A new media operation content recommendation system based on big data

Abstract

The application discloses a new media operation content recommendation system based on big data, and relates to the technical field of new media operation.The application collects cross-platform short-time behaviors in real time through edge computing, dynamically fuses multi-platform user interest features in combination with federated learning, breaks through the limitation of single-platform data, and accurately captures user instant interest shift; the application fills in cross-platform behavior breakpoints by using bidirectional LSTM and attention mechanism, dynamically adjusts prediction weights according to device types, and significantly improves the response speed of sudden interest; the application opens up multi-platform content resource pools through semantic association index and hotspot adjustment matrix, realizes semantic-level matching and real-time scheduling of cross-platform content; in addition, the application completes multi-platform cooperation under the premise of guaranteeing data security through a privacy compliance unit, and avoids legal risks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new media operation, and particularly relates to a new media operation content recommendation system based on big data. BACKGROUND

[0002] The current content recommendation system faces the double bottlenecks of cross-platform real-time interest capture and multi-terminal collaborative scheduling; the user behavior track has shifted from a single platform to fragmented switching of multiple scenes such as microblog, short video, information App, but the mainstream recommendation algorithm still adopts a static portrait mechanism, and relies on single-platform historical data to build a long-term interest model; this mode has two fatal defects: first, it cannot capture the instantaneous interest shift generated in the multi-terminal switching process of the user, for example, the immediate attention behavior triggered by a sudden news is often missed after the algorithm update cycle; second, the existing system excessively relies on explicit behavior signals such as clicks and collections, and lacks analysis ability for microsecond-level implicit behaviors such as sliding speed and reading pauses, resulting in a long delay of several hours in responding to hot events; even if part of the platform tries to introduce a real-time computing framework, it is still limited by single-point data dimensions, and it is difficult to build a complete instantaneous interest graph.

[0003] The data island and strategy fragmentation among multiple platforms further aggravate the recommendation misalignment problem; when the user generates a new interest on platform A, the recommendation engine of platform B / C still uses the old portrait for content matching due to the lack of cross-domain information sharing mechanism; although the existing technology attempts to realize cross-platform recommendation through user ID, it faces two major practical obstacles: first, the privacy compliance restrictions prevent the original behavior data from being circulated across platforms; second, the metadata standard difference of the content resource library of each platform makes it difficult to quickly schedule the associated content of a hot topic; this fragmented state directly leads to the failure of two typical scenarios: after the user triggers interest on a short video platform, the information platform cannot immediately push in-depth interpretation; when using cross-device, the behavior data generated on the mobile phone cannot real-time correct the recommendation strategy on the PC; the existing technical solutions generally fall into a dilemma when balancing real-time and collaboration. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] The present application provides a new media operation content recommendation system based on big data to solve the problems of traditional recommendation systems relying on single-platform static portraits, cross-platform real-time interest capture failure, data island leading to recommendation lag and collaboration loss.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The embodiment of the present application provides a new media operation content recommendation system based on big data, which comprises,

[0008] An edge behavior perception module deployed on a client or an edge computing node of a plurality of new media platforms, configured to collect user short-time high-frequency behavior data across platforms in real time and generate a time-series behavior vector;

[0009] A federal aggregation module connected to local modeling units of the platforms, configured to perform weighted fusion on user interest features output by the platforms through a federal learning mechanism, and generate a cross-platform user interest graph;

[0010] An interest transition prediction module configured to construct a time-series prediction model based on the interest graph, and output a prediction result of interest transition direction of a user in a future preset time period;

[0011] A cross-platform scheduling module configured to match a content resource pool of the platforms according to the prediction result, and push adaptive content to a target platform through a preset interface and configure a recommendation position triggering logic.

[0012] As a preferred scheme of the new media operation content recommendation system based on big data, the edge behavior perception module specifically includes:

[0013] A lightweight behavior analysis unit embedded in a client or a CDN edge server of each platform, configured to capture user swipe frequency, click event sequence, page dwell time and reading speed threshold breakthrough behavior in real time;

[0014] A timestamp marking unit configured to generate a discrete event stream with a timestamp according to an occurrence time of the collected original behavior data;

[0015] A vector encoding unit configured to convert the event stream into a multi-dimensional behavior vector within a fixed time window and attach a platform identifier;

[0016] The short-time high-frequency behavior data refers to operation events triggered by a user continuously within a preset time window, including behaviors such as swipe rate exceeding a threshold, single-page dwell time being lower than a set lower limit, or continuous click interval being shorter than a set time length;

[0017] The time-series behavior vector is generated through a sliding time window mechanism, each window contains normalized behavior indicators within a preset time period, including swipe acceleration mean, click event density, page switching frequency and reading speed fluctuation variance, and is spliced into a multi-dimensional tensor in timestamp order.

[0018] As a preferred scheme of the new media operation content recommendation system based on big data, the federal aggregation module performs the following operations:

[0019] A convolutional neural network is used in the local modeling unit to extract time-local features of user behavior within a platform;

[0020] Model parameters are added with noise through a differential privacy mechanism and then transmitted to a central server.

[0021] A dynamic weighting strategy is used for parameter aggregation, and the weight values ​​are dynamically adjusted based on the real-time data freshness and user activity of each platform.

[0022] In the dynamic weight allocation strategy, the data freshness weight is calculated inversely proportional to the difference between the platform data update timestamp and the current time, and the user activity weight is dynamically adjusted based on the user's real-time operation frequency on the current platform.

[0023] As a preferred embodiment of the new media operation content recommendation system based on big data described in this invention, the federated aggregation module employs a dynamic weight allocation strategy for parameter aggregation, including:

[0024] In federated aggregation, parameter weights are dynamically assigned based on the data freshness and user activity of each platform, for the first... Each platform calculates its data freshness:

[0025] ,

[0026] in, Indicates the current time and platform The difference in the time of the most recent data update This represents the time decay coefficient, ranging from 0.01 to 1.0.

[0027] Then measure user activity:

[0028] ,in, Indicates the platform within the preset time window. The total number of operation events on the device. This represents the total number of operation events across all platforms within the same time window. Indicates the total number of platforms;

[0029] Introducing the balance coefficient value range Construct a comprehensive score:

[0030] ,

[0031] in, The importance of controlling the ratio between freshness and activity. This is the overall score before integration;

[0032] The ratings from all platforms are normalized to obtain the final weights:

[0033] ,in, This represents the sum of the overall ratings across all platforms. That is, the first The platform dynamically allocates weights during federated aggregation.

[0034] As a preferred embodiment of the new media operation content recommendation system based on big data described in this invention, the interest transition prediction module includes:

[0035] A bidirectional LSTM network is used to process the temporal correlation features of cross-platform behavioral sequences.

[0036] Attention mechanism layer, identifies the contribution weight of behavioral data from different platforms to the current state of interest;

[0037] The output layer generates a multidimensional prediction vector that includes interest category, intensity value, and duration.

[0038] The bidirectional LSTM network introduces a platform gap filling mechanism: when a user switches platforms, a virtual filling vector is generated based on historical behavior patterns;

[0039] The attention mechanism layer dynamically adjusts the weight allocation strategy based on the user activity coefficient and the user device type, assigning higher attention weights to sudden behavioral data generated by the mobile platform.

[0040] As a preferred embodiment of the new media operation content recommendation system based on big data described in this invention, in the process of constructing a time-series prediction model based on interest graphs, the user's cross-platform interest node sequence is mapped to time-series input, capturing the dynamic evolution of interest state over time, including:

[0041] Let the sequence of user interest nodes within a preset time period be: It is mapped into a vector sequence through knowledge graph embedding technology. , Indicates the length of the interest sequence;

[0042] Based on this, a bidirectional long short-term memory network (Bi-LSTM) is used to extract temporal correlation features, described as follows:

[0043] ,

[0044] ,

[0045] ,

[0046] in, Indicates time Interest vector, and These are forward and backward LSTM units, respectively. and These are the corresponding hidden states, which are concatenated to obtain a bidirectional hidden state. ;

[0047] To fill the blank behavior segment of users when switching across platforms, define the switching interval with the threshold , generate a virtual padding vector:

[0048] ,

[0049] wherein represents the historical hidden state mean of the user when switching from platform to , represents the platform index where the current behavior occurs, is a linear mapping function, is a mapping weight parameter, is a padding vector inserted into the sequence, which is inserted into the corresponding position to ensure the continuity of the sequence;

[0050] Introduce an attention mechanism to dynamically adjust the contribution weight at different times according to the device type and user activity. First, embed the device type as , embed the user activity as , construct the context vector , and calculate the attention raw score:

[0051] ,

[0052] wherein and are hidden state and context mapping matrices respectively, is a bias vector, is an attention projection vector, is the score at time , and the attention weight is obtained by normalization:

[0053] ,

[0054] wherein is the normalized weight at time ;

[0055] Aggregate the weighted hidden state into the final context vector:

[0056] ,

[0057] And the output layer generates the probability distribution of the future interest transition direction:

[0058] ,

[0059] ​wherein, and are output mapping matrix and bias, respectively, is a multi-dimensional probability vector of interest transfer direction.

[0060] As a preferred scheme of the new media operation content recommendation system based on big data, the cross-platform scheduling module comprises:

[0061] a content metadata mapping unit, which establishes a semantic association index of multi-platform content resources;

[0062] a push strategy engine, which matches a content recommendation position specification of a target platform and a user equipment type according to a prediction result;

[0063] a preloading interface, which caches the recommended content to an edge node of the target platform through an API protocol between platforms;

[0064] The semantic association index is realized through a knowledge graph embedding technology, maps theme labels of multi-platform content to a unified vector space, and establishes a cross-platform content association relationship based on cosine similarity.

[0065] As a preferred scheme of the new media operation content recommendation system based on big data, the content metadata mapping unit performs:

[0066] extracting theme labels, sentiment tendencies and knowledge graph entities of content resources of each platform;

[0067] constructing a cross-platform semantic similarity matrix and establishing a conversion rule library between different content formats;

[0068] dynamically updating a content association strength coefficient of a hot event;

[0069] The semantic similarity matrix is generated through comprehensive calculation of a title keyword co-occurrence frequency of cross-platform content, a user comment sentiment polarity consistency and a knowledge graph entity association degree.

[0070] As a preferred scheme of the new media operation content recommendation system based on big data, the content metadata mapping unit constructs the cross-platform semantic similarity matrix in the following manner:

[0071] Let a cross-platform content set be , denote a total number of contents in the cross-platform content set, extract a knowledge graph embedding vector , a theme label set , a sentiment score and an entity set of each piece of content, respectively;

[0072] On this basis, define the initial semantic similarity matrix :

[0073] ,

[0074] wherein, is a normalized weight coefficient, , is a vector cosine similarity, is a label Jaccard similarity, is a sentiment polarity difference function, is an entity overlap degree;

[0075] Introduce a hot spot adjustment matrix H:

[0076] wherein, is a hot spot fusion coefficient, , indicates that the content and co-occurrence hotness in the current hot spot event, is the maximum value of all content pair hotness, traverse the index set of all content pairs , indicates the normalized hot spot co-occurrence intensity, and the value range is [0, 1].

[0077] As a preferred scheme of the new media operation content recommendation system based on big data, the application further comprises:

[0078] A privacy compliance verification unit performs homomorphic encryption processing on model parameters before federal aggregation;

[0079] A behavior data life cycle management unit sets an automatic erasure mechanism for short-time behavior data;

[0080] A cross-platform identifier desensitization unit converts user equipment fingerprint information by using an irreversible hash algorithm.

[0081] The application has the following advantages: the application collects cross-platform short-time behaviors in real time through edge computing, dynamically fuses multi-platform user interest features in combination with federal learning, breaks through the limitation of single-platform data, and accurately captures user immediate interest shifts; the application fills in cross-platform behavior breakpoints by using a bidirectional LSTM and an attention mechanism, and dynamically adjusts prediction weights according to device types, thereby significantly improving the response speed of sudden interest; the application connects multi-platform content resource pools by using semantic association indexes and a hot spot adjustment matrix, realizes semantic-level matching and real-time scheduling of cross-platform content; and the application completes multi-platform cooperation under the premise of ensuring data security by using a privacy compliance unit, thereby avoiding legal risks.

[0082] The application effectively solves the recommendation lag, cross-platform resource fragmentation and privacy compliance obstacles caused by data silos in traditional recommendation systems, and realizes closed-loop optimization of accurate perception, rapid prediction and collaborative push. BRIEF DESCRIPTION OF DRAWINGS

[0083] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0084] Figure 1 The figure is a schematic diagram of the framework of a new media operation content recommendation system based on big data in embodiment 1. DETAILED DESCRIPTION

[0085] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings of the specification.

[0086] In the following description, many specific details are set forth in order to provide a thorough understanding of the application, but the application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the application, therefore the application is not limited to the specific embodiments disclosed below.

[0087] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the application. In this specification, "in one embodiment" does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0088] Embodiment 1, reference Figure 1 The embodiment provides a new media operation content recommendation system based on big data, which comprises:

[0089] The edge behavior perception module is deployed on the client or edge computing node of the plurality of new media platforms, and is used for collecting short-time high-frequency behavior data of users across platforms in real time and generating a time sequence behavior vector;

[0090] The edge behavior perception module specifically comprises:

[0091] The lightweight behavior analysis unit is embedded in each platform client or CDN edge server, and is used for capturing user sliding frequency, click event sequence, page dwell time and reading speed threshold breakthrough behavior in real time;

[0092] A timestamp marking unit generates a timestamped discrete event stream from the collected raw behavior data according to the occurrence time;

[0093] A vector encoding unit converts the event stream into a multi-dimensional behavior vector within a fixed time window and appends a platform identifier;

[0094] Short-term high-frequency behavior data refers to user operations triggered continuously within a preset time window (such as 30 seconds to 5 minutes), including behaviors such as a sliding speed exceeding a threshold (such as a sliding frequency ≥ 2 times per second), a single-page stay duration below a set lower limit (such as ≤ 3 seconds), or a continuous click interval shorter than a set duration (such as ≤ 500 ms);

[0095] A time-series behavior vector is generated through a sliding time window mechanism, with each window containing normalized behavior indicators within a preset time period (such as 1 minute), including sliding acceleration mean, click event density, page switching frequency, and reading speed fluctuation variance, and spliced into a multi-dimensional tensor in timestamp order;

[0096] A federal aggregation module connects local modeling units of various platforms and performs weighted fusion on user interest features output by multiple platforms through a federal learning mechanism to generate a cross-platform user interest map;

[0097] The federal aggregation module performs the following operations:

[0098] A convolutional neural network is used in the local modeling unit to extract time-local features of user behavior within the platform;

[0099] Model parameters are added with noise through a differential privacy mechanism and then transmitted to the central server;

[0100] A dynamic weight allocation strategy is used for parameter aggregation, with weight values dynamically adjusted according to the freshness of real-time data and user activity levels of each platform;

[0101] In the dynamic weight allocation strategy, the data freshness weight is inversely proportional to the difference between the platform data update timestamp and the current time, and the user activity level weight is dynamically adjusted based on the real-time operation frequency of users on the current platform (such as the number of clicks per unit time);

[0102] In the federal aggregation module, a dynamic weight allocation strategy is used for parameter aggregation, including:

[0103] In federal aggregation, the parameter weight is dynamically allocated according to the data freshness and user activity level of each platform, and the parameter weight of the i-th platform is calculated as:

[0104] ,

[0105] wherein, represents the difference between the current time and the platform​ The difference in the time of the most recent data update This represents the time decay coefficient, ranging from 0.01 to 1.0.

[0106] Then measure user activity:

[0107] ,in, Indicates the platform within the preset time window. The total number of operation events on the device. This represents the total number of operation events across all platforms within the same time window. Indicates the total number of platforms;

[0108] Introducing the balance coefficient value range Construct a comprehensive score:

[0109] ,

[0110] in, The importance of controlling the ratio between freshness and activity. This is the overall score before integration;

[0111] The ratings from all platforms are normalized to obtain the final weights:

[0112] ,in, This represents the sum of the overall ratings across all platforms. That is, the first The platform dynamically allocates weights during federated aggregation;

[0113] Specifically, this strategy quantifies the timeliness of data across different platforms using an exponential decay model, giving higher weight to recently updated data to reflect the advantage of content freshness. Simultaneously, it measures activity using event frequency, reflecting actual user engagement across platforms, with a balanced coefficient. It provides a flexible way to adjust the relative influence of freshness and activity, making it easy to optimize based on business needs or offline experiment results;

[0114] The interest transition prediction module constructs a time-series prediction model based on the interest graph and outputs the prediction results of the user's interest transition direction in the future preset time period.

[0115] The interest transition prediction module includes:

[0116] A bidirectional LSTM network is used to process the temporal correlation features of cross-platform behavioral sequences.

[0117] Attention mechanism layer, identifies the contribution weight of behavioral data from different platforms to the current state of interest;

[0118] The output layer generates a multi-dimensional prediction vector containing interest categories, intensity values and duration;

[0119] The bidirectional LSTM network introduces an inter-platform gap filling mechanism: when the user switches platforms, a virtual filling vector is generated based on historical behavior patterns;

[0120] The attention mechanism layer dynamically adjusts the weight distribution strategy according to the user activity coefficient and the user device type (such as mobile / PC), giving higher attention weight to the burst behavior data generated by the mobile platform;

[0121] In the process of constructing the time series prediction model based on the interest graph, the user's cross-platform interest node sequence is mapped to the time series input, capturing the dynamic evolution of interest state over time, including:

[0122] Let the user's interest node sequence in the preset period be , which is mapped to the vector sequence by knowledge graph embedding technology, denotes the length of the interest sequence;

[0123] On this basis, the bidirectional long short-term memory network Bi-LSTM is used to extract time series correlation features, described as:

[0124] ,

[0125] ,

[0126] ,

[0127] where, denotes the interest vector at time , and are the forward and backward LSTM units respectively, and are the corresponding hidden states, and the bidirectional hidden state is obtained by concatenating them;

[0128] To fill in the blank behavior segment of the user when switching between platforms, define the switching interval and the threshold , when , generate a virtual filling vector:

[0129] ,

[0130] where, denotes the historical hidden state average of the user when switching from platform to , Platform index representing the current behavior occurrence, is a linear mapping function, is a mapping weight parameter, is a padding vector inserted into the sequence, which is inserted into the corresponding position to ensure the continuity of the sequence;

[0131] Attention mechanism is introduced to dynamically adjust the contribution weight at different times according to the device type and user activity. First, the device type is embedded as , the user activity is embedded as , the context vector is constructed, and the attention raw score is calculated:

[0132] ,

[0133] wherein, and are hidden state and context mapping matrices respectively, is a bias vector, is an attention projection vector, is the score at time , and the attention weight is obtained by normalization:

[0134] ,

[0135] wherein, is the normalized weight at time ;

[0136] The weighted hidden state is aggregated into the final context vector:

[0137] ,

[0138] and the probability distribution of the future interest transition direction is generated by the output layer:

[0139] ,

[0140] wherein, and are output mapping matrix and bias respectively, is a multi-dimensional probability vector of interest transition direction;

[0141] Specifically, the bidirectional time sequence features of the interest sequence are comprehensively captured by the Bi-LSTM, the filling mechanism during cross-platform switching is combined, information loss is avoided, and the robustness of the model to breakpoint behavior is improved. The attention layer dynamically allocates weights according to the device type and user activity, so that the model can respond more agilely to sudden interest changes on mobile terminals or high-activity platforms, improve the recognition ability of short-term interest changes, and finally the softmax output converts the interest transition direction into a probability distribution, which is convenient for subsequent recommendation strategies to make threshold decisions or sorting;

[0142] The cross-platform scheduling module matches the multi-platform content resource pool according to the prediction result, and pushes the adaptive content to the target platform through the preset interface and configures the recommendation position trigger logic;

[0143] The cross-platform scheduling module comprises:

[0144] The content metadata mapping unit establishes a semantic association index of multi-platform content resources;

[0145] The push strategy engine matches the content recommendation position specifications of the target platform and the user device type according to the prediction result;

[0146] The preloading interface caches the recommended content to the edge node of the target platform through the API protocol between platforms;

[0147] The semantic association index is realized by knowledge graph embedding technology, which maps the theme tags of multi-platform content to a unified vector space, and establishes a cross-platform content association relationship based on cosine similarity;

[0148] The content metadata mapping unit performs:

[0149] Extract the theme tags, sentiment orientation and knowledge graph entities of each platform content resource;

[0150] Construct a cross-platform semantic similarity matrix and establish a conversion rule library between different content formats;

[0151] Dynamically update the content association strength coefficient of hot events;

[0152] The semantic similarity matrix is generated by comprehensive calculation of the co-occurrence frequency of title keywords, user comment sentiment polarity consistency and knowledge graph entity association degree of cross-platform content;

[0153] The content metadata mapping unit constructs the cross-platform semantic similarity matrix in the following manner:

[0154] Let the cross-platform content set be , denote the total number of contents in the cross-platform content set, extract the knowledge graph embedding vector , theme tag set , sentiment score with entity set ;

[0155] On this basis, define the initial semantic similarity matrix :

[0156] ,

[0157] wherein, is a normalized weight coefficient, , is the vector cosine similarity, is the label Jaccard similarity, is the sentiment polarity difference function, is the entity overlap degree;

[0158] Introduce the hot spot adjustment matrix H:

[0159] ,

[0160] wherein, is a hot spot fusion coefficient, , indicates the co-occurrence heat of content and in the current hot spot event, is the maximum value of all content pair heat, traverse the index set of all content pairs , indicates the normalized hot spot co-occurrence intensity, the value range is [0, 1];

[0161] Specifically, this place comprehensively utilizes four types of features such as text embedding, label association, sentiment difference and entity co-occurrence to comprehensively measure the semantic similarity of cross-platform content, so that similar theme resources can be effectively connected in a unified matrix. After introducing the dynamic hot spot adjustment, the association strength of the content related to the current hot spot event can be timely improved, the sensitivity of the recommendation system to real-time events is enhanced, the multi-weight coefficient design facilitates offline or online tuning for different business needs, and the matrix calculation mainly depends on vector operation and set intersection and union, the calculation complexity and content size grow approximately square, which can be realized by block or approximate index technology Real-time update;

[0162] The new media operation content recommendation system also includes:

[0163] A privacy compliance verification unit performs homomorphic encryption processing on model parameters before federated aggregation;

[0164] A behavior data life cycle management unit sets an automatic erasure mechanism for short-time behavior data;

[0165] The cross-platform identification desensitization unit converts the user equipment fingerprint information using an irreversible hash algorithm.

[0166] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application, and all modifications and equivalent replacements should be covered in the scope of the claims of the present application.

Claims

1. A new media operation content recommendation system based on big data, characterized in that, include, The edge behavior perception module is deployed on the client or edge computing node of multiple new media platforms to collect short-term high-frequency behavior data of users across platforms in real time and generate time-series behavior vectors. The short-term high-frequency behavior data refers to the operation events continuously triggered by users within a preset time window, including behaviors such as scrolling speed exceeding a threshold, single-page dwell time being lower than a set lower limit, or continuous click interval being shorter than a set duration. The federated aggregation module connects the local modeling units of each platform and uses a federated learning mechanism to weightedly fuse user interest features output from multiple platforms to generate a cross-platform user interest graph. The interest transition prediction module constructs a time-series prediction model based on the interest graph and outputs the prediction results of the user's interest transition direction in a future preset time period. The cross-platform scheduling module matches content resource pools from multiple platforms based on prediction results, pushes adapted content to the target platform through a pre-defined interface, and configures recommendation trigger logic. The federated aggregation module performs the following operations: Convolutional neural networks are used in the local modeling unit to extract temporal local features of user behavior within the platform; Noise is added to the model parameters using a differential privacy mechanism before transmitting them to the central server. A dynamic weighting strategy is used for parameter aggregation, and the weight values ​​are dynamically adjusted based on the real-time data freshness and user activity of each platform. In the dynamic weight allocation strategy, the data freshness weight is calculated inversely proportional to the difference between the platform data update timestamp and the current time, and the user activity weight is dynamically adjusted based on the user's real-time operation frequency on the current platform.

2. The new media operation content recommendation system based on big data as described in claim 1, characterized in that, The edge behavior perception module specifically includes: A lightweight behavior analysis unit that can be embedded in clients of various platforms or CDN edge servers to capture user swipe frequency, click event sequences, page dwell time, and reading speed threshold exceeding behaviors in real time; The timestamp marking unit generates a discrete event stream with timestamps based on the time of occurrence of the collected raw behavioral data; The vector encoding unit transforms the event stream into a multi-dimensional behavior vector within a fixed time window and appends a platform identifier; The temporal behavior vector is generated through a sliding time window mechanism. Each window contains normalized behavior indicators within a preset time period, including the mean sliding acceleration, click event density, page switching frequency, and reading speed fluctuation variance, which are then concatenated into a multidimensional tensor in the order of timestamps.

3. The new media operation content recommendation system based on big data as described in claim 1, characterized in that, The federated aggregation module employs a dynamic weight allocation strategy for parameter aggregation, including: In federated aggregation, parameter weights are dynamically assigned based on the data freshness and user activity of each platform, for the first... Each platform calculates its data freshness: ， in, Indicates the current time and platform The difference in the time of the most recent data update This represents the time decay coefficient, ranging from 0.01 to 1.

0. Then measure user activity: ,in, Indicates the platform within the preset time window. The total number of operation events on the device. This represents the total number of operation events across all platforms within the same time window. Indicates the total number of platforms; Introducing the balance coefficient value range Construct a comprehensive score: ， in, The importance of controlling the ratio between freshness and activity. This is the overall score before integration; The ratings from all platforms are normalized to obtain the final weights: ,in, This represents the sum of the overall ratings across all platforms. That is, the first The platform dynamically allocates weights during federated aggregation.

4. The new media operation content recommendation system based on big data as described in claim 1, characterized in that, The interest transition prediction module includes: A bidirectional LSTM network is used to process the temporal correlation features of cross-platform behavioral sequences. Attention mechanism layer, identifies the contribution weight of behavioral data from different platforms to the current state of interest; The output layer generates a multidimensional prediction vector that includes interest category, intensity value, and duration. The bidirectional LSTM network introduces a platform gap filling mechanism: when a user switches platforms, a virtual filling vector is generated based on historical behavior patterns; The attention mechanism layer dynamically adjusts the weight allocation strategy based on the user activity coefficient and the user device type, assigning higher attention weights to sudden behavioral data generated by the mobile platform.

5. The new media operation content recommendation system based on big data as described in claim 4, characterized in that, In the process of building a time-series prediction model based on interest graphs, the user's cross-platform interest node sequence is mapped as a time-series input to capture the dynamic evolution of interest state over time, including: Let the sequence of user interest nodes within a preset time period be: It is mapped into a vector sequence through knowledge graph embedding technology. , Indicates the length of the interest sequence; Based on this, a bidirectional long short-term memory network (Bi-LSTM) is used to extract temporal correlation features, described as follows: ， ， ， in, Indicates time Interest vector, and These are forward and backward LSTM units, respectively. and These are the corresponding hidden states, which are concatenated to obtain a bidirectional hidden state. ; To fill the gaps in user behavior when switching between platforms, a switching interval is defined. With threshold ,when At that time, a virtual padding vector is generated: ， in, Indicates that the user is on the platform Switch to The historical hidden state mean, The platform index indicating where the current action occurred. It is a linear mapping function. For mapping weight parameters, The padding vector is inserted into the sequence at the appropriate position to ensure the continuity of the sequence. An attention mechanism is introduced to dynamically adjust the contribution weight at different times based on device type and user activity. First, the device type is embedded and represented as... User activity is embedded as a representation Construct context vectors And calculate the raw attention score: ， in, and These are the hidden state and context mapping matrices, respectively. For bias vectors, For attention projection vectors, For a moment The score is then normalized to obtain the attention weights: ， in, That is, at any moment Normalized weights; The weighted hidden states are aggregated into the final context vector: ， The output layer then generates a probability distribution for future interest transition directions. ， in, and These are the output mapping matrix and the bias, respectively. It is a multidimensional probability vector representing the direction of interest transfer.

6. The new media operation content recommendation system based on big data as described in claim 1, characterized in that, The cross-platform scheduling module includes: Content metadata mapping unit, establishing a semantic association index of content resources across multiple platforms; The push strategy engine matches the content recommendation slot specifications of the target platform and the user's device type based on the prediction results. The pre-loading interface caches recommended content to the edge nodes of the target platform via inter-platform API protocols. The semantic association index is implemented through knowledge graph embedding technology, which maps the topic tags of content from multiple platforms to a unified vector space and establishes cross-platform content association relationships based on cosine similarity.

7. A new media operation content recommendation system based on big data as described in claim 6, characterized in that, The content metadata mapping unit performs the following: Extract the topic tags, sentiment indicators, and knowledge graph entities from the content resources of each platform; Construct a cross-platform semantic similarity matrix and establish a conversion rule base between different content formats; Dynamically update the content relevance strength coefficient of trending events; The semantic similarity matrix is ​​generated by comprehensively calculating the co-occurrence frequency of title keywords in cross-platform content, the consistency of sentiment polarity in user comments, and the entity association degree of the knowledge graph.

8. The new media operation content recommendation system based on big data as described in claim 7, characterized in that, The content metadata mapping unit constructs a cross-platform semantic similarity matrix in the following way: Let the cross-platform content collection be... , This represents the total number of contents in the cross-platform content collection, and the knowledge graph embedding vector is extracted for each piece of content. Collection of topic tags Emotional score With entity set ; Based on this, define the initial semantic similarity matrix. : ， in, These are the normalized weighting coefficients. , For vector cosine similarity, For the similarity of the label Jaccard, For the difference in emotional polarity, For the degree of overlap between entities; Introducing a hotspot adjustment matrix H: ， in, The hotspot fusion coefficient, , Content and The co-occurrence of popularity in current trending events For all content, the maximum popularity. Iterate through all content pairs The set of indices This represents the normalized hotspot co-occurrence intensity, with a value range of [0,1].

9. A new media operation content recommendation system based on big data as described in claim 1, characterized in that, Also includes: The privacy compliance verification unit performs homomorphic encryption on the model parameters before federated aggregation; The behavioral data lifecycle management unit sets up an automatic erasure mechanism for short-term behavioral data. The cross-platform identifier desensitization unit uses an irreversible hash algorithm to convert user device fingerprint information.

Images