are calculated based on indicators such as viewing completion rate. The model parameters are dynamically adjusted to generate the target recommendation list, constructing a closed-loop optimization mechanism that enables the recommendation system to continuously self-optimize and adapt to changes in user emotional preferences. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of an embodiment of the short drama personalized content push method based on big data mining in this application.

[0017] Figure 2 This is a schematic diagram of an embodiment of the personalized short drama content push system based on big data mining in this application.

[0018] Figure 3 This is a schematic block diagram of the structure of the short drama personalized content push device based on big data mining in an embodiment of the present invention. Detailed Implementation

[0019] This application provides a method and system for personalized short drama content recommendation based on big data mining. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0020] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the personalized short drama content push method based on big data mining in this application includes:

[0021] Step S101: Extract features and construct tags for the short drama content to generate a short drama content feature vector;

[0022] Step S102: Collect and process user behavior data and sentiment analysis based on the short drama content feature vector to obtain user sentiment pattern data;

[0023] Step S103: Construct a user interest graph based on the short drama content feature vector and user emotional pattern data, and train the emotional resonance network on the user interest graph to obtain a content recommendation model;

[0024] Step S104: Apply the content recommendation model to the candidate content pool, and sort the candidate short dramas using the emotional rhythm planning algorithm to obtain the short drama recommendation list;

[0025] Step S105: Perform multimodal preview resource generation processing on the short drama content in the short drama recommendation list to obtain content preview data;

[0026] Step S106: Collect user interaction data with content preview data, and calculate sentiment matching score by combining user completion rate, number of repeated views and social sharing behavior. Adjust the sentiment weight coefficient in the content recommendation model according to the sentiment matching score to generate a target short drama recommendation list.

[0027] It is understood that the executing entity of this application can be a personalized short drama content recommendation system based on big data mining, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiment uses a server as an example for illustration.

[0028] Specifically, keyframes of the short drama videos are sampled, and image recognition algorithms are used to extract scene type and visual style features. For example, through keyframe analysis of a certain urban emotional short drama, visual features such as "coffee shop scene," "warm color tone," and "close-up shot" are identified. Simultaneously, word segmentation and semantic analysis are performed on the dialogue and subtitles of the short drama to extract thematic keywords and emotional tendency data, such as identifying keywords like "workplace," "struggle," and "emotional dilemma," and determining the overall emotional tone to be "inspirational and positive." Furthermore, statistical analysis is conducted on editing speed and transition methods to obtain rhythmic features and narrative structure data. For example, the average shot length of this short drama is 4 seconds, with fast editing accounting for 70%, reflecting the characteristic of "tight pacing." Audio features are extracted from sound elements to obtain musical tone and emotional rendering features. For example, the background music is mainly piano with a soothing rhythm, creating a "warm atmosphere." These multi-dimensional features are then fused to construct a label vector containing 50 dimensions, and dimensionality reduction is used to generate the final short drama content feature vector.

[0029] User viewing behavior data is collected through a real-time data stream processing framework, including the duration of user dwell time at different short plot points, repeated viewing of segments, and fast-forwarding / rewinding behaviors. For example, analysis of a user's viewing data shows that the user spends more time on segments containing "workplace competition" plots, while frequently fast-forwarding on "romantic entanglements" plots. Simultaneously, micro-expression recognition technology is used to collect changes in users' facial expressions during viewing; for example, slightly raised eyebrows and slightly moist eyes when watching inspirational scenes indicate positive emotional resonance. Analysis of user comment text content extracts emotional polarity and intensity; for example, 85% of positive words in user comments on inspirational content indicate strong identification. By aligning behavioral data with content feature vectors over time, an emotional response-content node mapping matrix is ​​constructed to accurately locate content feature points that trigger user emotional fluctuations. By integrating micro-expression responses (implicit emotions) and comment expressions (explicit emotions), a two-tiered emotional model is formed to comprehensively capture users' true emotional preferences.

[0030] When constructing a user interest graph and training an emotional resonance network based on short drama content feature vectors and user emotional pattern data, a knowledge graph technique is used to establish a user-centric interest graph. Nodes include content categories, specific short dramas, and emotional tags, with edge weights representing the strength of association. Graph neural networks are applied to this graph for representation learning, compressing the complex graph structure into low-dimensional dense vectors. A three-component neural network architecture is constructed, including an emotional trajectory extractor, a resonance matcher, and an emotional reinforcement predictor. The emotional trajectory extractor analyzes the emotional change curve of short drama content over time, identifying emotional climaxes and turning points; the resonance matcher calculates the similarity between user emotional patterns and content emotional trajectories, identifying the content segments that most resonate with users; and the emotional reinforcement predictor predicts the intensity of the emotional impact of specific content on users based on historical data. The network is trained using a contrastive learning method, with content that elicits a strong emotional response as positive samples and content with a lukewarm response as negative samples, ultimately forming a dual-channel recommendation model that simultaneously considers content semantic matching and emotional resonance.

[0031] When applying a content recommendation model to generate a recommendation list from a candidate content pool, the first step is to construct a candidate content pool containing newly released short dramas, popular short dramas, and short dramas similar to the user's historical preferences. For each short drama in the pool, a basic recommendation score is calculated using the content recommendation model. Based on emotional type, short dramas are categorized into three types: high-energy (e.g., inspiring, tense, funny), medium-energy (e.g., heartwarming, motivational), and low-energy (e.g., sad, calm). Multiple emotional sequence templates are constructed, such as "gradual climax type" (low-medium-high energy order) and "emotional fluctuation type" (high-low-high energy alternation). The emotional sequence preferences shown in the user's historical viewing behavior are analyzed, and the most suitable template is selected. For example, data shows that a user prefers to watch light content first on weekday evenings, then transitions to content with stronger emotional depth, exhibiting a clear "gradual climax type" preference. Based on the selected template and basic recommendation score, the candidate short dramas are dynamically sorted and combined to form a short drama recommendation list with a preset emotional rhythm.

[0032] When generating multimodal preview resources for the content in the short drama recommendation list, emotional intensity curve analysis is performed on the content of each short drama, and video clips with emotional intensity values ​​exceeding a preset threshold are extracted to generate dynamic preview materials; key sentences are extracted and prioritized in importance from the plot text, and converted into personalized voice summaries; the emotional value change data throughout the entire process is visualized to generate an emotional change graph with time on the horizontal axis and emotional intensity on the vertical axis; a directed graph structure is constructed based on character relationship data, and node positions are calculated using a force-directed layout algorithm to generate an interactive character relationship graph. These preview resources are adapted according to network bandwidth and device performance to form multi-level preview resource packages suitable for different terminal devices. Finally, the layout engine is used to arrange components and bind interactive events to generate structured content preview data. When collecting and optimizing user interaction with content preview data, the process involves event capture of the user's interaction with the preview content, recording interaction trajectories such as click areas, dwell time, and browsing paths. Statistical analysis is performed on user behavior metrics after watching short dramas, including completion rate, number of times segments are rewatched, frequency of interactive comments, and number of social sharing sessions. The interaction behavior is correlated with viewing metrics to construct a user-content interaction intensity matrix. An emotional satisfaction calculation model is applied to determine the user's preference for different emotional content types, generating an emotional matching score. Based on the emotional matching score, the weight coefficients of each emotional dimension in the recommendation model are optimized and adjusted, updating the emotional weight parameters. The updated parameters are then applied to the candidate content pool for re-sorting, generating a target short drama recommendation list that better matches the user's latest emotional preferences. For example, when a user shows a high completion rate and multiple social sharing behaviors for recommended heartwarming short dramas, the system calculates that the user's matching score for the "heartwarming" emotional dimension is 0.92, thereby increasing the weight coefficient of the "heartwarming" dimension in the recommendation model, increasing the proportion of heartwarming short dramas in subsequent recommendation results, and at the same time, considering more heartwarming content as the climax in the emotional sequence design, thereby continuously optimizing the personalized push effect.

[0033] In this embodiment, by extracting features and constructing tags from the short drama content, a content feature vector is generated, achieving accurate expression of multi-dimensional information about the short drama content. Based on the short drama content feature vector, user behavior data is collected and processed for sentiment analysis to obtain user sentiment pattern data. This breaks through the limitation of traditional recommendation systems that only focus on explicit behaviors. By using innovative technologies such as micro-expression recognition to capture users' true emotional preferences, the accuracy and depth of user profiles are significantly improved. A user interest graph is constructed based on the short drama content feature vector and user sentiment pattern data, and an emotional resonance network is trained to obtain a content recommendation model. This innovatively combines graph neural networks with emotional resonance enhancement networks, which can not only identify users' content type preferences but also accurately capture content features that can evoke emotional resonance in users, achieving a recommendation dimension upgrade from "content matching" to "emotional resonance." The recommendation model is applied to the candidate content pool and sorted using an emotional rhythm planning algorithm to obtain a short drama recommendation list. It creatively introduces the rhythm construction principle from music theory, transforming the recommendation results from isolated content sets into a sequence of content with artistic emotional rhythm, significantly improving the user's continuous viewing experience. Multimodal preview resource generation is performed on the short drama content in the recommendation list to obtain content preview data. This breaks through the traditional single preview mode of static covers and text descriptions, greatly improving the value delivery efficiency of short drama content through multimodal preview methods such as dynamic covers with emotional climaxes and emotional concentration heatmaps. User interaction with preview data is collected, and emotional matching scores are calculated based on indicators such as viewing completion rate. The model parameters are dynamically adjusted to generate the target recommendation list, constructing a closed-loop optimization mechanism that enables the recommendation system to continuously self-optimize and adapt to changes in user emotional preferences.

[0034] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0035] Keyframes are extracted from short drama videos and analyzed using image recognition algorithms to obtain scene type and visual style features;

[0036] The dialogue and subtitles of the short drama were segmented and semantically analyzed to obtain thematic keywords and sentiment data.

[0037] Statistical analysis was performed on the editing speed and transition methods of short dramas to obtain data on rhythm characteristics and narrative structure;

[0038] Audio feature extraction processing is performed on the sound elements of the short drama to obtain the musical tone and emotional rendering features;

[0039] The visual style features, sentiment data, narrative structure data, and emotion rendering features are fused together to construct a multi-dimensional label vector.

[0040] The multidimensional label vectors are standardized and dimensionality reduced to generate short drama content feature vectors.

[0041] Specifically, keyframes are extracted from short drama videos using a combination of uniform sampling and scene change perception, at a frequency of 1-2 frames per second. Keyframe extraction employs a scene boundary detection algorithm, calculating the color histogram differences and edge change rates between adjacent frames. When the difference exceeds a preset threshold, it is identified as a scene transition point, and the corresponding frame is extracted as a keyframe. The extracted keyframes are then used for scene recognition via a deep convolutional neural network pre-trained on a scene classification dataset, capable of classifying keyframes into scene types such as "coffee shop," "office," and "family living room." Visual style feature extraction involves calculating parameters such as the color distribution histogram, mean saturation, contrast, and texture complexity of the keyframes to analyze whether the color tone is warm or cool, whether the composition is symmetrical, and whether the depth of field is shallow or deep, thereby quantifying the visual style features of the short drama.

[0042] When processing the dialogue and subtitles of the short drama, the dialogue is first converted into text using speech recognition technology, and combined with the subtitle file to form complete text content. Chinese word segmentation technology is used to divide the continuous text into meaningful word units. After removing stop words, a word frequency-inverse document frequency (TF-IDF) matrix is ​​constructed, and high-weight words in the text are extracted as topic keywords. Simultaneously, sentiment lexicon matching and deep semantic analysis models are used to calculate the sentiment polarity and intensity of the text, determining whether the text content is positive, negative, or neutral, and whether the sentiment intensity is high, medium, or low, thus obtaining sentiment tendency data. The editing rate and transition methods of the short drama are processed through video frame analysis. First, the average duration of all shots in the short drama is calculated to obtain the editing rate index. Edge detection and optical flow estimation algorithms are used to identify transition types, classifying transitions into categories such as hard cut, fade in / fade out, wipe, and dissolve, and the frequency and time distribution of each type of transition are statistically analyzed. Combining the statistical data on shot length distribution and transition methods, a rhythm curve of the short drama is constructed to reflect the plot's ups and downs and changes in narrative rhythm, thus obtaining rhythmic features and narrative structure data.

[0043] When extracting audio features from the sound elements of a short drama, the audio signal is first analyzed in the time-frequency domain to separate three channels: vocals, background music, and ambient sound. For the background music, pitch analysis, beat detection, and chord recognition are performed to obtain parameters such as the key (major / minor), rhythmic characteristics (slow, medium, fast), and harmonic complexity. Combining volume variation curves and timbre characteristics, the role of music in emotional rendering is analyzed, such as whether it enhances tension, warmth, or humor, thus obtaining the music's key and emotional rendering features. When fusing the extracted visual style features, emotional tendency data, narrative structure data, and emotional rendering features, a multimodal feature fusion method is used. First, all features are normalized to convert feature data of different dimensions to a unified numerical range. Then, feature importance analysis is performed to calculate the contribution of each feature to the expression of the short drama's content features, and corresponding weights are set. Through weighted fusion or tensor fusion methods, the various feature data are merged into a fusion feature vector containing 100-200 dimensions, forming a multidimensional label vector.

[0044] When standardizing and reducing the dimensionality of the multidimensional label vector, the Z-score standardization method is first used to adjust the features of each dimension to a distribution with a mean of 0 and a standard deviation of 1. Then, dimensionality reduction algorithms such as Principal Component Analysis (PCA) or t-SNE are used to compress the high-dimensional feature space to about 50 dimensions, retaining key information while reducing data redundancy and computational complexity. The final generated short drama content feature vector is a compact and information-rich vector representation, used for subsequent similarity calculations and recommendation model training.

[0045] For example, approximately 1800 keyframes were extracted from the 25-minute short drama. Scene recognition identified "office" scenes accounting for 45%, "business meeting" scenes for 30%, and other scenes for 25%. Visual style analysis showed that the visuals were predominantly blue-gray, with a relatively regular and symmetrical composition and moderate depth of field. Dialogue and subtitle analysis extracted core keywords such as "workplace competition," "teamwork," and "career growth" from approximately 3000 words of text. Sentiment analysis showed that the overall emotional tone of the text was "positive and uplifting," but there were emotional fluctuations of "pressure" and "challenge." Editing analysis indicated that the average shot length of the short drama was 4.2 seconds, which is considered a medium editing rate, mainly using hard cuts and transitions, with a significantly faster pace during the climax. Audio analysis showed that the background music was mainly piano and strings, with a medium to fast tempo, and the volume and pace increased significantly during the climax to enhance the emotions of "tension" and "striving." After normalizing and weighting these feature data, an initial feature vector of 173 dimensions is obtained. After PCA dimensionality reduction, a short drama content feature vector of 48 dimensions is obtained. This vector accurately captures the content characteristics of this short drama: "urban workplace, positive and enterprising, and fast-paced".

[0046] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0047] The real-time data stream processing framework collects user viewing behavior from all dimensions, resulting in behavioral data including viewing duration, completion rate, number of pauses, fast forward and rewind behaviors, and repeated viewing segments.

[0048] Facial micro-expression data collected by the front-facing camera of the user's device is processed to obtain the emotional response characteristics of the user during the viewing process;

[0049] Natural language processing analysis was performed on the text content of user comments to obtain explicit sentiment expression data containing sentiment polarity and sentiment intensity;

[0050] The behavioral data and the feature vectors of short drama content are time-series aligned to construct an emotional response-content node mapping matrix.

[0051] Multi-source fusion processing of emotional response characteristics and explicit emotional expression data yields a two-level emotional pattern that includes implicit and explicit emotions;

[0052] An emotional fluctuation model is constructed based on a two-level emotional pattern and an emotional response-content node mapping matrix to obtain user emotional pattern data.

[0053] Specifically, a real-time data stream processing framework is used to collect user viewing behavior across all dimensions. This framework employs an event-driven architecture, embedding a data collection SDK on the user's end. Each time a user interacts with the short drama application, corresponding event collection is triggered. The collected data includes viewing duration data (recording the actual viewing time from start to finish), completion rate data (calculating the ratio of the complete viewing time to the total duration of the short drama), pause counts (recording the frequency of users actively clicking the pause button), fast-forward and rewind behaviors (recording the direction, duration, and frequency of users dragging the progress bar), and repeated viewing segment data (identifying the video segments and number of times users rewatched multiple times). This raw behavioral data undergoes preprocessing, including outlier filtering, missing value imputation, and time window segmentation, forming a standardized user behavior dataset. When processing facial micro-expression data collected by the user's front-facing camera, user authorization is first obtained. Then, a facial keypoint detection algorithm is used to extract 68 feature points from the user's face, tracking subtle changes in these points during the viewing process. Micro-expression recognition employs deep learning methods, inputting extracted facial feature sequences into a pre-trained expression recognition neural network. This network maps facial expression changes to basic emotion categories (such as happiness, sadness, surprise, anger, fear, disgust, and neutrality) and their intensity values. To improve recognition accuracy, ambient light sensor data from the user's device is used for illumination compensation, reducing the interference of ambient light changes on expression recognition. By analyzing micro-expression changes in real time as the user views different content segments, the system accurately captures the user's emotional response characteristics, such as a smile in a humorous scene or a slight frown in a touching scene, reflecting empathy.

[0054] When performing natural language processing analysis on user comment text, raw text data is collected from user comments on short dramas, bullet comments, and content shared on social media platforms. First, text preprocessing is performed, including word segmentation, stop word removal, and part-of-speech tagging. Then, sentiment analysis is conducted using a combination of a sentiment dictionary and deep learning. The sentiment dictionary is used to determine the polarity and intensity of sentiment words in the text, while pre-trained language models such as BERT are applied to capture contextual semantics, comprehensively judging the sentiment polarity (positive, negative, or neutral) and sentiment intensity (high, medium, or low). For complex expressions such as irony and metaphor, special processing is performed using a contextual semantic understanding model to ensure the accuracy of sentiment judgment. Finally, explicit sentiment expression data containing sentiment polarity and intensity is generated, directly reflecting the emotional attitudes expressed by users through text.

[0055] A crucial step is to align user behavior data with the feature vectors of short drama content in a time-series manner. First, a timeline mapping relationship is established, aligning the timestamps of user behavior data with the timeline of the short drama content. Using a sliding window technique, the short drama content is divided into fixed time windows (e.g., 5 seconds or 10 seconds), and a corresponding content feature sub-vector is assigned to each window. Simultaneously, user behavior data and micro-expression data are aggregated within the same time window. Through the aligned time windows, a correspondence between user emotional responses and short drama content nodes is established, constructing an emotional response-content node mapping matrix. The rows of this matrix represent different time windows, and the columns include the content features of that window and the corresponding user emotional response data, forming a content-response correspondence graph.

[0056] When performing multi-source fusion processing on emotional response features and explicit emotional expression data, a combination of weighted fusion and probabilistic models is employed. First, weight coefficients are assigned to emotional data from different sources, with weight values ​​determined based on data reliability and directness of expression. For example, micro-expression data (implicit emotion) may have a higher weight than text comments (explicit emotion) because micro-expressions are more difficult to fake and reflect a more authentic emotional state. Then, a Bayesian network model is used, with emotional data from different sources as observed variables and users' true emotional preferences as latent variables, to calculate the most probable user emotional state. Ultimately, a two-tiered emotional model is formed, encompassing implicit emotions (subconscious emotional responses reflected in micro-expressions) and explicit emotions (subjective emotional attitudes expressed in text comments), comprehensively capturing users' emotional preference characteristics.

[0057] When constructing an emotional fluctuation model based on a two-level emotional pattern and an emotional response-content node mapping matrix, a combination of time series pattern mining and machine learning is employed. First, time series analysis is applied to the emotional response-content node mapping matrix to extract the fluctuation patterns of user emotions as content changes, such as which types of content elicit strong emotional responses and which elicit neutral responses. Then, combined with historical viewing data, a sequence prediction model (such as a Long Short-Term Memory network, LSTM) is applied to learn the patterns of user emotional changes and predict the potential emotional impact of different types of content on the user. Through cluster analysis, the main types and characteristics of user emotional preferences are identified, such as "emotional resonance," "rational analysis," or "light entertainment." Finally, user emotional pattern data is generated, including user emotional preference characteristics, emotional fluctuation patterns, and emotional sensitivity to different content types.

[0058] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0059] Based on the short drama content feature vectors and user emotional pattern data, a knowledge graph structure with users as the central nodes is constructed to obtain an initial interest graph containing content categories, short drama entities, and emotional tags;

[0060] The initial interest map is processed by a graph neural network algorithm to obtain a low-dimensional dense vector that captures the structural features of user interests.

[0061] A three-component neural network architecture consisting of an emotion trajectory extractor, a resonance matcher, and an emotion enhancement predictor is constructed to obtain an emotion resonance enhancement network model.

[0062] The emotional change data of the short drama content is input into the emotional trajectory extractor for processing to obtain the emotional trajectory of the content that captures the emotional climax and turning point in the development of the plot.

[0063] The content emotional trajectory and user emotional pattern data are input into the resonance matcher for similarity calculation to obtain the user-content emotional resonance metric.

[0064] By performing deep feature fusion processing based on user-content sentiment resonance metric and low-dimensional dense vector, a content recommendation model integrating both semantic matching and sentiment resonance capabilities is obtained.

[0065] Specifically, a user-centric knowledge graph is constructed based on the short drama content feature vectors and user sentiment pattern data. This knowledge graph is a multi-relational directed graph structure, where nodes include user nodes, content category nodes, short drama entity nodes, and sentiment tag nodes. During construction, the user is set as the central node. By analyzing the user's historical interaction data, relational edges between the user and various types of nodes are established. The edge between the user and the content category node represents the intensity of interest, ranging from 0 to 1, calculated based on the proportion of each type of content in the user's viewing history and the completion rate of viewing. The edge between the user and the short drama entity node represents the intensity of interaction, calculated comprehensively based on indicators such as viewing duration, completion rate, and number of repeated views. The edge between the user and the sentiment tag represents the degree of sentiment preference, determined based on the intensity of the user's response to different sentiment types of content. Simultaneously, relational edges are also established between nodes, such as the "belonging" relationship between short drama entities and content categories, and the "triggering" relationship between short drama entities and sentiment tags, forming a complete initial interest graph containing content categories, short drama entities, and sentiment tags. When applying graph neural network algorithms to the initial interest graph for representation learning, a graph convolutional network (GCN) is used as the base model. A GCN is a deep learning model specifically designed for graph-structured data, capable of learning features from nodes in the graph through a message passing mechanism. Specifically, initial feature vectors are first assigned to various types of nodes in the graph. User nodes' initial features are derived from user sentiment pattern data, content category nodes from category semantic vectors, short drama entity nodes from short drama content feature vectors, and sentiment tag nodes from sentiment semantic vectors. Then, through multi-layer graph convolution operations, each node continuously updates its features by aggregating the feature information of its neighbors. After multiple iterations, the node features incorporate graph structure information, reflecting the semantic position and relational structure of the nodes within the entire interest network. Finally, the final feature vectors of user nodes are extracted and, after dimensionality reduction, transformed into low-dimensional dense vectors of less than 200 dimensions. These vectors effectively capture the structured features of user interests, including the diversity, focus, and conversion patterns of interests.

[0066] Constructing an emotional resonance enhancement network model is crucial for achieving emotion-oriented recommendations. This network comprises three core components: an emotion trajectory extractor, a resonance matcher, and an emotion reinforcement predictor. The emotion trajectory extractor is a temporal model based on a Long Short-Term Memory (LSTM) network, specifically designed to capture the emotional development trajectory of short drama content over time. It receives a sequence of emotion vectors from time slices of the short drama as input, with each time slice containing information on the emotion type and intensity. Through the gating mechanism and memory units of the LSTM, it captures the long-term and short-term dependencies of emotional changes, identifying emotional climaxes (sudden increases in emotional intensity) and emotional turning points (changes in emotion type). The resonance matcher is a neural network model based on an attention mechanism. It receives user emotional pattern data and content emotional trajectories as dual inputs. It extracts key features of user emotional preferences and content emotional trajectories through self-attention layers, and then calculates the correspondence between them through a cross-attention mechanism to identify the content segments most likely to evoke emotional resonance in the user. The emotion reinforcement predictor is a feedforward neural network that, based on the user's historical emotional response data, predicts the intensity of the emotional impact of specific content on the user, providing an emotional guidance basis for the final recommendation decision. The three components are trained end-to-end to form a unified emotional resonance enhancement network model.

[0067] When processing the emotional change data of a short drama content into an emotional trajectory extractor, the first step is to construct the emotional change data of the short drama. Through emotional analysis of the short drama content, including sentiment analysis of dialogue, background music, and visual tone analysis, emotional type and intensity data arranged in a time sequence are obtained. This data is then divided into fixed time windows (e.g., 5 seconds) to form an emotional vector sequence. Each vector contains the dominant emotional type (e.g., joy, tension, sadness) and its corresponding intensity value for that time window. This sequence data is input into the LSTM network of the emotional trajectory extractor. Through a forward propagation process, the LSTM unit progressively processes the emotional vector of each time segment while maintaining its internal state, capturing the temporal pattern of emotional changes. During processing, when the emotional intensity value suddenly increases or the emotional type changes significantly, a threshold is set to mark it as an emotional climax or turning point. The output emotional trajectory is an enhanced sequence containing the original emotional sequence and marked key points, accurately describing the emotional change curve of the short drama content over time.

[0068] When inputting content emotional trajectory and user emotional pattern data into the resonance matcher for similarity calculation, a multi-scale matching algorithm is employed. First, the user emotional pattern data is processed to extract user preference weights and emotional sensitivity parameters for different emotional types. Then, the matching degree between each time segment of the content emotional trajectory and the user's emotional preferences is calculated. For each time segment's emotional vector, its cosine similarity to the user's emotional preference vector is calculated to obtain a basic matching score. However, considering that emotional resonance is affected not only by a single moment but also by the emotional development process, a sliding window method is used to calculate the overall matching degree between the emotional sequence and user preferences under different window lengths. Finally, based on the importance of emotional climaxes and turning points, higher weights are assigned to the matching scores of these key moments, and a comprehensive calculation yields the user-content emotional resonance metric, which reflects the degree of fit between the content's emotional development process and the user's emotional preferences.

[0069] Deep feature fusion based on user-content sentiment resonance metrics and low-dimensional dense vectors is crucial for forming the final recommendation model. A dual-channel architecture is employed: on one hand, traditional collaborative filtering and content matching are used to calculate the user's interest matching degree with the content, primarily based on low-dimensional dense vectors obtained from graph neural networks; on the other hand, a sentiment resonance network calculates the user's emotional resonance with the content. These two channels yield semantic matching scores and sentiment resonance scores, respectively. Then, a feature fusion layer combines these two scores with other auxiliary features (such as content popularity and novelty) and inputs them into the final ranking model. The ranking model uses Gradient Boosting Decision Tree (GBDT) or a deep neural network, adaptively adjusting the weights of each feature by learning from historical recommendation performance, ultimately outputting a recommendation score. This entire process forms a content recommendation model integrating both semantic matching and sentiment resonance capabilities, considering both user content type preferences and emotional experience needs.

[0070] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0071] Construct a candidate content pool that includes newly launched short dramas, popular short dramas, and short dramas similar to users' historical preferences to obtain an initial set of candidate short dramas;

[0072] For each short drama in the initial candidate short drama set, a matching degree calculation is performed using a content recommendation model to obtain a basic recommendation score;

[0073] The short drama content was categorized by emotional type, resulting in an emotional classification that includes high-energy, medium-energy, and low-energy emotional types.

[0074] Based on the results of emotion classification, various emotion sequence templates were constructed to obtain an emotion sequence library containing progressive climax, emotional fluctuation, and soothing transition types.

[0075] The user's historical viewing behavior is analyzed for sentiment sequence preferences to obtain the most suitable sentiment sequence template for the user at present;

[0076] The initial candidate short drama set is dynamically sorted and combined based on the emotional sequence template and the basic recommendation score to obtain a short drama recommendation list with a preset emotional rhythm.

[0077] Specifically, a candidate content pool is constructed using a three-source fusion strategy, selecting content from the newly released short drama library, the popular short drama library, and the user-similar content library. Newly released short dramas are filtered using a time window, selecting those released within the last 7 days and arranged in reverse chronological order. Popular short dramas are selected based on user behavior data across the platform, calculating a popularity score for each drama. This score considers four dimensions: views, completion rate, interaction, and dissemination, selecting the top 100 most popular dramas. Short dramas similar to users' historical preferences are selected using a content collaborative filtering algorithm, calculating the similarity to other dramas in the library based on the content feature vectors of dramas already watched by the user, selecting the top 200 most similar dramas. After deduplication and merging of short dramas from the three sources, an initial candidate short drama set of approximately 300 dramas is formed, ensuring the diversity, timeliness, and personalization of the recommended content. When calculating the matching degree for each drama in the initial candidate set, the aforementioned content recommendation model is used as the scoring engine. For each candidate short drama, its content feature vector and emotional trajectory data are extracted and matched with the user's interest vector and emotional pattern data. The matching calculation consists of two dimensions: first, semantic content matching, which calculates the cosine similarity between the short drama's content feature vector and the user's interest vector to obtain a content matching score; second, emotional resonance matching, which utilizes an emotional resonance enhancement network to calculate the fit between the short drama's emotional trajectory and the user's emotional pattern to obtain an emotional resonance score. The two scores are then weighted and fused according to preset weights (typically content matching accounts for 0.6, and emotional resonance accounts for 0.4) to obtain a base recommendation score for each short drama. The score ranges from 0 to 1, with a higher score indicating that the short drama better matches the user's content preferences and emotional needs.

[0078] Categorizing short drama content by emotional type is the foundation of emotional pacing planning. Based on the emotional characteristics of the short drama content, all short dramas are divided into three emotional energy levels: high-energy emotional types include inspirational, tense, and humorous, which are typically fast-paced, emotionally rich, and easily evoke strong emotional responses; medium-energy emotional types include heartwarming, inspirational, and touching, which are mild in emotion, highly empathetic, and evoke positive but not intense emotional responses; low-energy emotional types include sadness, calmness, and contemplation, which are slow-paced, emotionally restrained, and guide users to emotional reflection. Emotional classification employs a multi-feature determination method, comprehensively considering factors such as the short drama's plot type, emotional tone, pacing, and musical style. An automatic classification is performed using a decision tree model to obtain an emotional energy type label for each short drama.

[0079] Constructing emotional sequence templates based on emotion classification results is an innovative step in achieving emotion rhythm perception recommendation. Emotional sequence templates refer to the sorting pattern of short dramas in the recommendation list; different templates correspond to different emotional experience flows. The progressive climax template arranges short dramas in a "low-medium-high energy" order, similar to the emotional progression in music or literature, allowing the user's emotional experience to gradually transition from calm to intensity. The emotional fluctuation template alternates between "high-energy-low-high energy," creating a viewing experience with emotional ups and downs, similar to a rollercoaster of emotions. The soothing transition template arranges short dramas in a "high-energy-medium-low energy" order, guiding the user from an excited state to gradually calming down, suitable for viewing before bed or during relaxation time. Each template includes specific positional ratio parameters; for example, in the progressive climax template, low-medium energy short dramas account for 40%, medium-energy short dramas for 30%, and high-energy short dramas for 30%, forming a complete emotional sequence library.

[0080] Analyzing user viewing history for sentiment sequence preferences is crucial for personalized sentiment rhythm planning. This study identifies preferred sentiment rhythm patterns by analyzing the sentiment sequence of short dramas watched by users over the past 30 days. Specifically, the analysis involves arranging the user's viewing history chronologically and extracting the sentiment energy type of each watched drama to form sentiment sequence samples. Then, a sequence pattern mining algorithm is used to extract frequently occurring sentiment transition patterns from multiple viewing sessions, such as a gradual progression from low to high energy or a fluctuating pattern of alternating high and low energy. Simultaneously, the study considers the time factor, analyzing the user's preferred sentiment sequence patterns at different times (e.g., morning, afternoon, evening) and the differences between weekdays and weekends. Through these analyses, the most suitable sentiment sequence template for the user's current state is identified; for example, a user might prefer a "gradual climax" type on weekday evenings and a "fluctuating sentiment" type on weekend mornings.

[0081] The core step in generating the final recommendation list is to dynamically sort and combine candidate short dramas based on emotional sequence templates and basic recommendation scores. First, the most suitable emotional sequence template is selected based on the user's current time period and date type. Then, according to the emotional type ratio requirements of the selected template, a corresponding number of high, medium, and low-energy short dramas are selected from the candidate short drama set. Within each emotional type, they are sorted according to their basic recommendation scores. Finally, based on the emotional order defined by the template, the different types of short dramas are combined into a complete recommendation sequence. To increase the personalization of the recommendations, an exploration factor is introduced, appropriately adding a certain proportion of new types of content while maintaining the overall emotional rhythm, helping users discover potential interests. The final generated short drama recommendation list not only contains highly matched content but also features a carefully designed emotional rhythm.

[0082] In one specific embodiment, the process of executing step S105 may specifically include the following steps:

[0083] The emotional intensity curve analysis of the short drama content is performed to extract video clips with emotional intensity values ​​exceeding a preset threshold, thus obtaining dynamic preview materials;

[0084] Key sentences are extracted and importance is ranked from the short drama plot text, and then converted into audio data through a speech synthesis engine to obtain a personalized plot voice summary.

[0085] Numerical mapping and color coding were performed on the emotional value change data throughout the short drama to generate an emotional change visual spectrum with time on the horizontal axis and emotional intensity on the vertical axis.

[0086] A directed graph structure is constructed based on the character relationship data in the short drama content, and the node positions are calculated using a force-directed layout algorithm to obtain an interactive character relationship graph;

[0087] Based on the current network bandwidth and device parameters, the resolution of the dynamic preview materials, personalized plot voice summaries, emotional change visualization charts, and interactive character relationship diagrams are adjusted and compressed to obtain multi-level preview resource packages.

[0088] The multi-level preview resource package is processed by the layout engine to arrange components and bind interactive events to obtain content preview data.

[0089] Specifically, the emotional intensity curve analysis of the short drama content is performed using a multi-dimensional feature fusion method. By processing the video frame sequence of the short drama, visual emotional features for each time period are extracted, including color saturation, contrast, and compositional complexity. Simultaneously, the audio track is processed to extract audio emotional features such as volume changes, tempo, and pitch. Combined with the emotional analysis results of the dialogue text, a comprehensive emotional intensity value is calculated for each time window (typically 3-5 seconds). The emotional intensity value is represented by a standardized numerical value of 0-1, with higher values ​​indicating stronger emotions. The emotional intensity data for the entire drama is arranged along a timeline to form an emotional intensity curve. Then, an emotional intensity threshold is set (typically 0.7 or higher), and consecutive segments exceeding the threshold are identified on the curve; these segments typically represent the emotional climax of the short drama. Two to three of the most representative segments are selected, each 5-10 seconds long, and the corresponding video clips are extracted as dynamic preview material. This material can quickly attract user attention and evoke emotional resonance. When extracting key sentences and ranking their importance from the short drama's plot text, the process begins by collecting complete text from the dialogue, subtitles, and synopsis. This text is then segmented into semantically coherent paragraphs. The TextRank algorithm is used to extract key sentences. Based on text network construction and importance propagation principles, this algorithm treats each sentence as a node in the network, with the similarity between sentences serving as edge weights. The importance score for each sentence is calculated iteratively. The top 5-10 sentences with the highest scores are selected as key sentences and reordered according to their position in the original text and their emotional significance, constructing a concise narrative that aligns with the story's development. The sorted key sentences are then processed by a deep neural network speech synthesis engine. Appropriate timbre and emotional tone are selected based on the short drama's type and style (e.g., a mature and stable timbre for workplace dramas, and a lively and cheerful timbre for youth dramas), generating fluent and natural speech data to create a personalized plot summary, providing users with an immersive auditory preview experience.

[0090] Numerical mapping and color coding of the emotional value changes throughout the short drama are performed to visually represent these emotional shifts. First, the acquired emotional intensity curves are smoothed to reduce noise interference. Then, emotions are categorized, typically into several basic types such as joy, sadness, anger, fear, and emotion. A mapping relationship between emotional type and color is established; for example, joy is mapped to yellow, sadness to blue, and anger to red. Emotional intensity is represented by the depth of color, with darker colors indicating higher intensity. The emotional value data is then plotted as an intuitive emotional change graph along a timeline, with the horizontal axis representing time, the vertical axis representing emotional intensity, and the curve color indicating the emotional type. Climaxes and turning points are also marked, creating a visual emotional change graph that allows users to intuitively grasp the emotional trajectory and fluctuations of the short drama.

[0091] Constructing a directed graph structure based on character relationship data from the short drama content is the foundation for creating an interactive character relationship graph. First, through natural language processing and scene analysis, the character entities in the short drama and their relationship types are extracted, including familial, romantic, friendly, adversarial, and superior-subordinate relationships. Each character is treated as a node in the graph, and the relationships between characters are represented as directed edges. The direction of the edge indicates the direction of the relationship, and the thickness or color of the edge indicates the strength or type of the relationship. The constructed directed graph is then optimized for spatial arrangement using a force-directed layout algorithm. This algorithm simulates the principles of repulsion and attraction in physics, treating nodes as charged particles and edges as springs. Through iterative calculations, it achieves mechanical equilibrium between nodes, resulting in an aesthetically pleasing and clearly structured layout. The algorithm starts from a random initial position and adjusts the node positions in each iteration, reducing the total energy of the system until a stable state is reached. The final generated interactive character relationship graph allows users to click on character nodes to view detailed information or click on relationship edges to learn about the interactions between characters, providing an interactive experience for exploring the character relationship network of the short drama.

[0092] Adapting preview resources based on current network bandwidth and device parameters is crucial for ensuring a good user experience. First, front-end JavaScript code obtains the user's device screen resolution, processor performance metrics, and current network conditions (such as bandwidth and latency). Based on these parameters, preview resources are categorized into high, medium, and low quality levels. Dynamic preview materials undergo transcoding and compression at different resolutions; typically, high quality uses the original resolution, medium quality uses 720p, and low quality uses 480p. Personalized storyline voice-over summaries are compressed at different bitrates: high quality uses 128kbps, medium quality uses 96kbps, and low quality uses 64kbps. Emotional change visualizations and interactive character relationship graphs are rendered with varying levels of complexity; high quality includes full animation effects and interactive functions, while low quality simplifies to static charts. Based on a combined assessment of device performance and network conditions, the most suitable resource quality level is selected and packaged into a multi-level preview resource package to ensure a smooth preview experience across various terminal conditions.

[0093] Arranging multi-level preview resource packages into components and binding interactive events using a layout engine is a crucial step in generating the final content preview data. Based on responsive design principles, the layout engine employs a flexible layout and grid system to dynamically adjust the size and position of each preview component according to the device's screen size. Component arrangement follows visual hierarchy principles, typically placing dynamic preview materials in the most prominent position, displaying emotional change visual graphs and character relationship diagrams side-by-side, and presenting voice summaries in the form of an audio player. Corresponding interactive event listeners are bound to each component; for example, clicking on a dynamic preview material plays the complete segment, clicking on a climax point on the emotional graph jumps to the corresponding segment, and clicking on a character node displays character information. Through the componentization mechanism of front-end frameworks (such as React or Vue), each preview element is encapsulated as an independent component, and data flow and interactive linkage between components are achieved through state management. Ultimately, structured content preview data is generated, providing users with a rich, intuitive, and highly interactive short drama content preview experience.

[0094] For example, an emotional intensity curve analysis of a 25-minute short drama identified three climactic scenes with emotional intensity values ​​exceeding 0.8: the surprise scene of the protagonist's successful job interview, the exhilarating scene of the team overcoming difficulties to complete a project, and the touching scene of the protagonist receiving recognition. Three approximately 8-second video clips were extracted from these scenes as dynamic preview material. Next, the drama text was analyzed, extracting eight key sentences from approximately 4,000 words. These sentences summarized the protagonist's storyline from entering the workplace from school, facing challenges, striving for growth, and ultimately gaining recognition. These sentences were rearranged according to the story's development order, and a 35-second personalized voice summary of the drama was generated using a sunny and energetic female voice. Simultaneously, the emotional changes throughout the drama were visualized as an emotional curve graph. The horizontal axis represents the 25-minute timeline, and the vertical axis represents the emotional intensity from 0 to 1. The curve color transitions from blue (tension) to yellow (joy) and then to red (exhilaration) to indicate changes in emotional type, clearly presenting the emotional development path of "pressure-effort-success." In addition, data on five main characters and their relationships were extracted from the plot to construct a character relationship diagram. The protagonist was positioned in the center, surrounded by mentors, rivals, team members, and clients. Different colored lines represented different types of relationships. The layout was optimized using a force-directed algorithm to form a clear and intuitive network of character relationships. Based on a detected network bandwidth of 4Mbps and a mid-range smartphone, a medium-quality preview resource package was selected. Dynamic preview materials were used at 720p resolution, the audio summary at a bitrate of 96kbps, and the visual graph and relationship diagram at a moderate level of interaction complexity. Finally, a responsive layout engine organized these elements into a unified preview interface. Dynamic preview materials played in a loop as the background, the audio summary could be listened to by clicking the play button, and the emotional graph and character relationship diagram were displayed side by side below. Users could interact with each element through clicks, swipes, and other actions, creating an immersive multimodal preview experience that effectively conveyed the core appeal and emotional value of the short drama, helping users quickly decide whether to watch the full content.

[0095] In one specific embodiment, the process of executing step S106 may specifically include the following steps:

[0096] The interaction process between the user and the content preview data is captured and processed to obtain interaction behavior trajectory data including clicked areas, dwell time, and browsing path;

[0097] Multidimensional statistical analysis was performed on user behavior metrics for watching short dramas to obtain a behavioral feature vector that includes viewing completion rate, repeated viewing segments, frequency of interactive comments, and number of social sharing sessions.

[0098] The interaction behavior trajectory data and behavior feature vectors are correlated and analyzed to construct a user-content interaction intensity matrix;

[0099] The user-content interaction intensity matrix is ​​processed using an emotional satisfaction calculation model to obtain an emotional matching score that reflects the degree of user preference for different emotional types of content.

[0100] Based on the sentiment matching score, the weight coefficients of each sentiment dimension in the content recommendation model are optimized by gradient descent to obtain the updated sentiment weight parameter set.

[0101] The updated set of sentiment weight parameters is applied to the candidate content pool for reordering, generating a target short drama recommendation list.

[0102] Specifically, the interaction process between the user and the content preview data is processed through event capture. Event listeners are embedded in the front-end application to track all user actions. Event capture covers various interaction types, including click events (recording the type of UI element clicked, its location coordinates, and timestamp), hover events (recording the duration the mouse stays on each element), swipe events (recording the direction and speed of finger or mouse swipes), and view switching events (recording the order in which the user browses different preview resources). After preprocessing, this raw event data is transformed into structured interaction behavior trajectory data, including click area data (recording the content elements the user focuses on, such as specific characters, emotional climaxes, etc.), dwell time data (recording the user's attention to different preview elements), and browsing path data (recording the order in which the user's attention shifts). This data can accurately depict the user's interaction process with the preview content, reflecting the user's focus of interest and attention distribution.

[0103] Multidimensional statistical analysis of user behavior metrics while watching short dramas is a key approach to understanding deeper user preferences. This involves collecting various behavioral data generated by users during actual viewing, including completion rate (the ratio of actual viewing time to the total duration of the short drama), repeated viewing segment data (identifying video segments and frequencies watched multiple times), interactive comment data (recording the content, timing, and sentiment of user comments), and social sharing behavior data (recording the sharing platform, sharing text, and sharing frequency). Cluster analysis and principal component analysis are applied to these multidimensional behavioral data to extract key feature dimensions. Normalization is then used to convert behavioral metrics of different dimensions into a unified numerical range. Finally, a multidimensional behavioral feature vector is constructed, where each dimension represents a type of behavioral feature, and the vector values ​​reflect the intensity of the behavior, comprehensively describing the user's interaction patterns and level of interest in the short drama content.

[0104] Correlation analysis of interactive behavior trajectory data with behavioral feature vectors is fundamental to constructing a user-content interaction intensity matrix. First, each short drama is categorized by content features and emotional type, forming a short drama-feature matrix. Then, user interactive behavior trajectory data (for the preview stage) and behavioral feature vectors (for the viewing stage) are time-series aligned and feature-mapped to analyze the correlation between preview-stage behavior and actual viewing-stage behavior. Collaborative filtering and association rule mining algorithms are used to identify which preview interactions are highly correlated with which viewing behaviors. Based on these correlations, a user-content interaction intensity matrix is ​​constructed. Rows in the matrix represent users, columns represent different types or features of content, and matrix values ​​represent the intensity of user interaction with that type of content, comprehensively reflecting users' content preference characteristics throughout the preview and viewing cycle.

[0105] Applying an emotional satisfaction calculation model to the user-content interaction intensity matrix is ​​a crucial step in accurately measuring user emotional preferences. This model is a multi-factor weighted model that uses the emotional type of content as an analytical dimension. By analyzing the intensity of user interaction with content of different emotional types, it infers user emotional preferences. The model first categorizes short drama content into multiple categories based on emotional characteristics, such as "inspirational growth," "heartwarming and touching," and "thrilling and exciting." Then, it calculates a weighted mean for the column vectors corresponding to each emotional type in the user-content interaction intensity matrix, using completion rate, repeat viewing frequency, and social sharing behavior as weighting factors. The weight values ​​are determined based on the indicative strength of these behaviors on emotional satisfaction; typically, completion rate is weighted at 0.4, repeat viewing at 0.3, and social sharing at 0.3. The calculated emotional matching score is a numerical vector between 0 and 1, with each element corresponding to an emotional type. Higher values ​​indicate a higher degree of user preference for that type of emotional content.

[0106] Gradient descent optimization of the sentiment dimension weights in a content recommendation model based on sentiment matching scores is the core mechanism for adaptive adjustment in recommendation systems. Content recommendation models contain multiple sentiment dimension weight parameters, which determine the proportion of different sentiment types in the recommendation results. Gradient descent optimization is an iterative optimization algorithm that calculates the gradient of the objective function with respect to the parameters and adjusts the parameters in the opposite direction of the gradient, gradually bringing the objective function to a local optimum. In this method, the objective function is defined as the mean squared error between the predicted user satisfaction and the actual observed sentiment matching score. In each iteration, the difference between the predicted recommendation effect and the actual sentiment matching score under the current weight parameters is calculated, and then the weight parameters are adjusted according to the gradient direction to gradually reduce the prediction error. After multiple iterations, an updated set of sentiment weight parameters is obtained. These parameters more accurately reflect the user's true sentiment preferences, providing a more precise decision-making basis for subsequent recommendations.

[0107] The final step in generating a personalized recommendation list is to apply the updated sentiment weight parameter set to the candidate content pool for re-sorting. First, a set of candidate short dramas matching the user's basic preferences is selected from the content library, including newly released, popular, and short dramas similar to the user's historical preferences. Then, a content matching score and a sentiment matching score are calculated for each candidate short drama. The content matching score is calculated based on the similarity between the content feature vector and the user's interest vector, while the sentiment matching score is calculated based on the dot product of the short drama's sentiment characteristic vector and the updated sentiment weight parameter set. The two scores are weighted and fused according to a preset ratio to obtain the final recommendation score. The candidate short dramas are then sorted according to the recommendation score, while considering the sentiment sequence template requirements to ensure that the recommendation list meets both personalized matching requirements and has reasonable emotional rhythm changes, ultimately generating a highly personalized target short drama recommendation list with optimized emotional experience.

[0108] Taking the short drama recommendation optimization process of a professional as an example, the entire data processing flow is illustrated: When browsing the recommended short drama previews, the user's interaction behavior trajectory data shows that he frequently clicked on keywords such as "teamwork" and "career advancement," spent a long time on dynamic preview clips showcasing workplace challenges, and quickly swiped past content involving emotional entanglements. This interaction data was recorded by an event capture system, forming a structured interaction behavior trajectory. Simultaneously, the user's short drama viewing behavior indicators show that his average completion rate for workplace growth-related short dramas reached 85%, he frequently rewatched key clips solving workplace problems, and shared inspirational workplace content on social media platforms, but rarely interacted with or shared content related to emotional dramas. Through correlation analysis, it was found that the user's high interest in workplace success scenarios during the preview stage was positively correlated with the high completion rate during the viewing stage. The constructed user-content interaction intensity matrix shows that the user had the highest interaction intensity values ​​for "workplace struggle" and "teamwork" content, at 0.89 and 0.83 respectively, while the interaction intensity for "emotional entanglement" content was only 0.32. Analysis using an emotional satisfaction calculation model revealed that the user scored 0.92 for "positive and uplifting" emotions, 0.78 for "tense and exciting" emotions, 0.65 for "warm and touching" emotions, and only 0.25 for "sad" emotions. Based on these emotional matching scores, the emotional dimension weights in the recommendation model were adjusted using a gradient descent optimization algorithm. The initial emotional weights ["positive and uplifting": 0.5," "tense and exciting": 0.3," "warm and touching": 0.1," "sad": 0.1] were adjusted to ["positive and uplifting": 0.64," "tense and exciting": 0.22," "warm and touching": 0.12," "sad": 0.02] after five rounds of iteration. These updated emotional weight parameters were applied to the ranking of 300 candidate short dramas. Combined with a "gradual climax" emotional sequence template, a short drama recommendation list was generated, primarily featuring workplace inspirational content with appropriate integration of team collaboration challenges. The recommendation results better matched the user's content preferences and emotional needs, achieving closed-loop optimization and precise personalization of the recommendation system.

[0109] In personalized short drama content recommendation methods based on big data mining, closed-loop feedback and adaptive iterative optimization are key to ensuring the continuous improvement of the recommendation system. First, the interaction process between the user and the content preview data is processed through event capture. Event listeners are embedded in the front-end application to track all user actions. Event capture covers various interaction types, including click events (recording the type of UI element clicked, its location coordinates, and timestamp), hover events (recording the duration of mouse hover over each element), swipe events (recording the direction and speed of finger or mouse swipes), and view switching events (recording the order in which the user browses different preview resources). This raw event data is preprocessed and transformed into structured interaction behavior trajectory data, including click area data (recording content elements the user focuses on, such as specific characters or emotional climaxes), dwell time data (recording the user's attention to different preview elements), and browsing path data (recording the order in which the user's attention shifts). This data accurately depicts the user's interaction process with the preview content, reflecting the user's focus of interest and attention distribution.

[0110] Multidimensional statistical analysis of user behavior metrics while watching short dramas is a key approach to understanding deeper user preferences. This involves collecting various behavioral data generated by users during actual viewing, including completion rate (the ratio of actual viewing time to the total duration of the short drama), repeated viewing segment data (identifying video segments and frequencies watched multiple times), interactive comment data (recording the content, timing, and sentiment of user comments), and social sharing behavior data (recording the sharing platform, sharing text, and sharing frequency). Cluster analysis and principal component analysis are applied to these multidimensional behavioral data to extract key feature dimensions. Normalization is then used to convert behavioral metrics of different dimensions into a unified numerical range. Finally, a multidimensional behavioral feature vector is constructed, where each dimension represents a type of behavioral feature, and the vector values ​​reflect the intensity of the behavior, comprehensively describing the user's interaction patterns and level of interest in the short drama content.

[0111] Correlation analysis of interactive behavior trajectory data with behavioral feature vectors is fundamental to constructing a user-content interaction intensity matrix. First, each short drama is categorized by content features and emotional type, forming a short drama-feature matrix. Then, user interactive behavior trajectory data (for the preview stage) and behavioral feature vectors (for the viewing stage) are time-series aligned and feature-mapped to analyze the correlation between preview-stage behavior and actual viewing-stage behavior. Collaborative filtering and association rule mining algorithms are used to identify which preview interactions are highly correlated with which viewing behaviors. Based on these correlations, a user-content interaction intensity matrix is ​​constructed. Rows in the matrix represent users, columns represent different types or features of content, and matrix values ​​represent the intensity of user interaction with that type of content, comprehensively reflecting users' content preference characteristics throughout the preview and viewing cycle.

[0112] Applying an emotional satisfaction calculation model to the user-content interaction intensity matrix is ​​a crucial step in accurately measuring user emotional preferences. This model is a multi-factor weighted model that uses the emotional type of content as an analytical dimension. By analyzing the intensity of user interaction with content of different emotional types, it infers user emotional preferences. The model first categorizes short drama content into multiple categories based on emotional characteristics, such as "inspirational growth," "heartwarming and touching," and "thrilling and exciting." Then, it calculates a weighted mean for the column vectors corresponding to each emotional type in the user-content interaction intensity matrix, using completion rate, repeat viewing frequency, and social sharing behavior as weighting factors. The weight values ​​are determined based on the indicative strength of these behaviors on emotional satisfaction; typically, completion rate is weighted at 0.4, repeat viewing at 0.3, and social sharing at 0.3. The calculated emotional matching score is a numerical vector between 0 and 1, with each element corresponding to an emotional type. Higher values ​​indicate a higher degree of user preference for that type of emotional content.

[0113] Gradient descent optimization of the sentiment dimension weights in a content recommendation model based on sentiment matching scores is the core mechanism for adaptive adjustment in recommendation systems. Content recommendation models contain multiple sentiment dimension weight parameters, which determine the proportion of different sentiment types in the recommendation results. Gradient descent optimization is an iterative optimization algorithm that calculates the gradient of the objective function with respect to the parameters and adjusts the parameters in the opposite direction of the gradient, gradually bringing the objective function to a local optimum. In this method, the objective function is defined as the mean squared error between the predicted user satisfaction and the actual observed sentiment matching score. In each iteration, the difference between the predicted recommendation effect and the actual sentiment matching score under the current weight parameters is calculated, and then the weight parameters are adjusted according to the gradient direction to gradually reduce the prediction error. After multiple iterations, an updated set of sentiment weight parameters is obtained. These parameters more accurately reflect the user's true sentiment preferences, providing a more precise decision-making basis for subsequent recommendations.

[0114] The final step in generating a personalized recommendation list is to apply the updated sentiment weight parameter set to the candidate content pool for re-sorting. First, a set of candidate short dramas matching the user's basic preferences is selected from the content library, including newly released, popular, and short dramas similar to the user's historical preferences. Then, a content matching score and a sentiment matching score are calculated for each candidate short drama. The content matching score is calculated based on the similarity between the content feature vector and the user's interest vector, while the sentiment matching score is calculated based on the dot product of the short drama's sentiment characteristic vector and the updated sentiment weight parameter set. The two scores are weighted and fused according to a preset ratio to obtain the final recommendation score. The candidate short dramas are then sorted according to the recommendation score, while considering the sentiment sequence template requirements to ensure that the recommendation list meets both personalized matching requirements and has reasonable emotional rhythm changes, ultimately generating a highly personalized target short drama recommendation list with optimized emotional experience.

[0115] Taking the short drama recommendation optimization process of a professional as an example, the entire data processing flow is illustrated: When browsing the recommended short drama previews, the user's interaction behavior trajectory data shows that he frequently clicked on keywords such as "teamwork" and "career advancement," spent a long time on dynamic preview clips showcasing workplace challenges, and quickly swiped past content involving emotional entanglements. This interaction data was recorded by an event capture system, forming a structured interaction behavior trajectory. Simultaneously, the user's short drama viewing behavior indicators show that his average completion rate for workplace growth-related short dramas reached 85%, he frequently rewatched key clips solving workplace problems, and shared inspirational workplace content on social media platforms, but rarely interacted with or shared content related to emotional dramas. Through correlation analysis, it was found that the user's high interest in workplace success scenarios during the preview stage was positively correlated with the high completion rate during the viewing stage. The constructed user-content interaction intensity matrix shows that the user had the highest interaction intensity values ​​for "workplace struggle" and "teamwork" content, at 0.89 and 0.83 respectively, while the interaction intensity for "emotional entanglement" content was only 0.32. Analysis using an emotional satisfaction calculation model revealed that the user scored 0.92 for "positive and uplifting" emotions, 0.78 for "tense and exciting" emotions, 0.65 for "warm and touching" emotions, and only 0.25 for "sad" emotions. Based on these emotional matching scores, the emotional dimension weights in the recommendation model were adjusted using a gradient descent optimization algorithm. The initial emotional weights ["positive and uplifting": 0.5," "tense and exciting": 0.3," "warm and touching": 0.1," "sad": 0.1] were adjusted to ["positive and uplifting": 0.64," "tense and exciting": 0.22," "warm and touching": 0.12," "sad": 0.02] after five rounds of iteration. These updated emotional weight parameters were applied to the ranking of 300 candidate short dramas. Combined with a "gradual climax" emotional sequence template, a short drama recommendation list was generated, primarily featuring workplace inspirational content with appropriate integration of team collaboration challenges. The recommendation results better matched the user's content preferences and emotional needs, achieving closed-loop optimization and precise personalization of the recommendation system.

[0116] The above describes the method for personalized short drama content push based on big data mining in the embodiments of this application. The following describes the system for personalized short drama content push based on big data mining in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the personalized short drama content push system based on big data mining in this application includes:

[0117] Module 201 is used to perform feature extraction and tag construction on the short drama content to generate a feature vector of the short drama content.

[0118] Analysis module 202 is used to collect and process user behavior data and sentiment analysis based on the feature vector of short drama content to obtain user sentiment pattern data;

[0119] Training module 203 is used to construct a user interest graph based on the short drama content feature vector and user emotional pattern data, and to train the emotional resonance network on the user interest graph to obtain a content recommendation model;

[0120] The sorting module 204 is used to apply the content recommendation model to the candidate content pool, and sort the candidate short dramas through the emotional rhythm planning algorithm to obtain a short drama recommendation list;

[0121] The generation module 205 is used to perform multimodal preview resource generation processing on the short drama content in the short drama recommendation list to obtain content preview data;

[0122] The recommendation module 206 is used to collect user interaction behavior data with the content preview data, and calculate the sentiment matching score by combining the user's viewing completion rate, repeated viewing times and social sharing behavior. Based on the sentiment matching score, the sentiment weight coefficient in the content recommendation model is adjusted to generate a target short drama recommendation list.

[0123] Through the collaborative efforts of the aforementioned components, feature extraction and tagging of short drama content generate content feature vectors, achieving accurate expression of multi-dimensional information about the short drama content. Based on these feature vectors, user behavior data is collected and processed for sentiment analysis, yielding user emotional pattern data. This breaks through the limitations of traditional recommendation systems that only focus on explicit behaviors. Innovative technologies such as micro-expression recognition capture users' true emotional preferences, significantly improving the accuracy and depth of user profiles. A user interest graph is constructed based on the short drama content feature vectors and user emotional pattern data, and an emotional resonance network is trained to obtain a content recommendation model. This innovative combination of graph neural networks and emotional resonance enhancement networks not only identifies users' content type preferences but also accurately captures content features that evoke emotional resonance, achieving a recommendation dimension upgrade from "content matching" to "emotional resonance." The content recommendation model is applied to the candidate content pool and sorted using an emotional rhythm planning algorithm to obtain a short drama recommendation list. It creatively introduces the rhythm construction principle from music theory, transforming the recommendation results from isolated content sets into content sequences with artistic emotional rhythm, significantly improving the user's continuous viewing experience. Multimodal preview resource generation is performed on the short drama content in the recommendation list to obtain content preview data. This breaks through the traditional single preview mode of static covers and text descriptions, greatly improving the value delivery efficiency of short drama content through multimodal preview methods such as dynamic covers with emotional climaxes and emotional concentration heatmaps. User interaction with preview data is collected, and emotional matching scores are calculated based on indicators such as viewing completion rate. The model parameters are dynamically adjusted to generate the target recommendation list, constructing a closed-loop optimization mechanism that enables the recommendation system to continuously self-optimize and adapt to changes in user emotional preferences.

[0124] above Figure 2 The short drama personalized content push system based on big data mining in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The short drama personalized content push device based on big data mining in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0125] Figure 3This is a schematic diagram of the structure of a short drama personalized content push device based on big data mining, provided by an embodiment of the present invention. This short drama personalized content push device 300 based on big data mining can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing applications 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the short drama personalized content push device 300 based on big data mining. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute a series of instruction operations in the storage media 330 on the short drama personalized content push device 300 based on big data mining to implement the steps of the aforementioned short drama personalized content push method based on big data mining.

[0126] The short drama personalized content push device 300 based on big data mining may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The illustrated structure of the short drama personalized content push device based on big data mining does not constitute a limitation on the short drama personalized content push device based on big data mining provided by the present invention. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0127] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the short drama personalized content push method based on big data mining.

[0128] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0129] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a short drama personalized content push device (which can be a personal computer, server, or network device, etc.) based on big data mining to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1.A method for short drama personalized content pushing based on big data mining, characterized in that, The method comprises: The short drama content is extracted and labeled to generate a short drama content feature vector; Based on the short drama content feature vector, user behavior data is collected and sentiment analysis is performed to obtain user sentiment pattern data; According to the short drama content feature vector and the user sentiment pattern data, a user interest graph is constructed, and a sentiment resonance network is trained to obtain a content recommendation model, including: based on the short drama content feature vector and the user sentiment pattern data, a knowledge graph structure with a user as a center node is constructed to obtain an initial interest graph containing content categories, short drama entities and emotional labels; the initial interest graph is applied to a graph neural network algorithm for representation learning processing to obtain a low-dimensional dense vector capturing user interest structure characteristics; a three-component neural network architecture including an emotional trajectory extractor, a resonance matcher and an emotional reinforcement predictor is constructed to obtain an emotional resonance enhancement network model; the emotional change data of the short drama content is input into the emotional trajectory extractor for processing to obtain a content emotional trajectory capturing emotional climax and turning points in the plot development; the content emotional trajectory and the user emotional pattern data are input into the resonance matcher for similarity calculation processing to obtain a user-content emotional resonance measurement value; based on the user-content emotional resonance measurement value and the low-dimensional dense vector, deep feature fusion processing is performed to obtain a content recommendation model with integrated semantic matching and emotional resonance dual capabilities; The content recommendation model is applied to a candidate content pool, and a candidate short drama is sorted by an emotional rhythm planning algorithm to obtain a short drama recommendation list; The short drama content in the short drama recommendation list is processed to generate a multi-modal preview resource to obtain content preview data; Interaction behavior data of the user with the content preview data is collected, and the emotional matching score is calculated in combination with the user viewing completion rate, the number of repeated views and the social sharing behavior, and the emotional weight coefficient in the content recommendation model is adjusted according to the emotional matching score to generate a target short drama recommendation list. 2.The short drama personalized content pushing method based on big data mining of claim 1, wherein, The short drama content is extracted and labeled to generate a short drama content feature vector, including: Key frames are extracted from short drama video samples and analyzed by image recognition algorithm to obtain scene type and visual style features; The short drama dialogue and subtitle content are processed by word segmentation and semantic analysis to obtain theme keywords and emotional tendency data; The editing rate and transition mode of the short drama are statistically analyzed to obtain rhythm features and narrative structure data; The sound elements of the short drama are processed by audio feature extraction to obtain music key and emotional rendering features; The visual style features, emotional tendency data, narrative structure data and emotional rendering features are processed by data fusion to construct a multi-dimensional label vector; The multi-dimensional label vector is standardized and reduced in dimension to generate a short drama content feature vector. 3.The short drama personalized content pushing method based on big data mining of claim 1, wherein, The user behavior data is collected and sentiment analysis is performed based on the short drama content feature vector to obtain user sentiment pattern data, including: The user viewing behavior is collected in full dimension through a real-time data stream processing framework to obtain behavior data including viewing time, completion rate, pause times, fast forward and fast backward behavior, and repeated viewing segments; Facial micro-expression data collected by the front camera of the user equipment are identified and processed to obtain emotional response features in the user viewing process; Natural language processing analysis is performed on the user comment text content to obtain explicit emotional expression data including emotional polarity and emotional intensity; The behavior data and the short drama content feature vector are time-aligned to construct an emotional response-content node mapping matrix; The emotional response features and the explicit emotional expression data are multi-source fused to obtain a two-level emotional model including implicit emotions and explicit emotions; Based on the two-level emotional model and the emotional response-content node mapping matrix, an emotional fluctuation model is constructed to obtain user emotional pattern data. 4.The short drama personalized content pushing method based on big data mining of claim 1, wherein, The content recommendation model is applied to the candidate content pool, and the candidate short dramas are sorted through an emotional rhythm planning algorithm to obtain a short drama recommendation list, including: A candidate content pool including newly released short dramas, popular short dramas, and short dramas similar to the user's historical preferences is constructed to obtain an initial candidate short drama set; The content recommendation model is used to calculate the matching degree of each short drama in the initial candidate short drama set to obtain a basic recommendation score; The short drama content is classified by emotional type to obtain emotional classification results including high-energy emotional type, medium-energy emotional type, and low-energy emotional type; Based on the emotional classification results, multiple emotional sequence templates are constructed to obtain an emotional sequence library including gradual climax type, emotional fluctuation type, and slow transition type; The user's historical viewing behavior is analyzed to obtain the most suitable emotional sequence template for the user; The initial candidate short drama set is dynamically sorted and combined based on the emotional sequence template and the basic recommendation score to obtain a short drama recommendation list with a preset emotional rhythm. 5.The short drama personalized content pushing method based on big data mining of claim 1, wherein, The short drama content in the short drama recommendation list is processed to generate multi-modal preview resources to obtain content preview data, including: The emotional intensity curve of the short drama content is analyzed to extract video segments with emotional intensity values exceeding a preset threshold to obtain dynamic preview materials; The key sentences of the short drama plot text content are extracted and sorted by importance, and the voice synthesis engine is used to convert them into audio data to obtain personalized plot voice summaries; The numerical mapping and color coding of the short drama overall mood value change data are processed to generate an emotional change visual map with time on the horizontal axis and mood intensity on the vertical axis; A directed graph structure is constructed based on the character relationship data in the short drama content, and the node positions are calculated through a force-directed layout algorithm to obtain an interactive role relationship graph; The dynamic preview materials, the personalized plot voice summaries, the emotional change visual map, and the interactive role relationship graph are resolution-adjusted and compressed to obtain multi-level preview resource packages based on the current network bandwidth and device parameters. The multi-level preview resource package is arranged and interacted by a layout engine to obtain content preview data. 6.The short drama personalized content pushing method based on big data mining of claim 1, wherein, The interactive behavior data of the user and the content preview data is collected, and the emotional matching score is calculated in combination with the user viewing completion rate, repeated viewing times, and social sharing behavior. The emotional weight coefficient in the content recommendation model is adjusted according to the emotional matching score, and a target short drama recommendation list is generated. Event capture processing is performed on the interactive process of the user and the content preview data to obtain interactive behavior trajectory data including a click area, a stay time, and a browsing path. Multi-dimensional statistical analysis processing is performed on the behavior indicators of the user watching the short drama to obtain a behavior feature vector including a viewing completion rate, repeated viewing segments, interactive comment frequency, and social sharing times. The interactive behavior trajectory data and the behavior feature vector are associated and analyzed to construct a user-content interaction intensity matrix. An emotional satisfaction calculation model is applied to the user-content interaction intensity matrix to obtain an emotional matching score reflecting the preference degree of the user for different emotional type content. The emotional weight parameter set after the update is applied to the candidate content pool for reordering processing to generate a target short drama recommendation list. The short drama personalized content pushing system based on big data mining for implementing the short drama personalized content pushing method based on big data mining as claimed in any one of claims 1-6 comprises: 7.A short play personalized content pushing system based on big data mining, characterized in that, A construction module for performing feature extraction and label construction processing on short drama content to generate a short drama content feature vector. An analysis module for performing collection and emotional analysis processing on user behavior data based on the short drama content feature vector to obtain user emotional pattern data. A training module for constructing a user interest graph based on the short drama content feature vector and the user emotional pattern data, and performing emotional resonance network training on the user interest graph to obtain a content recommendation model. A sorting module for applying the content recommendation model to a candidate content pool, and performing sorting processing on candidate short dramas by an emotional rhythm planning algorithm to obtain a short drama recommendation list. A generation module for performing multi-modal preview resource generation processing on short drama content in the short drama recommendation list to obtain content preview data. A recommendation module for collecting interactive behavior data of the user and the content preview data, and calculating an emotional matching score in combination with the user viewing completion rate, repeated viewing times, and social sharing behavior. The emotional weight coefficient in the content recommendation model is adjusted according to the emotional matching score, and a target short drama recommendation list is generated. A processor and a memory, the memory stores a computer program capable of running on the processor, and the processor implements the short drama personalized content pushing method based on big data mining as claimed in any one of claims 1-6 when executing the computer program. 8.A short play personalized content pushing device based on big data mining, characterized in that, ​ 9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, causes the processor to perform the short play personalized content pushing method based on big data mining as claimed in any one of claims 1 to 6.

Images