A new media content recommendation method and system based on ensemble learning
By combining multiple models and networks to process user behavior data through ensemble learning methods, the accuracy and personalization issues of new media recommendation systems when user behavior is scarce and interests are changing are solved, thus achieving efficient and personalized content recommendation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN DREAM AID PLAN TECHNOLOGY GROUP CO LTD
- Filing Date
- 2026-02-26
- Publication Date
- 2026-06-02
AI Technical Summary
Existing new media recommendation systems have limitations in accuracy and personalization when faced with scarce user behavior data, changing user interests, and cold start problems. Furthermore, deep learning methods are computationally expensive and struggle to respond quickly to changes in user preferences.
An ensemble learning approach is adopted, combining collaborative filtering models, content-based recommendation models, and deep learning models. Through a multi-level weighted fusion strategy, LSTM and CNN networks are used to process user behavior data, capture users' long-term interests and short-term preferences, and combine contextual information to perform recommendation ranking.
It significantly improves the accuracy and diversity of recommendations, can respond to changes in user behavior in real time, solves the cold start problem, and provides more accurate and personalized content recommendations.
Smart Images
Figure CN122132626A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ensemble learning technology, and in particular to a new media content recommendation method and system based on ensemble learning. Background Technology
[0002] With the rapid development of information technology, user-generated content (UGC) on new media platforms has grown dramatically, and users' demands for personalized content have become increasingly strong. To meet these diverse needs, recommendation systems, as a crucial component of new media platforms, have been widely adopted. Traditional recommendation system methods include content-based recommendation, collaborative filtering recommendation, and hybrid recommendation. Content-based recommendation methods primarily rely on the similarity between user and content features, making them suitable for scenarios where content types are fixed and user interests are relatively stable. However, when faced with massive amounts of content data, content-based recommendation methods suffer from the cold start problem, making it difficult to accurately capture users' potential interests. Collaborative filtering recommendation, by analyzing the similarity between users, can effectively recommend content that users may be interested in, but its limitation lies in its reliance on historical user behavior data; when user behavior data is scarce, the accuracy of the recommendation results is significantly reduced.
[0003] Currently, most recommender systems still rely on a single recommendation algorithm, leading to limitations in the accuracy and personalization of recommendation results. For example, collaborative filtering-based recommendation methods perform poorly when user behavior data is scarce, especially in the cold start problem, where the system struggles to provide effective recommendations for new users or new content. Furthermore, while content-based recommendation methods can recommend similar content to users relatively well, they neglect the dynamic changes in user interests and fail to fully consider user behavior sequences and contextual information, resulting in a lack of diversity and timeliness in recommendation results.
[0004] While some research has attempted to combine multiple recommendation methods for hybrid recommendation systems, existing hybrid recommendation algorithms primarily fuse the outputs of different models through simple weighting or voting, often failing to fully leverage the strengths of each model. Furthermore, although deep learning methods can model the complex non-linear relationships between user behavior and content, their reliance on large amounts of data and long training times results in high computational costs and makes it difficult to quickly respond to changes in user preferences.
[0005] Therefore, how to provide a new media content recommendation method and system based on ensemble learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] One objective of this invention is to propose a new media content recommendation method and system based on ensemble learning. This invention fully utilizes ensemble learning, LSTM networks, CNN networks, and NLP techniques, detailing how weighted fusion of multiple models, temporal processing of user behavior data, and deep perception of contextual information enhance the personalization and accuracy of the recommendation system. Specifically, this invention comprehensively analyzes multi-dimensional information such as user behavior characteristics, social interactions, and device type, and combines this with an ensemble learning framework to achieve deep integration of different recommendation algorithms, overcoming the shortcomings of traditional recommendation systems when facing scarce user behavior, changing user interests, and cold start problems.
[0007] A new media content recommendation method based on ensemble learning according to an embodiment of the present invention includes the following steps: S1. Extract multi-dimensional features from user behavior data, including user click history, browsing history, comment behavior, sharing behavior and like data. Through data cleaning and standardization, generate a behavior feature vector containing time, frequency and behavior type. S2. Use NLP technology to perform deep semantic analysis on the text in user interaction with new media content, identify keywords and topics through syntax parsing algorithms, and construct user interest semantic vectors. S3. An improved CNN network is used to jointly model behavioral feature vectors and semantic representation vectors, capture the non-linear relationship between user behavior and the semantics of new media content, and perform feature cross-fertilization and fusion. S4. Based on the ensemble learning framework, the recommendation results of the collaborative filtering model, the content recommendation model and the deep learning model are combined through a multi-level weighted fusion strategy to generate a candidate set of new media content recommendations. S5. Based on the improved LSTM network, perform time-series prediction on the recommendation candidate set, analyze users' long-term interests and short-term preferences, and adjust the recommendation ranking of new media content in combination with contextual information. S6 uses a Hadoop database for large-scale data storage and querying, leverages a cloud-network collaborative architecture to update recommendation rankings in real time, and provides recommendation services to users, pushing new media content to them.
[0008] Optionally, S1 specifically includes: S11. Collect user behavior data on new media platforms and extract user behavior characteristics, including: Click history: the time, page, and frequency of each click; Browsing history: Pages viewed each time, duration of stay, and browsing time period; Commenting behavior: comment content, comment duration, and comment sentiment; Sharing behavior: the content shared, the time of sharing, and the platform on which it is shared; Likes data: Liked content, liked time period, and liked frequency; S12. The above behavioral data is cleaned to remove invalid data and outliers, and missing data is filled by nearest neighbor interpolation. It is then standardized by timestamp, frequency and category label to convert it into a value between 0 and 1, removing the deviation caused by the difference in data volume, and forming a behavioral feature vector in a unified format.
[0009] Optionally, S2 specifically includes: S21. Collect the text of user interactions with new media content, including user comments, replies, shared content, dynamic posts, and message interactions; S22. Use the pre-trained language model BERT based on the Transformer architecture to perform word segmentation and syntactic analysis on the text to ensure that the semantic information of each word is fully captured. S23. Model the grammatical structure of the text using a dependency parsing algorithm to clarify the relationships between each word, including subject-verb-object structures, modification relationships, and clause relationships; S24. Based on this, the Skip-gram model is used to vectorize all words and map them to a high-dimensional vector space. The contextual semantic information of the words is obtained through training. S25. Use the K-means algorithm to cluster words, identify hidden keywords in the text, and optimize the weight of keywords through the TF-IDF algorithm to emphasize high-frequency and representative keywords. S26. Use the latent Dirichlet assignment algorithm to model topics for all texts, extract the topics that each user is interested in, and construct a user interest topic model based on the degree of correlation between the user and each topic. S27. Use an MLP network to fuse all extracted keywords, interest topic models, and user behavior data to generate user interest semantic vectors.
[0010] Optionally, the keyword identification process specifically includes: After segmenting all collected text and removing stop words, the Skip-gram model is used to vectorize each word, including: The Skip-gram model is adopted. During training, context word pairs are extracted from the text through a sliding window mechanism. Each time, the context words in the window are predicted based on the current word, and the vector representation of the words is adjusted by maximizing the accuracy of the prediction. For each text, a large amount of corpus data will be traversed during the training process, so that each word is mapped to a vector in a high-dimensional space, and the vector can reflect the semantic relationship between words; The K-means algorithm is used to cluster all word vectors. The value of K is selected according to the characteristics of the dataset. The Euclidean distance between each word and the cluster center is calculated. All words are assigned to the cluster closest to the cluster center, forming multiple word clusters with similar semantics. The TF-IDF algorithm is used to weight the words in each cluster, including: Calculate the inverse document frequency (IDF) value of each word in the entire text corpus. The higher the IDF value, the stronger the word's discriminative power in the text. By combining the frequency (TF) value of words in the current document, the TF-IDF value of each word is calculated to obtain keywords that are highly representative in each text.
[0011] Optionally, the topic modeling process of the potential Dirichlet assignment algorithm specifically includes: LDA assumes that each text consists of multiple topics, and each topic is represented by a probability distribution of words; Set the number of topics K in the LDA model, and use the variational inference algorithm to model the documents. Initially, the topic distribution of each document is randomly assigned, and the topic assignment of each word is randomly initialized. During the inference process, LDA adjusts the topic assignment probability of each word based on the probability of each word appearing and the overall content of the document; As the number of iterations increases, the LDA model will gradually converge, eventually outputting the topic distribution of each text and the vocabulary distribution of each topic; Based on the LDA model, user behavior features are mapped to the weights of each topic, and different priorities are assigned to each topic according to the frequency of user behavior and interaction. Based on the distribution of user interests, a correlation score between the user and each topic is generated to form a user interest topic model. The correlation score reflects the intensity of the user's interest in a specific topic.
[0012] Optionally, S3 specifically includes: S31. Standardize the user's behavior feature vector and semantic representation vector to ensure consistent data range and avoid affecting network training due to uneven feature values. S32. When performing joint modeling, the CNN network structure is improved by assigning user behavior features and new media content semantic features to different convolutional channels for processing, capturing the non-linear relationships within them. Specifically, this includes: Small-scale convolutional kernels process low-frequency user behavior data, including occasional clicks and short-term dwell times; Mesoscale convolutional kernels handle mid-frequency behaviors, including prolonged browsing and sustained liking. Large-scale convolutional kernels handle high-frequency behaviors, including frequent commenting and sharing. S33. By designing specific convolution kernels, the relationship between behavioral feature vectors and semantic representation vectors can be directly cross-captured during the convolution process; S34. When processing semantic representation vectors, three different scales of convolution kernels are used to extract local features of short text, long text and deep semantics respectively. S35. Each convolutional kernel extracts features at the corresponding scale through convolution operations, and then uses max pooling layers to compress the features, reducing the amount of computation. S36. After the convolutional layer, a multi-head self-attention mechanism is introduced. By calculating the similarity matrix between different convolutional channels, the key features of each channel are automatically selected for weighted fusion. S37. After feature fusion, the low-level convolutional features and the high-level fused features are directly passed to the next layer through cross-layer connections to avoid information loss.
[0013] Optionally, S4 specifically includes: S41. For the recommendation results of each model, an ensemble learning framework is used for deep fusion. The models include a collaborative filtering model, a content-based recommendation model, and a deep learning model, specifically including: The collaborative filtering model uses user-item similarity calculation and extracts potential user behavior patterns and item similarities from the user-item interaction matrix through matrix singular value decomposition, thereby recommending items with similar behaviors to each user; Content recommendation models rely on the characteristics of new media content, including text, images, and tags. By calculating the similarity between a user's historical behavior and the content of an item, the model recommends the most similar content to the user. The deep learning model uses an improved CNN network to take behavioral feature vectors and semantic representation vectors as input, extracts deep features through multi-layer convolution and pooling operations, and finally outputs personalized recommendation results through fully connected layers. S42. During the deep weighted fusion process, calculate the performance metrics of each model on the training set, including precision, recall, and F1 score. Assign different weight coefficients to each model based on its performance in different scenarios, including: When user activity is high, collaborative filtering models will receive higher weights, while when user behavior data is scarce, content-based recommendation models will receive higher weights. S43. Use a weighted voting mechanism to perform preliminary weighted fusion of the recommendation results of each model to obtain a preliminary recommendation candidate set; S44. Introduce the gradient boosting tree algorithm for secondary weighted fusion, specifically including: The initial recommendation candidate set of each model is transformed into a feature vector, representing the score of each content under different models; These features are learned using the gradient boosting tree algorithm to train a fusion model. The fusion model can automatically adjust the weighting coefficients of each model and output the final weighted recommendation score. S45. Optimize and combine the initial recommendation candidate set based on the weighted recommendation score to form a recommendation candidate set for new media content, and verify it through online A / B testing.
[0014] Optionally, S5 specifically includes: S51. Perform time-series processing on user behavior data, label each user behavior with its corresponding timestamp, and form a user behavior time series; S52. To explore users’ long-term interests and short-term preferences, the LSTM network is structurally improved. The forward LSTM captures the historical dependence of user behavior, while the backward LSTM can better understand the potential impact of the user’s current behavior on the future. S53. By inputting user behavior data into a bidirectional LSTM network, the input at each time step not only considers the user's past behavior, but also the future behavior trend. S54. To solve the gradient vanishing problem, an attention-based weight adjustment layer is introduced into the bidirectional LSTM network. The weight adjustment layer adjusts the weights of the time steps according to the user's behavior frequency, time interval and interaction intensity, so that short-term behavior occupies a larger proportion in the prediction, while long-term behavior is stably maintained through enhanced weights. S55. To address the changes in user preferences at different time periods, a context-aware layer is designed. The context information of time period, user device type, and social activities is used as additional input. The features are extracted by an improved CNN network, and these features are concatenated and fused with the output of a bidirectional LSTM network. S56. Based on the time series embedding mechanism and user behavior time series, the user behavior characteristics of each hour within a 24-hour period are labeled, and the weight of new media content is dynamically adjusted according to the time of the behavior. S57. In the prediction phase, the bidirectional LSTM network combines contextual information to output the user's long-term interests and short-term preferences, and recommends and ranks new media content.
[0015] A new media content recommendation system based on ensemble learning according to an embodiment of the present invention includes: The data processing module is used to extract multi-dimensional features from user behavior data, including users' click history, browsing history, comment behavior, sharing behavior and like data. Through data cleaning and standardization, it generates behavioral feature vectors that include time, frequency and behavior type. The semantic analysis module is used to perform deep semantic analysis on the text in user interactions with new media content using NLP technology. It identifies keywords and topics through syntax parsing algorithms and constructs semantic vectors of user interests. The feature modeling module is used to jointly model behavioral feature vectors and semantic representation vectors using an improved CNN network, capture the non-linear relationship between user behavior and the semantics of new media content, and perform feature cross-fertilization and fusion. The content recommendation module is used to combine the recommendation results of collaborative filtering model, content recommendation model and deep learning model based on ensemble learning framework and through multi-level weighted fusion strategy to generate a candidate set of new media content recommendations. The temporal prediction module is used to perform temporal prediction on the recommendation candidate set based on the improved LSTM network, analyze users' long-term interests and short-term preferences, and adjust the recommendation ranking of new media content in combination with contextual information. The storage update module is used for large-scale data storage and querying based on Hadoop databases, uses a cloud-network collaborative architecture to update the recommendation ranking in real time, and provides the recommendation service to the user end to push new media content to the user.
[0016] The beneficial effects of this invention are: First, this invention employs a strategy of integrating multiple models, combining collaborative filtering, content-based recommendation, and deep learning models to achieve weighted fusion of recommendation results. This method leverages the strengths of each model, effectively overcoming the limitations of single recommendation algorithms in the face of scarce user behavior data and cold-start problems, significantly improving the accuracy and diversity of recommendations.
[0017] Secondly, this invention, by processing user behavior data in a time-series manner and introducing a bidirectional LSTM network, can accurately capture changes in users' long-term interests and short-term preferences. This structural improvement enables the recommender system to respond to changes in user behavior in real time, improving the timeliness and personalization of recommendations, and avoiding the limitations of traditional recommender methods in dynamic user interest scenarios.
[0018] Finally, this invention designs a context-aware layer, which further enhances the diversity and relevance of recommendation results through in-depth analysis of contextual information such as user device type, social activities, and time periods. This innovative approach enables the recommendation system to more accurately understand user needs when dealing with the complexity and variability of user behavior, thereby providing more precise and personalized content recommendations. Attached Figure Description
[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a new media content recommendation method based on ensemble learning proposed in this invention; Figure 2 This is a schematic diagram of the user interest prediction process for a new media content recommendation method based on ensemble learning proposed in this invention. Figure 3 This is a module structure diagram of a new media content recommendation system based on ensemble learning proposed in this invention. Detailed Implementation
[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0021] refer to Figures 1-2 A new media content recommendation method based on ensemble learning includes the following steps: S1. Extract multi-dimensional features from user behavior data, including user click history, browsing history, comment behavior, sharing behavior and like data. Through data cleaning and standardization, generate a behavior feature vector containing time, frequency and behavior type. S2. Use NLP technology to perform deep semantic analysis on the text in user interaction with new media content, identify keywords and topics through syntax parsing algorithms, and construct user interest semantic vectors. S3. An improved CNN network is used to jointly model behavioral feature vectors and semantic representation vectors, capture the non-linear relationship between user behavior and the semantics of new media content, and perform feature cross-fertilization and fusion. S4. Based on the ensemble learning framework, the recommendation results of the collaborative filtering model, the content recommendation model and the deep learning model are combined through a multi-level weighted fusion strategy to generate a candidate set of new media content recommendations. S5. Based on the improved LSTM network, perform time-series prediction on the recommendation candidate set, analyze users' long-term interests and short-term preferences, and adjust the recommendation ranking of new media content in combination with contextual information. S6 uses a Hadoop database for large-scale data storage and querying, leverages a cloud-network collaborative architecture to update recommendation rankings in real time, and provides recommendation services to users, pushing new media content to them.
[0022] In this embodiment, S1 specifically includes: S11. Collect user behavior data on new media platforms and extract user behavior characteristics, including: Click history: the time, page, and frequency of each click; Browsing history: Pages viewed each time, duration of stay, and browsing time period; Commenting behavior: comment content, comment duration, and comment sentiment; Sharing behavior: the content shared, the time of sharing, and the platform on which it is shared; Likes data: Liked content, liked time period, and liked frequency; S12. The above behavioral data is cleaned to remove invalid data and outliers, and missing data is filled by nearest neighbor interpolation. It is then standardized by timestamp, frequency and category label to convert it into a value between 0 and 1, removing the deviation caused by the difference in data volume, and forming a behavioral feature vector in a unified format.
[0023] In this embodiment, S2 specifically includes: S21. Collect the text of user interactions with new media content, including user comments, replies, shared content, dynamic posts, and message interactions; S22. Use the pre-trained language model BERT based on the Transformer architecture to perform word segmentation and syntactic analysis on the text to ensure that the semantic information of each word is fully captured. S23. Model the grammatical structure of the text using a dependency parsing algorithm to clarify the relationships between each word, including subject-verb-object structures, modification relationships, and clause relationships; S24. Based on this, the Skip-gram model is used to vectorize all words and map them to a high-dimensional vector space. The contextual semantic information of the words is obtained through training. S25. Use the K-means algorithm to cluster words, identify hidden keywords in the text, and optimize the weight of keywords through the TF-IDF algorithm to emphasize high-frequency and representative keywords. S26. Use the latent Dirichlet assignment algorithm to model topics for all texts, extract the topics that each user is interested in, and construct a user interest topic model based on the degree of correlation between the user and each topic. S27. Use an MLP network to fuse all extracted keywords, interest topic models, and user behavior data to generate user interest semantic vectors.
[0024] In this embodiment, the keyword identification process specifically includes: After segmenting all collected text and removing stop words, the Skip-gram model is used to vectorize each word, including: The Skip-gram model is adopted. During training, context word pairs are extracted from the text through a sliding window mechanism. Each time, the context words in the window are predicted based on the current word, and the vector representation of the words is adjusted by maximizing the accuracy of the prediction. For each text, a large amount of corpus data will be traversed during the training process, so that each word is mapped to a vector in a high-dimensional space, and the vector can reflect the semantic relationship between words; The K-means algorithm is used to cluster all word vectors. The value of K is selected according to the characteristics of the dataset. The Euclidean distance between each word and the cluster center is calculated. All words are assigned to the cluster closest to the cluster center, forming multiple word clusters with similar semantics. The TF-IDF algorithm is used to weight the words in each cluster, including: Calculate the inverse document frequency (IDF) value of each word in the entire text corpus. The higher the IDF value, the stronger the word's discriminative power in the text. By combining the frequency (TF) value of words in the current document, the TF-IDF value of each word is calculated to obtain keywords that are highly representative in each text.
[0025] In this embodiment, the topic modeling process of the potential Dirichlet assignment algorithm specifically includes: LDA assumes that each text consists of multiple topics, and each topic is represented by a probability distribution of words; Set the number of topics K in the LDA model, and use the variational inference algorithm to model the documents. Initially, the topic distribution of each document is randomly assigned, and the topic assignment of each word is randomly initialized. During the inference process, LDA adjusts the topic assignment probability of each word based on the probability of each word appearing and the overall content of the document; As the number of iterations increases, the LDA model will gradually converge, eventually outputting the topic distribution of each text and the vocabulary distribution of each topic; Based on the LDA model, user behavior features are mapped to the weights of each topic, and different priorities are assigned to each topic according to the frequency of user behavior and interaction. Based on the distribution of user interests, a correlation score between the user and each topic is generated to form a user interest topic model. The correlation score reflects the intensity of the user's interest in a specific topic.
[0026] In this embodiment, S3 specifically includes: S31. Standardize the user's behavior feature vector and semantic representation vector to ensure consistent data range and avoid affecting network training due to uneven feature values. S32. When performing joint modeling, the CNN network structure is improved by assigning user behavior features and new media content semantic features to different convolutional channels for processing, capturing the non-linear relationships within them. Specifically, this includes: Small-scale convolutional kernels process low-frequency user behavior data, including occasional clicks and short-term dwell times; Mesoscale convolutional kernels handle mid-frequency behaviors, including prolonged browsing and sustained liking. Large-scale convolutional kernels handle high-frequency behaviors, including frequent commenting and sharing. S33. By designing specific convolution kernels, the relationship between behavioral feature vectors and semantic representation vectors can be directly cross-captured during the convolution process; S34. When processing semantic representation vectors, three different scales of convolution kernels are used to extract local features of short text, long text and deep semantics respectively. S35. Each convolutional kernel extracts features at the corresponding scale through convolution operations, and then uses max pooling layers to compress the features, reducing the amount of computation. S36. After the convolutional layer, a multi-head self-attention mechanism is introduced. By calculating the similarity matrix between different convolutional channels, the key features of each channel are automatically selected for weighted fusion. S37. After feature fusion, the low-level convolutional features and the high-level fused features are directly passed to the next layer through cross-layer connections to avoid information loss.
[0027] In this embodiment, S4 specifically includes: S41. For the recommendation results of each model, an ensemble learning framework is used for deep fusion. The models include a collaborative filtering model, a content-based recommendation model, and a deep learning model, specifically including: The collaborative filtering model uses user-item similarity calculation and extracts potential user behavior patterns and item similarities from the user-item interaction matrix through matrix singular value decomposition, thereby recommending items with similar behaviors to each user; Content recommendation models rely on the characteristics of new media content, including text, images, and tags. By calculating the similarity between a user's historical behavior and the content of an item, the model recommends the most similar content to the user. The deep learning model uses an improved CNN network to take behavioral feature vectors and semantic representation vectors as input, extracts deep features through multi-layer convolution and pooling operations, and finally outputs personalized recommendation results through fully connected layers. S42. During the deep weighted fusion process, calculate the performance metrics of each model on the training set, including precision, recall, and F1 score. Assign different weight coefficients to each model based on its performance in different scenarios, including: When user activity is high, collaborative filtering models will receive higher weights, while when user behavior data is scarce, content-based recommendation models will receive higher weights. S43. Use a weighted voting mechanism to perform preliminary weighted fusion of the recommendation results of each model to obtain a preliminary recommendation candidate set; S44. Introduce the gradient boosting tree algorithm for secondary weighted fusion, specifically including: The initial recommendation candidate set of each model is transformed into a feature vector, representing the score of each content under different models; These features are learned using the gradient boosting tree algorithm to train a fusion model. The fusion model can automatically adjust the weighting coefficients of each model and output the final weighted recommendation score. S45. Optimize and combine the initial recommendation candidate set based on the weighted recommendation score to form a recommendation candidate set for new media content, and verify it through online A / B testing.
[0028] In this embodiment, S5 specifically includes: S51. Perform time-series processing on user behavior data, label each user behavior with its corresponding timestamp, and form a user behavior time series; S52. To explore users’ long-term interests and short-term preferences, the LSTM network is structurally improved. The forward LSTM captures the historical dependence of user behavior, while the backward LSTM can better understand the potential impact of the user’s current behavior on the future. S53. By inputting user behavior data into a bidirectional LSTM network, the input at each time step not only considers the user's past behavior, but also the future behavior trend. S54. To solve the gradient vanishing problem, an attention-based weight adjustment layer is introduced into the bidirectional LSTM network. The weight adjustment layer adjusts the weights of the time steps according to the user's behavior frequency, time interval and interaction intensity, so that short-term behavior occupies a larger proportion in the prediction, while long-term behavior is stably maintained through enhanced weights. S55. To address user preference changes across different time periods, a context-aware layer is designed. This layer takes time period, user device type, and social activity context information as additional input, extracts features using an improved CNN network, and then concatenates and fuses these features with the output of a bidirectional LSTM network. Specifically, this includes: Extract device type information from user devices, and obtain device usage time and status through sensors and platform APIs; Extract social activity features from user interaction data on new media platforms, including user posts, discussions, likes, and shares; An improved CNN network is used to perform multi-scale convolutional operations on these device types, time periods, and social activity features; Convolutional kernels are designed in three types: short-term, medium-term, and long-term. Each convolutional kernel is responsible for extracting behavioral patterns at different time scales. After obtaining feature maps through the convolutional layers, the max pooling method is used to retain important time-related features and remove redundant information. The context feature map after convolution and pooling is concatenated and fused with the user behavior sequence features output from the bidirectional LSTM network. S56. Based on the time series embedding mechanism and user behavior time series, the user behavior characteristics of each hour within a 24-hour period are labeled, and the weight of new media content is dynamically adjusted according to the time of the behavior. S57. In the prediction phase, the bidirectional LSTM network combines contextual information to output the user's long-term interests and short-term preferences, and recommends and ranks new media content.
[0029] refer to Figure 3 A new media content recommendation system based on ensemble learning includes: The data processing module is used to extract multi-dimensional features from user behavior data, including users' click history, browsing history, comment behavior, sharing behavior and like data. Through data cleaning and standardization, it generates behavioral feature vectors that include time, frequency and behavior type. The semantic analysis module is used to perform deep semantic analysis on the text in user interactions with new media content using NLP technology. It identifies keywords and topics through syntax parsing algorithms and constructs semantic vectors of user interests. The feature modeling module is used to jointly model behavioral feature vectors and semantic representation vectors using an improved CNN network, capture the non-linear relationship between user behavior and the semantics of new media content, and perform feature cross-fertilization and fusion. The content recommendation module is used to combine the recommendation results of collaborative filtering model, content recommendation model and deep learning model based on ensemble learning framework and through multi-level weighted fusion strategy to generate a candidate set of new media content recommendations. The temporal prediction module is used to perform temporal prediction on the recommendation candidate set based on the improved LSTM network, analyze users' long-term interests and short-term preferences, and adjust the recommendation ranking of new media content in combination with contextual information. The storage update module is used for large-scale data storage and querying based on Hadoop databases, uses a cloud-network collaborative architecture to update the recommendation ranking in real time, and provides the recommendation service to the user end to push new media content to the user.
[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a content recommendation system on a social platform. This platform has a large user base, diverse content, and users' interests change frequently, resulting in numerous cold start problems for the recommendation system. Traditional recommendation algorithms, unable to quickly respond to changes in user interests, suffer from low accuracy and diversity in recommendation results, leading to a poor user experience. In such a scenario, the application of ensemble learning methods can effectively address the shortcomings of existing recommendation systems.
[0031] In this scenario, user behavior data includes click history, browsing history, comments, sharing, and likes. Traditional methods typically rely on a single model (such as collaborative filtering or content-based recommendation models), but these models perform poorly when data is scarce and cannot meet dynamically changing user needs. To address this issue, we employ the ensemble learning method described in this invention. By combining the advantages of multiple models, the system can provide more accurate and diverse recommendation results in different contexts.
[0032] In the implementation of this invention, multi-dimensional feature extraction is first performed on user behavior data. For example, user clicks, comments, and likes are converted into corresponding feature vectors, and after data cleaning and standardization, user behavior features in a unified format are generated. Next, the system uses NLP technology to perform deep semantic analysis on user comments, dynamic posts, and other text data, constructing user interest semantic vectors. Through convolutional neural networks (CNN) and long short-term memory networks (LSTM), the system can process time-series data of user behavior and capture changes in users' long-term interests and short-term preferences.
[0033] Subsequently, through an ensemble learning framework, the system combines the recommendation results of collaborative filtering, content-based recommendation, and deep learning models. The results of each model are weighted and fused based on their performance in different contexts to obtain a comprehensive recommendation candidate set. Finally, an improved LSTM network is used to perform temporal prediction on the recommendation candidate set, and the recommendation results are ranked based on contextual information (such as time period and device type) before being output to the user.
[0034] To verify the effectiveness of this invention, the system was compared with traditional single models (such as collaborative filtering models and content-based recommendation models) in terms of recommendation accuracy and diversity during implementation. Table 1 shows that after introducing the ensemble learning method, the recommendation accuracy improved by 15%, and the recommendation diversity improved by 12%. Especially when dealing with the cold start problem, the ensemble learning method exhibits stronger robustness and can provide more accurate recommendations for new users or new content.
[0035] Table 1. Comparison of the effects of ensemble learning methods in new media content recommendation.
[0036] As shown in Table 1, the ensemble learning method significantly improves recommendation accuracy, precision, recall, and diversity compared to traditional collaborative filtering and content-based recommendation models. Furthermore, the ensemble learning method demonstrates a clear advantage in handling the cold start problem, particularly in recommending new users and new content, effectively providing more accurate and personalized recommendations.
[0037] In summary, this invention not only improves the accuracy and diversity of new media content recommendation systems, but also effectively solves the cold start problem faced by traditional recommendation algorithms, demonstrating good practical application results.
[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A new media content recommendation method based on ensemble learning, characterized in that, Includes the following steps: S1. Extract multi-dimensional features from user behavior data, including user click history, browsing history, comment behavior, sharing behavior and like data. Through data cleaning and standardization, generate a behavior feature vector containing time, frequency and behavior type. S2. Use NLP technology to perform deep semantic analysis on the text in user interaction with new media content, identify keywords and topics through syntax parsing algorithms, and construct user interest semantic vectors. S3. An improved CNN network is used to jointly model behavioral feature vectors and semantic representation vectors, capture the non-linear relationship between user behavior and the semantics of new media content, and perform feature cross-fertilization and fusion. S4. Based on the ensemble learning framework, the recommendation results of the collaborative filtering model, the content recommendation model and the deep learning model are combined through a multi-level weighted fusion strategy to generate a candidate set of new media content recommendations. S5. Based on the improved LSTM network, perform time-series prediction on the recommendation candidate set, analyze users' long-term interests and short-term preferences, and adjust the recommendation ranking of new media content in combination with contextual information. S6 uses a Hadoop database for large-scale data storage and querying, leverages a cloud-network collaborative architecture to update recommendation rankings in real time, and provides recommendation services to users, pushing new media content to them.
2. The new media content recommendation method based on ensemble learning according to claim 1, characterized in that, S1 specifically includes: S11. Collect user behavior data on new media platforms and extract user behavior characteristics, including: Click history: the time, page, and frequency of each click; Browsing history: Pages viewed each time, duration of stay, and browsing time period; Commenting behavior: comment content, comment duration, and comment sentiment; Sharing behavior: the content shared, the time of sharing, and the platform on which it is shared; Likes data: Liked content, liked time period, and liked frequency; S12. The above behavioral data is cleaned to remove invalid data and outliers, and missing data is filled by nearest neighbor interpolation. It is then standardized by timestamp, frequency and category label to convert it into a value between 0 and 1, removing the deviation caused by the difference in data volume, and forming a behavioral feature vector in a unified format.
3. The new media content recommendation method based on ensemble learning according to claim 1, characterized in that, S2 specifically includes: S21. Collect the text of user interactions with new media content, including user comments, replies, shared content, dynamic posts, and message interactions; S22. Use the pre-trained language model BERT based on the Transformer architecture to perform word segmentation and syntactic analysis on the text to ensure that the semantic information of each word is fully captured. S23. Model the grammatical structure of the text using a dependency parsing algorithm to clarify the relationships between each word, including subject-verb-object structures, modification relationships, and clause relationships; S24. Based on this, the Skip-gram model is used to vectorize all words and map them to a high-dimensional vector space. The contextual semantic information of the words is obtained through training. S25. Use the K-means algorithm to cluster words, identify hidden keywords in the text, and optimize the weight of keywords through the TF-IDF algorithm to emphasize high-frequency and representative keywords. S26. Use the latent Dirichlet assignment algorithm to model topics for all texts, extract the topics that each user is interested in, and construct a user interest topic model based on the degree of correlation between the user and each topic. S27. Use an MLP network to fuse all extracted keywords, interest topic models, and user behavior data to generate user interest semantic vectors.
4. The new media content recommendation method based on ensemble learning according to claim 3, characterized in that, The keyword identification process specifically includes: After segmenting all collected text and removing stop words, the Skip-gram model is used to vectorize each word, including: The Skip-gram model is adopted. During training, context word pairs are extracted from the text through a sliding window mechanism. Each time, the context words in the window are predicted based on the current word, and the vector representation of the words is adjusted by maximizing the accuracy of the prediction. For each text, a large amount of corpus data will be traversed during the training process, so that each word is mapped to a vector in a high-dimensional space, and the vector can reflect the semantic relationship between words; The K-means algorithm is used to cluster all word vectors. The value of K is selected according to the characteristics of the dataset. The Euclidean distance between each word and the cluster center is calculated. All words are assigned to the cluster closest to the cluster center, forming multiple word clusters with similar semantics. The TF-IDF algorithm is used to weight the words in each cluster, including: Calculate the inverse document frequency (IDF) value of each word in the entire text corpus. The higher the IDF value, the stronger the word's discriminative power in the text. By combining the frequency (TF) value of words in the current document, the TF-IDF value of each word is calculated to obtain keywords that are highly representative in each text.
5. The new media content recommendation method based on ensemble learning according to claim 3, characterized in that, The topic modeling process of the potential Dirichlet assignment algorithm specifically includes: LDA assumes that each text consists of multiple topics, and each topic is represented by a probability distribution of words; Set the number of topics K in the LDA model, and use the variational inference algorithm to model the documents. Initially, the topic distribution of each document is randomly assigned, and the topic assignment of each word is randomly initialized. During the inference process, LDA adjusts the topic assignment probability of each word based on the probability of each word appearing and the overall content of the document; As the number of iterations increases, the LDA model will gradually converge, eventually outputting the topic distribution of each text and the vocabulary distribution of each topic; Based on the LDA model, user behavior features are mapped to the weights of each topic, and different priorities are assigned to each topic according to the frequency of user behavior and interaction. Based on the distribution of user interests, a correlation score between the user and each topic is generated to form a user interest topic model. The correlation score reflects the intensity of the user's interest in a specific topic.
6. The new media content recommendation method based on ensemble learning according to claim 1, characterized in that, S3 specifically includes: S31. Standardize the user's behavior feature vector and semantic representation vector to ensure consistent data range and avoid affecting network training due to uneven feature values. S32. When performing joint modeling, the CNN network structure is improved by assigning user behavior features and new media content semantic features to different convolutional channels for processing, capturing the non-linear relationships within them. Specifically, this includes: Small-scale convolutional kernels process low-frequency user behavior data, including occasional clicks and short-term dwell times; Mesoscale convolutional kernels handle mid-frequency behaviors, including prolonged browsing and sustained liking. Large-scale convolutional kernels handle high-frequency behaviors, including frequent commenting and sharing. S33. By designing specific convolution kernels, the relationship between behavioral feature vectors and semantic representation vectors can be directly cross-captured during the convolution process; S34. When processing semantic representation vectors, three different scales of convolution kernels are used to extract local features of short text, long text and deep semantics respectively. S35. Each convolutional kernel extracts features at the corresponding scale through convolution operations, and then uses max pooling layers to compress the features, reducing the amount of computation. S36. After the convolutional layer, a multi-head self-attention mechanism is introduced. By calculating the similarity matrix between different convolutional channels, the key features of each channel are automatically selected for weighted fusion. S37. After feature fusion, the low-level convolutional features and the high-level fused features are directly passed to the next layer through cross-layer connections to avoid information loss.
7. The new media content recommendation method based on ensemble learning according to claim 1, characterized in that, S4 specifically includes: S41. For the recommendation results of each model, an ensemble learning framework is used for deep fusion. The models include a collaborative filtering model, a content-based recommendation model, and a deep learning model, specifically including: The collaborative filtering model uses user-item similarity calculation and extracts potential user behavior patterns and item similarities from the user-item interaction matrix through matrix singular value decomposition, thereby recommending items with similar behaviors to each user; Content recommendation models rely on the characteristics of new media content, including text, images, and tags. By calculating the similarity between a user's historical behavior and the content of an item, the model recommends the most similar content to the user. The deep learning model uses an improved CNN network to take behavioral feature vectors and semantic representation vectors as input, extracts deep features through multi-layer convolution and pooling operations, and finally outputs personalized recommendation results through fully connected layers. S42. During the deep weighted fusion process, calculate the performance metrics of each model on the training set, including precision, recall, and F1 score. Assign different weight coefficients to each model based on its performance in different scenarios, including: When user activity is high, collaborative filtering models will receive higher weights, while when user behavior data is scarce, content-based recommendation models will receive higher weights. S43. Use a weighted voting mechanism to perform preliminary weighted fusion of the recommendation results of each model to obtain a preliminary recommendation candidate set; S44. Introduce the gradient boosting tree algorithm for secondary weighted fusion, specifically including: The initial recommendation candidate set of each model is transformed into a feature vector, representing the score of each content under different models; These features are learned using the gradient boosting tree algorithm to train a fusion model. The fusion model can automatically adjust the weighting coefficients of each model and output the final weighted recommendation score. S45. Optimize and combine the initial recommendation candidate set based on the weighted recommendation score to form a recommendation candidate set for new media content, and verify it through online A / B testing.
8. The new media content recommendation method based on ensemble learning according to claim 1, characterized in that, S5 specifically includes: S51. Perform time-series processing on user behavior data, label each user behavior with its corresponding timestamp, and form a user behavior time series; S52. To explore users’ long-term interests and short-term preferences, the LSTM network is structurally improved. The forward LSTM captures the historical dependence of user behavior, while the backward LSTM can better understand the potential impact of the user’s current behavior on the future. S53. By inputting user behavior data into a bidirectional LSTM network, the input at each time step not only considers the user's past behavior, but also the future behavior trend. S54. To solve the gradient vanishing problem, an attention-based weight adjustment layer is introduced into the bidirectional LSTM network. The weight adjustment layer adjusts the weights of the time steps according to the user's behavior frequency, time interval and interaction intensity, so that short-term behavior occupies a larger proportion in the prediction, while long-term behavior is stably maintained through enhanced weights. S55. To address the changes in user preferences at different time periods, a context-aware layer is designed. The context information of time period, user device type, and social activities is used as additional input. The features are extracted by an improved CNN network, and these features are concatenated and fused with the output of a bidirectional LSTM network. S56. Based on the time series embedding mechanism and user behavior time series, the user behavior characteristics of each hour within a 24-hour period are labeled, and the weight of new media content is dynamically adjusted according to the time of the behavior. S57. In the prediction phase, the bidirectional LSTM network combines contextual information to output the user's long-term interests and short-term preferences, and recommends and ranks new media content.
9. A new media content recommendation system based on ensemble learning, comprising executing the new media content recommendation method based on ensemble learning as described in any one of claims 1 to 8, characterized in that, include: The data processing module is used to extract multi-dimensional features from user behavior data, including users' click history, browsing history, comment behavior, sharing behavior and like data. Through data cleaning and standardization, it generates behavioral feature vectors that include time, frequency and behavior type. The semantic analysis module is used to perform deep semantic analysis on the text in user interactions with new media content using NLP technology. It identifies keywords and topics through syntax parsing algorithms and constructs semantic vectors of user interests. The feature modeling module is used to jointly model behavioral feature vectors and semantic representation vectors using an improved CNN network, capture the non-linear relationship between user behavior and the semantics of new media content, and perform feature cross-fertilization and fusion. The content recommendation module is used to combine the recommendation results of collaborative filtering model, content recommendation model and deep learning model based on ensemble learning framework and through multi-level weighted fusion strategy to generate a candidate set of new media content recommendations. The temporal prediction module is used to perform temporal prediction on the recommendation candidate set based on the improved LSTM network, analyze users' long-term interests and short-term preferences, and adjust the recommendation ranking of new media content in combination with contextual information. The storage update module is used for large-scale data storage and querying based on Hadoop databases, uses a cloud-network collaborative architecture to update the recommendation ranking in real time, and provides the recommendation service to the user end to push new media content to the user.