New media operation content recommendation system based on big data

Abstract

The invention discloses a new media operation content recommendation system based on big data, and relates to the technical field of new media operation, cross-platform short-time behaviors are collected in real time through edge calculation, multi-platform user interest characteristics are dynamically fused in combination with federal learning, the limitation of single platform data is broken through, and instant interest transfer of users is accurately captured; cross-platform behavior breakpoints are filled by using bidirectional LSTM and an attention mechanism, and prediction weights are dynamically adjusted according to equipment types, so that the response speed of sudden interests is remarkably improved; a multi-platform content resource pool is opened through a semantic association index and a hotspot adjustment matrix, and semantic-level matching and real-time scheduling of cross-platform content are achieved; in addition, multi-platform collaboration is completed through the privacy compliance unit on the premise of guaranteeing data security, and legal risks are avoided.

Description

Technical Field

[0001] The present invention relates to the field of new media operation technology, and in particular to a new media operation content recommendation system based on big data. Background Art

[0002] Current content recommendation systems face the dual bottlenecks of real-time interest capture across platforms and coordinated scheduling across multiple terminals. User behavior trajectories have shifted from a single platform to fragmented switching across multiple scenarios such as Weibo, short videos, and information apps, but mainstream recommendation algorithms still use a static portrait mechanism and rely on historical data from a single platform to build long-term interest models. This model has two fatal flaws: First, it is unable to capture the instantaneous interest shifts generated by users in the process of switching between multiple terminals. For example, the immediate attention behavior triggered by breaking news is often missed after the algorithm update cycle. Second, the existing system relies too much on explicit behavioral signals such as clicks and favorites, and lacks the ability to analyze microsecond-level implicit behaviors such as sliding speed and reading pauses, resulting in delays of up to several hours in responding to hot events. Even if some platforms try to introduce real-time computing frameworks, they are still limited by the single-point data dimension and find it difficult to build a complete instantaneous interest map.

[0003] The data silos and strategic fragmentation between multiple platforms further exacerbate the recommendation misalignment problem; when a user develops a new interest on platform A, the recommendation engine of platform B / C still uses outdated portraits for content matching due to the lack of a cross-domain information sharing mechanism; although existing technologies attempt to achieve cross-platform recommendations through user IDs, they face two major practical obstacles: first, privacy compliance restrictions prevent the original behavioral data from flowing across platforms; second, the differences in metadata standards of the content resource libraries of each platform make it difficult to quickly dispatch related content of hot topics; this fragmented state directly leads to the failure of two typical scenarios: after a user triggers an interest on a short video platform, the information platform cannot immediately push in-depth interpretations; when used across devices, the behavioral data generated on the mobile phone cannot modify the PC recommendation strategy in real time; existing technical solutions generally fall into a dilemma when balancing real-time and collaboration. Summary of the Invention

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

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

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

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

[0008] The edge behavior perception module is deployed on the clients or edge computing nodes of multiple new media platforms to collect users' short-term and high-frequency cross-platform behavior data in real time and generate time-series behavior vectors.

[0009] The federated aggregation module connects the local modeling units of each platform and uses the federated learning mechanism to perform weighted fusion of user interest features output by multiple platforms to generate a cross-platform user interest graph;

[0010] An interest transition prediction module builds a time series prediction model based on the interest graph and outputs a prediction result of the user's interest transition direction in a preset time period in the future;

[0011] The cross-platform scheduling module matches the multi-platform content resource pool based on the prediction results, pushes adapted content to the target platform through the preset interface, and configures the recommendation position trigger logic.

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

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

[0014] The timestamp marking unit generates a discrete event stream with a timestamp according to the time of occurrence of the collected original behavior data;

[0015] Vector encoding unit, which converts the event stream into a multi-dimensional behavior vector within a fixed time window and appends a platform identifier;

[0016] The short-term high-frequency behavior data refers to the operation events continuously triggered by the user within a preset time window, including the behavior of the sliding rate exceeding the threshold, the single page stay time being lower than the set lower limit, or the continuous click interval being shorter than the set time.

[0017] The temporal behavior vector is generated through a sliding time window mechanism. Each window contains normalized behavioral indicators within a preset time period, including the mean sliding acceleration, click event density, page switching frequency, and reading speed fluctuation variance, and is spliced ​​into a multidimensional tensor in timestamp order.

[0018] As a preferred solution of the new media operation content recommendation system based on big data described in the present invention, the federation aggregation module performs the following operations:

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

[0020] The model parameters are added with noise through the differential privacy mechanism and then transmitted to the central server;

[0021] A dynamic weight allocation strategy is used for parameter aggregation, and the weight value is 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 in inverse proportion 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 solution of the new media operation content recommendation system based on big data described in the present invention, wherein: in the federation aggregation module, a dynamic weight allocation strategy is used to perform 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 i-th platform, its data freshness is calculated as follows:

[0025]

[0026] Where, Δt i It represents the difference between the current time and the most recent data update time of platform i, and γ represents the time decay coefficient, which ranges from 0.01 to 1.0;

[0027] Then measure platform activity:

[0028] Among them, n i represents the total number of operation events on platform i within the preset time window, represents the total number of operation events on all platforms in the same time window, and N represents the total number of platforms;

[0029] The balance coefficient λ is introduced with a value range of 0≤λ≤1 to construct a comprehensive score:

[0030] s i =λd i +(1-λ)a i ,

[0031] Among them, λ controls the importance ratio of freshness to activity, s i is the comprehensive score before fusion;

[0032] Normalize the scores of all platforms to get the final weight:

[0033] in, represents the sum of the comprehensive scores of all platforms, w i That is, the dynamic allocation weight of the i-th platform during federation aggregation.

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

[0035] Bidirectional LSTM network to process temporal correlation features of cross-platform behavior sequences;

[0036] The attention mechanism layer identifies the contribution weight of different platform behavior data to the current interest state;

[0037] The output layer generates a multidimensional prediction vector containing interest categories, intensity values, and durations;

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

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

[0040] As a preferred solution of the new media operation content recommendation system based on big data described in the present invention, in the process of building a time series prediction model based on the interest graph, the user's cross-platform interest node sequence is mapped into a time series input to capture the dynamic evolution of interest status over time, including:

[0041] Assume that the sequence of nodes of interest to the user in the preset period is {v1,v2,…,v T}, and mapped it into a vector sequence {x1,x2,…,x T}, T represents the length of the sequence of interest;

[0042] On this basis, the bidirectional long short-term memory network Bi-LSTM is used to extract temporal correlation features, which can be described as follows:

[0043]

[0044] Among them, x t Represents the interest vector at time t, LSTM fw With LSTM bw are forward and backward LSTM units, respectively. and They correspond to hidden states respectively, and after splicing, the bidirectional hidden state h is obtained t ;

[0045] In order to fill the blank behavior segments when users switch across platforms, the switching interval Δτ and the threshold τ are defined. When Δτ>τ, a virtual filling vector is generated:

[0046]

[0047] in, represents the historical hidden state mean of the user switching from platform p to q, q represents the platform index where the current behavior occurs, φ is the linear mapping function, β is the mapping weight parameter, and g k To insert a filling vector into the sequence, insert the vector into the corresponding position to ensure the continuity of the sequence;

[0048] The attention mechanism is introduced to dynamically adjust the contribution weights at different times according to the device type and platform activity. The device type is first embedded as e d , embedding the platform activity into e p , construct the context vector c = [e d ;e p ], and calculate the raw attention score:

[0049]

[0050] Among them, W h With W c are the hidden state and context mapping matrices, b is the bias vector, v is the attention projection vector, e t is the score at time t, and then the attention weight is obtained by normalization:

[0051]

[0052] Among them, α t That is, the normalized weight at time t;

[0053] Aggregate the weighted hidden states into the final context vector:

[0054]

[0055] The output layer generates the probability distribution of future interest transition directions:

[0056] y=softmax(W o c * +b o ),

[0057] Among them, W o and b o are the output mapping matrix and bias respectively, and y is the multidimensional probability vector of the interest transfer direction.

[0058] As a preferred solution of the new media operation content recommendation system based on big data described in the present invention, the cross-platform scheduling module includes:

[0059] Content metadata mapping unit, which establishes semantic association indexes of multi-platform content resources;

[0060] Push strategy engine, matching the target platform's content recommendation specifications and user device type based on prediction results;

[0061] Preloading interface, caching recommended content to the edge node of the target platform through the inter-platform API protocol;

[0062] The semantic association index is implemented through knowledge graph embedding technology, mapping the topic tags of multi-platform content to a unified vector space, and establishing cross-platform content association relationships based on cosine similarity.

[0063] As a preferred solution of the new media operation content recommendation system based on big data described in the present invention, the content metadata mapping unit performs:

[0064] Extract topic tags, sentiment tendencies, and knowledge graph entities from content resources on each platform;

[0065] Build a cross-platform semantic similarity matrix and establish a conversion rule library between different content formats;

[0066] Dynamically update the content relevance strength coefficient of hot events;

[0067] The semantic similarity matrix is ​​generated by comprehensively calculating the co-occurrence frequency of keywords in titles of cross-platform content, the consistency of sentiment polarity of user comments, and the relevance of knowledge graph entities.

[0068] As a preferred solution of the new media operation content recommendation system based on big data described in the present invention, the content metadata mapping unit constructs a cross-platform semantic similarity matrix in the following manner:

[0069] Assume that the cross-platform content set is {c1,c2,…,c M}, M represents the total number of contents in the cross-platform content collection, and the knowledge graph embedding vector {e i}、Topic tag collection{T i}, sentiment score {σ i} and entity collection

[0070] On this basis, the initial semantic similarity matrix S∈R is defined M×M :

[0071]

[0072] Among them, α c ,α t ,α s ,α e is the normalized weight coefficient, α c ,α t ,αs ,α e ≥0, and α c +α t +α s + is the vector cosine similarity, is the label Jaccard similarity, exp(-|σ i -σ j |) is the sentiment polarity difference function, is the entity overlap;

[0073] Introduce the hotspot adjustment matrix H:

[0074] Among them, ρ is the hotspot fusion coefficient, ρ∈[0,1], δ ij Indicates content c i with c j Co-occurrence heat in current hot events, For the maximum value of heat of all content pairs, p,q traverses all content pairs (c p ,c q ), H ij Indicates the normalized hotspot co-occurrence intensity, with a value range of [0,1].

[0075] As a preferred solution of the new media operation content recommendation system based on big data described in the present invention, it also includes:

[0076] Privacy compliance verification unit, which performs homomorphic encryption on model parameters before federation aggregation;

[0077] Behavioral data lifecycle management unit, which sets the automatic erasure mechanism for short-term behavioral data;

[0078] The cross-platform identification desensitization unit uses an irreversible hash algorithm to convert user device fingerprint information.

[0079] The beneficial effects of the present invention are as follows: the present invention uses edge computing to collect short-term cross-platform behaviors in real time, combines federated learning to dynamically integrate multi-platform user interest characteristics, breaks through the data limitations of a single platform, and accurately captures users' immediate interest shifts; uses bidirectional LSTM and attention mechanisms to fill cross-platform behavior breakpoints, and dynamically adjusts prediction weights based on device types, significantly improving the response speed of sudden interests; through semantic association indexing and hotspot adjustment matrices, it connects multi-platform content resource pools and achieves semantic-level matching and real-time scheduling of cross-platform content; and through a privacy compliance unit, it completes multi-platform collaboration while ensuring data security, avoiding legal risks.

[0080] This invention effectively solves the problems of recommendation lag, cross-platform resource fragmentation and privacy compliance barriers caused by data silos in traditional recommendation systems, and achieves closed-loop optimization of accurate perception, rapid prediction and collaborative push. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0082] Figure 1 This is a schematic diagram of the framework of a new media operation content recommendation system based on big data in Example 1. DETAILED DESCRIPTION

[0083] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0084] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0085] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0086] Example 1, reference Figure 1 This embodiment provides a new media operation content recommendation system based on big data, including:

[0087] The edge behavior perception module is deployed on the clients or edge computing nodes of multiple new media platforms to collect users' short-term and high-frequency cross-platform behavior data in real time and generate time-series behavior vectors.

[0088] The edge behavior perception module specifically includes:

[0089] A lightweight behavior analysis unit, embedded in each platform client or CDN edge server, captures user scrolling frequency, click event sequence, page dwell time, and reading speed threshold breakthrough behavior in real time;

[0090] The timestamp marking unit generates a discrete event stream with a timestamp according to the time of occurrence of the collected original behavior data;

[0091] Vector encoding unit, which converts the event stream into a multi-dimensional behavior vector within a fixed time window and appends a platform identifier;

[0092] Short-term, high-frequency behavior data refers to user-triggered action events that occur continuously within a preset time window (e.g., 30 seconds to 5 minutes). This includes actions where the sliding rate exceeds a threshold (e.g., ≥2 slides per second), the duration of a single page stay falls below a set lower limit (e.g., ≤3 seconds), or the interval between consecutive clicks is shorter than a set time (e.g., ≤500ms).

[0093] The time series behavior vector is generated through a sliding time window mechanism. Each window contains normalized behavior indicators within a preset time period (e.g., 1 minute), including the mean sliding acceleration, click event density, page switching frequency, and reading speed fluctuation variance. These indicators are then concatenated into a multidimensional tensor in timestamp order.

[0094] The federated aggregation module connects the local modeling units of each platform and uses the federated learning mechanism to perform weighted fusion of user interest features output by multiple platforms to generate a cross-platform user interest graph;

[0095] The federated aggregation module performs the following operations:

[0096] A convolutional neural network is used in the local modeling unit to extract the temporal local features of user behavior within the platform;

[0097] The model parameters are added with noise through the differential privacy mechanism and then transmitted to the central server;

[0098] A dynamic weight allocation strategy is used for parameter aggregation, and the weight value is dynamically adjusted based on the real-time data freshness and user activity of each platform;

[0099] In the dynamic weight allocation strategy, the data freshness weight is calculated in inverse proportion 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 (such as the number of clicks per unit time);

[0100] In the federated aggregation module, a dynamic weight distribution strategy is used for parameter aggregation, including:

[0101] In federated aggregation, parameter weights are dynamically assigned based on the data freshness and user activity of each platform. For the i-th platform, its data freshness is calculated as follows:

[0102]

[0103] Where, Δt iIt represents the difference between the current time and the most recent data update time of platform i, and γ represents the time decay coefficient, which ranges from 0.01 to 1.0;

[0104] Then measure platform activity:

[0105] Among them, n i represents the total number of operation events on platform i within the preset time window, represents the total number of operation events on all platforms in the same time window, and N represents the total number of platforms;

[0106] The balance coefficient λ is introduced with a value range of 0≤λ≤1 to construct a comprehensive score:

[0107] s i =λd i +(1-λ)a i ,

[0108] Among them, λ controls the importance ratio of freshness to activity, s i is the comprehensive score before fusion;

[0109] Normalize the scores of all platforms to get the final weight:

[0110] in, represents the sum of the comprehensive scores of all platforms, w i That is, the dynamic allocation weight of the i-th platform during federation aggregation;

[0111] Specifically, this strategy quantifies the timeliness of data from different platforms through an exponential decay model, giving more weight to recently updated data, reflecting the freshness of the content. It also measures activity by event frequency, reflecting actual user engagement on each platform. The balance coefficient λ provides a flexible way to adjust the relative influence of freshness and activity, facilitating optimization based on business needs or offline experimental results.

[0112] The interest transition prediction module builds a time series prediction model based on the interest graph and outputs the prediction results of the user's interest transfer direction in the future preset period;

[0113] The interest transition prediction module includes:

[0114] Bidirectional LSTM network to process temporal correlation features of cross-platform behavior sequences;

[0115] The attention mechanism layer identifies the contribution weight of different platform behavior data to the current interest state;

[0116] The output layer generates a multidimensional prediction vector containing interest categories, intensity values, and durations;

[0117] The bidirectional LSTM network introduces a gap-filling mechanism between platforms: when users switch platforms, a virtual filling vector is generated based on historical behavior patterns;

[0118] The attention mechanism layer dynamically adjusts the weight distribution strategy based on the platform activity coefficient and the user device type (e.g., mobile / PC), giving higher attention weight to the sudden behavior data generated by the mobile platform;

[0119] In the process of building a time series prediction model based on the interest graph, the user's cross-platform interest node sequence is mapped into a time series input to capture the dynamic evolution of interest status over time, including:

[0120] Assume that the sequence of nodes of interest to the user in the preset period is {v1,v2,…,v T}, and mapped it into a vector sequence {x1,x2,…,x T}, T represents the length of the sequence of interest;

[0121] On this basis, the bidirectional long short-term memory network Bi-LSTM is used to extract temporal correlation features, which can be described as follows:

[0122]

[0123] Among them, x t Represents the interest vector at time t, LSTM fw With LSTM bw are forward and backward LSTM units, respectively. and They correspond to hidden states respectively, and after splicing, the bidirectional hidden state h is obtained t ;

[0124] In order to fill the blank behavior segments when users switch across platforms, the switching interval Δτ and the threshold τ are defined. When Δτ>τ, a virtual filling vector is generated:

[0125]

[0126] in, represents the historical hidden state mean of the user switching from platform p to q, q represents the platform index where the current behavior occurs, φ is the linear mapping function, β is the mapping weight parameter, and g k To insert a filling vector into the sequence, insert the vector into the corresponding position to ensure the continuity of the sequence;

[0127] The attention mechanism is introduced to dynamically adjust the contribution weights at different times according to the device type and platform activity. The device type is first embedded as e d , embedding the platform activity into e p , construct the context vector c = [ed ;e p ], and calculate the raw attention score:

[0128]

[0129] Among them, W h With W c are the hidden state and context mapping matrices, b is the bias vector, v is the attention projection vector, e t is the score at time t, and then the attention weight is obtained by normalization:

[0130]

[0131] Among them, α t That is, the normalized weight at time t;

[0132] Aggregate the weighted hidden states into the final context vector:

[0133]

[0134] The output layer generates the probability distribution of future interest transition directions:

[0135] y=softmax(W o c * +b o ),

[0136] Among them, W o and b o are the output mapping matrix and bias respectively, and y is the multidimensional probability vector of the direction of interest transfer;

[0137] Specifically, the Bi-LSTM method comprehensively captures the bidirectional temporal features of interest sequences. Combined with a fill-in mechanism during cross-platform switching, it avoids information loss and improves the model's robustness to breakpoint behavior. The attention layer dynamically assigns weights based on device type and platform activity, enabling the model to more agilely respond to sudden changes in interest on mobile devices or highly active platforms, improving its ability to identify short-term interest shifts. The final softmax output converts the direction of interest transitions into a probability distribution, facilitating threshold decisions or ranking for subsequent recommendation strategies.

[0138] The cross-platform scheduling module matches the multi-platform content resource pool based on the prediction results, pushes the adapted content to the target platform through the preset interface, and configures the recommendation trigger logic;

[0139] The cross-platform scheduling module includes:

[0140] Content metadata mapping unit, which establishes semantic association indexes of multi-platform content resources;

[0141] Push strategy engine, matching the target platform's content recommendation specifications and user device type based on prediction results;

[0142] Preloading interface, caching recommended content to the edge node of the target platform through the inter-platform API protocol;

[0143] Semantic association indexing is achieved through knowledge graph embedding technology, which maps the topic tags of multi-platform content into a unified vector space and establishes cross-platform content association relationships based on cosine similarity;

[0144] The content metadata mapping unit performs:

[0145] Extract topic tags, sentiment tendencies, and knowledge graph entities from content resources on each platform;

[0146] Build a cross-platform semantic similarity matrix and establish a conversion rule library between different content formats;

[0147] Dynamically update the content relevance strength coefficient of hot events;

[0148] The semantic similarity matrix is ​​generated by comprehensively calculating the co-occurrence frequency of keywords in titles of cross-platform content, the consistency of sentiment polarity of user comments, and the relevance of knowledge graph entities;

[0149] The content metadata mapping unit constructs a cross-platform semantic similarity matrix in the following way:

[0150] Assume that the cross-platform content set is {c1,c2,…,c M}, M represents the total number of contents in the cross-platform content collection, and the knowledge graph embedding vector {e i}、Topic tag collection{T i}, sentiment score {σ i} and entity collection

[0151] On this basis, the initial semantic similarity matrix S∈R is defined M×M :

[0152]

[0153] Among them, α c ,α t ,α s ,α e is the normalized weight coefficient, α c ,α t ,α s ,α e ≥0, and α c +α t +α s +α e =1, is the vector cosine similarity, is the label Jaccard similarity, exp(-|σ i -σ j |) is the sentiment polarity difference function, is the entity overlap;

[0154] Introduce the hotspot adjustment matrix H:

[0155]

[0156] Among them, ρ is the hotspot fusion coefficient, ρ∈[0,1], δ ij Indicates content c i with c j Co-occurrence heat in current hot events, For the maximum value of heat of all content pairs, p,q traverses all content pairs (c p ,c q ), H ij Indicates the normalized hotspot co-occurrence intensity, with a value range of [0,1];

[0157] Specifically, this paper comprehensively utilizes four types of features: text embedding, tag association, sentiment difference, and entity co-occurrence, to comprehensively measure the semantic similarity of cross-platform content, so that similar topic resources can be effectively connected in a unified matrix. After the introduction of dynamic hot spot adjustment, the correlation strength of content related to current hot events can be timely improved, and the recommendation system's sensitivity to real-time events can be enhanced. The multi-weight coefficient design facilitates offline or online tuning for different business needs. Matrix calculations mainly rely on vector operations and set intersections. The computational complexity grows approximately quadratically with the content size. Real-time updates can be achieved through block or approximate indexing technology.

[0158] The new media operation content recommendation system also includes:

[0159] Privacy compliance verification unit, which performs homomorphic encryption on model parameters before federation aggregation;

[0160] Behavioral data lifecycle management unit, which sets the automatic erasure mechanism for short-term behavioral data;

[0161] The cross-platform identification desensitization unit uses an irreversible hash algorithm to convert user device fingerprint information.

[0162] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention 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 invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A new media operation content recommendation system based on big data, characterized by: include, The edge behavior perception module is deployed on the clients or edge computing nodes of multiple new media platforms to collect users' short-term and high-frequency cross-platform behavior data in real time and generate time-series behavior vectors. The federated aggregation module connects the local modeling units of each platform and uses the federated learning mechanism to perform weighted fusion of user interest features output by multiple platforms to generate a cross-platform user interest graph; An interest transition prediction module builds a time series prediction model based on the interest graph and outputs a prediction result of the user's interest transition direction in a preset time period in the future; The cross-platform scheduling module matches the multi-platform content resource pool based on the prediction results, pushes adapted content to the target platform through the preset interface, and configures the recommendation position trigger logic.

2. A new media operation content recommendation system based on big data according to claim 1, characterized in that: The edge behavior perception module specifically includes: A lightweight behavior analysis unit, embedded in each platform client or CDN edge server, captures user scrolling frequency, click event sequence, page dwell time, and reading speed threshold breakthrough behavior in real time; The timestamp marking unit generates a discrete event stream with a timestamp according to the time of occurrence of the collected original behavior data; Vector encoding unit, which converts the event stream into a multi-dimensional behavior vector within a fixed time window and appends a platform identifier; The short-term high-frequency behavior data refers to the operation events continuously triggered by the user within a preset time window, including the behavior of the sliding rate exceeding the threshold, the single page stay time being lower than the set lower limit, or the continuous click interval being shorter than the set time. The temporal behavior vector is generated through a sliding time window mechanism. Each window contains normalized behavioral indicators within a preset time period, including the mean sliding acceleration, click event density, page switching frequency, and reading speed fluctuation variance, and is spliced ​​into a multidimensional tensor in timestamp order.

3. The new media operation content recommendation system based on big data according to claim 1, characterized in that: The federation aggregation module performs the following operations: A convolutional neural network is used in the local modeling unit to extract the temporal local features of user behavior within the platform; The model parameters are added with noise through the differential privacy mechanism and then transmitted to the central server; A dynamic weight allocation strategy is used for parameter aggregation, and the weight value is 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 in inverse proportion 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.

4. A new media operation content recommendation system based on big data as claimed in claim 3, characterized in that: In the federated aggregation module, a dynamic weight allocation strategy is used to aggregate parameters, including: In federated aggregation, parameter weights are dynamically assigned based on the data freshness and user activity of each platform. For the i-th platform, its data freshness is calculated as follows: Where Δt i It represents the difference between the current time and the most recent data update time of platform i, and γ represents the time decay coefficient, which ranges from 0.01 to 1.0; Then measure platform activity: Among them, n i represents the total number of operation events on platform i within the preset time window, represents the total number of operation events on all platforms in the same time window, and N represents the total number of platforms; The balance coefficient λ is introduced with a value range of 0≤λ≤1 to construct a comprehensive score: s i =λd i +(1-λ)a i , Among them, λ controls the importance ratio of freshness to activity, s i is the comprehensive score before fusion; Normalize the scores of all platforms to get the final weight: in, represents the sum of the comprehensive scores of all platforms, w i That is, the dynamic allocation weight of the i-th platform during federation aggregation.

5. The new media operation content recommendation system based on big data according to claim 1, characterized in that: The interest transition prediction module includes: Bidirectional LSTM network to process temporal correlation features of cross-platform behavior sequences; The attention mechanism layer identifies the contribution weight of different platform behavior data to the current interest state; The output layer generates a multidimensional prediction vector containing interest categories, intensity values, and durations; The bidirectional LSTM network introduces a gap filling mechanism between platforms: when users switch platforms, a virtual filling vector is generated based on historical behavior patterns; The attention mechanism layer dynamically adjusts the weight allocation strategy according to the platform activity coefficient and the user device type, and assigns higher attention weights to the sudden behavior data generated by the mobile platform.

6. A new media operation content recommendation system based on big data as claimed in claim 5, characterized in that: In the process of building a time series prediction model based on the interest graph, the user's cross-platform interest node sequence is mapped into a time series input to capture the dynamic evolution of interest status over time, including: Assume that the sequence of nodes of interest to the user in the preset period is {v1,v2,…,v T }, and mapped it into a vector sequence {x1,x2,…,x T }, T represents the length of the sequence of interest; On this basis, the bidirectional long short-term memory network Bi-LSTM is used to extract temporal correlation features, which can be described as follows: Among them, x t Represents the interest vector at time t, LSTM fw With LSTM bw are forward and backward LSTM units, respectively. and They correspond to hidden states respectively, and after splicing, the bidirectional hidden state h is obtained t ; In order to fill the blank behavior segments when users switch across platforms, the switching interval Δτ and the threshold τ are defined. When Δτ>τ, a virtual filling vector is generated: in, represents the historical hidden state mean of the user switching from platform p to q, q represents the platform index where the current behavior occurs, φ is the linear mapping function, β is the mapping weight parameter, and g k To insert a filling vector into the sequence, insert the vector into the corresponding position to ensure the continuity of the sequence; The attention mechanism is introduced to dynamically adjust the contribution weights at different times according to the device type and platform activity. The device type is first embedded as e d , embedding the platform activity into e p , construct the context vector c = [e d ;e p ], and calculate the raw attention score: Among them, W h With W c are the hidden state and context mapping matrices, b is the bias vector, v is the attention projection vector, e t is the score at time t, and then the attention weight is obtained by normalization: Among them, α t That is, the normalized weight at time t; Aggregate the weighted hidden states into the final context vector: The output layer generates the probability distribution of future interest transition directions: y =softmax(W o c * +b o ), Among them, W o and b o are the output mapping matrix and bias respectively, and y is the multidimensional probability vector of the interest transfer direction.

7. The new media operation content recommendation system based on big data according to claim 1, characterized in that: The cross-platform scheduling module includes: Content metadata mapping unit, which establishes semantic association indexes of multi-platform content resources; Push strategy engine, matching the target platform's content recommendation specifications and user device type based on prediction results; Preloading interface, caching recommended content to the edge node of the target platform through the inter-platform API protocol; The semantic association index is implemented through knowledge graph embedding technology, mapping the topic tags of multi-platform content to a unified vector space, and establishing cross-platform content association relationships based on cosine similarity.

8. A new media operation content recommendation system based on big data according to claim 7, characterized in that: The content metadata mapping unit performs: Extract topic tags, sentiment tendencies, and knowledge graph entities from content resources on each platform; Build a cross-platform semantic similarity matrix and establish a conversion rule library between different content formats; Dynamically update the content relevance strength coefficient of hot events; The semantic similarity matrix is ​​generated by comprehensively calculating the co-occurrence frequency of keywords in titles of cross-platform content, the consistency of sentiment polarity of user comments, and the relevance of knowledge graph entities.

9. A new media operation content recommendation system based on big data according to claim 8, characterized in that: The content metadata mapping unit constructs a cross-platform semantic similarity matrix in the following manner: Assume that the cross-platform content set is {c1,c2,…,c M }, M represents the total number of contents in the cross-platform content collection, and the knowledge graph embedding vector {e i }, topic tag collection {T i }, sentiment score {σ i } and entity collection On this basis, the initial semantic similarity matrix S∈R is defined M×M : Among them, α c ,α t ,α s ,α e is the normalized weight coefficient, α c ,α t ,α s ,α e ≥0, and α c +α t +α s +α e =1, is the vector cosine similarity, is the label Jaccard similarity, exp(-|σ i -σ j |) is the sentiment polarity difference function, is the entity overlap; Introduce the hotspot adjustment matrix H: Among them, ρ is the hotspot fusion coefficient, ρ∈[0,1], δ ij Indicates content c i with c j Co-occurrence heat in current hot events, For the maximum value of heat of all content pairs, p,q traverses all content pairs (c p ,c q ), H ij Indicates the normalized hotspot co-occurrence intensity, with a value range of [0,1].

10. The new media operation content recommendation system based on big data according to claim 1, characterized in that: Also includes: Privacy compliance verification unit, which performs homomorphic encryption on model parameters before federation aggregation; Behavioral data lifecycle management unit, which sets the automatic erasure mechanism for short-term behavioral data; The cross-platform identification desensitization unit uses an irreversible hash algorithm to convert user device fingerprint information.

Images