Big data mining-based method and system for pushing personalized content of short episodes

Abstract

The invention relates to the technical field of data processing, and discloses a method and a system for pushing personalized content of a short episode based on big data mining. The method comprises the following steps: analyzing a short episode feature to generate a content vector; analyzing a user behavior according to the content vector to obtain an emotion mode; constructing an interest map and training an emotion resonance network to form a recommendation model; applying an emotion rhythm algorithm to sort and generate a recommendation list; generating a multi-mode preview resource; user interaction behaviors are analyzed, the emotion weight is adjusted, and the pushing effect is optimized. According to the method, the problems that the emotion experience dimension is neglected due to excessive attention to content type matching, the emotion sequence overall planning is lacked and the content preview mode is single in an existing short play recommendation system are solved, and the emotion matching precision and the user satisfaction degree of recommended content are improved by constructing a recommendation mechanism and a multi-mode preview system based on emotion resonance.

Description

Technical Field

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

[0002] The rapid development of short video platforms has spawned a vast amount of short drama content. Faced with this vast amount 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. By analyzing historical user behavior data and content features, they calculate the degree of match between the user and the content, generating personalized recommendation lists. These technologies have achieved some success in improving content distribution efficiency and user experience, enabling users to more conveniently access interesting short drama content and enabling platforms to more accurately push content to potential audiences.

[0003] However, traditional recommendation algorithms overly focus on matching content type and subject matter, ignoring users' emotional experience needs. This results in a single emotional dimension in recommended content, failing to meet users' expectations for a rich and diverse emotional experience. Second, existing recommendation systems generally lack the ability to capture and analyze the emotional characteristics of content and users' emotional responses in a refined manner, failing to accurately identify which content elements truly resonate with users. Furthermore, traditional recommendation systems lack a holistic approach to organizing recommended content, simply pursuing the degree of match between individual content and users while ignoring the impact of the content sequence as a whole on the user's emotional experience. Furthermore, existing content preview methods are overly simplistic, failing to effectively convey the emotional value and core appeal of a skit, impacting the efficiency and accuracy of users' content selection. These issues have severely constrained the user experience and content value delivery efficiency of skit platforms. Summary of the Invention

[0004] This application provides a method and system for pushing personalized content for short dramas based on big data mining, which is used to solve the problems in existing short drama recommendation systems, such as excessive focus on content type matching while ignoring the emotional experience dimension, lack of overall planning of emotional sequences, and single content preview method. By constructing a recommendation mechanism based on emotional resonance and a multimodal preview system, the emotional matching accuracy and user satisfaction of recommended content are improved.

[0005] In the first aspect, the present application provides a method for pushing personalized content of short dramas based on big data mining, and the method for pushing personalized content of short dramas based on big data mining includes: extracting features and constructing labels for short drama content to generate a short drama content feature vector; collecting and sentiment analysis of user behavior data based on the short drama content feature vector to obtain user sentiment pattern data; constructing a user interest map based on the short drama content feature vector and the user sentiment pattern data, and performing sentiment resonance network training on the user interest map to obtain a content recommendation model; applying the content recommendation model to a candidate content pool, sorting the candidate short dramas through an emotional rhythm planning algorithm to obtain a short drama recommendation list; performing multimodal preview resource generation processing on the short drama content in the short drama recommendation list to obtain content preview data; collecting interactive behavior data between users and the content preview data, and calculating the sentiment matching score in combination with the user's viewing completion rate, number of repeated viewings and social sharing behavior, adjusting the sentiment weight coefficient in the content recommendation model according to the sentiment matching score, and generating a target short drama recommendation list.

[0006] In a second aspect, the present application provides a short play personalized content push system based on big data mining, the short play personalized content push system based on big data mining comprising: a construction module for performing feature extraction and label construction processing on the short play content to generate a short play content feature vector;

[0007] An analysis module, configured to collect and perform sentiment analysis on user behavior data based on the feature vector of the skit content to obtain user sentiment pattern data;

[0008] A training module, configured to construct a user interest graph based on the feature vector of the skit content and the user emotion pattern data, and perform emotion resonance network training on the user interest graph to obtain a content recommendation model;

[0009] A sorting module is used to apply the content recommendation model to the candidate content pool, sort the candidate short plays using the emotional rhythm planning algorithm, and obtain a short play recommendation list;

[0010] A generation module, configured to perform multimodal preview resource generation processing on the short play content in the short play recommendation list to obtain content preview data;

[0011] The recommendation module is used to collect the user's interactive behavior data with the content preview data, and calculate the emotional matching score based on the user's viewing completion rate, number of repeated viewings and social sharing behavior, adjust the emotional weight coefficient in the content recommendation model according to the emotional matching score, and generate a target short drama recommendation list.

[0012] On the third aspect, 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 enable the short drama personalized content push device based on big data mining to execute the above-mentioned short drama personalized content push method based on big data mining.

[0013] In a fourth aspect, a computer-readable storage medium is provided, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned method for pushing personalized content of short dramas based on big data mining.

[0014] In the technical solution provided by this application, by extracting features and constructing labels for the short play content, a content feature vector is generated, which realizes the accurate expression of the multi-dimensional information of the short play content. Based on the short play content feature vector, the user behavior data is collected and sentiment analysis is processed to obtain the user emotion pattern data, breaking through the limitation of the traditional recommendation system that only focuses on explicit behavior. The user's real emotional preferences are captured through innovative technologies such as micro-expression recognition, which greatly improves the accuracy and depth of user portraits; based on the short play content feature vector and the user's emotion pattern data, a user interest map is constructed and an emotion resonance network is trained to obtain a content recommendation model, which innovatively combines the graph neural network with the emotion resonance enhancement network, which can not only identify the user's content type preference, but also accurately capture the content features that can trigger the user's emotional resonance, realizing the upgrade of the recommendation dimension from "content matching" to "emotional resonance"; The recommendation model is applied to the candidate content pool, and the short drama recommendation list is obtained through sorting and processing by the emotional rhythm planning algorithm. The rhythm construction principle in music theory is creatively introduced, so that the recommendation results are no longer an isolated content collection, but a content sequence with artistic emotional rhythm, which significantly improves the user's continuous viewing experience; the short drama content in the short drama recommendation list is processed through multimodal preview resource generation to obtain content preview data, breaking through the traditional single preview mode of static cover plus text introduction, and greatly improving the value delivery efficiency of short drama content through multimodal preview methods such as emotional climax dynamic cover and emotional concentration heat map; the interaction behavior of users with preview data is collected, and the emotional matching score is calculated in combination with indicators such as viewing completion rate, and the model parameters are dynamically adjusted to generate the target recommendation list, and a closed-loop optimization mechanism is constructed to enable the recommendation system to continuously optimize itself and adapt to changes in user emotional preferences. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 This is a schematic diagram of an embodiment of a method for pushing personalized content of short dramas based on big data mining in an embodiment of the present application;

[0017] Figure 2 This is a schematic diagram of an embodiment of a system for pushing personalized content of short plays based on big data mining in an embodiment of the present application;

[0018] Figure 3 It is a schematic block diagram of the structure of a short play personalized content push device based on big data mining in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The embodiments of the present application provide a method and system for pushing personalized content of short dramas based on big data mining. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0020] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of a method for pushing personalized content of short dramas based on big data mining includes:

[0021] Step S101: extracting features and constructing labels for the skit content to generate a feature vector for the skit content;

[0022] Step S102: collecting and sentimentally analyzing user behavior data based on the skit content feature vector to obtain user sentiment pattern data;

[0023] Step S103: constructing a user interest graph based on the short play content feature vector and the user emotion pattern data, and performing emotion resonance network training on the user interest graph to obtain a content recommendation model;

[0024] Step S104: Apply the content recommendation model to the candidate content pool, sort the candidate skits using the emotional rhythm planning algorithm, and obtain a recommended skit list;

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

[0026] Step S106: Collect the interactive behavior data between the user and the content preview data, and calculate the emotional matching score based on the user's viewing completion rate, number of repeated viewings and social sharing behavior. Adjust the emotional weight coefficient in the content recommendation model according to the emotional matching score to generate a target short drama recommendation list.

[0027] It is understandable that the execution subject of this application can be a short play personalized content push system based on big data mining, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0028] Specifically, keyframes from short drama videos were sampled and image recognition algorithms were used to extract scene type and visual style features. For example, by analyzing the keyframes of a short urban romance drama, visual features such as "café scene," "warm tones," and "close-up shots" were identified. Furthermore, word segmentation and semantic analysis were performed on the dialogue and subtitles to extract thematic keywords and sentiment data. For example, keywords such as "workplace," "struggle," and "relationship dilemma" were identified, and the overall emotional tone was determined to be "inspirational and positive." Furthermore, statistical analysis of editing rate and transition patterns was performed to obtain rhythmic characteristics and narrative structure data. For example, the average shot length of the short drama was 4 seconds, with 70% of fast cuts, reflecting a "compact rhythm." Audio feature extraction was performed on sound elements to obtain musical tone and emotional rendering characteristics. For example, the background music was mainly piano with a slow rhythm, creating a "warm atmosphere." These multi-dimensional features were then fused to construct a 50-dimensional label vector, which was then subjected to dimensionality reduction to generate the final feature vector for the short drama content.

[0029] A real-time data stream processing framework collects user viewing behavior data, including user dwell time at different short plot points, repeated viewing of segments, and fast-forward and rewind behavior. For example, analysis of a user's viewing data reveals that they spend more time on segments depicting "workplace competition" and frequently fast-forward through segments depicting "romantic entanglements." Micro-expression recognition technology is also used to capture changes in users' facial expressions during viewing, such as a slight raise of eyebrows and a slight moistening of the corners of the eyes when watching inspirational content, indicating positive emotional resonance. User comment text is analyzed to extract sentiment polarity and intensity. For example, 85% of user comments on inspirational content contain positive words, indicating strong identification. By aligning behavioral data with the time series of content feature vectors, an emotional response-content node mapping matrix is ​​constructed to precisely locate content features that trigger user emotional fluctuations. By integrating micro-expression responses (implicit emotions) with comment expressions (explicit emotions), a dual-level emotional model is formed to comprehensively capture users' true emotional preferences.

[0030] To construct a user interest graph and train an emotional resonance network based on skit content feature vectors and user emotion pattern data, knowledge graph technology was used to create a user-centric interest graph. Nodes included content categories, specific skits, and emotional tags, and edge weights represented the strength of associations. Graph neural networks were applied to this graph for representation learning, compressing complex graph structures into low-dimensional dense vectors. A three-component neural network architecture was constructed, consisting of an emotion trajectory extractor, a resonance matcher, and an emotion reinforcement predictor. The emotion trajectory extractor analyzed the emotional evolution of the skit content over time, identifying emotional climaxes and turning points. The resonance matcher calculated the similarity between user emotion patterns and the emotional trajectory of the content, identifying the content segments that resonated most with users. The emotion reinforcement predictor used historical data to predict the emotional impact of specific content on users. The network was trained using a contrastive learning method, using content that users expressed strong emotional reactions to as positive samples and content that had weaker reactions to as negative samples. This ultimately resulted in a dual-channel recommendation model that simultaneously considered content semantic matching and emotional resonance.

[0031] When applying the content recommendation model to a candidate content pool to generate a recommendation list, a candidate content pool is first constructed, including newly released short dramas, popular short dramas, and short dramas similar to the user's historical preferences. The content recommendation model then calculates a basic recommendation score for each short drama in the pool. Based on emotional type, the short dramas are categorized into three categories: high energy (e.g., inspiring, tense, funny), medium energy (e.g., heartwarming, inspirational), and low energy (e.g., sad, calm). Various emotional sequence templates are constructed, such as "progressive climax" (low-medium-high energy sequence) and "emotional fluctuation" (high-low-high energy alternation). The user's emotional sequence preferences, as demonstrated in their historical viewing behavior, are analyzed to select the most appropriate template. For example, data shows that a user prefers to watch light-hearted content first on weekday evenings, then transition to content with stronger emotional depth, demonstrating a clear preference for "progressive climax." Based on the selected template and basic recommendation score, the candidate short dramas are dynamically ranked and combined to form a recommended list of short dramas with a pre-set emotional rhythm.

[0032] When generating multimodal preview resources for the content in the recommended short play list, an emotional intensity curve analysis is performed on each short play, and video clips with emotional intensity values ​​exceeding a preset threshold are extracted to generate dynamic preview materials. Key sentences in the plot text are extracted and ranked by importance, converting them into personalized voice summaries. The emotional value change data throughout the entire process is visualized, generating a visual map of emotional changes with time as the horizontal axis and emotional intensity as 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 diagram. These preview resources are adapted based on network bandwidth and device performance, forming a multi-level preview resource package suitable for different terminal devices. Finally, the layout engine arranges components and binds interactive events to generate structured content preview data. When collecting and optimizing the interactive behavior between users and content preview data, events are captured during the interactive process between users and the preview content, and interactive behavior trajectories such as click areas, dwell time, and browsing paths are recorded; statistical analysis is performed on the behavioral indicators of users after watching the short drama, including viewing completion rate, number of repeated viewing clips, frequency of interactive comments, and number of social sharing; the interactive behavior is correlated with the viewing indicators to construct a user-content interaction intensity matrix; the emotional satisfaction calculation model is applied to derive the user's preference for content of different emotional types and form an emotional matching score; the weight coefficients of each emotional dimension in the recommendation model are optimized and adjusted based on the emotional matching score, and the emotional weight parameters are updated; the updated parameters are applied to the candidate content pool for re-sorting to generate a target short drama recommendation list that better meets the user's latest emotional preferences. For example, when a user shows a high completion rate and multiple social sharing behaviors for the recommended heartwarming short drama, the system calculates that the user's matching score for the "warmth" emotional dimension is 0.92, thereby increasing the weight coefficient of the "warmth" dimension in the recommendation model, thereby increasing the proportion of heartwarming short dramas in subsequent recommendation results. At the same time, in the design of emotional sequences, more consideration is given to using heartwarming content as the climax point, thereby continuously optimizing the personalized push effect.

[0033] In the embodiment of the present application, by extracting features and constructing labels for the short play content, a content feature vector is generated, which realizes the accurate expression of the multi-dimensional information of the short play content, and collects and processes the user behavior data based on the short play content feature vector and performs sentiment analysis to obtain the user emotion pattern data, breaking through the limitation of the traditional recommendation system that only focuses on explicit behavior, and captures the user's real emotional preferences through innovative technologies such as micro-expression recognition, greatly improving the accuracy and depth of user portraits; constructing a user interest map based on the short play content feature vector and the user emotion pattern data and training the emotion resonance network to obtain a content recommendation model, which innovatively combines the graph neural network with the emotion resonance enhancement network, which can not only identify the user's content type preference, but also accurately capture the content features that can trigger the user's emotional resonance, realizing the upgrade of the recommendation dimension from "content matching" to "emotional resonance"; The recommendation model is applied to the candidate content pool, and the short drama recommendation list is obtained through sorting and processing by the emotional rhythm planning algorithm. The rhythm construction principle in music theory is creatively introduced, so that the recommendation results are no longer an isolated content collection, but a content sequence with artistic emotional rhythm, which significantly improves the user's continuous viewing experience; the short drama content in the short drama recommendation list is processed through multimodal preview resource generation to obtain content preview data, breaking through the traditional single preview mode of static cover plus text introduction, and greatly improving the value delivery efficiency of short drama content through multimodal preview methods such as emotional climax dynamic cover and emotional concentration heat map; the interaction behavior of users with preview data is collected, and the emotional matching score is calculated in combination with indicators such as viewing completion rate, and the model parameters are dynamically adjusted to generate the target recommendation list, and a closed-loop optimization mechanism is constructed to enable the recommendation system to continuously optimize itself and adapt to changes in user emotional preferences.

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

[0035] Key frames are extracted from the short drama videos and analyzed using image recognition algorithms to obtain scene type and visual style features.

[0036] Perform word segmentation and semantic analysis on the dialogue and subtitles of the skit to obtain topic keywords and sentiment tendency data;

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

[0038] Perform audio feature extraction on the sound elements of the skit to obtain the musical tone and emotional rendering features;

[0039] Perform data fusion processing on visual style features, emotional tendency data, narrative structure data and emotional rendering features to construct a multi-dimensional label vector;

[0040] The multi-dimensional label vector is normalized and dimensionally reduced to generate the feature vector of the skit content.

[0041] Specifically, key frames are extracted from the skits by sampling them. A method combining uniform sampling and scene change perception is used to extract key frames at a frequency of 1-2 frames per second. Key frame extraction utilizes 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 identifies a scene transition point and extracts the corresponding frame as a key frame. The extracted key frames are then used for scene recognition using a deep convolutional neural network. This network, pre-trained on a scene classification dataset, can classify key frames into scene types such as "cafe," "office," and "living room." Visual style feature extraction calculates parameters such as the color distribution histogram, saturation mean, contrast, and texture complexity of the key frames to analyze whether the image tones are warm or cool, whether the composition is symmetrical, and whether the depth of field is shallow or deep, thereby quantifying the visual style characteristics of the skits.

[0042] To process the skit's dialogue and subtitles, speech recognition technology was first used to convert the dialogue into text, which was then combined with the subtitles to form the complete text. Chinese word segmentation technology was then used to segment the continuous text into meaningful word units. After removing stop words, a term frequency-inverse document frequency (TF-IDF) matrix was constructed, and high-weighted words in the text were extracted as topic keywords. Simultaneously, sentiment lexicon matching and deep semantic analysis models were used to calculate the text's sentiment polarity and intensity, determining whether the content was positive, negative, or neutral, and whether the sentiment intensity was high, medium, or low, thereby generating sentiment trend data. The skit's editing rate and transition patterns were analyzed through video frame analysis. The average duration of all shots in the skit was first calculated to obtain the editing rate metric. Edge detection and optical flow estimation algorithms were used to identify transition types, categorizing them into hard cuts, fades, wipes, and dissolves. The frequency and temporal distribution of each transition type were then calculated. Combining statistical data on shot length distribution and transition patterns, a rhythm curve was constructed for the skit, reflecting the plot's ups and downs and changes in narrative rhythm, thereby generating rhythmic characteristics and narrative structure data.

[0043] To extract audio features from the skit's sound elements, the audio signal is first analyzed in the time-frequency domain to separate the three channels: vocals, background music, and ambient sound. Tonal analysis, beat detection, and chord recognition are then performed on the background music to determine parameters such as the musical key (e.g., major, minor), rhythmic characteristics (e.g., slow, medium, fast), and harmonic complexity. Combining volume curves and timbre characteristics, the music's emotional impact is analyzed, such as whether it creates tension, creates warmth, or enhances humor, thereby capturing the musical key and emotional impact characteristics. To fuse the extracted visual style features, emotional tendency data, narrative structure data, and emotional impact features, a multimodal feature fusion method is employed. First, each feature is normalized to convert feature data of different dimensions to a uniform numerical range. Feature importance analysis is then performed to calculate the contribution of each feature to the representation of the skit's content and assign corresponding weights. Using weighted fusion or tensor fusion methods, these feature data are combined into a fused feature vector containing 100-200 dimensions, forming a multidimensional label vector.

[0044] When normalizing and reducing the dimensionality of multidimensional label vectors, we first use the Z-score normalization method to adjust the features of each dimension to a distribution with a mean of 0 and a standard deviation of 1. Dimensionality reduction algorithms such as principal component analysis (PCA) or t-SNE are then used to compress the high-dimensional feature space to approximately 50 dimensions, preserving key information while reducing data redundancy and computational complexity. The resulting feature vector of the skit content is a compact and information-rich vector representation used for subsequent similarity calculations and recommendation model training.

[0045] For example, approximately 1,800 keyframes were extracted from a 25-minute short. Scene recognition revealed that "office" scenes accounted for 45%, "business meeting" scenes for 30%, and other scenes for 25%. Visual style analysis revealed a predominantly blue-gray hue, with regular, symmetrical compositions and a moderate depth of field. Dialogue and subtitle analysis extracted key keywords such as "workplace competition," "teamwork," and "career growth" from the approximately 3,000-word text. Sentiment analysis revealed a positive overall emotional tone, with fluctuations in feelings of "pressure" and "challenge." Editing analysis revealed that the average shot length of the short was 4.2 seconds, indicating a medium editing rate. Cuts and transitions were predominant, with a noticeable increase in tempo during climaxes. Audio analysis revealed that the background music, primarily piano and string music, was of a medium-to-fast tempo, with a noticeable increase in volume and tempo at climaxes, fostering a sense of tension and striving. After normalizing and weighted fusion of these feature data, a 173-dimensional initial feature vector was obtained. After PCA dimensionality reduction processing, a 48-dimensional short drama content feature vector was obtained, which accurately captures the content characteristics of this short drama: "urban workplace, positive progress, and brisk pace."

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

[0047] Through the real-time data stream processing framework, we collect all dimensions of user viewing behavior, including viewing time, completion rate, number of pauses, fast forward and rewind behavior, and repeated viewing of segments.

[0048] Recognize and process facial micro-expression data collected by the front camera of the user's device to obtain the emotional response characteristics of the user during the viewing process;

[0049] Perform natural language processing analysis on the text content of user comments to obtain explicit emotional expression data including emotional polarity and emotional intensity;

[0050] Perform temporal alignment processing on behavioral data and skit content feature vectors to construct an emotional response-content node mapping matrix;

[0051] The emotional response characteristics and explicit emotional expression data are processed through multi-source fusion to obtain a two-level emotional model containing implicit emotions and explicit emotions;

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

[0053] Specifically, a real-time data stream processing framework collects comprehensive data on user viewing behavior. This framework utilizes an event-driven architecture and embeds a data collection SDK on the user side. Whenever a user interacts with the skit app, corresponding events are collected. This data includes viewing duration (recording the user's actual viewing time from start to finish), completion rate (calculating the ratio of the full viewing time to the total duration of the skit), pause counts (recording the frequency of users actively clicking the pause button), fast-forward and rewind behavior (recording the direction, duration, and frequency of users dragging the progress bar), and replayed segments (identifying the video segments and the number of times users have replayed multiple times). This raw behavioral data undergoes preprocessing, including outlier filtering, missing value interpolation, and time windowing, to form a standardized user behavior dataset. Processing facial micro-expression data captured by the user's device's front-facing camera requires user authorization. A facial landmark detection algorithm is then used to extract 68 facial feature points, tracking subtle changes in these points during viewing. Micro-expression recognition utilizes deep learning methods, extracting sequences of facial features and feeding them into a pre-trained expression recognition neural network. This network maps facial expression variations into basic emotion categories (such as happiness, sadness, surprise, anger, fear, disgust, and neutrality) and their intensity values. To improve recognition accuracy, the system incorporates data from the user's device's ambient light sensor to compensate for illumination changes, minimizing the impact of ambient light variations on expression recognition. By analyzing micro-expressions in real time as users watch different content clips, the system accurately captures the user's emotional responses, such as a smile that rises in humorous scenes or a slight frown that conveys empathy in touching ones.

[0054] When performing natural language processing analysis on user comment text, users' comments on short dramas, comments on bullet screens, and content shared on social platforms are collected as raw text data. First, text preprocessing is performed, including word segmentation, stop word removal, and part-of-speech tagging. Then, sentiment analysis is performed using a combination of sentiment dictionaries and deep learning. The sentiment dictionaries are used to determine the polarity and assign intensity to sentiment words in the text. At the same time, pre-trained language models such as BERT are applied to capture contextual semantics and comprehensively determine the sentiment polarity (positive, negative, or neutral) and sentiment intensity (high, medium, or low) of the text. Complex expressions such as irony and metaphors are specially processed using a contextual semantic understanding model to ensure the accuracy of sentiment judgment. Ultimately, explicit sentiment expression data containing sentiment polarity and sentiment intensity is generated, directly reflecting the emotional attitude expressed by the user through text.

[0055] The key step is to perform time-series alignment processing on user behavior data and skit content feature vectors. First, a timeline mapping relationship is established to align the timestamp of the user behavior data with the timeline of the skit content. Through the sliding window technology, the skit content is divided into fixed time windows (such as 5 seconds or 10 seconds), and a corresponding content feature sub-vector is assigned to each window. At the same time, according to the same time window, the user's behavior data and micro-expression data within the window are aggregated. Through the aligned time windows, the correspondence between the user's emotional response and the skit content nodes is established, and an emotional response-content node mapping matrix is ​​constructed. The rows of the matrix represent different time windows, and the columns include the content features of the window and the corresponding user emotional response data, forming a content-response correspondence map.

[0056] When fusing emotional response features and explicit emotional expression data from multiple sources, a method combining weighted fusion and probabilistic modeling is employed. First, weight coefficients are assigned to emotional data from different sources, determined based on data reliability and directness of expression. For example, micro-expression data (implicit emotions) may be given a higher weight than textual comments (explicit emotions) because micro-expressions are more difficult to disguise and reflect a more authentic emotional state. Then, using a Bayesian network model, the emotional data from different sources is used as observed variables, and the user's true emotional preferences are used as latent variables to calculate the most likely user emotional state. Ultimately, a two-level emotional model is formed, encompassing implicit emotions (subconscious emotional reactions reflected by micro-expressions) and explicit emotions (subjective emotional attitudes expressed in textual comments), comprehensively capturing the user's emotional preferences.

[0057] The emotional fluctuation model, based on a two-level emotional pattern and an emotional response-content node mapping matrix, is constructed using a combination of time series pattern mining and machine learning. First, time series analysis is applied to the emotional response-content node mapping matrix to extract patterns in user emotional fluctuations as they change with content, such as which types of content trigger strong and which types of content trigger neutral responses. Then, combined with historical viewing data, a sequence prediction model (such as a long short-term memory (LSTM) network) is applied to learn patterns in user emotional fluctuations and predict the likely emotional impact of different types of content on the user. Cluster analysis is used to identify the main types and characteristics of user emotional preferences, such as "emotional resonance," "rational analysis," or "lighthearted entertainment." Ultimately, user emotional pattern data is generated, encompassing user emotional preference characteristics, emotional fluctuation patterns, and emotional sensitivity to different content types.

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

[0059] Based on the feature vectors of skit content 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, skit entities, and emotional labels;

[0060] Applying graph neural network algorithms to the initial interest graph for representation learning, we obtain a low-dimensional dense vector that captures the structural features of user interests.

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

[0062] The emotional change data of the skit 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 points in the plot development;

[0063] The content emotion trajectory and user emotion pattern data are input into the resonance matcher for similarity calculation and processing to obtain the user-content emotion resonance measurement value;

[0064] Based on the user-content sentiment resonance metric and low-dimensional dense vector, deep feature fusion processing is performed to obtain a content recommendation model that integrates the dual capabilities of semantic matching and sentiment resonance.

[0065] Specifically, a user-centric knowledge graph is constructed based on skit content feature vectors and user emotion pattern data. This knowledge graph is a multi-relational directed graph structure, with nodes including user nodes, content category nodes, skit entity nodes, and emotion tag nodes. During the construction process, the user is set as the central node, and by analyzing historical user interaction data, association edges are established between the user and various nodes. Edges between users and content category nodes represent interest intensity, with a value ranging from 0 to 1, calculated based on the proportion of each type of content in the user's viewing history and the degree of viewing completeness. Edges between users and skit entity nodes represent interaction intensity, calculated based on metrics such as viewing time, completion rate, and repeat viewing. Edges between users and emotion tags represent emotional preference, determined by the intensity of users' reactions to content of different emotion types. Furthermore, association edges are established between nodes, such as the "belongs to" relationship between skit entities and content categories, and the "initiates" relationship between skit entities and emotion tags, forming a complete initial interest graph encompassing content categories, skit entities, and emotion tags. When applying a graph neural network algorithm to the initial interest graph for representation learning, a graph convolutional network (GCN) is used as the underlying model. A GCN is a deep learning model specialized for processing graph-structured data, capable of learning features from nodes in the graph through a message-passing mechanism. Specifically, an initial feature vector is assigned to each node in the graph. The initial features of user nodes are derived from user sentiment pattern data, the initial features of content category nodes are derived from category semantic vectors, the initial features of skit entity nodes are derived from skit content feature vectors, and the initial features of emotion label nodes are derived from emotion semantic vectors. Then, through multi-layer graph convolution operations, each node continuously updates its own features by aggregating feature information from its neighboring nodes. After multiple rounds of iteration, the node features are integrated with graph structural information, reflecting the node's semantic position and relational structure within the entire interest network. Finally, the final feature vector of the user node is extracted and, after dimensionality reduction, converted into a low-dimensional dense vector of less than 200 dimensions. This vector effectively captures the structural characteristics of user interests, including interest diversity, focus, and transition patterns.

[0066] Building an emotional resonance enhancement network model is key to achieving emotion-guided recommendations. This network consists of three core components: an emotion trajectory extractor, an emotional resonance matcher, and an emotional reinforcement predictor. The emotion trajectory extractor is a time series model based on a long short-term memory (LSTM) network, specifically designed to capture the emotional trajectory of short drama content over time. The extractor receives as input a sequence of time-sliced ​​emotion vectors from the short drama. Each time segment contains information about the emotional type and intensity of the segment. Using the LSTM's gating mechanism and memory units, it captures the long-term and short-term dependencies of emotional changes and identifies emotional climaxes (sudden increases in emotional intensity) and emotional turning points (shifts in emotional type). The emotional resonance matcher is a neural network model based on an attention mechanism. It receives as dual input user emotional pattern data and content emotional trajectory. Using a self-attention layer, it extracts key features of user emotional preferences and content emotional trajectory, respectively. Then, using a cross-attention mechanism, it calculates the correspondence between the two and identifies the content segments most likely to elicit emotional resonance in the user. The emotional reinforcement predictor is a feedforward neural network that, based on a user's historical emotional response data, predicts the likely emotional impact of a particular piece of content on that user, providing an emotionally guided basis for the final recommendation decision. The three components form a unified emotional resonance enhancement network model through end-to-end training.

[0067] When processing the emotional change data of a skit's content into the emotion trajectory extractor, the first step is to construct the data. Through sentiment analysis of the skit's content, including analysis of the dialogue, background music, and visual imagery, we obtain emotion type and intensity data arranged in a time series. This data is then segmented into fixed time windows (e.g., 5 seconds), forming a sequence of emotion vectors. Each vector contains the dominant emotion type (e.g., joy, tension, sadness) and its corresponding intensity value for that time window. This sequence data is then fed into the LSTM network of the emotion trajectory extractor. Through forward propagation, the LSTM units progressively process the emotion vectors for each time segment while maintaining their internal state to capture the temporal pattern of emotional change. During this processing, when the emotion intensity value suddenly increases or the emotion type changes significantly, thresholds are set to identify these as emotional climaxes or turning points. The output content emotion trajectory is a sequence consisting of the original emotion sequence and an enhanced sequence marked with key points, accurately depicting the emotional changes of the skit's content over time.

[0068] A multi-scale matching algorithm is used to input content emotion trajectory and user emotion pattern data into the resonance matcher for similarity calculation. First, the user emotion pattern data is processed to extract the user's preference weights for different emotion types and the emotion sensitivity parameters. The matching degree between each time segment in the content emotion trajectory and the user's emotion preferences is then calculated. For each time segment, the cosine similarity between the emotion vector and the user's emotion preference vector is calculated to obtain a basic matching score. However, considering that emotional resonance is influenced 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 emotion sequence and the user's preferences over windows of different lengths. Finally, based on the importance of emotional climaxes and turning points, the matching scores of these key moments are given higher weights. The user-content emotion resonance metric is then calculated comprehensively, reflecting the degree of fit between the emotional development process of the content and the user's emotional preferences.

[0069] Deep feature fusion based on user-content sentiment resonance metrics and low-dimensional dense vectors is key to forming the final recommendation model. Using a dual-channel architecture, traditional collaborative filtering and content matching are used to calculate the user's interest in content, primarily based on low-dimensional dense vectors derived from graph neural networks. Furthermore, a sentiment resonance network is used to calculate the user's sentiment resonance with the content, yielding semantic matching scores and sentiment resonance scores, respectively. These two scores are then combined with other auxiliary features (such as content popularity and novelty) through a feature fusion layer and input into the final ranking model. The ranking model utilizes a gradient boosted decision tree (GBDT) or a deep neural network, learning from historical recommendation results to adaptively adjust the weights of each feature and ultimately output a recommendation score. This entire process forms a content recommendation model that integrates the dual capabilities of semantic matching and sentiment resonance, taking into account both user preferences for content type and their emotional experience needs.

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

[0071] Build a candidate content pool that includes the latest online short dramas, popular short dramas, and short dramas similar to the user's historical preferences to obtain an initial set of candidate short dramas;

[0072] Calculate the matching degree of each short play in the initial candidate short play set through the content recommendation model to obtain a basic recommendation score;

[0073] The content of the short play was classified according to the emotional type, and the emotional classification results including high-energy emotional type, medium-energy emotional type and low-energy emotional type were obtained;

[0074] Based on the emotion classification results, a variety of emotion sequence templates are constructed to obtain an emotion sequence library including progressive climax type, emotion fluctuation type and soothing transition type;

[0075] Perform emotional sequence preference analysis on the user's historical viewing behavior to obtain the most suitable emotional sequence template for the user;

[0076] The initial candidate skit set is dynamically sorted and combined according to the emotional sequence template and the basic recommendation score to obtain a recommended list of skits with a preset emotional rhythm.

[0077] Specifically, the candidate content pool is constructed using a three-source fusion strategy, filtering content from a library of recently released short dramas, a library of popular short dramas, and a library of user-similar content. Recently released short dramas are filtered using a time window, selecting short dramas released within the past seven days and sorting them in reverse chronological order. Popular short dramas are calculated based on user behavior data across the entire platform, with a popularity score for each short drama. This score takes into account four dimensions: play volume, completion rate, interaction volume, and dissemination volume, and the top 100 most popular short dramas are selected. Short dramas similar to users' historical preferences are then selected using a collaborative content filtering algorithm. Based on the content feature vectors of the short dramas the user has already watched, the similarity with other short dramas in the content library is calculated, and the top 200 short dramas are selected. After de-duplication and merging the short dramas from these three sources, an initial candidate set of approximately 300 short dramas is formed, ensuring the diversity, timeliness, and personalization of the recommended content. The matching degree of each short drama in the initial candidate set is calculated using the previously constructed content recommendation model as the scoring engine. For each candidate short play, 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 is divided into two dimensions: the first is content semantic matching, which calculates the cosine similarity between the short play's content feature vector and the user's interest vector to obtain a content matching score; the second is emotional resonance matching, which uses the emotional resonance enhancement network to calculate the degree of fit between the short play's emotional trajectory and the user's emotional pattern to obtain an emotional resonance score. The two scores are weighted and fused according to preset weights (usually 0.6 for content matching and 0.4 for emotional resonance) to obtain a basic recommendation score for each short play, ranging from 0 to 1. The higher the score, the more it meets the user's content preferences and emotional needs.

[0078] Categorizing skit content by emotional type is the foundation of emotional pacing planning. Based on the emotional characteristics of the skits' content, all skits are divided into three emotional energy levels: High-energy emotional types, such as motivational, tense, and funny, are typically fast-paced, emotionally charged, and prone to strong emotional reactions; Medium-energy emotional types, such as warmth, inspirational, and moving, are gentle and empathetic, evoking positive but not intense emotional reactions; Low-energy emotional types, such as sadness, calmness, and reflection, are slow-paced, restrained, and guide users to emotionally process. Emotional classification utilizes a multi-feature approach, comprehensively considering factors such as the skit's plot type, emotional tone, tempo, and musical style. Automatic classification is performed using a decision tree model to determine the emotional energy type label for each skit.

[0079] Constructing emotional sequence templates based on emotion classification results is an innovative step in achieving emotional rhythm-aware recommendations. Emotional sequence templates refer to the order in which skits are ranked in the recommended list, with different templates corresponding to different emotional experience flows. The Progressive Climax template arranges skits in a "low-medium-high" order, similar to the emotional progression in music or literature, allowing the user's emotional experience to gradually transition from calm to intense. The Emotional Fluctuation template alternates between "high-low-high" and "high" energy, creating a rollercoaster-like emotional experience. The Smooth Transition template arranges skits in a "high-medium-low" order, guiding users from excitement to a gradual calm, making it suitable for viewing before bed or during relaxation. Each template includes specific placement parameters, such as 40% of the Progressive Climax skits are low-medium, 30% are medium-energy, and 30% are high-energy, forming a comprehensive emotional sequence library.

[0080] Analyzing user sentiment sequence preferences based on historical viewing behavior is key to achieving personalized sentiment rhythm planning. By analyzing the sentiment type sequences of short dramas watched by users over the past 30 days, we can identify the user's preferred sentiment rhythm patterns. The specific analysis method involves chronologically sorting the user's viewing history and extracting the sentiment energy type of each short drama, forming sentiment sequence samples. A sequential pattern mining algorithm is then used to extract frequently occurring sentiment transition patterns from multiple viewing sessions, such as a gradual transition from low to high energy or a fluctuating pattern of alternating highs and lows. Furthermore, by factoring in time, we analyze the user's preferred sentiment sequence patterns across different time periods (e.g., morning, afternoon, and evening), as well as differences between weekdays and weekends. Through this analysis, we can identify the most suitable sentiment sequence template for the user at the moment. For example, a user may prefer a "gradual climax" pattern on weekday evenings, but a "fluctuating" pattern on weekend mornings.

[0081] Dynamically sorting and combining candidate skits based on the emotional sequence template and basic recommendation score is the core step in ultimately forming the recommendation list. First, the most appropriate 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 skits are screened from the candidate skit collection. Within each emotional type, the basic recommendation score is sorted. Finally, based on the emotional sequence defined by the template, the different types of skits are combined into a complete recommendation sequence. To increase the level of personalization of recommendations, an exploration factor is introduced. While maintaining the overall emotional rhythm, a certain proportion of new types of content is appropriately added to help users discover potential interests. The resulting skit recommendation list not only contains highly matching content but also has a carefully designed emotional rhythm.

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

[0083] Perform emotional intensity curve analysis on the skit content, extract video clips whose emotional intensity exceeds a preset threshold, and obtain dynamic preview materials;

[0084] Extract key sentences from the short play's plot text, rank them by importance, and convert them into audio data using a speech synthesis engine to produce a personalized plot summary.

[0085] The data on emotional value changes throughout the skit are numerically mapped and color-coded to generate a visual graph of emotional changes with time as the horizontal axis and emotional intensity as the vertical axis.

[0086] A directed graph structure is constructed based on the character relationship data in the skit 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 dynamic preview material, personalized plot voice summary, emotional change visual map, and interactive character relationship diagram are subjected to resolution adjustment and compression encoding processing to obtain a multi-level preview resource package;

[0088] The multi-level preview resource package is arranged into components and interactive event binding is processed through the layout engine to obtain content preview data.

[0089] Specifically, the skit content is analyzed and processed using a multidimensional feature fusion method to generate an emotional intensity curve. By processing the skit's video frame sequence, visual emotional features for each time segment are extracted, including color saturation, light-dark contrast, and composition complexity. Simultaneously, the audio track is processed to extract audio emotional features such as volume variation, 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). Emotional intensity values ​​are represented on a standardized scale of 0-1, with higher values ​​indicating stronger emotions. The emotional intensity data for the entire film is arranged along a timeline to form an emotional intensity curve. Then, an emotional intensity threshold is set (typically 0.7 or higher). The curve identifies consecutive segments that exceed the threshold. These segments typically represent the emotional climaxes of the skit. Two to three representative segments, each 5-10 seconds long, are selected and extracted as dynamic preview materials. These materials are designed to quickly capture user attention and elicit emotional resonance. To extract key sentences and rank their importance from the skit's plot text, we first collect the full text from the skit's dialogue, subtitles, and plot summary. We then segment the text into semantically coherent paragraphs using text chunking. The TextRank algorithm, based on the principles of text network construction and importance propagation, treats each sentence as a node in the network, uses the similarity between sentences as the weight of edges, and iteratively calculates the importance score for each sentence. The top 5-10 sentences are selected as key sentences and re-ranked based on their position in the original text and their emotional significance, constructing a concise narrative that aligns with the story's developmental logic. The ranked key sentences are then processed by a deep neural network speech synthesis engine, which selects appropriate timbre and emotional overtones based on the skit's genre and style (e.g., mature and steady timbre for workplace dramas, lively and bright timbre for youth dramas). This generates smooth and natural speech data, creating a personalized audio summary of the plot and providing users with an immersive auditory preview experience.

[0090] The purpose of numerically mapping and color-coding the emotional value change data throughout the skit is to intuitively display emotional changes. First, the emotional intensity curve obtained above is smoothed to reduce noise interference. Then, the emotional types are classified. Emotions are usually divided into multiple basic categories, such as joy, sadness, anger, fear, and emotion. A mapping relationship between emotional types and colors is established. For example, joyful emotions are mapped to yellow, sad emotions are mapped to blue, and angry emotions are mapped to red. The intensity of emotions is represented by the depth of the color, with the higher the intensity, the darker the color. The emotional value data is plotted into an intuitive emotional change map along the timeline. The horizontal axis represents the time process, the vertical axis represents the emotional intensity, and the curve color represents the emotional type. At the same time, the emotional climax and turning points are marked to form a visual emotional change map, allowing users to intuitively grasp the emotional context and ups and downs of the skit.

[0091] Constructing a directed graph structure based on the character relationship data within the skits is the foundation for creating an interactive character relationship graph. First, through natural language processing and scene analysis, the character entities and their relationship types within the skits are extracted. These relationship types include family, romantic, friendship, hostility, and superior-subordinate relationships. Each character is represented as a node in the graph, and the relationships between them are represented as directed edges. The direction of the edge indicates the direction of the relationship, while the thickness or color of the edge indicates the strength or type of the relationship. The constructed character relationship directed graph is then spatially optimized using a force-directed layout algorithm. This algorithm simulates the principles of repulsion and attraction in physics, treating nodes in the graph 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. Starting from random initial positions, the algorithm adjusts node positions with each iteration to reduce the total energy of the system until a stable state is reached. The resulting interactive character relationship graph allows users to click on a character node to view detailed information or on a relationship edge to explore the interactions between characters, providing an interactive experience for exploring the character relationship network of the skits.

[0092] Adapting preview resources based on current network bandwidth and device parameters is key to ensuring a positive user experience. First, front-end JavaScript code retrieves the user's device's screen resolution, processor performance, and current network conditions (such as bandwidth and latency). Based on these parameters, preview resources are categorized into three quality levels: high, medium, and low. Dynamic preview assets are transcoded and compressed at different resolutions, typically using the original resolution for high quality, 720p for medium quality, and 480p for low quality. Audio compression for personalized storyline audio summaries is performed at different bitrates, with 128kbps for high quality, 96kbps for medium quality, and 64kbps for low quality. Rendering settings for the emotional visualization and interactive character relationship diagrams vary in complexity, with high quality including full animation effects and interactive features, while low quality simplifies to static charts. Based on a combined assessment of device performance and network conditions, the most appropriate resource quality level is selected and packaged into multi-level preview resource packages to ensure a smooth preview experience across various device conditions.

[0093] Arranging components and binding interactive events within the multi-level preview asset package through the layout engine is a key 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 based on the device screen size. Component arrangement adheres to visual hierarchy, typically placing dynamic preview assets in the most prominent position. A visual map of emotional transitions and a character relationship diagram are displayed side by side, and a speech overview is presented as an audio player. Each component is bound to a corresponding interactive event listener. For example, clicking a dynamic preview asset plays the full clip, clicking a climax on the emotional map jumps to the corresponding clip, and clicking a character node displays character information. Using the componentization mechanism of front-end frameworks (such as React or Vue), each preview element is encapsulated as an independent component, and state management enables data flow and interaction between components. The resulting structured content preview data provides users with a rich, intuitive, and highly interactive skit content preview experience.

[0094] For example, an emotional intensity curve analysis of a 25-minute short play identified three climaxes with emotional intensity values ​​exceeding 0.8: the protagonist's surprise at a successful interview, the team's exhilarating moment of overcoming difficulties to complete a project, and the touching scene of the protagonist receiving recognition. Three 8-second video clips were then extracted from these clips as dynamic preview material. The plot text was then analyzed, with eight key sentences extracted from the approximately 4,000-word text, summarizing the protagonist's journey from campus to the workplace, facing challenges, striving for growth, and ultimately achieving recognition. These sentences were then rearranged according to the story's developmental sequence and voice synthesised using a cheerful and energetic female voice to generate a personalized 35-second audio summary of the plot. The emotional changes throughout the play were also visualized as an emotional curve graph, with the horizontal axis representing the 25-minute timeline and the vertical axis representing emotional intensity on a scale of 0-1. The curve's color gradient, from blue (tension) to yellow (joy) to red (excitement), represents the changing emotional types, clearly depicting the emotional progression from "pressure-effort-success." In addition, data on the five main characters and their relationships was extracted from the plot to construct a character relationship diagram, with the protagonist centered and mentors, rivals, teammates, and clients distributed around them. Different colored lines represent different types of relationships, and a force-directed algorithm was used to optimize the layout, resulting in a clear and intuitive character relationship network. Based on the detected network bandwidth of 4Mbps and the device being a mid-range smartphone, a medium-quality preview resource package was selected, with a 720p resolution for the dynamic preview material, a 96kbps bitrate for the audio overview, and a moderate interaction complexity for the visual map and relationship diagram. Finally, a responsive layout engine organized these elements into a unified preview interface. The dynamic preview material loops in the background, the audio overview can be listened to by clicking the play button, and the emotion map and character relationship diagram are displayed side by side below. Users can interact with each element through clicks, swipes, and other actions, creating an immersive, multimodal preview experience that effectively conveys the core appeal and emotional value of the short, helping users quickly decide whether to watch the entire episode.

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

[0096] Capture and process events during the user's interaction with content preview data to obtain interactive behavior trajectory data including click area, dwell time, and browsing path;

[0097] Perform multi-dimensional statistical analysis on users' viewing behavior indicators of short dramas to obtain behavioral feature vectors including viewing completion rate, repeated viewing segments, interactive comment frequency, and social sharing number;

[0098] Perform correlation analysis on the interaction behavior trajectory data and the behavior feature vector to construct the user-content interaction intensity matrix;

[0099] Applying the emotional satisfaction calculation model to the user-content interaction intensity matrix, we obtain the emotional matching score that reflects the user's preference for content of different emotional types.

[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 emotional weight parameter set is applied to the candidate content pool for re-ranking to generate a target short drama recommendation list.

[0102] Specifically, the interaction process between the user and the content preview data is captured and processed by events, and an event listener is embedded in the front-end application to track all user operations. Event capture covers a variety of interaction types, including click events (recording the type of UI element clicked by the user, location coordinates and timestamp), hover events (recording the length of time the mouse stays on each element), sliding events (recording the sliding direction and speed of the finger or mouse) and view switching events (recording the order in which the user browses different preview resources). After preprocessing, these raw event data are converted into structured interaction behavior trajectory data, including click area data (recording the content elements that the user pays attention to, 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). These data can accurately depict the user's interaction process with the preview content and reflect the user's focus of interest and attention distribution.

[0103] Conducting multidimensional statistical analysis on the behavioral indicators of users watching short dramas is a way to obtain deep-seated user preferences. Various behavioral data generated by users in the actual process of watching short dramas are collected, including viewing completion rate (calculating the ratio of actual viewing time to the total length of the short drama), repeated viewing segment data (identifying the video intervals and frequency of multiple replays by users), interactive comment data (recording the content, time points and emotional tendencies of user comments) and social sharing behavior data (recording sharing platforms, sharing texts and sharing frequencies). Cluster analysis and principal component analysis are applied to these multidimensional behavioral data to extract key feature dimensions, and through normalization, behavioral indicators of different dimensions are converted into a unified numerical range. Finally, a multidimensional behavioral feature vector is constructed, in which each dimension represents a type of behavioral feature, and the vector value reflects the intensity of the behavior, which fully describes the user's interaction pattern and interest in the content of the short drama.

[0104] Correlating interactive behavior trajectory data with behavioral feature vectors forms the basis for constructing a user-content interaction intensity matrix. First, each skit is categorized by content characteristics and emotional type to form a skit-feature matrix. Then, the user's interactive behavior trajectory data (for the preview phase) is time-aligned and feature-mapped with the behavioral feature vectors (for the viewing phase), analyzing the correlation between preview-phase behavior and actual viewing-phase 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. The rows of the matrix represent users, and the columns represent content of different types or characteristics. The matrix values ​​represent the user's interaction intensity with that type of content, comprehensively reflecting the user's content preferences throughout the preview and viewing cycle.

[0105] Applying the emotional satisfaction calculation model to the user-content interaction intensity matrix is ​​a key step in accurately measuring user emotional preferences. The emotional satisfaction calculation model is a multi-factor weighted model that uses the emotional type of content as an analytical dimension. It infers user emotional preferences by analyzing the intensity of user interaction with content of different emotional types. The model first categorizes the short drama content into multiple categories based on its emotional characteristics, such as "inspirational and growth-oriented," "warm and touching," and "thrilling and exciting." It then calculates the weighted mean of the column vectors corresponding to each emotional type in the user-content interaction intensity matrix, using completion rate, repeat viewing, and social sharing as weighting factors. Weights are determined based on the strength of the behavior's indicator of emotional satisfaction; typically, completion rate is weighted 0.4, repeat viewing 0.3, and social sharing 0.3. The resulting emotional match score is a numeric vector between 0 and 1, with each element corresponding to an emotional type. Higher values ​​indicate a higher user preference for content of that type.

[0106] Gradient descent optimization of the weight coefficients of each sentiment dimension in the content recommendation model based on the sentiment matching score is the core mechanism for adaptive adjustment of the recommendation system. The content recommendation model contains multiple sentiment dimension weight parameters, which determine the proportion of content 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 to gradually reach 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 under the current weight parameters and the actual sentiment matching score is calculated. The weight parameters are then adjusted in the direction of the gradient to gradually reduce the prediction error. After multiple rounds of iteration, an updated set of sentiment weight parameters is obtained. These parameters more accurately reflect the user's true emotional preferences and provide a more precise decision-making basis for subsequent recommendations.

[0107] Applying the updated emotional weight parameter set to the candidate content pool for re-ranking is the final step in generating a personalized recommendation list. First, a set of candidate short dramas that meet the user's basic preferences are screened from the content library, including the latest online, popular, and similar short dramas to the user's historical preferences. Then, the content matching score and emotional matching score are calculated for each candidate short drama. The content matching score is based on the similarity between the content feature vector and the user's interest vector, and the emotional matching score is based on the dot product calculation of the emotional characteristic vector of the short drama and the updated emotional 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 sorted according to the recommendation score, while taking into account the requirements of the emotional sequence template to ensure that the recommendation list meets the personalized matching requirements and has reasonable emotional rhythm changes. Finally, a target short drama recommendation list with high personalization and optimized emotional experience is generated.

[0108] The entire data processing process is illustrated using the example of a professional's short drama recommendation optimization process. While browsing recommended short drama previews, this user's interaction behavior trajectory data shows that they frequently click on keywords such as "teamwork" and "career advancement," spend extended time on dynamic preview clips depicting workplace challenges, and quickly scroll past content depicting emotional entanglements. This interaction data is recorded via an event capture system, forming a structured interaction behavior trajectory. Furthermore, this user's viewing behavior indicators reveal an average completion rate of 85% for short dramas about career growth. They frequently rewatch key clips addressing workplace problems and share inspirational workplace content on social platforms, but rarely interact with or share emotional dramas. Correlation analysis reveals that this user's high interest in career success scenes during the preview phase is positively correlated with a high completion rate during the viewing phase. A user-content interaction intensity matrix shows that this user's interaction intensity for content depicting "career struggle" and "teamwork" is highest, at 0.89 and 0.83, respectively, while their interaction intensity for content depicting "emotional entanglements" is only 0.32. Analysis using the emotional satisfaction calculation model revealed that the user's match score for the "positive" emotional category was 0.92, for the "exciting" category was 0.78, for the "warm and touching" category was 0.65, and for the "sad" category was only 0.25. Based on these emotional match scores, the weights of the emotional dimensions in the recommendation model were adjusted using a gradient descent optimization algorithm. The initial emotional weights ["positive": 0.5, "exciting": 0.3, "warm and touching": 0.1, "sad": 0.1] were adjusted to ["positive": 0.64, "exciting": 0.22, "warm and touching": 0.12, "sad": 0.02] after five rounds of iterative optimization. This updated set of emotional weight parameters was applied to the ranking of 300 candidate short dramas. Combined with the "progressive climax" emotional sequence template, a list of recommended short dramas was generated that primarily focused on workplace motivation and incorporated appropriate teamwork challenges. The recommendations better aligned with the user's content preferences and emotional needs, achieving closed-loop optimization and precise personalization of the recommendation system.

[0109] In a method for personalized short drama content push based on big data mining, closed-loop feedback and adaptive iterative optimization are key components for ensuring the continuous improvement of the recommendation system. First, events are captured during user interactions with content preview data. 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 the mouse hovering over each element), swipe events (recording the direction and speed of finger or mouse swipes), and view switch events (recording the order in which users browse through preview resources). After preprocessing, this raw event data is converted into structured interaction trajectory data, including click area data (recording the content elements that users focused on, such as specific characters or emotional climaxes), dwell time data (recording the user's attention to different preview elements), and navigation path data (recording the order in which users' attention shifted). This data accurately depicts the user's interaction with the preview content, reflecting the user's focus and attention distribution.

[0110] Conducting multidimensional statistical analysis on the behavioral indicators of users watching short dramas is a way to obtain deep-seated user preferences. Various behavioral data generated by users in the actual process of watching short dramas are collected, including viewing completion rate (calculating the ratio of actual viewing time to the total length of the short drama), repeated viewing segment data (identifying the video intervals and frequency of multiple replays by users), interactive comment data (recording the content, time points and emotional tendencies of user comments) and social sharing behavior data (recording sharing platforms, sharing texts and sharing frequencies). Cluster analysis and principal component analysis are applied to these multidimensional behavioral data to extract key feature dimensions, and through normalization, behavioral indicators of different dimensions are converted into a unified numerical range. Finally, a multidimensional behavioral feature vector is constructed, in which each dimension represents a type of behavioral feature, and the vector value reflects the intensity of the behavior, which fully describes the user's interaction pattern and interest in the content of the short drama.

[0111] Correlating interactive behavior trajectory data with behavioral feature vectors forms the basis for constructing a user-content interaction intensity matrix. First, each skit is categorized by content characteristics and emotional type to form a skit-feature matrix. Then, the user's interactive behavior trajectory data (for the preview phase) is time-aligned and feature-mapped with the behavioral feature vectors (for the viewing phase), analyzing the correlation between preview-phase behavior and actual viewing-phase 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. The rows of the matrix represent users, and the columns represent content of different types or characteristics. The matrix values ​​represent the user's interaction intensity with that type of content, comprehensively reflecting the user's content preferences throughout the preview and viewing cycle.

[0112] Applying the emotional satisfaction calculation model to the user-content interaction intensity matrix is ​​a key step in accurately measuring user emotional preferences. The emotional satisfaction calculation model is a multi-factor weighted model that uses the emotional type of content as an analytical dimension. It infers user emotional preferences by analyzing the intensity of user interaction with content of different emotional types. The model first categorizes the short drama content into multiple categories based on its emotional characteristics, such as "inspirational and growth-oriented," "warm and touching," and "thrilling and exciting." It then calculates the weighted mean of the column vectors corresponding to each emotional type in the user-content interaction intensity matrix, using completion rate, repeat viewing, and social sharing as weighting factors. Weights are determined based on the strength of the behavior's indicator of emotional satisfaction; typically, completion rate is weighted 0.4, repeat viewing 0.3, and social sharing 0.3. The resulting emotional match score is a numeric vector between 0 and 1, with each element corresponding to an emotional type. Higher values ​​indicate a higher user preference for content of that type.

[0113] Gradient descent optimization of the weight coefficients of each sentiment dimension in the content recommendation model based on the sentiment matching score is the core mechanism for adaptive adjustment of the recommendation system. The content recommendation model contains multiple sentiment dimension weight parameters, which determine the proportion of content 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 to gradually reach 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 under the current weight parameters and the actual sentiment matching score is calculated. The weight parameters are then adjusted in the direction of the gradient to gradually reduce the prediction error. After multiple rounds of iteration, an updated set of sentiment weight parameters is obtained. These parameters more accurately reflect the user's true emotional preferences and provide a more precise decision-making basis for subsequent recommendations.

[0114] Applying the updated emotional weight parameter set to the candidate content pool for re-ranking is the final step in generating a personalized recommendation list. First, a set of candidate short dramas that meet the user's basic preferences are screened from the content library, including the latest online, popular, and similar short dramas to the user's historical preferences. Then, the content matching score and emotional matching score are calculated for each candidate short drama. The content matching score is based on the similarity between the content feature vector and the user's interest vector, and the emotional matching score is based on the dot product calculation of the emotional characteristic vector of the short drama and the updated emotional 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 sorted according to the recommendation score, while taking into account the requirements of the emotional sequence template to ensure that the recommendation list meets the personalized matching requirements and has reasonable emotional rhythm changes. Finally, a target short drama recommendation list with high personalization and optimized emotional experience is generated.

[0115] The entire data processing process is illustrated using the example of a professional's short drama recommendation optimization process. While browsing recommended short drama previews, this user's interaction behavior trajectory data shows that they frequently click on keywords such as "teamwork" and "career advancement," spend extended time on dynamic preview clips depicting workplace challenges, and quickly scroll past content depicting emotional entanglements. This interaction data is recorded via an event capture system, forming a structured interaction behavior trajectory. Furthermore, this user's viewing behavior indicators reveal an average completion rate of 85% for short dramas about career growth. They frequently rewatch key clips addressing workplace problems and share inspirational workplace content on social platforms, but rarely interact with or share emotional dramas. Correlation analysis reveals that this user's high interest in career success scenes during the preview phase is positively correlated with a high completion rate during the viewing phase. A user-content interaction intensity matrix shows that this user's interaction intensity for content depicting "career struggle" and "teamwork" is highest, at 0.89 and 0.83, respectively, while their interaction intensity for content depicting "emotional entanglements" is only 0.32. Analysis using the emotional satisfaction calculation model revealed that the user's match score for the "positive" emotional category was 0.92, for the "exciting" category was 0.78, for the "warm and touching" category was 0.65, and for the "sad" category was only 0.25. Based on these emotional match scores, the weights of the emotional dimensions in the recommendation model were adjusted using a gradient descent optimization algorithm. The initial emotional weights ["positive": 0.5, "exciting": 0.3, "warm and touching": 0.1, "sad": 0.1] were adjusted to ["positive": 0.64, "exciting": 0.22, "warm and touching": 0.12, "sad": 0.02] after five rounds of iterative optimization. This updated set of emotional weight parameters was applied to the ranking of 300 candidate short dramas. Combined with the "progressive climax" emotional sequence template, a list of recommended short dramas was generated that primarily focused on workplace motivation and incorporated appropriate teamwork challenges. The recommendations better aligned with 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 pushing personalized content of short plays based on big data mining in the embodiment of the present application. The following describes the system for pushing personalized content of short plays based on big data mining in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of a short play personalized content push system based on big data mining includes:

[0117] Construction module 201, for performing feature extraction and label construction processing on the skit content to generate a feature vector of the skit content;

[0118] Analysis module 202, for collecting and sentimentally analyzing user behavior data based on the skit content feature vector to obtain user sentiment pattern data;

[0119] The training module 203 is used to construct a user interest map based on the feature vector of the short play content and the user emotion pattern data, and perform emotion resonance network training on the user interest map to obtain a content recommendation model;

[0120] A sorting module 204 is configured to apply the content recommendation model to the candidate content pool, sort the candidate skits using an emotional rhythm planning algorithm, and obtain a recommended skit list;

[0121] A generating module 205 is configured to generate multimodal preview resources for the short play contents in the short play recommendation list to obtain content preview data;

[0122] The recommendation module 206 is used to collect the interactive behavior data of the user and the content preview data, and calculate the emotional matching score based on the user's viewing completion rate, number of repeated viewings and social sharing behavior, and adjust the emotional weight coefficient in the content recommendation model according to the emotional matching score to generate a target short drama recommendation list.

[0123] Through the collaborative cooperation of the above components, the feature extraction and label construction of the short play content are carried out to generate the content feature vector, which realizes the accurate expression of the multi-dimensional information of the short play content. Based on the feature vector of the short play content, the user behavior data is collected and the sentiment analysis is processed to obtain the user emotion pattern data, breaking through the limitation of the traditional recommendation system that only focuses on explicit behavior. Through innovative technologies such as micro-expression recognition, the user's real emotional preferences are captured, which greatly improves the accuracy and depth of user portraits; based on the feature vector of the short play content and the user's emotion pattern data, the user interest map is constructed and the emotion resonance network is trained to obtain the content recommendation model, which innovatively combines the graph neural network with the emotion resonance enhancement network, which can not only identify the user's content type preference, but also accurately capture the content features that can trigger the user's emotional resonance, realizing the upgrade of the recommendation dimension from "content matching" to "emotional resonance"; The content recommendation model is applied to the candidate content pool, and is sorted and processed through the emotional rhythm planning algorithm to obtain a list of short drama recommendations. The rhythm construction principle in music theory is creatively introduced, so that the recommendation results are no longer an isolated content collection, but a content sequence with artistic emotional rhythm, which significantly improves the user's continuous viewing experience; multimodal preview resource generation and processing are performed on the short drama content in the short drama recommendation list to obtain content preview data, breaking through the traditional single preview mode of static cover plus text introduction, and greatly improving the value delivery efficiency of short drama content through multimodal preview methods such as emotional climax dynamic cover and emotional concentration heat map; the interaction behavior of users with preview data is collected, and the emotional matching score is calculated in combination with indicators such as viewing completion rate, and the model parameters are dynamically adjusted to generate the target recommendation list, and a closed-loop optimization mechanism is constructed to enable the recommendation system to continuously optimize itself 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 the embodiment of the present invention is described in detail from the perspective of modular functional entities. The short drama personalized content push device based on big data mining in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0125] Figure 3: This is a schematic diagram of the structure of a device for pushing personalized content for short plays based on big data mining provided by an embodiment of the present invention. The device 300 for pushing personalized content for short plays based on big data mining may have relatively large differences due to different configurations or performances. It may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage devices) for storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be temporary storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), each of which may include a series of instruction operations in the device 300 for pushing personalized content for short plays based on big data mining. Furthermore, the processor 310 can be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the device 300 for pushing personalized content for short plays based on big data mining to implement the steps of the above-mentioned method for pushing personalized content for short plays based on big data mining.

[0126] The short play 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 and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the short play personalized content push device based on big data mining shown does not constitute a limitation of the short play personalized content push device based on big data mining provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[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. The computer-readable storage medium stores instructions. When the instructions are run on a computer, the computer executes the steps of the method for pushing personalized content of short dramas based on big data mining.

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

[0129] If the integrated unit is implemented in the form of 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 is essentially 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. The computer software product is stored in a storage medium, including several instructions for enabling a short drama personalized content push device based on big data mining (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk and other media that can store program code.

[0130] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for pushing personalized content of short dramas based on big data mining, characterized in that: The method comprises: Perform feature extraction and label construction on the skit content to generate a feature vector of the skit content; Collecting and sentiment analysis of user behavior data based on the short play content feature vector to obtain user sentiment pattern data; Constructing a user interest graph based on the short play content feature vector and the user emotion pattern data, and performing emotion resonance network training on the user interest graph to obtain a content recommendation model; Applying the content recommendation model to the candidate content pool, sorting the candidate short plays using the emotional rhythm planning algorithm, and obtaining a short play recommendation list; Performing multimodal preview resource generation processing on the short drama content in the short drama recommendation list to obtain content preview data; Collect the user's interactive behavior data with the content preview data, and calculate the emotional matching score based on the user's viewing completion rate, number of repeated viewings and social sharing behavior. Adjust the emotional weight coefficient in the content recommendation model according to the emotional matching score to generate a target short drama recommendation list.

2. The method for pushing personalized content of short dramas based on big data mining according to claim 1 is characterized in that: The feature extraction and label construction processing of the skit content to generate the feature vector of the skit content includes: Key frames are extracted from the short drama videos and analyzed using image recognition algorithms to obtain scene type and visual style features. Perform word segmentation and semantic analysis on the dialogue and subtitles of the skit to obtain topic keywords and sentiment tendency data; Statistical analysis was performed on the editing rate and transition methods of the skits to obtain rhythm characteristics and narrative structure data; Perform audio feature extraction on the sound elements of the skit to obtain the musical tone and emotional rendering features; Performing data fusion processing on the visual style features, the emotional tendency data, the narrative structure data, and the emotional rendering features to construct a multidimensional label vector; The multi-dimensional label vector is normalized and dimensionally reduced to generate a feature vector of the short play content.

3. The method for pushing personalized content of short dramas based on big data mining according to claim 1 is characterized in that: The user behavior data is collected and sentiment analysis is performed based on the feature vector of the short play content to obtain user sentiment pattern data, including: Through the real-time data stream processing framework, we collect all dimensions of user viewing behavior, including viewing time, completion rate, number of pauses, fast forward and rewind behavior, and repeated viewing of segments. Recognize and process facial micro-expression data collected by the front camera of the user's device to obtain the emotional response characteristics of the user during the viewing process; Perform natural language processing analysis on the text content of user comments to obtain explicit emotional expression data including emotional polarity and emotional intensity; Performing time-series alignment processing on the behavior data and the feature vector of the skit content to construct an emotion response-content node mapping matrix; Performing multi-source fusion processing on the emotional response characteristics and the explicit emotional expression data to obtain a two-level emotional model including implicit emotions and explicit emotions; An emotion fluctuation model is constructed based on the dual-level emotion pattern and the emotion response-content node mapping matrix to obtain user emotion pattern data.

4. The method for pushing personalized content of short dramas based on big data mining according to claim 1 is characterized in that: The step of constructing a user interest graph based on the short play content feature vector and the user emotion pattern data, and performing emotion resonance network training on the user interest graph to obtain a content recommendation model includes: Based on the skit content feature vector and the user emotion pattern data, a knowledge graph structure with the user as the central node is constructed to obtain an initial interest graph including content categories, skit entities, and emotion labels; Applying a graph neural network algorithm to the initial interest graph to perform representation learning processing to obtain a low-dimensional dense vector that captures the structural characteristics of the user's interests; Construct a three-component neural network architecture consisting of an emotion trajectory extractor, a resonance matcher, and an emotion enhancement predictor to obtain an emotion resonance enhancement network model; Inputting the emotional change data of the short play content into the emotional trajectory extractor for processing to obtain the emotional trajectory of the content that captures the emotional climax and turning points in the plot development; Inputting the content emotion trajectory and the user emotion pattern data into the resonance matcher for similarity calculation processing to obtain a user-content emotion resonance metric value; Based on the user-content emotion resonance metric and the low-dimensional dense vector, deep feature fusion processing is performed to obtain a content recommendation model that integrates the dual capabilities of semantic matching and emotion resonance.

5. The method for pushing personalized content of short dramas based on big data mining according to claim 1 is characterized in that: The content recommendation model is applied to the candidate content pool, and the candidate short plays are sorted by the emotional rhythm planning algorithm to obtain a short play recommendation list, including: Build a candidate content pool that includes the latest online short dramas, popular short dramas, and short dramas similar to the user's historical preferences to obtain an initial set of candidate short dramas; Performing a matching calculation on each short play in the initial candidate short play set using the content recommendation model to obtain a basic recommendation score; The content of the short play was classified according to the emotional type, and the emotional classification results including high-energy emotional type, medium-energy emotional type and low-energy emotional type were obtained; Based on the emotion classification results, a plurality of emotion sequence templates are constructed to obtain an emotion sequence library including a progressive climax type, an emotion fluctuation type, and a soothing transition type; Perform emotional sequence preference analysis on the user's historical viewing behavior to obtain the most suitable emotional sequence template for the user; The initial candidate skit set is dynamically sorted and combined according to the emotional sequence template and the basic recommendation score to obtain a recommended list of skits with a preset emotional rhythm.

6. The method for pushing personalized content of short dramas based on big data mining according to claim 1 is characterized in that: The step of performing multimodal preview resource generation processing on the short play content in the short play recommendation list to obtain content preview data includes: Perform emotional intensity curve analysis on the skit content, extract video clips whose emotional intensity exceeds a preset threshold, and obtain dynamic preview materials; Extract key sentences from the short play's plot text, rank them by importance, and convert them into audio data using a speech synthesis engine to produce a personalized plot summary. The data on emotional value changes throughout the skit are numerically mapped and color-coded to generate a visual graph of emotional changes with time as the horizontal axis and emotional intensity as the vertical axis. A directed graph structure is constructed based on the character relationship data in the skit content, and the node positions are calculated using a force-directed layout algorithm to obtain an interactive character relationship graph. The dynamic preview material, the personalized plot voice summary, the emotional change visual map, and the interactive character relationship diagram are subjected to resolution adjustment and compression encoding processing according to the current network bandwidth and device parameters to obtain a multi-level preview resource package; The multi-level preview resource package is subjected to component arrangement and interaction event binding processing by a layout engine to obtain content preview data.

7. The method for pushing personalized content of short dramas based on big data mining according to claim 1 is characterized in that: The method collects interactive behavior data between the user and the content preview data, calculates an emotion matching score based on the user's viewing completion rate, number of repeated viewings, and social sharing behavior, and adjusts the emotion weight coefficient in the content recommendation model according to the emotion matching score to generate a target short drama recommendation list, including: Capturing and processing events during the user's interaction with the content preview data to obtain interactive behavior trajectory data including clicked areas, dwell time, and browsing paths; Perform multi-dimensional statistical analysis on users' viewing behavior indicators of short dramas to obtain behavioral feature vectors including viewing completion rate, repeated viewing segments, interactive comment frequency, and social sharing number; Performing correlation analysis on the interaction behavior trajectory data and the behavior feature vector to construct a user-content interaction intensity matrix; Applying an emotion satisfaction calculation model to the user-content interaction intensity matrix to obtain an emotion matching score reflecting the user's preference for content of different emotion types; Performing gradient descent optimization on the weight coefficients of each emotional dimension in the content recommendation model based on the emotional matching score to obtain an updated emotional weight parameter set; The updated emotion weight parameter set is applied to the candidate content pool for re-ranking to generate a target short play recommendation list.

8. A short play personalized content push system based on big data mining, characterized by: A method for pushing personalized content of short plays based on big data mining according to any one of claims 1 to 7 is implemented, wherein the system for pushing personalized content of short plays based on big data mining comprises: A construction module is used to extract features and construct labels for the skit content to generate a feature vector of the skit content; An analysis module, configured to collect and perform sentiment analysis on user behavior data based on the feature vector of the skit content to obtain user sentiment pattern data; A training module, configured to construct a user interest graph based on the feature vector of the skit content and the user emotion pattern data, and perform emotion resonance network training on the user interest graph to obtain a content recommendation model; A sorting module is used to apply the content recommendation model to the candidate content pool, sort the candidate short plays using the emotional rhythm planning algorithm, and obtain a short play recommendation list; A generation module, configured to perform multimodal preview resource generation processing on the short play content in the short play recommendation list to obtain content preview data; The recommendation module is used to collect the user's interactive behavior data with the content preview data, and calculate the emotional matching score based on the user's viewing completion rate, number of repeated viewings and social sharing behavior, adjust the emotional weight coefficient in the content recommendation model according to the emotional matching score, and generate a target short drama recommendation list.

9. A device for pushing personalized content of short dramas based on big data mining, characterized in that: It includes a memory and a processor, the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, it implements the method for pushing personalized content of short dramas based on big data mining as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor executes the method for pushing personalized content of short dramas based on big data mining as described in any one of claims 1 to 7.

Images