An intelligent personalized content pushing system based on big data

Abstract

This invention belongs to the field of intelligent push, specifically a personalized content intelligent push system based on big data. It includes a data acquisition module, a dynamic multi-granularity user profile construction module, an intelligent content understanding and representation learning module, and a deep fusion recommendation module. This invention captures instantaneous interest fluctuations through dynamic multi-granularity user modeling, models periodic patterns, stores stable interest preferences, and enables features to automatically update and evolve with user behavior. For new or sparse users, cross-domain transfer learning is introduced to alleviate the cold start problem and improve initial recommendation quality. Through a spatiotemporal attention mechanism, dynamic weighting automatically learns interest weights in different scenarios, achieving scenario adaptation. A dynamic gating mechanism is introduced to avoid contextual interference, pushing appropriate content in appropriate scenarios. Through a deep cross-network, high-order feature interactions are explicitly modeled, and through multi-head attention and multi-view matching, the matching degree is comprehensively evaluated to achieve better content push.

Description

Technical Field

[0001] This invention belongs to the field of intelligent push, specifically a personalized content intelligent push system based on big data. Background Technology

[0002] With the widespread adoption of mobile internet and the exponential growth of digital content, we live in an era of information overload. Statistics show that over 250 million TB of data is generated globally every day, while users' average daily attention span is only 3-4 hours. This huge gap between supply and demand has spurred the emergence of intelligent push systems. However, existing intelligent push systems suffer from a contradiction between static characteristics and dynamic needs, and their contextual awareness of content delivery is relatively superficial. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a personalized content intelligent push system based on big data. Addressing the contradiction between static features and dynamic needs, this invention employs dynamic multi-granularity user modeling, using temporal convolutional networks to capture instantaneous interest fluctuations, gated recurrent units to model periodic patterns, and memory networks to store stable interest preferences. This enables features to automatically update and evolve with user behavior. Furthermore, for new or sparse users, cross-domain transfer learning is introduced to effectively alleviate the cold start problem and improve initial recommendation quality. To address the issue of shallow context awareness in content push, a spatiotemporal attention mechanism is used for dynamic weighting, allowing the system to automatically learn interest weights in different scenarios, achieving scenario adaptation. A dynamic gating mechanism is introduced to avoid contextual interference, pushing appropriate content in appropriate scenarios. A deep cross-network is used to explicitly model high-order feature interactions, and multi-head attention and multi-view matching are used to comprehensively evaluate the matching degree, avoiding single-view bias, ultimately achieving superior content push.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides a personalized content intelligent push system based on big data, including a data acquisition module, a dynamic multi-granular user profile construction module, an intelligent content understanding and representation learning module, and a deep fusion recommendation module;

[0005] The data acquisition module collects user behavior data, which includes behavior data from different platforms, social networks, real-time interactive streams, and explicit feedback. The content of the behavior data is divided into text and visual.

[0006] The dynamic multi-granularity user profile construction module constructs a dynamic multi-granularity user profile based on multiple scales from behavioral data to obtain behavioral representations.

[0007] The intelligent content understanding and representation learning module understands user behavior data according to content and interacts with the behavior representation to obtain cross features and deep abstract representations.

[0008] The deep fusion recommendation module calculates the matching score between user i and content j based on cross features and deep abstract representation through a multi-head attention mechanism. ;

[0009] In the formula, Represents multi-head attention mechanism processing, Represents multilayer perceptron processing, This represents the match score between user i and content j;

[0010] Personalized content recommendations are given to users based on their matching score.

[0011] In the dynamic multi-granularity user profile construction module, the construction of dynamic multi-granularity user profiles based on multi-scale behavioral data specifically includes the following steps:

[0012] Step S1: Short-term interest modeling, which involves embedding and encoding features into short-term user behavior data, and using a convolutional network to extract spatiotemporal features from the embedded features to obtain the user's short-term interest vector.

[0013] Step S2: Mid-term interest modeling. Gated recurrent units are used to model the user's mid-term behavioral data to obtain the user's mid-term interest vector.

[0014] Step S3: Long-term interest modeling, using a memory network to store and retrieve long-term interests from behavioral data, resulting in the user's long-term interest vector;

[0015] Step S4: Cross-domain interest transfer. For new or sparse users, cross-domain transfer learning is introduced. Cross-domain features are extracted using user behavior data across different platforms. ;

[0016] In the formula, This represents the user representation of the source domain, i.e., the previous platform. The user representation includes short-term interest vectors, medium-term interest vectors, and long-term interest vectors. User representations that represent the contextual information of behavioral data on the current platform. Represents cross-domain migration networks. Represents preset parameters. Represents cross-domain characteristics;

[0017] Step S5: Behavioral representation construction, which involves fusing short-term interest vectors, medium-term interest vectors, and long-term interest vectors with cross-domain features to obtain behavioral representations.

[0018] In the intelligent content understanding and representation learning module, the calculation of cross features and deep abstract representation is as follows:

[0019] Step Q1: Semantic representation learning, using a BERT pre-trained model to extract deep semantics from text in user behavior data: ;

[0020] In the formula, Text representing behavioral data, Represents text segmentation and encoding operations. Represents the BERT pre-trained model. Represents semantic meaning;

[0021] Step Q2: Visual learning representation and cross-modal alignment. The visual representation in the user behavior data is encoded using a visual Transformer encoder to obtain the visual representation. The CLIP model is used to align the visual representation and the semantic representation.

[0022] Step Q3: Knowledge graph augmentation representation. By fusing external knowledge graphs, the visual and semantic representations are enhanced to obtain the augmented representation: ; ;

[0023] In the formula, Content representation representing knowledge enhancement Representative graph neural networks, Represents the initial entity embedding. Represents the adjacency matrix of a knowledge graph. This represents the adjacency matrix of a knowledge graph with added self-connections. represent The degree matrix, Represents the learnable weight matrix. Represents the entity embedding at layer l. Represents the activation function;

[0024] Step Q4: Content understanding construction, which involves fusing semantic representation, visual representation, and knowledge-enhanced content representation to obtain a content representation;

[0025] Step Q5: Model higher-order interactions between features. Through deep cross-networks, model the higher-order interactions between behavioral representations and content representations to obtain cross-features: ; ;

[0026] In the formula, Represents the initial behavioral and content representations. Represents the cross features of the l-th layer. The learning weights and biases of layer l represent the learningable weights and biases of layer l, and L represents the depth of the cross-network. Represents cross-features;

[0027] Step Q6: Model deep nonlinear relationships. Through deep networks, model deep abstract representations of behavioral and content representations. .

[0028] The beneficial results achieved by the present invention using the above solution are as follows:

[0029] (1) In view of the contradiction between static features and dynamic needs, this invention uses dynamic multi-granularity user modeling, uses temporal convolutional networks to capture instantaneous interest fluctuations, uses gated recurrent units to model periodic patterns, and uses memory networks to store stable interest preferences, so as to realize the automatic updating and evolution of features with user behavior. At the same time, for new users or sparse users, cross-domain transfer learning is introduced to effectively alleviate the cold start problem of new users and improve the quality of initial recommendations.

[0030] (2) To address the issue of shallow context awareness in content push, the spatiotemporal attention mechanism is used to dynamically weight the system, allowing it to automatically learn the interest weights in different scenarios and achieve scenario adaptation. A dynamic gating mechanism is introduced to avoid interference from context information and push appropriate content in appropriate scenarios. High-order feature interaction is explicitly modeled through deep cross-networks, and the matching degree is comprehensively evaluated through multi-head attention and multi-view matching to avoid single-view bias and ultimately achieve better content push. Attached Figure Description

[0031] Figure 1 This invention provides a module diagram of a personalized content intelligent push system based on big data.

[0032] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] Example 1, see Figure 1The present invention provides a personalized content intelligent push system based on big data, including a data acquisition module, a dynamic multi-granular user profile construction module, an intelligent content understanding and representation learning module, and a deep fusion recommendation module;

[0035] The data acquisition module collects user behavior data, which includes behavior data from different platforms, social networks, real-time interactive streams, and explicit feedback. The content of the behavior data is divided into text and visual.

[0036] The dynamic multi-granularity user profile construction module constructs a dynamic multi-granularity user profile based on multiple scales from behavioral data to obtain behavioral representations.

[0037] The intelligent content understanding and representation learning module understands user behavior data according to content and interacts with the behavior representation to obtain cross features and deep abstract representations.

[0038] The deep fusion recommendation module calculates the matching score between user i and content j based on cross features and deep abstract representation through a multi-head attention mechanism. ;

[0039] In the formula, Represents multi-head attention mechanism processing, Represents multilayer perceptron processing, This represents the match score between user i and content j;

[0040] Personalized content recommendations are given to users based on their matching score.

[0041] Example 2: Based on the above examples, this example describes the construction of a dynamic multi-granularity user profile based on multiple scales within the dynamic multi-granularity user profile construction module. Specifically, it includes the following steps:

[0042] Step S1: Short-term interest modeling, which involves embedding and encoding features into short-term user behavior data, and using a convolutional network to extract spatiotemporal features from the embedded features to obtain the user's short-term interest vector.

[0043] Step S2: Mid-term interest modeling. Gated recurrent units are used to model the user's mid-term behavioral data to obtain the user's mid-term interest vector.

[0044] Step S3: Long-term interest modeling, using a memory network to store and retrieve long-term interests from behavioral data, resulting in the user's long-term interest vector;

[0045] Step S4: Cross-domain interest transfer. For new or sparse users, cross-domain transfer learning is introduced. Cross-domain features are extracted using user behavior data across different platforms. ;

[0046] In the formula, This represents the user representation of the source domain, i.e., the previous platform. The user representation includes short-term interest vectors, medium-term interest vectors, and long-term interest vectors. User representations that represent the contextual information of behavioral data on the current platform. Represents cross-domain migration networks. Represents preset parameters. Represents cross-domain characteristics;

[0047] Step S5: Behavioral representation construction, which involves fusing short-term interest vectors, medium-term interest vectors, and long-term interest vectors with cross-domain features to obtain behavioral representations.

[0048] Example 3: Based on the above examples, this example calculates the cross features and deep abstract representations in the intelligent content understanding and representation learning module as follows:

[0049] Step Q1: Semantic representation learning, using a BERT pre-trained model to extract deep semantics from text in user behavior data: ;

[0050] In the formula, Text representing behavioral data, Represents text segmentation and encoding operations. Represents the BERT pre-trained model. Represents semantic meaning;

[0051] Step Q2: Visual learning representation and cross-modal alignment. The visual representation in the user behavior data is encoded using a visual Transformer encoder to obtain the visual representation. The CLIP model is used to align the visual representation and the semantic representation.

[0052] Step Q3: Knowledge graph augmentation representation. By fusing external knowledge graphs, the visual and semantic representations are enhanced to obtain the augmented representation: ; ;

[0053] In the formula, Content representation representing knowledge enhancement Representative graph neural networks, Represents the initial entity embedding. Represents the adjacency matrix of a knowledge graph. This represents the adjacency matrix of a knowledge graph with added self-connections. represent The degree matrix, Represents the learnable weight matrix. Represents the entity embedding at layer l. Represents the activation function;

[0054] Step Q4: Content understanding construction, which involves fusing semantic representation, visual representation, and knowledge-enhanced content representation to obtain a content representation;

[0055] Step Q5: Model higher-order interactions between features. Through deep cross-networks, model the higher-order interactions between behavioral representations and content representations to obtain cross-features: ; ;

[0056] In the formula, Represents the initial behavioral and content representations. Represents the cross features of the l-th layer. The learning weights and biases of layer l represent the learningable weights and biases of layer l, and L represents the depth of the cross-network. Represents cross-features;

[0057] Step Q6: Model deep nonlinear relationships. Through deep networks, model deep abstract representations of behavioral and content representations. .

[0058] Example 4: Applying the above solution to the intelligent recommendation system of a short video platform:

[0059] Application scenarios

[0060] A short video platform is facing problems such as declining daily active user time and low new user retention rate. It has adopted this system to optimize personalized content recommendations.

[0061] Technical Implementation Plan

[0062] Data acquisition module implementation:

[0063] Cross-platform behavioral data: Collecting users' public behaviors and interest tags on platforms such as Weibo, WeChat, and Taobao;

[0064] Real-time interactive stream: real-time actions such as user dwell time, likes, comments, shares, and swipes;

[0065] Explicit feedback: Users actively mark "not interested" or "add to list";

[0066] Social networks: following relationships, interaction frequency, and watching friends' posts together;

[0067] Dynamic multi-granular user profile construction:

[0068] (1) Short-term interest modeling:

[0069] Use temporal convolutional networks to analyze a user's viewing sequence over the last 10 minutes;

[0070] Capture fleeting interests: After watching three cooking videos consecutively, the short-term interest vector strengthens the "cooking" tag;

[0071] Actual result: When a user switches from "funny" content to "science" content, the system updates the user's interests within 30 seconds;

[0072] (2) Mid-term interest modeling:

[0073] Gated loop unit analyzes the trend of user interest changes over 7 days;

[0074] Discover periodic patterns: For example, users prefer movie and TV show commentary between 8-10 pm and lighthearted and humorous content during their lunch break;

[0075] Practical application: On Wednesday, the system detected that users were starting to pay attention to "weekend travel" content, and on Thursday, it began recommending travel guides for nearby destinations.

[0076] (3) Long-term interest modeling:

[0077] Memory network stores stable user preferences;

[0078] Avoid being overly influenced by short-term trends;

[0079] Cross-domain migration: When a new user registers for the first time, their initial interests are migrated from WeChat's public data.

[0080] Intelligent content understanding:

[0081] (1) Multimodal alignment:

[0082] Video content: Visual Transformer extracts image features;

[0083] Audio content: speech-to-text + sentiment analysis;

[0084] Text description: BERT extracts semantics from titles and tags;

[0085] CLIP model alignment with text and images: Ensures that visual features are consistent with text labels;

[0086] (2) Knowledge graph enhancement:

[0087] Constructing a knowledge graph of video content: "Fitness videos" are linked to "protein supplementation" and "sports equipment";

[0088] Practical application: After watching "fat loss training" videos, the system understands the user's underlying needs and recommends "low-calorie recipes" instead of simply showing more fitness videos;

[0089] Deep integration recommendation

[0090] (1) Context awareness:

[0091] Spatiotemporal attention: During commutes (movement + morning and evening rush hours), videos shorter than 1 minute are recommended;

[0092] For weekend evenings (in a home Wi-Fi environment), we recommend watching movies or TV shows with commentary of 45 minutes or more.

[0093] Dynamic gating: Reduce the weight of entertainment content and increase the weight of knowledge-based content during weekdays;

[0094] (2) Multi-source fusion:

[0095] Deep cross-networks discover higher-order features;

[0096] (3) Multi-head attention and multi-perspective assessment:

[0097] Perspective 1: Content quality (image quality, sound quality, production level);

[0098] Perspective 2: Topic matching degree (user's historical interests);

[0099] Perspective 3: Novelty (Has the user seen similar content before?);

[0100] Perspective 4: Social impact (friend interaction).

[0101] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0103] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A personalized content intelligent push system based on big data, characterized in that: It includes a data acquisition module, a dynamic multi-granular user profile construction module, an intelligent content understanding and representation learning module, and a deep fusion recommendation module; The data acquisition module collects user behavior data, which includes behavior data from different platforms, social networks, real-time interactive streams, and explicit feedback. The content of the behavior data is divided into text and visual. The dynamic multi-granularity user profile construction module constructs a dynamic multi-granularity user profile based on multiple scales from behavioral data to obtain behavioral representations. The intelligent content understanding and representation learning module understands user behavior data according to content and interacts with the behavior representation to obtain cross features and deep abstract representations. The deep fusion recommendation module calculates the matching score between user i and content j based on cross features and deep abstract representation through a multi-head attention mechanism. ； In the formula, Represents multi-head attention mechanism processing, Represents multilayer perceptron processing, This represents the match score between user i and content j; Personalized content recommendations are given to users based on their matching score.

2. The personalized content intelligent push system based on big data according to claim 1, characterized in that: In the dynamic multi-granularity user profile construction module, the construction of dynamic multi-granularity user profiles based on multi-scale behavioral data specifically includes the following steps: Step S1: Short-term interest modeling, which involves embedding and encoding features into short-term user behavior data, and using a convolutional network to extract spatiotemporal features from the embedded features to obtain the user's short-term interest vector. Step S2: Mid-term interest modeling. Gated recurrent units are used to model the user's mid-term behavioral data to obtain the user's mid-term interest vector. Step S3: Long-term interest modeling, using a memory network to store and retrieve long-term interests from behavioral data, resulting in the user's long-term interest vector; Step S4: Cross-domain interest transfer. For new or sparse users, cross-domain transfer learning is introduced. Cross-domain features are extracted using user behavior data across different platforms. ； In the formula, This represents the user representation of the source domain, i.e., the previous platform. The user representation includes short-term interest vectors, medium-term interest vectors, and long-term interest vectors. User representations that represent the contextual information of behavioral data on the current platform. Represents cross-domain migration networks. Represents preset parameters. Represents cross-domain characteristics; Step S5: Behavioral representation construction, which involves fusing short-term interest vectors, medium-term interest vectors, and long-term interest vectors with cross-domain features to obtain behavioral representations.

3. The personalized content intelligent push system based on big data according to claim 2, characterized in that: In the intelligent content understanding and representation learning module, the calculation of cross features and deep abstract representation is as follows: Step Q1: Semantic representation learning, using a BERT pre-trained model to extract deep semantics from text in user behavior data: ； In the formula, Text representing behavioral data, Represents text segmentation and encoding operations. Represents the BERT pre-trained model. Represents semantic meaning; Step Q2: Visual learning representation and cross-modal alignment. The visual representation in the user behavior data is encoded using a visual Transformer encoder to obtain the visual representation. The CLIP model is used to align the visual representation and the semantic representation. Step Q3: Knowledge graph augmentation representation. By fusing external knowledge graphs, visual and semantic representations are augmented to obtain enhanced representations. Step Q4: Content understanding construction, which involves fusing semantic representation, visual representation, and knowledge-enhanced content representation to obtain a content representation; Step Q5: Model higher-order interactions between features. Through deep cross-networks, model higher-order interactions between behavioral representations and content representations to obtain cross-features. Step Q6: Model deep nonlinear relationships. Through deep networks, model deep abstract representations of behavioral and content representations. .

Images