An intention-guided push method and system based on geographic location information and multimedia interaction

Abstract

The present invention provides an intention-guided push method and system based on geographic location information and multimedia interaction, by constructing an interest model based on geographic fence technology and deep neural network; generating user activity intention prediction results based on the interest model, using Bayesian network and time series prediction algorithm based on the user activity intention prediction results to obtain behavior rhythm analysis results, based on determining the best push timing of the behavior rhythm analysis results, using graph theory algorithm to construct a user and location relationship map to generate a time sensitivity optimization plan; based on the time sensitivity optimization plan, generating customized content, using sentiment analysis method, collaborative filtering algorithm, autoencoder and multi-head attention mechanism to process the customized content to obtain final content, adjusting the final content through an instant feedback mechanism, and generating a multimedia push package; the present invention improves the performance and service quality of the intelligent push system, and provides users with a more accurate and personalized recommendation experience.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of multimedia technology, and in particular to an intention-guided push method and system based on geographic location information and multimedia interaction. Background Art

[0002] In modern intelligent recommendation systems, the intent-guided push method based on geographic location information and multimedia interaction has a wide range of application scenarios. For example, in the fields of e-commerce platforms, social media applications, and travel navigation services, personalized recommendations for different users can significantly improve user experience and service quality.

[0003] At present, there are many recommendation systems based on geographic location and user behavior on the market. The recommendation algorithms in existing solutions perform simple content matching and push according to preset rules and user tags. The existing solution technology focuses on simple identification of geographic location and lacks deep environmental perception, resulting in inaccurate recommendations. Most systems use fixed recommendation strategies and cannot be dynamically adjusted to adapt to users' immediate needs and changes in external factors. Recommended content often lacks personalized customization and is difficult to meet the diverse needs and emotional preferences of different users. Summary of the invention

[0004] The embodiments of the present invention provide an intention-guided push method and system based on geographic location information and multimedia interaction, so as to solve the problems of inaccurate recommendations, poor flexibility and insufficient customization in the prior art.

[0005] In a first aspect, an embodiment of the present invention provides an intention-guided push method based on geographic location information and multimedia interaction, comprising:

[0006] Based on geo-fencing technology and deep neural networks, the specific geographical location of the user and the surrounding environment characteristics of the specific geographical location are identified, and based on the specific geographical location, the surrounding environment characteristics and user behavior data, an interest model is constructed, and the user activity intention is predicted using the interest model, and the influence of external factors is calculated using context awareness to obtain the external factor influence result, and the user activity intention is optimized based on the external factor influence result to generate a user activity intention prediction result;

[0007] Based on the prediction result of the user activity intention, multiple target information is selected using a Bayesian network, and the multiple target information is fused to obtain fused information. Based on the fused information, a time series prediction algorithm is used to analyze the user's behavior rhythm to obtain a behavior rhythm analysis result. Based on the behavior rhythm analysis result, an optimal push timing is determined. Based on the optimal push timing, a graph theory algorithm is used to construct a user and location relationship map, and a time sensitivity optimization plan is generated based on the user and location relationship map.

[0008] Based on the time-sensitivity optimization scheme, natural language processing technology and generative adversarial networks are used to generate customized content, sentiment analysis is used to adjust sentiment adaptability of the customized content to obtain content to be optimized, collaborative filtering algorithm is used to analyze feedback information of similar users to obtain feedback information analysis results, and the content to be optimized is adjusted based on the feedback information analysis results to obtain optimized content;

[0009] The optimized content is processed by using an autoencoder and a multi-head attention mechanism to obtain target content, and the target multimedia element and the target content are combined to obtain final content. Feedback suggestions from users on the final content are collected through an instant feedback mechanism, and a multimedia push package is generated based on the feedback suggestions and the final content.

[0010] Optionally, based on geo-fencing technology and deep neural network, a specific geographical location of the user and surrounding environment characteristics of the specific geographical location are identified; based on the specific geographical location, the surrounding environment characteristics and user behavior data, an interest model is constructed; the user activity intention is predicted using the interest model; the influence of external factors is calculated using context awareness to obtain an external factor influence result; the user activity intention is optimized based on the external factor influence result to generate a user activity intention prediction result, including:

[0011] Using geo-fencing technology combined with a spatiotemporal clustering algorithm to detect a user's behavior of entering or leaving a preset geographic area, and recording the user's geographic location information and a timestamp corresponding to the geographic location information, the preset geographic area includes home, workplace, concert venue, and sports field;

[0012] Based on the geographic location information and the timestamp, using a deep convolutional neural network to identify the specific geographic location of the user and the surrounding environment characteristics of the specific geographic location, the surrounding environment characteristics including facility type characteristics, crowd density characteristics and traffic flow characteristics, the specific geographic location is one or more locations within a preset geographic area;

[0013] Combining the user behavior data and the user multimedia interaction data, using a graph neural network to construct a user's interest model to be optimized, optimizing the interest model to be optimized based on the specific geographic location and the surrounding environment characteristics to obtain an interest model, and predicting the user's activity intention based on the interest model;

[0014] Acquire external factors, apply situational awareness to calculate the impact of the external factors on the user's activity intention, and obtain external factor impact results, wherein the external factors include weather, traffic conditions, and time;

[0015] Based on the external factors, the user activity intention is optimized using a multi-objective optimization algorithm to generate a user activity intention prediction result. Multiple optimization objectives are considered during the optimization process, and the optimization objectives include time sensitivity, location relevance, and user satisfaction.

[0016] Optionally, based on the geographic location information and the timestamp, a deep convolutional neural network is used to identify a specific geographic location where the user is located and surrounding environment characteristics of the specific geographic location, wherein the surrounding environment characteristics include facility type characteristics, crowd density characteristics, and traffic flow characteristics, and the specific geographic location is one or more locations within a preset geographic area, including:

[0017] Based on the geographic location information and the timestamp, a spatiotemporal attention mechanism and a deep convolutional neural network are used to perform multi-scale feature extraction on the user's location and dynamic features of the user's location that change over time, so as to obtain a user location feature map; based on the user location feature map, a semantic segmentation algorithm is applied to analyze the map data, extract environmental information, and generate a comprehensive environmental feature description;

[0018] Based on the comprehensive environmental feature description and the timestamp, a time series prediction model is applied to analyze the difference in pedestrian density or traffic flow changes in different time periods to obtain environmental feature prediction results.

[0019] Optionally, based on the geographic location information and the timestamp, a spatiotemporal attention mechanism and a deep convolutional neural network are used to perform multi-scale feature extraction on the user's location and the dynamic features of the user's location that change over time, so as to obtain a user location feature map. According to the user location feature map, a semantic segmentation algorithm is applied to analyze the map data, extract environmental information, and generate a comprehensive environmental feature description, including:

[0020] Based on the geographic location information and the timestamp, an input tensor is generated, based on the input tensor, an importance weight of each spatiotemporal point is calculated using a spatiotemporal attention mechanism, and a key area that has an impact on user behavior is determined based on the importance weight, and a spatiotemporal attention weight map is constructed based on the key area and the importance weight;

[0021] Based on the spatiotemporal attention weight map, a convolution operation is performed on the input tensor through a plurality of convolution kernels of different scales in a deep convolutional neural network to obtain different levels of features from local to global, and a multi-scale feature map is generated based on the different levels of features;

[0022] The multi-scale feature map is subjected to feature fusion through residual connection or skip connection to generate a fused feature map, a user position feature map is generated based on the fused feature map, semantic information in the user position feature map is extracted using a semantic segmentation network with an encoder-decoder structure in a semantic segmentation algorithm, and a semantic segmentation map is generated based on the semantic information, wherein the semantic information includes map data and environmental information;

[0023] The softmax layer in the semantic segmentation algorithm is used to classify each pixel in the semantic segmentation map to identify and classify different elements in the environment, and a comprehensive environmental feature description is generated based on the different elements. The different elements include buildings, roads, green spaces and waters. The comprehensive environmental feature description contains static geographic information and dynamic pedestrian and vehicle flow conditions.

[0024] Optionally, the optimized content is processed by an autoencoder and a multi-head attention mechanism to obtain target content, a target multimedia element and the target content are combined to obtain final content, feedback suggestions of users on the final content are collected through an instant feedback mechanism, and a multimedia push package is generated based on the feedback suggestions and the final content, including:

[0025] Using a variational autoencoder to compress and decode the optimized content to obtain a core content representation, and combining a multi-head attention mechanism to evaluate the importance of different parts of the core content representation to strengthen the key content parts and generate an enhanced core content representation;

[0026] Based on the enhanced core content representation, user interest points and the user behavior data, applying a conditional generative adversarial network to generate target content;

[0027] Based on the target content, using deep reinforcement learning to select the best multimedia form, based on the best multimedia form, generating a multimedia form selection strategy, based on the multimedia form selection strategy, generating preliminary multimedia elements, introducing style transfer technology to adjust the preliminary multimedia elements, and obtaining target multimedia elements;

[0028] Based on the target content and the target multimedia elements, a graph neural network is applied to construct a content graph to optimize the connection mode of the target multimedia elements to obtain optimized target multimedia elements, and the optimized target multimedia elements and mixed reality elements are combined to generate final content, wherein the mixed reality elements include AR filters;

[0029] Based on the final content, the user's emotional response is collected through an instant feedback mechanism, and the user behavior data is analyzed in real time to obtain a user behavior analysis result; based on the emotional response and the user behavior analysis result, feedback suggestions are generated; based on the feedback suggestions and the final content, a federated learning framework and causal inference method are used to generate a multimedia push package.

[0030] Optionally, based on the target content, using deep reinforcement learning to select an optimal multimedia form, based on the optimal multimedia form, generating a multimedia form selection strategy, generating preliminary multimedia elements based on the multimedia form selection strategy, introducing style transfer technology to adjust the preliminary multimedia elements, and obtaining target multimedia elements, including:

[0031] Using a deep reinforcement learning algorithm to state-encode the user's situation, the target content, and the user behavior data to construct a target vector, construct a dynamic reward function based on the target vector, user historical feedback, and user multimedia interaction data, and determine the optimal multimedia format using a policy gradient method based on the dynamic reward function, wherein the user's situation includes time, location, and activity;

[0032] Based on the optimal multimedia form, construct a context-aware decision tree, and dynamically update the context-aware decision tree using an online learning algorithm to obtain an updated context-aware decision tree, and generate a multimedia form selection strategy based on the updated context-aware decision tree and multiple influencing factors, wherein the influencing factors include user preferences, historical behavior patterns, and environmental conditions of the user;

[0033] Selecting a target template corresponding to the multimedia form selection strategy by using template matching and generation technology, and adding key information of the target content to the target template to generate preliminary multimedia elements, wherein the key information includes text summary, keywords and tags;

[0034] Based on the preliminary multimedia elements and user preferences, a target aesthetic preference model is constructed, and based on the target aesthetic preference model, personalized style transfer processing is performed on the preliminary multimedia elements to obtain target multimedia elements.

[0035] Optionally, based on the preliminary multimedia elements and user preferences, a target aesthetic preference model is constructed, and based on the target aesthetic preference model, personalized style transfer processing is performed on the preliminary multimedia elements to obtain target multimedia elements, including:

[0036] Processing the user multimedia interaction data using deep learning technology to obtain user feature representation, analyzing existing user comments using natural language processing technology to extract text features, optimizing the user feature representation based on the text features and user preferences to obtain aesthetic preference features, wherein the user preferences include the user's preference for color, shape, and texture;

[0037] Based on the preliminary multimedia elements and the aesthetic preference features, construct a preliminary aesthetic preference model of the user, based on the preliminary aesthetic preference model, introduce an online learning algorithm, update the preliminary aesthetic preference model, and obtain a target aesthetic preference model;

[0038] Based on the target aesthetic preference model, a user style model with a degree of adaptation higher than a preset threshold is selected, and based on the user style model, a style transfer algorithm is applied to perform personalized adjustment on the preliminary multimedia elements to obtain preliminary adjusted multimedia elements, wherein the personalized adjustment includes changing the color palette of the image, adjusting the visual effects of the video, and modifying the sound effects of the audio;

[0039] The preliminarily adjusted multimedia elements are evaluated using existing image or audio quality evaluation indicators to obtain quality evaluation results. Based on the quality evaluation results, the key parameters of the target aesthetic preference model and the user style model are optimized and adjusted using a Bayesian optimization method to optimize the preliminarily adjusted multimedia elements and generate target multimedia elements.

[0040] In a second aspect, an embodiment of the present invention provides an intention-guided push system based on geographic location information and multimedia interaction, including:

[0041] an identification module, for identifying the specific geographic location of the user and the surrounding environment characteristics of the specific geographic location based on geo-fencing technology and deep neural networks, constructing an interest model based on the specific geographic location, the surrounding environment characteristics and user behavior data, predicting the user's activity intention using the interest model, calculating the influence of external factors using context awareness, obtaining the external factor influence result, optimizing the user's activity intention based on the external factor influence result, and generating a user activity intention prediction result;

[0042] An analysis module, for selecting multiple target information using a Bayesian network based on the prediction result of the user's activity intention, fusing the multiple target information to obtain fused information, analyzing the user's behavior rhythm based on the fused information using a time series prediction algorithm to obtain a behavior rhythm analysis result, determining an optimal push timing based on the behavior rhythm analysis result, constructing a user and location relationship map based on the optimal push timing using a graph theory algorithm, and generating a time sensitivity optimization plan based on the user and location relationship map;

[0043] An optimization module is used to generate customized content based on the time-sensitivity optimization scheme by using natural language processing technology and generative adversarial networks, perform sentiment adaptive adjustment on the customized content by using sentiment analysis to obtain content to be optimized, analyze feedback information of similar users by using collaborative filtering algorithm to obtain feedback information analysis results, and adjust the content to be optimized based on the feedback information analysis results to obtain optimized content;

[0044] A generation module is used to process the optimized content using an autoencoder and a multi-head attention mechanism to obtain target content, combine target multimedia elements with the target content to obtain final content, collect user feedback suggestions on the final content through an instant feedback mechanism, and generate a multimedia push package based on the feedback suggestions and the final content.

[0045] In a third aspect, an embodiment of the present invention provides a computing device, comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute an intention-guided push method based on geographic location information and multimedia interaction as described in any one of the first aspects.

[0046] In a fourth aspect, an embodiment of the present invention provides a computer storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement an intention-guided push method based on geographic location information and multimedia interaction as described in any one of the first aspects.

[0047] In an embodiment of the present invention, a specific geographical location of a user and surrounding environment characteristics of the specific geographical location are identified based on geo-fencing technology and deep neural networks; an interest model is constructed based on the specific geographical location, the surrounding environment characteristics and user behavior data; the interest model is used to predict the user's activity intention; context awareness is used to calculate the influence of external factors to obtain external influence results; the user's activity intention is optimized based on the external influence results to generate a user activity intention prediction result; based on the user activity intention prediction result, a Bayesian network is used to select multiple target information, the multiple target information is fused to obtain fused information; based on the fused information, a time series prediction algorithm is used to analyze the user's behavior rhythm to obtain a behavior rhythm analysis result; based on the behavior rhythm analysis result, the best push timing is determined; based on the To determine the best time to push content, use graph theory algorithms to build a user and location relationship graph, and generate a time-sensitivity optimization plan based on the user and location relationship graph; based on the time-sensitivity optimization plan, use natural language processing technology and generative adversarial networks to generate customized content, use sentiment analysis to adjust the customized content for sentiment adaptability, and obtain content to be optimized; use collaborative filtering algorithms to analyze feedback information from similar users to obtain feedback information analysis results, and adjust the content to be optimized based on the feedback information analysis results to obtain optimized content; use autoencoders and multi-head attention mechanisms to process the optimized content to obtain target content, combine target multimedia elements with the target content to obtain final content, collect user feedback and suggestions on the final content through an instant feedback mechanism, and generate a multimedia push package based on the feedback suggestions and the final content. The technical solution provided by the present invention improves the accuracy of geographic location and environmental feature recognition through deep neural networks and context-aware computing, making the recommended content more in line with the user's immediate needs; improves the user's acceptance and satisfaction with the recommended content through time sensitivity optimization and emotional adaptability adjustment; utilizes Bayesian networks, graph theory algorithms, collaborative filtering and other technologies to achieve multi-dimensional information fusion and personalized recommendations, ensuring that the recommended content is both diverse and targeted; and forms a closed-loop optimization mechanism through instant feedback mechanisms and collaborative filtering algorithms, ensuring that the system can continuously learn and adapt to user changes, and provide long-term stable and high-quality services.Among them, by combining the spatiotemporal attention mechanism and deep convolutional neural network, it is possible to capture the dynamic characteristics of the user's location and its changes over time, thereby improving the accuracy of location recognition and environmental analysis; applying the semantic segmentation algorithm to extract detailed environmental information enables the system to more comprehensively understand the specific situation of the user and provide services that are more in line with the needs; through multi-scale feature extraction and feature fusion, it is possible to capture different levels of features from local to global, enhancing the understanding and prediction capabilities of complex environments; by generating a comprehensive description of environmental features, the recommendations and services provided are more intelligent and personalized, significantly improving user satisfaction and engagement; by combining user preferences and text features to optimize user feature representation, the generated aesthetic preference features are more in line with the user's unique needs, significantly improving the effect of personalized recommendations; by dynamically updating the preliminary aesthetic preference model through the online learning algorithm, it is possible to continuously optimize the recommendation strategy to adapt to the long-term changes and immediate needs of users; using the style transfer algorithm and image or audio quality evaluation indicators, it is ensured that the adjusted multimedia content is not only diverse in form, but also of high quality and in line with the user's aesthetic standards;.

[0048] These and other aspects of the present invention will become more apparent from the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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 paying creative work.

[0050] Figure 1 A flowchart of an intention-guided push method based on geographic location information and multimedia interaction provided by an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of the structure of an intention-guided push system based on geographic location information and multimedia interaction provided by an embodiment of the present invention;

[0052] Figure 3 A schematic diagram of the structure of a computing device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0054] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0056] Figure 1 A flowchart of an intention-guided push method based on geographic location information and multimedia interaction is provided for an embodiment of the present invention, such as Figure 1 As shown, the method includes:

[0057] In today's digital and mobile era, users' demand for personalized and intelligent services is growing. Although traditional recommendation systems can meet users' static needs to a certain extent, they still have many shortcomings in terms of accuracy, real-time and personalization. In order to overcome these limitations and provide services that better meet users' immediate needs, the present invention provides an intention-guided push method based on geographic location information and multimedia interaction, such as Figure 1 ,include:

[0058] Step 101: Based on geo-fencing technology and deep neural networks, identify the specific geographical location of the user and the surrounding environment characteristics of the specific geographical location, build an interest model based on the specific geographical location, the surrounding environment characteristics and user behavior data, use the interest model to predict the user's activity intention, apply context awareness to calculate the influence of external factors, obtain the external factor influence result, optimize the user's activity intention based on the external factor influence result, and generate the user's activity intention prediction result;

[0059] In this step, the specific geographic location refers to the location of the user in a specific preset geographic area, such as the specific address or room of the home. The surrounding environment characteristics refer to information describing the user's location and its surrounding environment, including facility type, crowd density, traffic flow, etc. The interest model refers to a mathematical model built based on user behavior data and environmental characteristics, which is used to predict the user's activity intention.

[0060] For example, user Xiao Li is attending a concert. Geofencing technology detects that Xiao Li has entered the preset geographic area of ​​the concert venue, and records his geographic location information and timestamp. At the same time, a deep convolutional neural network (CNN) is used to analyze the environmental characteristics around the concert venue, such as the facility type is a concert hall, the crowd density is high, and the traffic flow is medium, to obtain Xiao Li's current specific geographic location and surrounding environmental characteristics. Combining Xiao Li's historical behavior data, current geographic location, and environmental characteristics, Xiao Li's historical behavior data such as past concert preferences and browsing history, a graph neural network (GNN) is used to build an interest model, which predicts that Xiao Li may have activity intentions such as taking photos, sharing on social platforms, and buying souvenirs during the concert; then, context awareness is applied to calculate the impact of external factors, such as clear weather and smooth traffic, to optimize these activity intentions and generate the final user activity intention prediction results.

[0061] Step 102: Based on the prediction result of the user activity intention, multiple target information is selected using a Bayesian network, and the multiple target information is fused to obtain fused information. Based on the fused information, a time series prediction algorithm is used to analyze the user's behavior rhythm to obtain a behavior rhythm analysis result. Based on the behavior rhythm analysis result, an optimal push timing is determined. Based on the optimal push timing, a user and location relationship map is constructed using a graph theory algorithm, and a time sensitivity optimization plan is generated based on the user and location relationship map.

[0062] For example, based on the prediction results of user activity intention, a Bayesian network is used to select multiple target information, such as concert photos, peripheral product recommendations, nearby restaurants, etc. Then, these target information are fused to form comprehensive fusion information to ensure the diversity and comprehensiveness of the recommended content; the time series prediction algorithm is used to analyze Xiao Li's past behavior rhythm and find that he usually checks his phone during the intermission of the performance. Therefore, the best time to push is determined to be during the intermission to avoid disturbing his viewing experience; based on the best push time, a graph theory algorithm is used to construct a relationship map between Xiao Li and his location, and a time-sensitive optimization plan is generated. For example, nearby restaurant information is pushed during the intermission, and the location of the souvenir shop is pushed after the performance.

[0063] Step 103: Based on the time-sensitivity optimization scheme, natural language processing technology and generative adversarial networks are used to generate customized content, sentiment analysis is used to adjust sentiment adaptability of the customized content to obtain content to be optimized, collaborative filtering algorithm is used to analyze feedback information of similar users to obtain feedback information analysis results, and the content to be optimized is adjusted based on the feedback information analysis results to obtain optimized content;

[0064] For example, based on the time-sensitivity optimization solution, we use natural language processing technology and generative adversarial networks to generate customized content, such as Dear audience, you may want to visit our souvenir shop during the intermission! We use sentiment analysis to adjust the emotional adaptability of the content to ensure that the tone is friendly and in line with the atmosphere of the occasion; we use collaborative filtering algorithms to analyze the feedback information of users similar to Xiao Li and find that they generally like push notifications with discounts; therefore, we adjust the customized content and add limited-time discount information, and the optimized content is Dear audience, you may want to visit our souvenir shop during the intermission! There is a 20% discount now!

[0065] Step 104: using an autoencoder and a multi-head attention mechanism to process the optimized content to obtain target content, combining the target multimedia element and the target content to obtain final content, collecting user feedback suggestions on the final content through an instant feedback mechanism, and generating a multimedia push package based on the feedback suggestions and the final content;

[0066] For example, the optimized content is processed using autoencoders and multi-head attention mechanisms to ensure that the content is concise and highlights the key points. Combined with the target multimedia elements, such as concert photos and video clips, the final content is [Limited Time Offer] Dear audience, you may want to visit our souvenir shop during the intermission! There is a 20% discount now! [Attached with beautiful concert photos]; collect Xiao Li's feedback and suggestions on the final content through the instant feedback mechanism, such as likes and comments; based on these feedbacks, continuously optimize the push strategy to ensure that the next push is more accurate. Finally, a multimedia push package is generated, including customized content, discount information, and related multimedia elements.

[0067] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0068] By deeply integrating geo-fencing technology, deep neural networks, and context-aware computing, the accuracy of identifying geographic locations and environmental features is improved, making the recommended content more in line with the user's immediate needs;

[0069] By optimizing time sensitivity and adjusting emotional adaptability, users’ acceptance and satisfaction with recommended content are improved;

[0070] Using advanced technologies such as Bayesian networks, graph theory algorithms, and collaborative filtering, we have achieved multi-dimensional information fusion and personalized recommendations, ensuring that the recommended content is both diverse and targeted.

[0071] Through instant feedback mechanism and closed-loop optimization mechanism, the system can continuously learn and adapt to user changes, providing long-term stable and high-quality services.

[0072] In an intelligent recommendation system, accurately capturing the user's immediate needs and environmental changes is the key to improving the user experience. Traditional recommendation methods are often limited to static data and simple geographic location identification, which makes it difficult to meet the diverse and dynamic needs of users. Based on this, the present invention provides a specific embodiment, wherein step 101 identifies the specific geographic location of the user and the surrounding environmental characteristics of the specific geographic location based on geographic fencing technology and deep neural networks, constructs an interest model based on the specific geographic location, the surrounding environmental characteristics, and user behavior data, uses the interest model to predict the user's activity intention, applies context-awareness to calculate the impact of external factors, obtains the external factor impact result, optimizes the user's activity intention based on the external factor impact result, and generates a user activity intention prediction result, which specifically includes the following steps:

[0073] Step 111: using geo-fencing technology combined with a spatiotemporal clustering algorithm to detect the user's behavior of entering or leaving a preset geographic area, and recording the user's geographic location information and a timestamp corresponding to the geographic location information, wherein the preset geographic area includes a home, a workplace, a concert venue, and a sports field;

[0074] In this step, the preset geographic area refers to a specific area defined in advance, such as a home, a workplace, a concert venue, and a sports field.

[0075] For example, user Xiao Zhang passes through two preset geographic areas, home and workplace, on his way to and from work every day. Using geo-fencing technology and spatiotemporal clustering algorithms, it is detected that Xiao Zhang entered the geo-fence near his home at 7:30 in the morning, and the geographic location information and timestamp of this moment are recorded. Similarly, it is detected that Xiao Zhang left the geo-fence of his workplace at 6:00 in the evening, and the relevant information is recorded. The preset geographic areas also include concert venues and sports fields that he may go to on weekends.

[0076] Step 112: Based on the geographic location information and the timestamp, using a deep convolutional neural network to identify the specific geographic location of the user and the surrounding environment characteristics of the specific geographic location, the surrounding environment characteristics including facility type characteristics, crowd density characteristics and traffic flow characteristics, the specific geographic location is one or more locations within a preset geographic area;

[0077] In this step, the specific geographic location refers to the location of the user in a specific preset geographic area, such as a specific address or room of a home. The surrounding environment characteristics refer to information describing the user's location and its surrounding environment, including facility types, crowd density, traffic flow, etc.

[0078] For example, based on the location information and timestamp near Xiao Zhang's home, a deep convolutional neural network (CNN) is used to analyze the surrounding environmental features; it is identified that the facility type in the environment near Xiao Zhang's home is a residential area, with low density of people and moderate traffic flow. Similarly, for the workplace, the facility type is identified as an office building, with high density of people and large traffic flow. These features help the system understand Xiao Zhang's current specific location and its surrounding environment more accurately.

[0079] Step 113: combining the user behavior data and the user multimedia interaction data, using a graph neural network to construct a user's interest model to be optimized, optimizing the interest model to be optimized based on the specific geographic location and the surrounding environment characteristics to obtain an interest model, and predicting the user's activity intention based on the interest model;

[0080] For example, combining Xiao Zhang’s historical behavior data and multimedia interaction data, Xiao Zhang’s historical behavior data such as past travel habits and shopping preferences, Xiao Zhang’s multimedia interaction data such as browsed news and watched videos, use a graph neural network (GNN) to build an interest model to be optimized, which not only takes into account Xiao Zhang’s personal preferences, but also combines his location and the characteristics of his surrounding environment. For example, it is detected that Xiao Zhang pays more attention to family entertainment content when at home, and pays more attention to career development information at work. Based on this information, optimize the interest model to be optimized to ensure that the recommended content is more in line with Xiao Zhang’s needs.

[0081] Step 114: Acquire external factors, apply context awareness to calculate the impact of the external factors on the user's activity intention, and obtain external factor impact results, where the external factors include weather, traffic conditions, and time;

[0082] For example, external factors are obtained, such as whether the weather on that day is sunny, the traffic conditions are normal, and the specific time is morning; context-aware computing is applied to calculate the impact of these external factors on Xiao Zhang's activity intention. For example, a sunny day may increase the possibility of Xiao Zhang going out for a walk, while normal traffic conditions mean that he is more likely to arrive at work on time; based on the results of these external factors, the prediction of Xiao Zhang's activity intention is further optimized.

[0083] Step 115: Based on the external factor influence result, the user activity intention is optimized using a multi-objective optimization algorithm to generate a user activity intention prediction result, and multiple optimization objectives are considered in the optimization process, and the optimization objectives include time sensitivity, location relevance, and user satisfaction;

[0084] For example, based on the results of external factors, a multi-objective optimization algorithm is used to optimize Xiao Zhang's activity intention. Multiple objectives are considered during the optimization process, such as time sensitivity, location relevance, and user satisfaction. Time sensitivity, such as morning is suitable for recommending fast breakfast places, location relevance, such as supermarkets or gyms near home, and user satisfaction, such as the recommended content must be in line with Xiao Zhang's interests; finally, detailed user activity intention prediction results are generated, for example, at 7:30 in the morning, a nearby breakfast shop or cafe is recommended; at 6:00 in the afternoon, a relaxing activity after get off work, such as a gym or a cinema, is recommended.

[0085] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0086] By deeply integrating geo-fencing technology, spatiotemporal clustering algorithms, and deep convolutional neural networks, the accuracy of identifying geographic locations and environmental features is improved, making the recommended content more in line with the user's immediate needs;

[0087] Combining user behavior data and multimedia interaction data, using graph neural networks to build interest models, and applying context-awareness to calculate the impact of external factors, significantly improves user acceptance and satisfaction with recommended content;

[0088] A multi-objective optimization algorithm is used to comprehensively consider time sensitivity, location relevance, and user satisfaction, achieving more intelligent and personalized recommendations, ensuring that the recommended content is both diverse and targeted;

[0089] By obtaining external factors and user feedback in real time, the system can continuously optimize recommendation strategies and provide long-term stable and high-quality services.

[0090] In the intelligent recommendation system, accurately identifying the user's geographic location and the characteristics of its surrounding environment is the key to achieving personalized services. Traditional positioning technology can often only provide rough location information, and it is difficult to capture the specific situation and environmental changes of the user. Based on this, the present invention provides a specific embodiment, the step 112, based on the geographic location information and the timestamp, uses a deep convolutional neural network to identify the specific geographic location of the user and the surrounding environment characteristics of the specific geographic location, the surrounding environment characteristics include facility type characteristics, crowd density characteristics and traffic flow characteristics, the specific geographic location is one or more locations in a preset geographic area, specifically including the following steps:

[0091] Step 121: Based on the geographic location information and the timestamp, a spatiotemporal attention mechanism and a deep convolutional neural network are used to perform multi-scale feature extraction on the user's location and the dynamic features of the user's location that change over time, so as to obtain a user location feature map; based on the user location feature map, a semantic segmentation algorithm is applied to analyze the map data, extract environmental information, and generate a comprehensive environmental feature description;

[0092] In this step, geographic location information refers to the precise location coordinates of the user's device, usually including latitude and longitude information. Timestamp refers to the specific time when an event is recorded, which is used to track the user's activity time and location changes. User location feature map refers to a map generated by feature extraction, which describes the user's location and its dynamic characteristics over time.

[0093] Assume that user Xiao Li is attending a concert. Based on Xiao Li's geographic location information and the corresponding timestamp, Xiao Li's geographic location information such as longitude and latitude, Xiao Li's geographic location information and the corresponding timestamp such as 7:30 pm, the spatiotemporal attention mechanism and deep convolutional neural network are used to extract multi-scale features of Xiao Li's location. For example, the spatiotemporal attention mechanism is used to dynamically adjust the attention weight according to Xiao Li's historical behavior data and current timestamp, Xiao Li's historical behavior data such as the behavior pattern in the same time period in the past few months, and focus on the area and time period related to the concert; the deep convolutional neural network is used to extract features from different scales, such as local details and global scenes, to capture the dynamic changes inside and outside the concert venue, such as crowd gathering, vehicle flow, etc. Through these operations, a detailed user location feature map is generated, which not only describes Xiao Li's current specific location, but also reflects the dynamic features that change over time; based on the user location feature map, the system applies a semantic segmentation algorithm to analyze the map data and extract environmental information. For example, the semantic segmentation algorithm is used to classify each pixel in the feature map into different semantic categories, such as roads, buildings, pedestrians, vehicles, etc. In this way, the types of facilities, crowd density and traffic flow around the concert venue are identified. The types of facilities around the concert venue include parking lots and dining areas. The crowd density around the concert venue includes dense entrances and sparse interiors. The traffic flow around the concert venue includes congestion on surrounding roads. Finally, a comprehensive environmental feature description is generated, which comprehensively reflects the detailed information of Xiao Li's location and his surrounding environment.

[0094] Step 122: Based on the comprehensive environmental feature description and the timestamp, a time series prediction model is applied to analyze the difference in crowd density or traffic flow changes in different time periods to obtain an environmental feature prediction result;

[0095] This step is based on the comprehensive description of environmental features and timestamps, and applies time series prediction models, such as LSTM, to analyze differences in crowd density or changes in traffic flow in different time periods. For example, a time series prediction model is used to analyze data from the same time period in the past few weeks to predict environmental changes in the next few hours; for a concert scene, it is predicted that there will be a peak in crowd flow around 9:00 p.m., and the surrounding traffic flow will also increase significantly; through these prediction results, it is possible to plan recommended content and services in advance, such as suggesting users avoid peak hours or providing nearby alternative routes.

[0096] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0097] By combining the spatiotemporal attention mechanism and deep convolutional neural network, the system can capture the user's location and its dynamic characteristics over time, improving the accuracy of location recognition and environmental analysis;

[0098] Applying semantic segmentation algorithms to extract detailed environmental information enables the system to more fully understand the user's specific context and provide services that better meet their needs;

[0099] By using the time series prediction model to analyze changes in environmental characteristics, the system can predict future environmental conditions in advance and help users make better decisions;

[0100] Through multi-scale feature extraction and comprehensive environmental analysis, the recommendations and services provided by the system are more intelligent and personalized, significantly improving user satisfaction and engagement.

[0101] In an intelligent recommendation system, accurately capturing the specific location of the user and its dynamic features that change over time is the key to providing personalized services. Traditional positioning and environmental perception technologies can often only provide static information and are difficult to fully reflect the user's real-time context. Based on this, the present invention provides a specific embodiment, wherein step 121, based on the geographic location information and the timestamp, uses a spatiotemporal attention mechanism and a deep convolutional neural network to perform multi-scale feature extraction on the user's location and the dynamic features of the user's location that change over time to obtain a user location feature map, and according to the user location feature map, applies a semantic segmentation algorithm to analyze the map data, extracts environmental information, and generates a comprehensive environmental feature description, specifically comprising the following steps:

[0102] Step 131: Based on the geographic location information and the timestamp, an input tensor is generated, based on the input tensor, the importance weight of each spatiotemporal point is calculated using a spatiotemporal attention mechanism, and based on the importance weight, a key area that has an impact on user behavior is determined, and a spatiotemporal attention weight map is constructed based on the key area and the importance weight;

[0103] In this step, the input tensor refers to a multidimensional data structure consisting of geographic location information and timestamps, which is used as the input of the model. Importance weight refers to the relative importance of each spatiotemporal point in the current context, which is used to determine the key areas. The spatiotemporal attention weight map refers to a map generated based on the importance weights, which describes the importance distribution of each spatiotemporal point.

[0104] Assume that user Xiao Li is attending a concert. Based on Xiao Li’s geographic location information and the corresponding timestamp, an input tensor is generated. Then, the importance weight of each spatiotemporal point is calculated using the spatiotemporal attention mechanism. For example, Xiao Li’s location coordinates and time information are combined into a multidimensional tensor as the basis for subsequent processing. According to Xiao Li’s historical behavior data and current timestamp, the attention weight is dynamically adjusted, focusing on the areas and time periods related to the concert to generate a spatiotemporal attention weight map, which shows the importance distribution of each spatiotemporal point.

[0105] Step 132: Based on the spatiotemporal attention weight map, convolution operation is performed on the input tensor through convolution kernels of multiple different scales in a deep convolutional neural network to obtain different levels of features from local to global, and a multi-scale feature map is generated based on the different levels of features;

[0106] In this step, the multi-scale feature map refers to the feature map generated by convolving the input tensor with convolution kernels of multiple scales, which contains different levels of features from local to global.

[0107] This step is based on the spatiotemporal attention weight map, and convolution operations are performed on the input tensor through multiple convolution kernels of different scales in the deep convolutional neural network to obtain different levels of features from local to global. For example, multiple convolution kernels of different scales, such as 3x3, 5x5, and 7x7, are used to convolve the input tensor to extract local details, medium-range features, and global scenes, respectively. Local details such as a single pedestrian, medium-range features such as vehicle flow, and global scenes such as overall traffic conditions; these features are integrated into a multi-scale feature map to fully describe the dynamic changes of Xiao Li's location and his surrounding environment.

[0108] Step 133: fusing the multi-scale feature map through residual connection or skip connection to generate a fused feature map, generating a user position feature map based on the fused feature map, extracting semantic information from the user position feature map using a semantic segmentation network with an encoder-decoder structure in a semantic segmentation algorithm, and generating a semantic segmentation map based on the semantic information, wherein the semantic information includes map data and environmental information;

[0109] In this step, the multi-scale feature maps are fused through residual connections or skip connections to generate a fused feature map; based on the fused feature map, a user location feature map is generated. For example, residual connections are used to fuse feature maps of different scales to ensure that the information of each scale can be fully utilized. The user location feature map finally generated not only contains detailed local features, but also reflects the changes in the global scene, providing a solid foundation for subsequent analysis; the semantic segmentation network with an encoder-decoder structure in the semantic segmentation algorithm is used to extract semantic information from the user location feature map and generate a semantic segmentation map. For example, the encoder extracts high-level features in the feature map, and the decoder restores the spatial resolution, and finally generates a semantic segmentation map in which each pixel is classified into a specific semantic category, such as buildings, roads, green spaces, and waters.

[0110] Step 134: using the softmax layer in the semantic segmentation algorithm to classify each pixel in the semantic segmentation map to identify and classify different elements in the environment, and generating a comprehensive environmental feature description based on the different elements, wherein the different elements include buildings, roads, green spaces, and water areas, and the comprehensive environmental feature description includes static geographic information, dynamic pedestrian and vehicle flow conditions;

[0111] This step uses the softmax layer in the semantic segmentation algorithm to classify each pixel in the semantic segmentation map, identifying and classifying different elements in the environment. For example, each pixel is classified into different semantic categories, such as buildings, roads, green spaces, and waters, through the softmax layer. At the same time, dynamic elements such as pedestrian density and vehicle flow are identified and classified; the final generated comprehensive environmental feature description includes not only static geographic information, but also dynamic pedestrian and vehicle flow conditions, fully reflecting the detailed situation of Xiao Li's location and its surroundings.

[0112] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0113] By combining the spatiotemporal attention mechanism and deep convolutional neural network, it can capture the user's location and its dynamic characteristics over time, improving the accuracy of location recognition and environmental analysis;

[0114] Applying semantic segmentation algorithms to extract detailed environmental information enables the system to more fully understand the user's specific context and provide services that better meet their needs;

[0115] Through multi-scale feature extraction and feature fusion, it is possible to capture features at different levels from local to global, thus enhancing the ability to understand and predict complex environments;

[0116] By generating comprehensive environmental feature descriptions, the recommendations and services provided are more intelligent and personalized, significantly improving user satisfaction and engagement.

[0117] In order to further improve the accuracy and personalization level of the content, generate a highly customized multimedia push package, realize closed-loop optimization of content push, and continuously iterate to provide better services. Based on this, the present invention provides a specific embodiment, in which the step 104 uses an autoencoder and a multi-head attention mechanism to process the optimized content to obtain the target content, combines the target multimedia element and the target content to obtain the final content, collects user feedback suggestions on the final content through an instant feedback mechanism, and generates a multimedia push package based on the feedback suggestions and the final content, specifically including the following steps:

[0118] Step 401: compressing and decoding the optimized content using a variational autoencoder to obtain a core content representation, and evaluating the importance of different parts of the core content representation using a multi-head attention mechanism to strengthen the key content parts and generate an enhanced core content representation;

[0119] In this step, the enhanced core content representation refers to a more focused content representation obtained by strengthening the key content parts through a multi-head attention mechanism.

[0120] Assume that there is an optimized news article, and use the variational autoencoder to compress and decode the article to extract its core content representation. For example, the variational autoencoder compresses the article into a low-dimensional vector, namely the core content representation, which retains the main information of the article, such as the topic, key events, and emotional tendencies; combined with the multi-head attention mechanism, it evaluates the importance of different parts in the core content representation, identifies the key paragraphs or sentences that best reflect the main idea of ​​the article, and strengthens these parts to generate an enhanced core content representation.

[0121] Step 402: Based on the enhanced core content representation, user interest points and the user behavior data, a conditional generative adversarial network is applied to generate target content;

[0122] This step is based on the enhanced core content representation, user Xiao Li's interests, such as technology news, and his behavior data, such as the types of articles he has read in the past, and applies the conditional generative adversarial network to generate the target content. For example, the conditional generative adversarial network is applied to generate a new article that matches Xiao Li's interests and behavior data, ensuring that the content is both novel and in line with his reading habits.

[0123] Step 403: Based on the target content, select the best multimedia form using deep reinforcement learning, generate a multimedia form selection strategy based on the best multimedia form, generate preliminary multimedia elements based on the multimedia form selection strategy, introduce style transfer technology to adjust the preliminary multimedia elements, and obtain target multimedia elements;

[0124] In this step, the preliminary multimedia elements refer to initial multimedia contents generated according to the multimedia form selection strategy.

[0125] This step uses deep reinforcement learning to select the best multimedia form, such as pictures, videos, and audio, based on the generated target content. For example, the deep reinforcement learning algorithm is applied to select the most suitable form, such as a short video, based on the current context and external factors, such as the current context of Xiao Li commuting and external factors such as traffic conditions; based on the selected best multimedia form, preliminary multimedia elements are generated, such as a short news video clip; for further personalization, style transfer technology is introduced to adjust the preliminary multimedia elements. For example, based on Xiao Li's personal preferences, such as liking modern art style, the color palette and visual effects of the video are adjusted to make it more in line with Xiao Li's aesthetics.

[0126] Step 404: Based on the target content and the target multimedia elements, a graph neural network is applied to construct a content graph to optimize the connection mode of the target multimedia elements to obtain optimized target multimedia elements, and the optimized target multimedia elements are combined with mixed reality elements to generate final content, wherein the mixed reality elements include AR filters;

[0127] In this step, mixed reality elements include augmented reality (AR) filters, etc., to enhance the user experience.

[0128] This step builds a content graph based on the target content and target multimedia elements using graph neural networks to optimize the way multimedia elements are connected. For example, news articles are associated with key elements in video clips, such as people and places, to form a content graph; the connection of multimedia elements is optimized through the content graph to ensure the logic and coherence between the elements, and the optimized target multimedia elements are generated; the optimized target multimedia elements are combined with mixed reality elements, such as AR filters, to generate the final content. For example, AR filters, such as virtual annotations or interactive elements, are added to video clips to make the content more vivid and interesting, providing an immersive reading experience.

[0129] Step 405: Based on the final content, collect the user's emotional response through an instant feedback mechanism, and analyze the user behavior data in real time to obtain a user behavior analysis result, generate feedback suggestions based on the emotional response and the user behavior analysis result, and generate a multimedia push package based on the feedback suggestions and the final content by applying a federated learning framework and a causal inference method;

[0130] This step is based on the final content. Through the instant feedback mechanism, we collect Xiao Li's emotional response and analyze his behavioral data in real time. Xiao Li's emotional response, such as likes and comments, and Xiao Li's behavioral data, such as dwell time and scrolling speed, obtain the user behavior analysis results. For example, we collect Xiao Li's emotional response and behavioral data to the final content in real time, analyze his satisfaction and engagement, and generate feedback suggestions based on these feedback opinions and behavioral data analysis results, such as adjusting the push frequency or content type; we apply the federated learning framework and causal inference method, combine the user's feedback suggestions and the final content, and generate a highly customized multimedia push package to ensure that each push is more accurate and personalized.

[0131] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0132] Through the variational autoencoder and multi-head attention mechanism, the system can efficiently extract and strengthen the core content, improving the accuracy and relevance of content generation;

[0133] Combining user interests and behavior data, the target content generated by the conditional generative adversarial network is more in line with the user's personalized needs;

[0134] It uses deep reinforcement learning to select the best multimedia format and adjusts the content style through style transfer technology, which improves the user's visual and auditory experience.

[0135] By building a content graph through a graph neural network, the connection method of multimedia elements is optimized, which enhances the logic and coherence of the content.

[0136] Combined with mixed reality elements, the final content generated provides an immersive reading experience, significantly improving user engagement and satisfaction;

[0137] Through the instant feedback mechanism and federated learning framework, the system can continuously optimize the recommendation strategy and provide long-term stable and high-quality services.

[0138] In the intelligent content recommendation system, choosing the right multimedia format is crucial to improving the user experience. Traditional multimedia format selection is often based on static rules, which are difficult to adapt to the personalized needs and dynamic situations of users. Based on this, the present invention provides a specific embodiment, wherein step 403, based on the target content, uses deep reinforcement learning to select the best multimedia format, based on the best multimedia format, generates a multimedia format selection strategy, generates preliminary multimedia elements based on the multimedia format selection strategy, introduces style transfer technology to adjust the preliminary multimedia elements, and obtains the target multimedia elements, specifically including the following steps:

[0139] Step 411: using a deep reinforcement learning algorithm to perform state encoding on the user's situation, the target content, and the user behavior data to construct a target vector, constructing a dynamic reward function based on the target vector, user historical feedback, and user multimedia interaction data, and determining the optimal multimedia format using a policy gradient method based on the dynamic reward function, wherein the user's situation includes time, location, and activity;

[0140] In this step, state encoding refers to converting input data, such as user context, target content, and behavior data, into vector representations as input to the model. The target vector refers to a multidimensional vector obtained through state encoding, which comprehensively describes the user's current context and needs. The dynamic reward function refers to a reward mechanism built based on user historical feedback and multimedia interaction data, which is used to evaluate the effects of different actions.

[0141] Assume that user Xiao Li is preparing dinner at home at 7pm. The deep reinforcement learning algorithm is used to state encode Xiao Li's situation, target content, and user behavior data to construct a target vector. Xiao Li's situation includes time, place, and activity, target content includes recipe recommendations, and Xiao Li's user behavior data includes past cooking habits; for example, information such as 7pm, home, and preparing dinner is encoded into a multidimensional vector, and the target content and user behavior data are also encoded into corresponding vector parts, such as encoding pasta recipes and liking to try new dishes, and finally a target vector that comprehensively describes Xiao Li's current situation is obtained; based on the above target vector, Xiao Li's historical feedback, and multimedia interaction data, Xiao Li's historical feedback such as the recipes he liked, and Xiao Li's multimedia interaction data such as the content he browsed, a dynamic reward function is constructed, and the policy gradient method is used to determine the best multimedia format. For example, based on Xiao Li’s past behavior patterns and feedback, a dynamic reward function was constructed, which would give different reward values ​​according to different multimedia forms. Through the policy gradient method, a variety of possible multimedia forms were simulated, such as video tutorials and graphic descriptions. Finally, video tutorials were selected as the best form because videos are more intuitive and suitable for novice chefs.

[0142] Step 412: constructing a context-aware decision tree based on the optimal multimedia format, and dynamically updating the context-aware decision tree using an online learning algorithm to obtain an updated context-aware decision tree, and generating a multimedia format selection strategy based on the updated context-aware decision tree and multiple influencing factors, wherein the influencing factors include user preferences, historical behavior patterns, and environmental conditions of the user;

[0143] This step builds a context-aware decision tree based on the determined optimal multimedia format, and dynamically updates the decision tree using an online learning algorithm. For example, based on factors such as Xiao Li's time, location, and activities, a context-aware decision tree is built, which can recommend appropriate multimedia formats according to different scenarios; as Xiao Li uses the application more, the system continuously updates the decision tree through an online learning algorithm to make it more accurately reflect Xiao Li's needs and preferences; based on the updated context-aware decision tree and multiple influencing factors such as user preferences, historical behavior patterns, and environmental conditions, a multimedia format selection strategy is generated. For example, based on Xiao Li's habit of watching videos, past behavior patterns, and current home environmental conditions, Xiao Li's past behavior patterns such as frequent searches for Italian recipes, and Xiao Li's current home environmental conditions such as a well-equipped kitchen, a clear multimedia format selection strategy is generated, that is, between 7 and 9 p.m., when cooking at home, video tutorials are recommended first.

[0144] In addition, the best multimedia format is the optimal multimedia type determined by the deep reinforcement learning algorithm, which is the basis of the multimedia format selection strategy. The multimedia format selection strategy is a context-aware decision tree constructed on the basis of the best multimedia format, and is continuously updated through the online learning algorithm to form a set of dynamically adjusted decision rules. This strategy not only considers the selection results of the best multimedia format, but also combines multiple influencing factors to ensure that the form of the recommended content meets the user's immediate needs and adapts to their long-term preferences and environmental changes.

[0145] Step 413: Select a target template corresponding to the multimedia form selection strategy using template matching and generation technology, and add key information of the target content to the target template to generate preliminary multimedia elements, wherein the key information includes text summary, keywords and tags;

[0146] In this step, the multimedia format selection strategy refers to the selection rules generated based on the context-aware decision tree and multiple influencing factors to guide the presentation format of the multimedia content.

[0147] This step selects the target template for the video tutorial based on the multimedia form selection strategy, using template matching and generation technology, and adds key information of the target content, such as text summary, keywords, and tags, to the template to generate preliminary multimedia elements. For example, a template suitable for a video tutorial is selected, which includes a title, introduction, and step-by-step breakdown; key information of a pasta recipe, such as an easy-to-learn pasta making method and required materials such as pasta, tomato sauce, and basil leaves, is added to the template to generate a complete video tutorial.

[0148] Step 414: constructing a target aesthetic preference model based on the preliminary multimedia elements and user preferences, and performing personalized style transfer processing on the preliminary multimedia elements based on the target aesthetic preference model to obtain target multimedia elements;

[0149] This step builds a target aesthetic preference model based on the preliminary multimedia elements and Xiao Li's preferences, such as modern minimalist style, and performs personalized style transfer processing on the preliminary multimedia elements to finally obtain the target multimedia elements. For example, based on Xiao Li's aesthetic preferences, such as simple pictures and soft tones, a target aesthetic preference model is built; style transfer technology is applied to adjust the visual effects of the video tutorial to a modern minimalist style, making the video interface more concise and clear, and the colors soft and comfortable, thereby improving Xiao Li's viewing experience.

[0150] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0151] Through deep reinforcement learning algorithms for state encoding and dynamic reward function construction, the system can efficiently capture the specific context and needs of users, improving the accuracy and relevance of multimedia format selection;

[0152] Combining user preferences, historical behavior patterns, and environmental conditions, the generated multimedia format selection strategy is more in line with the user's personalized needs, significantly improving the user experience;

[0153] By dynamically updating the context-aware decision tree through online learning algorithms, the system can continuously optimize the recommendation strategy to adapt to the long-term changes and immediate needs of users;

[0154] Template matching and generation techniques are used to ensure the integrity of content structure, and personalized style transfer is performed through the target aesthetic preference model, which improves the quality and visual appeal of multimedia content.

[0155] Through the comprehensive application of a series of advanced technologies and algorithms, the multimedia content provided by the system is not only diverse in form but also highly personalized, which significantly improves user satisfaction and engagement.

[0156] In the intelligent content recommendation system, personalization and aesthetic preference are important factors in improving user experience. Traditional multimedia content generation often ignores the unique aesthetic needs of users and is difficult to provide a highly customized experience. Based on this, the present invention provides a specific embodiment, wherein step 414, based on the preliminary multimedia elements and user preferences, constructs a target aesthetic preference model, and performs personalized style transfer processing on the preliminary multimedia elements based on the target aesthetic preference model to obtain the target multimedia elements, specifically includes the following steps:

[0157] Step 421: Processing the user multimedia interaction data using deep learning technology to obtain user feature representation, analyzing existing user comments using natural language processing technology to extract text features, optimizing the user feature representation based on the text features and user preferences to obtain aesthetic preference features, wherein the user preferences include the user's preference for color, shape, and texture;

[0158] In this step, user feature representation refers to the multidimensional vector obtained by processing user multimedia interaction data, which describes the user's preferences and behavior patterns. Text features refer to key information extracted from user comments, such as sentiment, theme, etc. Aesthetic preference features refer to the user feature representation optimized by combining text features and user preferences, reflecting the user's aesthetic needs.

[0159] Assume that user Xiao Li often browses and comments on food pictures. Use deep learning technology to process his multimedia interaction data, such as likes, comments, and sharing records, to obtain user feature representation. For example, through the convolutional neural network (CNN) to analyze the pictures and videos that Xiao Li has browsed, extract his interests in different types of content, such as dessert pictures and cooking videos; these interests are encoded into a multidimensional vector, describing Xiao Li's multimedia preferences and behavior patterns; natural language processing technology is used to analyze Xiao Li's historical comments, extract text features, and optimize the user feature representation based on these features and his known preferences, such as liking warm colors and circular designs, to obtain aesthetic preference features. For example, through sentiment analysis and topic modeling, identify the colors, shapes, and textures that Xiao Li often mentions in his comments, such as warm red, shapes such as rounded designs, and textures such as smooth surfaces; combine the above text features and known preferences to optimize the user feature representation and generate a new vector, namely the aesthetic preference feature, which more accurately reflects Xiao Li's aesthetic needs.

[0160] Step 422: constructing a preliminary aesthetic preference model of the user based on the preliminary multimedia elements and the aesthetic preference features, introducing an online learning algorithm based on the preliminary aesthetic preference model, updating the preliminary aesthetic preference model, and obtaining a target aesthetic preference model;

[0161] This step is based on the preliminary multimedia elements and the above aesthetic preference features. The preliminary multimedia elements are such as a picture of food. The user's preliminary aesthetic preference model is constructed, and the online learning algorithm is introduced to dynamically update the model to obtain the target aesthetic preference model. For example, based on Xiao Li's aesthetic preference features, a preliminary aesthetic preference model is constructed, which can predict his reaction to different types of multimedia content; as Xiao Li uses the application more, the system continuously updates the model through the online learning algorithm to obtain the target aesthetic preference model, so that it can more accurately reflect Xiao Li's real-time needs and changes.

[0162] Step 423: Based on the target aesthetic preference model, a user style model with a degree of adaptation higher than a preset threshold is selected; based on the user style model, a style transfer algorithm is applied to perform personalized adjustment on the preliminary multimedia element to obtain a preliminary adjusted multimedia element, wherein the personalized adjustment includes changing the color palette of the image, adjusting the visual effect of the video, and modifying the sound effect of the audio;

[0163] In this step, the user style model refers to a model that describes the user's preferences for specific aspects such as color, shape, texture, etc.

[0164] This step selects user style models with a degree of fit higher than a preset threshold based on the target aesthetic preference model, and applies a style transfer algorithm to personalize the preliminary multimedia elements to obtain preliminary adjusted multimedia elements. For example, the user style model with the highest degree of fit is selected, which reflects Xiao Li's preference for warm tones and circular designs; the style transfer algorithm is applied to change the color palette of the image, such as changing cold tones to warm tones; the visual effects of the video are adjusted, such as adding a soft filter; and the sound effects of the audio are modified, such as adding relaxing background music, to generate preliminary adjusted multimedia elements.

[0165] Step 424: using an existing image or audio quality evaluation index to evaluate the preliminarily adjusted multimedia element to obtain a quality evaluation result, and based on the quality evaluation result, using a Bayesian optimization method to optimize and adjust key parameters of the target aesthetic preference model and the user style model to optimize the preliminarily adjusted multimedia element and generate a target multimedia element;

[0166] In this step, the image / audio quality evaluation index refers to the standard used to measure the quality of multimedia elements. The image quality evaluation index includes peak signal-to-noise ratio, structural similarity index and visual information fidelity. The audio quality evaluation index includes perceptual evaluation of speech quality, root mean square error and short-time objective.

[0167] This step uses existing image or audio quality evaluation indicators to evaluate the preliminarily adjusted multimedia elements, obtains quality evaluation results, and uses Bayesian optimization methods to optimize and adjust the key parameters of the target aesthetic preference model and user style model to generate the final target multimedia elements. For example, the quality of the preliminarily adjusted multimedia elements is evaluated by indicators such as peak signal-to-noise ratio to ensure that the adjusted content is both beautiful and in line with Xiao Li's aesthetic preferences; based on the quality evaluation results, the Bayesian optimization method is used to adjust the key parameters of the target aesthetic preference model and user style model, and further optimize the preliminarily adjusted multimedia elements to ensure that the quality of the final generated content reaches the best state.

[0168] More specifically, the present invention provides a calculation formula for a target multimedia element, which is calculated and generated by a Bayesian optimization method. The specific calculation formula is as follows:

[0169] ;

[0170] in, represents the target multimedia element; represents the Bayesian optimization method, which is used to optimize model parameters; Indicates the quality assessment result; Represents the target aesthetic preference model; Represents the user style model; represents a set of key parameters, including parameters of the target aesthetic preference model and the user style model; Indicates that in all possible In, find The biggest one ; Indicates the expected value, which means that given the parameter The expected quality assessment results are as follows; Representing the target aesthetic preference model At the given parameters The output of the following; represents the user style model At the given parameters The output of the following;

[0171] The target multimedia element formula uses the Bayesian optimization method to adjust key parameters to ensure the optimal quality of the generated content. By maximizing the expected quality evaluation results, it ensures that the generated content not only meets user preferences but also has high quality. Through Bayesian optimization, it ensures that the generated target multimedia elements meet the highest quality standards in terms of vision, hearing, etc., which are not only beautiful but also meet the user's aesthetic preferences, significantly improving user participation and satisfaction; through the instant feedback mechanism, users' feedback and suggestions on the final content are collected, and the model and content generation strategy are further optimized to form a virtuous circle.

[0172] Among them, the calculation formula of the user style model is as follows:

[0173] ;

[0174] in, Represents the user style model; Represents the target aesthetic preference model; Represents the user style model library, which contains models of various styles; represents the fitness score, which measures how well the model matches the user's preferences; Indicates the preset threshold, only models with a fitness score higher than this threshold will be selected; Represents finding the cosine similarity The largest user style model ; Represents the cosine similarity function, which is used to compare the similarity of two vectors.

[0175] The user style model formula ensures that the generated content is highly consistent with the user's personal aesthetic preferences by selecting a user style model with a degree of fit higher than a preset threshold; at the same time, the cosine similarity function is used to evaluate the degree of fit between the model and the user's aesthetics to ensure that the selected model is the most suitable for the user. By selecting the most suitable user style model, highly personalized multimedia content is generated to improve user experience, ensure that the generated content is consistent in style, avoid incoordination or abruptness, and improve user satisfaction.

[0176] Among them, the calculation formula of the target aesthetic preference model is as follows:

[0177] ;

[0178] in, Represents the target aesthetic preference model; represents an online learning algorithm for dynamically updating the preliminary aesthetic preference model; Indicates the search for The minimized one ; represents a preliminary aesthetic preference model; Indicates the length of the time series; Representing a preliminary aesthetic preference model In time The prediction results Indicates at time User feedback; Represents the loss function, which is used to evaluate the difference between the model prediction and the actual user feedback.

[0179] The target aesthetic preference model formula introduces an online learning algorithm, which enables the model to continuously learn and adapt to the user's latest behavior and feedback, ensuring that the model always reflects the latest user preferences; multiple optimization objectives are considered at the same time to comprehensively evaluate and optimize the user activity intention prediction results. Through online learning and multi-objective optimization, the model can not only perform stably over a long period of time, but also quickly respond to short-term changes in users; the optimized target aesthetic preference model can more accurately predict the user's behavioral intentions, thereby improving the relevance and attractiveness of the recommended content.

[0180] Among them, the calculation formula of the preliminary aesthetic preference model is as follows:

[0181] ;

[0182] in; represents a preliminary aesthetic preference model; Represents preliminary multimedia elements; Representing aesthetic preference features, combining user behavior data and text features extracted by natural language processing; represents the function for constructing a preliminary aesthetic preference model; represents the weight matrix that is dynamically adjusted according to the aesthetic preference characteristics; represents the Hadamard product, that is and The element-wise product of ; Represents the bias vector.

[0183] The formula of the preliminary aesthetic preference model is constructed based on the user's geographic location information, behavioral data and multimedia interaction data, aiming to capture the user's preliminary interests and preferences; through multi-source data fusion, it can more accurately depict the user's preliminary interests and preferences, providing a solid foundation for subsequent optimization; at the same time, it can be updated according to the user's real-time behavior to maintain sensitivity and accuracy to user preferences.

[0184] In summary, the preliminary aesthetic preference model provides a basic description of user behavior and provides data support for subsequent optimization; the target aesthetic preference model ensures the long-term stability and short-term responsiveness of the model through online learning and multi-objective optimization; the user style model ensures that the generated content is highly personalized, enhancing the user's immersion and satisfaction; the target multimedia elements ensure the quality of the content and the effective use of user feedback through Bayesian optimization and instant feedback mechanism.

[0185] The embodiments of the present invention can achieve the following beneficial effects through the above steps:

[0186] Through deep learning technology and natural language processing technology, the system can efficiently extract key information from users' multimedia interaction data and comments, improving the accuracy and relevance of user feature representation;

[0187] By combining user preferences and text features to optimize user feature representation, the generated aesthetic preference features are more in line with the unique needs of users, significantly improving the effect of personalized recommendations.

[0188] Dynamically updating the preliminary aesthetic preference model through online learning algorithms can continuously optimize the recommendation strategy to adapt to the long-term changes and immediate needs of users;

[0189] Using style transfer algorithms and image or audio quality assessment metrics to ensure that the adjusted multimedia content is not only diverse in form but also high in quality and meets the user's aesthetic standards;

[0190] Through the comprehensive application of a series of advanced technologies and algorithms, the multimedia content provided is highly personalized and high-quality, significantly improving user satisfaction and engagement.

[0191] Figure 2 A schematic diagram of the structure of an intention-guided push system based on geographic location information and multimedia interaction is provided for an embodiment of the present invention. Figure 2 As shown, the system includes:

[0192] An identification module 21 is used to identify the specific geographical location of the user and the surrounding environment characteristics of the specific geographical location based on the geo-fencing technology and the deep neural network, build an interest model based on the specific geographical location, the surrounding environment characteristics and the user behavior data, use the interest model to predict the user's activity intention, apply context awareness to calculate the influence of external factors, obtain the external factor influence result, optimize the user's activity intention based on the external factor influence result, and generate a user activity intention prediction result;

[0193] The analysis module 22 is used to select multiple target information using a Bayesian network based on the user activity intention prediction result, fuse the multiple target information to obtain fused information, analyze the user's behavior rhythm based on the fused information using a time series prediction algorithm to obtain a behavior rhythm analysis result, determine the best push timing based on the behavior rhythm analysis result, construct a user and location relationship map based on the best push timing using a graph theory algorithm, and generate a time sensitivity optimization plan based on the user and location relationship map;

[0194] The optimization module 23 is used to generate customized content based on the time sensitivity optimization scheme by using natural language processing technology and generative adversarial networks, perform sentiment adaptive adjustment on the customized content by using sentiment analysis to obtain content to be optimized, analyze feedback information of similar users by using collaborative filtering algorithm to obtain feedback information analysis results, and adjust the content to be optimized based on the feedback information analysis results to obtain optimized content;

[0195] The generation module 24 is used to process the optimized content using an autoencoder and a multi-head attention mechanism to obtain target content, combine the target multimedia element and the target content to obtain final content, collect user feedback and suggestions on the final content through an instant feedback mechanism, and generate a multimedia push package based on the feedback suggestions and the final content.

[0196] Figure 2 The intention-guided push system based on geographic location information and multimedia interaction can be executed Figure 1 The implementation principle and technical effect of the intention-guided push method based on geographic location information and multimedia interaction described in the illustrated embodiment will not be repeated. The specific manner in which each module and unit performs operations in the intention-guided push system based on geographic location information and multimedia interaction in the above embodiment has been described in detail in the embodiment of the method, and will not be elaborated here.

[0197] In one possible design, Figure 2 The intention-guided push system based on geographic location information and multimedia interaction of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0198] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0199] The processing component 32 is used to: identify the specific geographical location of the user and the surrounding environment characteristics of the specific geographical location based on geo-fencing technology and deep neural networks, build an interest model based on the specific geographical location, the surrounding environment characteristics and user behavior data, use the interest model to predict the user's activity intention, apply context awareness to calculate the impact of external factors, obtain the external factor impact result, optimize the user's activity intention based on the external factor impact result, and generate a user activity intention prediction result; based on the user activity intention prediction result, use a Bayesian network to select multiple target information, fuse the multiple target information to obtain fused information, based on the fused information, use a time series prediction algorithm to analyze the user's behavior rhythm to obtain a behavior rhythm analysis result, based on the behavior rhythm analysis result, determine the best push time, based on the user's activity intention prediction result. The best push timing is determined, a graph theory algorithm is used to construct a user and location relationship graph, and a time-sensitivity optimization plan is generated based on the user and location relationship graph; based on the time-sensitivity optimization plan, customized content is generated using natural language processing technology and a generative adversarial network, and a sentiment analysis method is used to perform sentiment adaptability adjustment on the customized content to obtain content to be optimized, and a collaborative filtering algorithm is used to analyze feedback information of similar users to obtain a feedback information analysis result, and the content to be optimized is adjusted based on the feedback information analysis result to obtain optimized content; the optimized content is processed using an autoencoder and a multi-head attention mechanism to obtain target content, and the target multimedia element and the target content are combined to obtain the final content, and feedback suggestions from users on the final content are collected through an instant feedback mechanism, and a multimedia push package is generated based on the feedback suggestions and the final content.

[0200] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0201] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0202] Of course, the computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0203] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc.

[0204] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0205] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0206] The embodiment of the present invention further provides a computer storage medium storing a computer program, which can achieve the above-mentioned Figure 1 The illustrated embodiment is an intent-guided push method based on geographic location information and multimedia interaction.

[0207] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0208] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0209] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0210] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. 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 embodiments of the present invention.

Claims

1. An intention-guided push method based on geographic location information and multimedia interaction, characterized in that: include: Based on geo-fencing technology and deep neural networks, the specific geographical location of the user and the surrounding environment characteristics of the specific geographical location are identified, and based on the specific geographical location, the surrounding environment characteristics and user behavior data, an interest model is constructed, and the user activity intention is predicted using the interest model, and the influence of external factors is calculated using context awareness to obtain the external factor influence result, and the user activity intention is optimized based on the external factor influence result to generate a user activity intention prediction result; Based on the prediction result of the user activity intention, multiple target information is selected using a Bayesian network, and the multiple target information is fused to obtain fused information. Based on the fused information, a time series prediction algorithm is used to analyze the user's behavior rhythm to obtain a behavior rhythm analysis result. Based on the behavior rhythm analysis result, an optimal push timing is determined. Based on the optimal push timing, a graph theory algorithm is used to construct a user and location relationship map, and a time sensitivity optimization plan is generated based on the user and location relationship map. Based on the time-sensitivity optimization scheme, natural language processing technology and generative adversarial networks are used to generate customized content, sentiment analysis is used to adjust sentiment adaptability of the customized content to obtain content to be optimized, collaborative filtering algorithm is used to analyze feedback information of similar users to obtain feedback information analysis results, and the content to be optimized is adjusted based on the feedback information analysis results to obtain optimized content; The optimized content is processed by using an autoencoder and a multi-head attention mechanism to obtain target content, and the target multimedia element and the target content are combined to obtain final content. Feedback suggestions from users on the final content are collected through an instant feedback mechanism, and a multimedia push package is generated based on the feedback suggestions and the final content.

2. The method according to claim 1, characterized in that Based on the geo-fencing technology and deep neural network, the specific geographical location of the user and the surrounding environment characteristics of the specific geographical location are identified; based on the specific geographical location, the surrounding environment characteristics and user behavior data, an interest model is constructed; the user activity intention is predicted using the interest model; the influence of external factors is calculated using context awareness to obtain the external factor influence result; the user activity intention is optimized based on the external factor influence result, and the user activity intention prediction result is generated, including: Using geo-fencing technology combined with a spatiotemporal clustering algorithm to detect a user's behavior of entering or leaving a preset geographic area, and recording the user's geographic location information and a timestamp corresponding to the geographic location information, the preset geographic area includes home, workplace, concert venue, and sports field; Based on the geographic location information and the timestamp, using a deep convolutional neural network to identify the specific geographic location of the user and the surrounding environment characteristics of the specific geographic location, the surrounding environment characteristics including facility type characteristics, crowd density characteristics and traffic flow characteristics, the specific geographic location is one or more locations within a preset geographic area; Combining the user behavior data and the user multimedia interaction data, using a graph neural network to construct a user's interest model to be optimized, optimizing the interest model to be optimized based on the specific geographic location and the surrounding environment characteristics to obtain an interest model, and predicting the user's activity intention based on the interest model; Acquire external factors, apply situational awareness to calculate the impact of the external factors on the user's activity intention, and obtain external factor impact results, wherein the external factors include weather, traffic conditions, and time; Based on the external factors, the user activity intention is optimized using a multi-objective optimization algorithm to generate a user activity intention prediction result. Multiple optimization objectives are considered during the optimization process, and the optimization objectives include time sensitivity, location relevance, and user satisfaction.

3. The method according to claim 2, characterized in that Based on the geographic location information and the timestamp, a deep convolutional neural network is used to identify the specific geographic location of the user and the surrounding environment characteristics of the specific geographic location, wherein the surrounding environment characteristics include facility type characteristics, crowd density characteristics, and traffic flow characteristics, and the specific geographic location is one or more locations within a preset geographic area, including: Based on the geographic location information and the timestamp, a spatiotemporal attention mechanism and a deep convolutional neural network are used to perform multi-scale feature extraction on the user's location and dynamic features of the user's location that change over time, so as to obtain a user location feature map; based on the user location feature map, a semantic segmentation algorithm is applied to analyze the map data, extract environmental information, and generate a comprehensive environmental feature description; Based on the comprehensive environmental feature description and the timestamp, a time series prediction model is applied to analyze the difference in pedestrian density or traffic flow changes in different time periods to obtain environmental feature prediction results.

4. The method according to claim 3, characterized in that Based on the geographic location information and the timestamp, a spatiotemporal attention mechanism and a deep convolutional neural network are used to perform multi-scale feature extraction on the user's location and the dynamic features of the user's location that change over time, so as to obtain a user location feature map. According to the user location feature map, a semantic segmentation algorithm is applied to analyze the map data, extract environmental information, and generate a comprehensive environmental feature description, including: Based on the geographic location information and the timestamp, an input tensor is generated, based on the input tensor, an importance weight of each spatiotemporal point is calculated using a spatiotemporal attention mechanism, and a key area that has an impact on user behavior is determined based on the importance weight, and a spatiotemporal attention weight map is constructed based on the key area and the importance weight; Based on the spatiotemporal attention weight map, a convolution operation is performed on the input tensor through a plurality of convolution kernels of different scales in a deep convolutional neural network to obtain different levels of features from local to global, and a multi-scale feature map is generated based on the different levels of features; The multi-scale feature map is subjected to feature fusion through residual connection or skip connection to generate a fused feature map, a user position feature map is generated based on the fused feature map, semantic information in the user position feature map is extracted using a semantic segmentation network with an encoder-decoder structure in a semantic segmentation algorithm, and a semantic segmentation map is generated based on the semantic information, wherein the semantic information includes map data and environmental information; The softmax layer in the semantic segmentation algorithm is used to classify each pixel in the semantic segmentation map to identify and classify different elements in the environment, and a comprehensive environmental feature description is generated based on the different elements. The different elements include buildings, roads, green spaces and waters. The comprehensive environmental feature description contains static geographic information and dynamic pedestrian and vehicle flow conditions.

5. The method according to claim 1, characterized in that: The optimized content is processed by using an autoencoder and a multi-head attention mechanism to obtain target content, a target multimedia element and the target content are combined to obtain final content, feedback suggestions of users on the final content are collected through an instant feedback mechanism, and a multimedia push package is generated based on the feedback suggestions and the final content, including: Using a variational autoencoder to compress and decode the optimized content to obtain a core content representation, and combining a multi-head attention mechanism to evaluate the importance of different parts of the core content representation to strengthen the key content parts and generate an enhanced core content representation; Based on the enhanced core content representation, user interest points and the user behavior data, applying a conditional generative adversarial network to generate target content; Based on the target content, using deep reinforcement learning to select the best multimedia form, based on the best multimedia form, generating a multimedia form selection strategy, based on the multimedia form selection strategy, generating preliminary multimedia elements, introducing style transfer technology to adjust the preliminary multimedia elements, and obtaining target multimedia elements; Based on the target content and the target multimedia elements, a graph neural network is applied to construct a content graph to optimize the connection mode of the target multimedia elements to obtain optimized target multimedia elements, and the optimized target multimedia elements and mixed reality elements are combined to generate final content, wherein the mixed reality elements include AR filters; Based on the final content, the user's emotional response is collected through an instant feedback mechanism, and the user behavior data is analyzed in real time to obtain a user behavior analysis result; based on the emotional response and the user behavior analysis result, feedback suggestions are generated; based on the feedback suggestions and the final content, a federated learning framework and causal inference method are used to generate a multimedia push package.

6. The method according to claim 5, characterized in that Based on the target content, using deep reinforcement learning to select the best multimedia form, based on the best multimedia form, generating a multimedia form selection strategy, generating preliminary multimedia elements based on the multimedia form selection strategy, introducing style transfer technology to adjust the preliminary multimedia elements, and obtaining target multimedia elements, including: Using a deep reinforcement learning algorithm to state-encode the user's situation, the target content, and the user behavior data to construct a target vector, construct a dynamic reward function based on the target vector, user historical feedback, and user multimedia interaction data, and determine the optimal multimedia format using a policy gradient method based on the dynamic reward function, wherein the user's situation includes time, location, and activity; Based on the optimal multimedia form, construct a context-aware decision tree, and dynamically update the context-aware decision tree using an online learning algorithm to obtain an updated context-aware decision tree, and generate a multimedia form selection strategy based on the updated context-aware decision tree and multiple influencing factors, wherein the influencing factors include user preferences, historical behavior patterns, and environmental conditions of the user; Selecting a target template corresponding to the multimedia form selection strategy by using template matching and generation technology, and adding key information of the target content to the target template to generate preliminary multimedia elements, wherein the key information includes text summary, keywords and tags; Based on the preliminary multimedia elements and user preferences, a target aesthetic preference model is constructed, and based on the target aesthetic preference model, personalized style transfer processing is performed on the preliminary multimedia elements to obtain target multimedia elements.

7. The method according to claim 6, characterized in that Based on the preliminary multimedia elements and user preferences, a target aesthetic preference model is constructed, and based on the target aesthetic preference model, personalized style transfer processing is performed on the preliminary multimedia elements to obtain target multimedia elements, including: Processing the user multimedia interaction data using deep learning technology to obtain user feature representation, analyzing existing user comments using natural language processing technology to extract text features, optimizing the user feature representation based on the text features and user preferences to obtain aesthetic preference features, wherein the user preferences include the user's preference for color, shape, and texture; Based on the preliminary multimedia elements and the aesthetic preference features, construct a preliminary aesthetic preference model of the user, based on the preliminary aesthetic preference model, introduce an online learning algorithm, update the preliminary aesthetic preference model, and obtain a target aesthetic preference model; Based on the target aesthetic preference model, a user style model with a degree of adaptation higher than a preset threshold is selected, and based on the user style model, a style transfer algorithm is applied to perform personalized adjustment on the preliminary multimedia elements to obtain preliminary adjusted multimedia elements, wherein the personalized adjustment includes changing the color palette of the image, adjusting the visual effects of the video, and modifying the sound effects of the audio; The preliminarily adjusted multimedia elements are evaluated using existing image or audio quality evaluation indicators to obtain quality evaluation results. Based on the quality evaluation results, the key parameters of the target aesthetic preference model and the user style model are optimized and adjusted using a Bayesian optimization method to optimize the preliminarily adjusted multimedia elements and generate target multimedia elements.

8. An intention-guided push system based on geographic location information and multimedia interaction, characterized in that: include: an identification module, for identifying the specific geographic location of the user and the surrounding environment characteristics of the specific geographic location based on geo-fencing technology and deep neural networks, constructing an interest model based on the specific geographic location, the surrounding environment characteristics and user behavior data, predicting the user's activity intention using the interest model, calculating the influence of external factors using context awareness, obtaining the external factor influence result, optimizing the user's activity intention based on the external factor influence result, and generating a user activity intention prediction result; An analysis module, for selecting multiple target information using a Bayesian network based on the prediction result of the user's activity intention, fusing the multiple target information to obtain fused information, analyzing the user's behavior rhythm based on the fused information using a time series prediction algorithm to obtain a behavior rhythm analysis result, determining an optimal push timing based on the behavior rhythm analysis result, constructing a user and location relationship map based on the optimal push timing using a graph theory algorithm, and generating a time sensitivity optimization plan based on the user and location relationship map; An optimization module is used to generate customized content based on the time-sensitivity optimization scheme by using natural language processing technology and generative adversarial networks, perform sentiment adaptive adjustment on the customized content by using sentiment analysis to obtain content to be optimized, analyze feedback information of similar users by using collaborative filtering algorithm to obtain feedback information analysis results, and adjust the content to be optimized based on the feedback information analysis results to obtain optimized content; A generation module is used to process the optimized content using an autoencoder and a multi-head attention mechanism to obtain target content, combine target multimedia elements with the target content to obtain final content, collect user feedback suggestions on the final content through an instant feedback mechanism, and generate a multimedia push package based on the feedback suggestions and the final content.

9. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement an intention-guided push method based on geographic location information and multimedia interaction as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, an intention-guided push method based on geographic location information and multimedia interaction as described in any one of claims 1 to 7 is implemented.

Images