Content precision marketing pushing system and method based on user behavior deep mining

Abstract

The invention provides a content precision marketing pushing system and method based on user behavior deep mining, and the system comprises a data preprocessing module which is used for carrying out the cleaning, conversion and feature extraction of original data, and obtaining processed data; the user behavior deep analysis module is used for performing clustering analysis on user behaviors based on the processing data and constructing a user interest model; the content feature extraction module is used for performing classification, tagging and feature vector representation on the marketing content to obtain a content feature vector; the precise pushing module is used for matching the user with the content according to the user interest model and the content feature vector, and formulating and executing a personalized pushing strategy; and the feedback and optimization module is used for collecting user feedback data and optimizing the user interest model. According to the technical scheme, accurate marketing content pushing is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet marketing, and particularly relates to a content precise marketing push system and method based on deep mining of user behavior. BACKGROUND

[0002] In today's digital marketing era, enterprises are facing the dual challenges of massive users and massive content. Traditional marketing push methods often use a "broad cast" mode to push the same content to a large number of users. The disadvantages of this method are obvious: on the one hand, a large number of users will receive content that is not related to their own interests, resulting in poor user experience and even aversion; on the other hand, enterprises cannot effectively improve the conversion rate of marketing content, resulting in waste of marketing resources.

[0003] With the development of big data and artificial intelligence technology, marketing push based on user behavior data has gradually become a research hotspot. However, most of the existing related technologies only perform simple statistical analysis on user behavior, lack of deep mining of user behavior behind the interests and needs, and are difficult to achieve truly precise marketing push. Therefore, it is of great practical significance to develop a system and method that can deeply mine user behavior data and achieve precise content marketing push. SUMMARY

[0004] The present application provides a content precise marketing push system and method based on deep mining of user behavior, which realizes precise content marketing push through deep analysis of user behavior data.

[0005] In a first aspect, a content precise marketing push system based on deep mining of user behavior is provided, comprising:

[0006] A data collection module is configured to collect user behavior data and obtain raw data.

[0007] A data preprocessing module is configured to clean, convert and feature extract the raw data to obtain processed data.

[0008] A user behavior deep analysis module is configured to perform cluster analysis on user behavior based on the processed data and construct a user interest model. The user interest model includes preference degree information of users for different types of content, interest persistence information and interest trend information.

[0009] A content feature extraction module is configured to classify, tag and feature vector represent marketing content to obtain content feature vectors. The marketing content includes product information, service information, article information and video content. The content feature vector includes the topic category of the content, the keyword appearance frequency and the sentiment tendency.

[0010] The precision pushing module is configured to match users and contents according to the user interest model and the content feature vector, and formulate and execute a personalized pushing strategy; the matching process includes: obtaining user interest preference information for different types of contents by using the user interest model; meanwhile, the similarity or matching degree between the user interest model and the content feature vector is calculated by using a similarity calculation algorithm according to the content feature vector; the pushing strategy includes pushing time, pushing frequency, pushing channel, pushing content and personalized content customization;

[0011] The feedback and optimization module is configured to collect user feedback data and optimize the user interest model.

[0012] In a specific implementation, the data collection module includes:

[0013] The multi-source data access unit is configured to collect user behavior data from multiple channels; the user behavior data includes user operation records of an enterprise self-owned platform, user interaction data on a social media platform, and user portrait data provided by a third-party data service provider.

[0014] The real-time data collection unit is configured to collect user behavior data with time effectiveness requirements in real time; the user behavior data with time effectiveness requirements includes real-time click data and purchase operation data of the user.

[0015] In a specific implementation, the data preprocessing module includes:

[0016] The data cleaning unit is configured to clean the raw data.

[0017] The data conversion and feature extraction unit is configured to convert the cleaned raw data and extract key features of user behavior.

[0018] In a specific implementation, the user behavior deep analysis module includes:

[0019] The user behavior clustering unit is configured to perform clustering analysis on user behavior by using an improved K-means clustering algorithm to obtain a clustering result; the clustering result includes classifying users with similar behavior patterns into a category to form different user groups.

[0020] The user interest modeling unit is configured to construct the user interest model by using a long short-term memory network combined with an attention mechanism based on the clustering result and the key features of user behavior; the user interest model is configured to process time series data, capture the change trend of user interest over time, and automatically focus on user interest behavior and features.

[0021] In one specific implementation, the content feature extraction module comprises:

[0022] a content classification and labeling unit for classifying and labeling marketing content; a deep learning-based text classification model is used to classify text content, and artificial annotation and semi-automatic annotation are combined to add detailed labels to the content; the labels include theme labels, style labels, and sentiment labels;

[0023] a content feature vector representation unit for converting the classified and labeled marketing content into a content feature vector; the content feature vector includes the theme category of the content, the keyword appearance frequency, and the sentiment tendency.

[0024] In one specific implementation, the content classification and labeling unit uses a deep learning-based text classification model to classify text content, and combines artificial annotation and semi-automatic annotation to add labels to marketing content.

[0025] In one specific implementation, the content feature vector representation unit uses a word embedding algorithm to convert text content in the marketing content into a word vector, and then combines the word vector into the content feature vector through average pooling or maximum pooling.

[0026] In one specific implementation, the precise pushing module comprises:

[0027] a user-content matching unit for matching users and content using a method combining a weighted fusion-based collaborative filtering algorithm and a content-based recommendation algorithm; including:

[0028] using a collaborative filtering algorithm to find marketing content that other users similar to the target user's interests like; using a content-based recommendation algorithm to calculate the similarity between users and content according to the user interest model and the content feature vector; and dynamically adjusting the weights of the weighted fusion-based collaborative filtering algorithm and the content-based recommendation algorithm to obtain the matching result of user-content;

[0029] a pushing strategy formulation and execution unit for formulating personalized pushing strategies according to the user-content matching result and the user's historical behavior data;

[0030] The pushing strategy includes: according to the active time law of the user, selecting the time period when the user is most likely to view information for pushing;

[0031] According to the user's acceptance of content and feedback, reasonably control the pushing frequency;

[0032] According to the user's usage habits and preferences, select appropriate pushing channels;

[0033] According to the user interest model, a plurality of related contents are combined and pushed;

[0034] The marketing content is customized and pushed according to the interest characteristics of different users.

[0035] In a specific embodiment, the feedback and optimization module comprises:

[0036] A user feedback collection unit is configured to collect feedback information of the user on the pushed content; the feedback information comprises a click rate, a conversion rate, a stay time and a user evaluation of the user;

[0037] A model optimization unit is configured to optimize the user interest model according to the user feedback data, comprising dynamically adjusting parameters of the user interest model and clustering weights of the content feature vector.

[0038] In a second aspect, a content precision marketing pushing method based on deep mining of user behavior is provided, comprising the following steps:

[0039] A data collection module is used to collect user behavior data to obtain raw data;

[0040] A data preprocessing module is used to clean, convert and extract features of the raw data to obtain processed data;

[0041] A user behavior deep analysis module is used to perform clustering analysis on user behavior based on the processed data, and a user interest model is constructed;

[0042] A content feature extraction module is used to classify, label and represent feature vectors of marketing content to obtain content feature vectors;

[0043] A precision pushing module is used to match users and content according to the user interest model and the content feature vector, and to develop and execute personalized pushing strategies;

[0044] A feedback and optimization module is used to collect user feedback data and optimize the user interest model.

[0045] In the technical solution, the data collection module is configured to collect user behavior data to obtain original data; the data preprocessing module is configured to clean, convert and extract features of the original data to obtain processed data; the user behavior deep analysis module is configured to cluster analyze user behavior based on the processed data and construct a user interest model; the content feature extraction module is configured to classify, tag and represent a feature vector of marketing content to obtain a content feature vector; the precise push module is configured to match users and content according to the user interest model and the content feature vector, and formulate and execute a personalized push strategy; and the feedback and optimization module is configured to collect user feedback data and optimize the user interest model, thereby realizing precise marketing content push. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A structure block diagram of a content precise marketing push system based on user behavior deep mining provided by an embodiment of the present application is provided.

[0047] Figure 2 A flow block diagram of a content precise marketing push method based on user behavior deep mining provided by an embodiment of the present application is provided. DETAILED DESCRIPTION

[0048] The present application will be further described in detail by the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present application will become more apparent.

[0049] The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any implementation described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other implementations. Unless specifically stated otherwise, the drawings are not drawn to scale and are shown for purposes of explanation only.

[0050] Furthermore, the technical features involved in different embodiments of the present application described below can be combined with each other as long as there is no conflict.

[0051] To facilitate the understanding of the content precise marketing push system and method based on user behavior deep mining provided by the embodiments of the present application, the application scenarios thereof are first described. The content precise marketing push system and method based on user behavior deep mining provided by the embodiments of the present application are used to realize precise content marketing push through deep analysis of user behavior data. In today's digital marketing era, enterprises are facing the dual challenges of massive users and massive content. The traditional marketing push mode often adopts a “broad cast” mode to push the same content to a large number of users. The disadvantages of this mode are obvious: on the one hand, a large number of users will receive content that is not related to their own interests, resulting in poor user experience and even aversion; on the other hand, enterprises cannot effectively improve the conversion rate of marketing content, causing waste of marketing resources. With the development of big data and artificial intelligence technology, marketing push based on user behavior data has gradually become a research hotspot. However, the existing related technologies mostly only perform simple statistical analysis on user behavior, lack of deep mining of the underlying interests and needs behind user behavior, and are difficult to realize truly precise marketing push. Therefore, it is of important practical significance to develop a system and method capable of deep mining of user behavior data and realizing content precise marketing push. For this purpose, the embodiments of the present application provide a content precise marketing push system and method based on user behavior deep mining to realize precise content marketing push through deep analysis of user behavior data. The embodiments thereof will be described in detail below with reference to specific drawings.

[0052] Reference Figure 1 and Figure 2 , Figure 1 The structural block diagram of the content precise marketing push system based on user behavior deep mining provided by the embodiments of the present application is shown in FIG. 1. Figure 2 The flowchart of the content precise marketing push method based on user behavior deep mining provided by the embodiments of the present application is shown in FIG. 2.

[0053] In Figure 1 , the embodiments of the present application provide a content precise marketing push system based on user behavior deep mining, which comprises:

[0054] A data collection module is configured to collect user behavior data to obtain raw data.

[0055] A data preprocessing module is configured to clean, convert and feature extract the raw data to obtain processed data.

[0056] A user behavior deep analysis module is configured to perform cluster analysis on user behavior based on the processed data and construct a user interest model. The user interest model comprises preference degree information of users for different types of content, persistence information of interests and change trend information of interests.

[0057] a content feature extraction module configured to classify, label, and represent a feature vector of marketing content, to obtain a content feature vector, wherein the marketing content comprises product information, service information, article information, and video content, and wherein the content feature vector comprises a topic category of the content, a keyword appearance frequency, and an emotional tendency;

[0058] a precise push module configured to match users and content based on the user interest model and the content feature vector, and to formulate and execute a personalized push strategy, wherein the matching process comprises obtaining user interest preference information for different types of content by using the user interest model, and simultaneously calculating a similarity or a matching degree between the user interest model and the content feature vector by using a similarity calculation algorithm based on the content feature vector, and wherein the push strategy comprises a push time, a push frequency, a push channel, a push content, and a personalized content customization;

[0059] a feedback and optimization module configured to collect user feedback data and optimize the user interest model.

[0060] In the above technical solution, the data acquisition module is configured to collect user behavior data to obtain raw data, the data preprocessing module is configured to clean, convert, and extract features of the raw data to obtain processed data, the user behavior deep analysis module is configured to cluster analyze user behavior based on the processed data and construct a user interest model, the content feature extraction module is configured to classify, label, and represent a feature vector of marketing content to obtain a content feature vector, the precise push module is configured to match users and content based on the user interest model and the content feature vector, and to formulate and execute a personalized push strategy, and the feedback and optimization module is configured to collect user feedback data and optimize the user interest model, thereby achieving precise marketing content push.

[0061] Specifically, the beneficial effects of the content precise marketing push system based on user behavior deep mining include:

[0062] I. Beneficial effects of individual functions of the modules

[0063] Data acquisition module

[0064] Global behavior coverage: user behavior data (such as clicks, browsing time, purchase records, social interactions) is collected through multiple channels to cover user behavior throughout their life cycle and avoid one-sidedness of a single data source.

[0065] Real-time and large-scale: real-time data stream collection (such as user real-time click events) and offline batch processing (such as historical order analysis) are supported to balance timeliness and data integrity.

[0066] Data preprocessing module

[0067] Data Quality Assurance: Enhance data usability and reduce subsequent analysis errors through cleaning (de-duplication, de-noising, missing value filling) and transformation (normalization, encoding). For example, convert user dwell time from a string to a numerical feature for model calculation.

[0068] Feature Engineering Support: Extract high-value features (such as user active period, preference category combination) to provide a foundation for in-depth analysis. For example, extract keyword features from user search records using the TF-IDF algorithm to reflect user potential interests.

[0069] User Behavior Deep Analysis Module

[0070] Precise User Segmentation: Divide users into subgroups (such as "price-sensitive" and "quality-pursuing") through clustering analysis (such as K-means and DBSCAN) to support differentiated marketing strategies.

[0071] Dynamic Interest Modeling: Build interest models based on user behavior sequences (such as "browse → add to cart → purchase") to capture dynamic changes in user interests (such as seasonal preference migration) and avoid the lag of static labels.

[0072] Content Feature Extraction Module

[0073] Structured Content Representation: Convert unstructured content into machine-understandable numerical vectors through classification (such as news, promotions, tutorials), labeling (such as "technology" and "maternity and baby"), and feature vector representation (such as Word2Vec and BERT embedding) to support efficient matching.

[0074] Multi-modal Support: Compatible with text, images, videos and other multi-type content, unified representation is achieved through cross-modal feature extraction (such as CLIP model) to expand application scenarios.

[0075] Precise Push Module

[0076] Personalized Matching: Based on the similarity calculation (such as cosine similarity) between user interest models and content feature vectors, realize "thousand people with different faces" push, improve click-through rate and conversion rate. For example, push new product review content to "technology enthusiasts" group.

[0077] Strategy Dynamic Adjustment: Support multi-dimensional rule combination (such as time window, frequency control, A / B test) to balance push effect and user experience. For example, push high-priority content during user active period and limit daily push frequency.

[0078] Feedback and Optimization Module

[0079] Closed-loop optimization mechanism: Quantify the push effect through user feedback data (such as click rate, dwell time, conversion rate, negative reviews), and drive user interest model iteration. For example, if the click rate of a certain type of content continues to be below the threshold, adjust the user interest weight or content label.

[0080] Cold start problem mitigation: Through user implicit feedback (such as browsing non-click content) and collaborative filtering algorithm, generate an initial interest model for new users or low active users, reduce the impact of data sparsity.

[0081] II. The beneficial effects of module synergy

[0082] Data-driven precise matching closed loop

[0083] The data collection module and the preprocessing module provide high-quality input for the analysis module, and the user interest model generated by the analysis module and the output of the content feature extraction module are jointly input into the precise push module to realize the "data-model-strategy" closed loop. For example, after a user clicks on a brand advertisement, the system automatically extracts brand keywords, updates the user interest model, and preferentially recommends brand-related content in subsequent push.

[0084] Dynamic interest and content co-evolution

[0085] The real-time interest update of the user behavior depth analysis module and the label dynamic adjustment of the content feature extraction module work together. For example, if the conversion rate of a certain type of content (such as "outdoor sports equipment") suddenly rises, the system will not only push more similar content to related interest users, but also extract common labels (such as "lightweight" and "waterproof") of this type of content through the content feature extraction module, further optimizing the push strategy.

[0086] Feedback-driven model self-optimization

[0087] The feedback and optimization module reversely injects user feedback data (such as "not interested" label) into the user behavior depth analysis module to adjust clustering parameters or interest weights. For example, if a user repeatedly marks "maternal and infant" content as not interested, the system will reduce the user's sensitivity to the maternal and infant label and reduce the content in subsequent push.

[0088] Multi-objective balance and strategy optimization

[0089] When the precise push module combines user interest model and content feature vector, it can optimize multiple objectives (such as click rate, conversion rate, and user retention) at the same time. For example, through multi-objective optimization algorithm (such as Pareto frontier), a push strategy that takes into account both short-term revenue and long-term value is generated, avoiding excessive push leading to user fatigue.

[0090] III. Overall benefits of the system

[0091] Marketing efficiency improvement

[0092] Reduce ineffective exposure and lower marketing costs through precise matching. For example, after an e-commerce platform applies this system, the ad click-through rate increases by 30%, the conversion rate increases by 15%, and the user complaint rate decreases by 20%.

[0093] User experience optimization

[0094] Personalized push reduces information overload and improves user satisfaction. For example, after a news client filters low-relevance content through a user interest model, the user's daily reading time increases by 25%.

[0095] Business decision support

[0096] User behavior insights (such as interest shift trends and content preference distribution) output by the system can provide data support for product optimization and inventory management. For example, a retail enterprise adjusts its product packaging strategy based on the increasing attention to "environmentally friendly packaging," leading to growth in related product categories.

[0097] System robustness and scalability

[0098] Modular design supports flexible expansion (such as adding data sources and algorithm models) to adapt to business changes. For example, when an enterprise adds a live streaming e-commerce scenario, it can quickly integrate live streaming interaction data into the data collection module and adjust the content feature extraction module to support video content analysis.

[0099] This system realizes a full-link closed loop from user behavior insight to marketing strategy implementation through the synergy of data collection, preprocessing, deep analysis, content representation, precise push, and feedback optimization. Its core advantages are:

[0100] Precision: Through multi-dimensional behavior analysis and content feature extraction, ensure that the pushed content is highly matched with user interest;

[0101] Dynamic: Real-time update of user interest model and content label to adapt to rapid changes in market and user demand;

[0102] Sustainability: Closed-loop feedback mechanism drives continuous system optimization, avoiding strategy ossification.

[0103] Finally, this system can significantly improve marketing ROI, enhance user stickiness, and provide data-driven decision support for enterprises.

[0104] In a specific implementable scheme, the data collection module comprises:

[0105] A multi-source data access unit for collecting user behavior data from multiple channels;

[0106] Real-time data collection unit for collecting user behavior data with time-sensitive requirements in real time.

[0107] In particular, the beneficial effects include:

[0108] I. Beneficial effects of individual functions of modules

[0109] Multi-source data access unit

[0110] Data comprehensiveness improvement: By integrating multi-channel data (such as APP clicks, web browsing, offline store code scanning, social media interaction), a user behavior panoramic view is constructed to avoid the bias of a single data source, providing rich materials for accurate analysis.

[0111] Cross-scenario user insight: Support cross-platform behavior correlation (such as "online search → offline purchase"), mine user complete consumption path, and help enterprises identify high-value scenarios.

[0112] Real-time data collection unit

[0113] Timeliness guarantee: Real-time capture of high timeliness behavior (such as live interaction, time-limited purchase participation), ensure marketing strategy response speed (such as second-level push of preferential information), and grasp user transient demand.

[0114] Dynamic strategy support: Real-time data flow provides immediate input for subsequent modules (such as user interest model update, push strategy adjustment), supports rapid iteration of "data-analysis-decision" closed loop.

[0115] II. Beneficial effects of module synergy

[0116] Complementarity of full volume and real-time data

[0117] The multi-source data access unit provides historical behavior and static features (such as user portrait), and the real-time data collection unit supplements dynamic behavior (such as real-time location, current browsing goods), which together form a three-dimensional user portrait of "long-term interest + transient demand". For example, the system identifies a user as a "long-term follower of baby products" through multi-source data, and then combines real-time data to find that the user is browsing a certain brand of milk powder, and immediately pushes related preferential information.

[0118] Balance between data quality and efficiency

[0119] The batch processing capability of the multi-source access unit reduces the pressure of real-time collection, while the lightweight design (such as Kafka streaming processing) of the real-time unit ensures that key data is not lost, and the two work together to achieve "low cost and high coverage" data collection.

[0120] The combination of multi-source data access and real-time collection unit not only ensures the comprehensiveness and depth of user behavior data, but also takes into account the timeliness requirement, providing high-quality, multi-dimensional, and high-timeliness data foundation for subsequent modules (such as user interest modeling and precise pushing), significantly improving the response speed and decision-making accuracy of the marketing system.

[0121] In a specific implementable embodiment, the data preprocessing module comprises:

[0122] a data cleaning unit for cleaning the raw data;

[0123] a data conversion and feature extraction unit for converting the cleaned raw data and extracting key features of user behavior.

[0124] Specifically, the beneficial effects of the data preprocessing module include:

[0125] I. The beneficial effects of the functions of the modules

[0126] Data cleaning unit

[0127] Improve data quality: by removing duplicates, removing noise (such as filtering crawlers or error logs), and filling missing values (such as mean / median completion), eliminate noise and redundancy in raw data, reduce subsequent analysis errors. For example, after removing duplicate click records, user behavior statistics are more accurate.

[0128] Standardize data format: unify the format of fields such as timestamp and user ID (such as converting "2023-1-1" and "01 / 01 / 2023" to standard format), facilitate cross-channel data fusion and calculation.

[0129] Data conversion and feature extraction unit

[0130] Structured data processing: convert unstructured data (such as text, logs) into numerical features (such as user dwell time → seconds, product category → one-hot encoding), adapt to machine learning model input requirements.

[0131] Key feature mining: through dimensionality reduction (such as PCA), statistical indicators (such as user average browsing frequency), or sequence analysis (such as user behavior path), extract high-value features, reduce redundant information, and improve model efficiency.

[0132] II. The beneficial effects of the synergistic action of the modules

[0133] Closed-loop optimization of cleaning and conversion

[0134] The output (clean data) of the data cleaning unit is directly input into the conversion and feature extraction unit, avoiding the interference of dirty data on feature effectiveness. For example, if repeated click records are not cleaned, the user activity feature may be overestimated, leading to subsequent interest model bias.

[0135] The feature extraction result can guide the optimization of the cleaning rule in reverse. For example, if it is found that the missing rate of a certain field is too high (such as 90% of users not filling in the age), the cleaning strategy can be adjusted to directly delete the field or complete it through associated data.

[0136] Dual improvement of efficiency and accuracy

[0137] The lightweight processing (such as regular expression filtering) of the cleaning unit reduces the data volume and reduces the calculation cost of conversion and feature extraction; and the accurate modeling (such as feature engineering based on user behavior sequence) of the feature extraction unit further refines the data value, and the two work together to achieve "low cost and high precision" preprocessing.

[0138] The combination of data cleaning and conversion and feature extraction unit not only guarantees the integrity and standardization of data, but also mines the deep patterns of user behavior through feature engineering, providing a high-quality, low-noise, high-information-density data foundation for subsequent user interest modeling and accurate pushing, significantly improving the reliability and efficiency of system decision-making.

[0139] In one specific implementable embodiment, the user behavior deep analysis module comprises:

[0140] a user behavior clustering unit for performing clustering analysis on user behavior using an improved K-means clustering algorithm to obtain a clustering result;

[0141] a user interest modeling unit for constructing the user interest model based on the clustering result and key features of the user behavior using a long short-term memory network combined with an attention mechanism.

[0142] Specifically, the accurate pushing module is configured to match users and content according to the user interest model and the content feature vector, and formulate and execute personalized pushing strategies; the matching process includes: obtaining user interest preference information for different types of content using the user interest model; at the same time, according to the content feature vector interest preference information, calculating the similarity or matching degree between the user interest model and the content feature vector through a similarity calculation algorithm; the pushing strategies include pushing time, pushing frequency, pushing channel, pushing content, and personalized content customization.

[0143] In one specific implementable embodiment, the pushing process of the accurate pushing module comprises:

[0144] 1. User interest model and content feature vector definition

[0145] User Interest Model:

[0146] Interest Preference: User has a preference weight of 0.9 (out of 1) for "football" type content, with a long duration of interest (interest duration > 1 year), and an interest change trend of "no change" (stable preference).

[0147] Extended Features: User also has a moderate preference (weight 0.6) for "sports news" and "player transfers" subtopics.

[0148] Content Feature Vector:

[0149] Topic Category: Content belongs to the "football - match analysis" category.

[0150] Keyword Frequency: High frequency keywords include "football", "goal" (football-related), and low frequency keywords "movie", "music" (irrelevant).

[0151] Sentiment Tendency: Positive sentiment (exciting match descriptions).

[0152] 2. Matching Degree Calculation Process

[0153] Step 1: Vector Representation

[0154] User Interest Vector:

[0155] U = [football (0.9), sports news (0.6), player transfers (0.6), others (0)];

[0156] Content Feature Vector:

[0157] C = [football - match analysis (1.0), goal (0.8), football match (0.7), movie (0.1), sentiment tendency (positive)];

[0158] Step 2: Similarity Calculation (Cosine Similarity)

[0159] Align dimensions: Map user interests and content keywords to the same space (e.g. "football" as the root topic, "Messi" as a subtopic).

[0160] Weighted calculation:

[0161] User weight for "football" is 0.9, content weight for "football - match analysis" is 1.0, matching score: 0.9 x 1.0 = 0.9;

[0162] User weight for "player transfers" is 0.6, content weight for "player name" (player-related) is 0.8, matching score: 0.6 x 0.8 = 0.48;

[0163] The irrelevant word "movie" has a weight of 0.1, and the user is not interested, so the score is: 0 x 0.1 = 0;

[0164] Step 3: Emotional Tendency Matching

[0165] The user has a historical preference for positive emotional content (such as victory reports), and the current content emotion is positive, so additional points are added (such as +0.1).

[0166] Final matching degree: 0.85 + 0.1 = 0.95 (high matching).

[0167] 3. Push Strategy Formulation

[0168] Push time: Push during the user's active period (such as 8 pm).

[0169] Push frequency: 3 football depth contents per week (avoid excessive pushing).

[0170] Channel: App push + email summary (users often use email).

[0171] Content customization: Preferentially display "player name" and "football match" related analysis, with a link to the match highlights.

[0172] 4. Dynamic Adjustment:

[0173] If the user clicks more on "player transfer" content in a certain week, the model will dynamically increase the weight of this sub-theme, and push more transfer news in the future.

[0174] Specifically, the beneficial effects include:

[0175] I. Beneficial effects of individual functions of the module

[0176] User Behavior Clustering Unit

[0177] Efficient user clustering: Improved K-means algorithm (such as introducing dynamic initial center point selection or distance metric optimization) improves clustering stability and accuracy, quickly dividing users into similar behavior groups (such as "highly active shoppers" and "low-frequency browsers"), supporting targeted marketing strategy formulation.

[0178] Interpretability: Clustering results intuitively show user behavior commonalities (such as some users preferring to shop at night), facilitating business personnel understanding and application.

[0179] User Interest Modeling Unit

[0180] Dynamic interest capture: Long Short-Term Memory Network (LSTM) combined with attention mechanism can model the long-term dependence (such as seasonal preferences) and short-term changes (such as temporary interests during promotional activities) of user interests, improving the model's fitting ability for complex behaviors.

[0181] Feature weight adaptation: attention mechanism automatically assigns key behavior feature weights (e.g., "add-to-cart" behavior is more important than "browsing"), enhancing the model's recognition of user's true intentions.

[0182] II. Benefits of module synergy

[0183] Clustering results guide interest modeling

[0184] The user group labels output by the user behavior clustering unit serve as prior knowledge for interest modeling, helping the LSTM model converge quickly. For example, for the "highly active shopper" group, the model can focus on learning high-frequency behavior features (such as daily browsing categories), improving modeling efficiency.

[0185] Group features in the clustering results (such as a certain type of user preferring electronic products) can constrain the learning direction of the attention mechanism, avoiding the model from paying excessive attention to noise behaviors (such as accidental clicks).

[0186] Interest model feedback optimizes clustering

[0187] The interest distribution output by the user interest modeling unit (such as the user's interest weight for "sports shoes") can reverse optimize the similarity measure of the clustering algorithm (such as adding interest weight as a distance calculation dimension), making the clustering results more suitable for business needs.

[0188] The combination of user behavior clustering and interest modeling units, through hierarchical analysis of "grouping first, then modeling", not only utilizes clustering to efficiently divide user groups, but also mines individual interest dynamics through deep learning models, achieving precise characterization from macro-grouping to micro-interest, significantly improving the accuracy of user portraits and the adaptability of marketing strategies.

[0189] In one specific implementation, the content feature extraction module includes:

[0190] A content classification and labeling unit for classifying and labeling marketing content;

[0191] A content feature vector representation unit for converting the classified and labeled marketing content into a content feature vector.

[0192] Specifically, the benefits include:

[0193] I. Benefits of module individual functions

[0194] Content classification and labeling unit

[0195] Structured content semantics: By pre-defined taxonomy (e.g. product categories, article topics) and tag library (e.g. "promotion", "new", "tutorial"), unstructured content is transformed into computable semantic labels, supporting quick retrieval and matching. For example, an article "Summer Sneaker Recommendation" is labeled with tags like "sports equipment", "summer", "discount".

[0196] Business rule compatibility: Tagging can incorporate business rules (e.g. "high-margin product" label) to assist marketing strategy formulation (e.g. prioritizing high-margin content).

[0197] Content feature vector representation unit

[0198] Machine readability enhancement: Convert classification labels and text content into numerical vectors (e.g. TF-IDF, Word2Vec or BERT embedding) to meet machine learning model input requirements, supporting efficient similarity calculation (e.g. cosine similarity).

[0199] Multimodal scalability: Support unified representation of multiple types of content such as text, pictures, videos (e.g. extract cross-modal vectors through CLIP model), expand application scenarios.

[0200] II. Benefits of module synergy

[0201] Complementarity of semantic labels and vector representation

[0202] The explicit labels of the classification and tagging unit (e.g. "sports equipment") provide interpretable semantic anchors for feature vectors, while the implicit semantics of the feature vector representation unit (e.g. the potential association between "sneakers" and "running") supplement the coverage blind spots of labels. For example, when the user interest model matches the "sports equipment" label, the vector similarity can be used to further filter content containing keywords such as "lightweight" and "breathable".

[0203] Label-guided vector optimization

[0204] Tagging results can constrain the learning direction of feature vectors. For example, if a content is labeled as "time-limited promotion", the feature vector representation unit can strengthen the weight of time-sensitive features (such as countdown, urgency words) to improve the timeliness of the push.

[0205] Cold start content quick adaptation

[0206] When new content is online, the classification and tagging unit can quickly generate initial labels, and then generate vectors through the feature vector representation unit, shortening the time period from content storage to pushable.

[0207] The combination of the content classification and labeling unit and the feature vector representation unit through the dual coding of "explicit semantics + implicit representation" ensures the accuracy and interpretability of content understanding, improves the efficiency and scalability of machine processing, provides high-quality, multi-dimensional, and easily matched content feature input for the precise pushing module, and significantly improves the matching accuracy of content and user interest.

[0208] In a specific and implementable embodiment, the content classification and labeling unit classifies the text content using a deep learning-based text classification model, and adds labels to marketing content by combining manual annotation and semi-automatic annotation methods.

[0209] Specifically, the beneficial effects include:

[0210] I. Beneficial effects of individual functions of modules

[0211] Deep learning text classification model

[0212] High-precision semantic recognition: Based on BERT, TextCNN, and other models, automatically extract deep semantic of text (such as recognizing that "summer clearance" belongs to "promotion" and is also related to "seasonal goods"), reduce the limitations of artificial rules, and adapt to complex and variable marketing language.

[0213] Efficient and scalable processing: Support batch automatic classification, quickly process massive content (such as thousands of new product descriptions per day), and reduce labor costs.

[0214] Combination of manual and semi-automatic annotation

[0215] Business rules accurately land: Manual annotation can incorporate business knowledge (such as labeling "high single-price goods" as "VIP exclusive"), ensuring that labels meet marketing strategy requirements.

[0216] Semi-automatic efficiency improvement: Through active learning (Active Learning), high-uncertainty samples are selected for manual review, reducing redundant annotation volume, and using manual annotation results to iteratively optimize the model, forming a "annotation-optimization" closed loop.

[0217] II. Beneficial effects of module synergy

[0218] Complementary optimization of model and manual

[0219] The deep learning model provides a basic classification framework, manual annotation corrects model bias (such as correcting the misclassification of "limited edition" as "ordinary goods"), and semi-automatic annotation feeds back human experience to the model, improving the classification accuracy of long-tail scenarios (such as niche categories).

[0220] The artificially labeled labels can be used as weak supervision signals for model training, and combined with a small amount of strongly labeled data (such as expert labeling) to improve the generalization ability of the model.

[0221] Dynamic label system expansion

[0222] The deep learning model can quickly adapt to new label requirements (such as adding a new "environmentally friendly material" label) by fine-tuning the model through transfer learning; artificial labeling is responsible for verifying the rationality of the new label to avoid the model generating meaningless labels (such as over-generalized "high-quality" labels).

[0223] The combination of deep learning text classification and artificial / semi-automatic labeling, through the collaborative mechanism of "model generalization + artificial calibration", not only guarantees the efficiency and accuracy of content classification, but also ensures that the label system meets business needs, while achieving a balance between cost and quality through semi-automatic labeling, providing high-quality, interpretable semantic labels for subsequent content feature vector representation and accurate pushing.

[0224] In a specific implementable embodiment, the content feature vector representation unit converts the text content into a word vector using a word embedding algorithm, and then combines the word vectors into the content feature vector through average pooling or maximum pooling methods.

[0225] Specifically, the beneficial effects include:

[0226] I. Beneficial effects of the functions of individual modules

[0227] Word embedding algorithm

[0228] Semantic dense representation: convert words in text into low-dimensional dense vectors through algorithms such as Word2Vec, GloVe, or BERT, capture semantic relationships between words (such as "athletic shoes" and "running shoes" vectors are similar), overcome the sparsity and semantic loss of traditional one-hot encoding.

[0229] Context awareness: word vectors generated by pre-trained models (such as BERT) can dynamically reflect the meaning of words in context (such as "apple" in "fruit" and "phone" contexts), improving feature accuracy.

[0230] Pooling combination method

[0231] Global information extraction: average pooling retains overall semantics (such as calculating the mean of all word vectors to reflect content themes), and maximum pooling highlights key features (such as extracting the highest activated word vector to capture core keywords), both methods can be used alone or in combination to adapt to different scene requirements.

[0232] Dimension controllability: the pooling operation converts indefinite length text into a fixed dimension vector, adapting to subsequent similarity calculation and model input requirements.

[0233] II. Benefits of module synergy

[0234] Complementary optimization of word vectors and pooling

[0235] Word vectors generated by word embedding algorithms already contain local semantic information, and the pooling operation further refines global features. For example, by averaging pooling, the word vectors of shoe keywords such as "lightweight", "breathable", and "cushioning" are combined to generate an overall feature vector representing "athletic shoe functionality".

[0236] Max pooling can strengthen significant information in the content (such as "limited time 5 off" in a promotion), making up for the key signals that may be weakened by average pooling.

[0237] Synergy with the classification tagging unit

[0238] The output of the word vector representation unit can be combined with the label of the classification tagging unit to form a multi-modal feature of "explicit label + implicit vector". For example, certain content is labeled as the "sports equipment" label, and its feature vector is highly similar to the vector of the "sports enthusiasts" group in the user interest model, and double verification improves the reliability of matching.

[0239] The combination of word embedding algorithms and pooling methods, through the hierarchical processing of "local semantic encoding + global feature refinement", not only retains the semantic richness of the text, but also generates structured feature vectors, providing efficient and computable semantic representations for the precision push module. After synergizing with the classification tagging unit, it further enhances the depth and breadth of content understanding, significantly improving the matching accuracy of content and user interest.

[0240] In a specific implementation, the precision push module includes:

[0241] A user-content matching unit for user-content matching using a method combining weighted fusion-based collaborative filtering algorithms and content-based recommendation algorithms;

[0242] A push strategy formulation and execution unit for formulating personalized push strategies based on user-content matching results and user historical behavior data.

[0243] Specifically, the benefits include:

[0244] I. Benefits of module individual functions

[0245] User-content matching unit

[0246] Algorithm complementarity enhances matching accuracy: Collaborative filtering algorithms (e.g., based on user similarity or item similarity) mine implicit preferences (e.g., "users who have purchased A also like B"), while content recommendation algorithms (e.g., based on feature vector similarity) capture explicit needs (e.g., "users prefer content with the'sports' tag"). A weighted fusion of both can balance "group behavior" and "individual characteristics," avoiding the cold start or long tail problems of a single algorithm.

[0247] Dynamic weight adaptation scenarios: By adjusting the weights of collaborative filtering and content algorithms (e.g., new users focus on content algorithms, and old users focus on collaborative filtering), the matching effect at different user stages can be improved.

[0248] Push strategy formulation and execution unit

[0249] Personalized strategy optimization: Combine matching results with user historical behavior (e.g., click-through rate, conversion rate) to dynamically adjust push frequency (e.g., increase push for high-active users), channels (e.g., APP push + SMS), and timing (e.g., trigger within 1 hour after user browsing), to improve user experience and conversion efficiency.

[0250] Strategy interpretability: Generate push reasons based on user behavior data (e.g., "recommended according to your recent browsing of'sports shoes'"), enhancing user trust.

[0251] II. Beneficial effects of module synergy

[0252] Matching results drive strategy optimization

[0253] The candidate content list and matching score output by the user-content matching unit provide priority basis for strategy formulation. For example, content with a matching score higher than the threshold is pushed first, or low-efficiency content is filtered in combination with user historical behavior data (e.g., "users often ignore promotional information").

[0254] Algorithm weight distribution in matching results (e.g., collaborative filtering contributes 60%, content algorithm contributes 40%) can guide strategy differentiation. For example, when collaborative filtering dominates, focus on social association (e.g., "friends purchase"), and when content algorithm dominates, focus on semantic matching (e.g., "title contains user interest keywords").

[0255] Strategy feedback iterates matching model

[0256] User feedback (e.g., clicks, dwell time) after push strategy execution can optimize the user-content matching algorithm in reverse. For example, if the click-through rate of content recommended by collaborative filtering is low, reduce its weight and strengthen the contribution of the content algorithm.

[0257] The combination of user-content matching and push strategy formulation unit, through the closed-loop design of "precise matching + dynamic strategy", not only utilizes hybrid algorithm to improve the matching degree of content and user, but also improves the push effect and user experience through personalized strategy. The two work together to realize the whole-link optimization from "candidate content generation" to "push execution", significantly improving marketing ROI and user retention rate.

[0258] In a specific implementable embodiment, the feedback and optimization module comprises:

[0259] a user feedback collection unit for collecting feedback information of the user on the pushed content;

[0260] a model optimization unit for optimizing the user interest model according to the user feedback data.

[0261] Specifically, the beneficial effects include:

[0262] I. Beneficial effects of the functions of the modules

[0263] User feedback collection unit

[0264] Multi-dimensional data collection: Through explicit feedback (such as likes, ratings, and clicks on "not interested") and implicit feedback (such as dwell time, completion rate, and sharing behavior), user preferences are captured comprehensively, avoiding bias from a single data source.

[0265] Real-time and coverage: Support for collecting feedback immediately after pushing, quickly responding to changes in user needs, while covering all user behaviors, reducing sample bias.

[0266] Model optimization unit

[0267] Dynamic interest update: Based on feedback data, fine-tune the user interest model (such as adjusting the "sports shoes" interest weight), ensuring that the model is synchronized with the user's real-time preferences, and alleviating the interest drift problem.

[0268] Long-tail interest mining: Through feedback data, discover potential user interests (such as users frequently ignoring "electronic products" but occasionally clicking on "outdoor equipment"), expanding the model coverage.

[0269] II. Beneficial effects of the synergistic action of the modules

[0270] Feedback-driven model iteration

[0271] The real-time data of the user feedback collection unit provides training materials for the model optimization unit. For example, if the user has marked "sports shoes" content as "not interested" multiple times, the model optimization unit can reduce the weight of this interest dimension, or even trigger model retraining.

[0272] Negative samples in the feedback data (such as content that users did not click on) can enhance the robustness of the model and avoid the "information cocoon" caused by over-recommendation of similar content.

[0273] Closed-loop optimization improves performance

[0274] The optimized user interest model can directly influence the content pushed in the next round, and user feedback on the new content flows back into the optimization unit, forming a closed loop of "feedback-optimization-push". For example, after the optimized model reduces the recommendation of "sports shoes", if users start actively searching for related content, the feedback data will guide the model to restore some of the interest weights.

[0275] The combination of user feedback collection and model optimization units, through a closed-loop mechanism of "real-time feedback + dynamic adjustment," enables the system to have self-correction capabilities, significantly improving the accuracy of user interest models and the adaptability of pushed content, ultimately achieving continuous improvement in user experience and business metrics (such as click-through rate and conversion rate).

[0276] In one specific implementation scheme, the content-based precision marketing push system based on in-depth user behavior mining includes:

[0277] 1. The data acquisition module includes:

[0278] Multi-source data access unit: Responsible for collecting user behavior data from multiple channels, including but not limited to user operation records from the enterprise's own platforms (such as websites and apps), user interaction data on social media platforms, and user profile data provided by third-party data service providers. It achieves automatic data collection and synchronization by writing interface programs adapted to different data sources.

[0279] Real-time data acquisition unit: For user behavior data with high timeliness requirements, such as real-time user clicks and purchases, real-time data acquisition technology is used to ensure that the data can be entered into the system for processing in a timely and accurate manner.

[0280] 2. Data Preprocessing Module

[0281] Data cleaning unit: Cleans the collected raw data, removing duplicate, erroneous, and noisy data. For example, it filters browsing duration records that are obviously illogical and fills in or deletes missing key fields.

[0282] The data transformation and feature extraction unit converts the cleaned data into a format suitable for subsequent analysis and extracts key features of user behavior. For example, it converts user browsing history into time series data and extracts features such as browsing frequency, dwell time, and click-through rate; it also performs word segmentation and part-of-speech tagging on user search keywords to extract user interest keyword features.

[0283] 3. User behavior deep analysis module

[0284] User behavior clustering unit: Improved K-means clustering algorithm is used for clustering analysis of user behavior. Traditional K-means algorithm is sensitive to initial clustering center, which can easily lead to unstable clustering results. The present invention improves the K-means algorithm, and determines the initial clustering center based on density, improving the accuracy and stability of clustering. Users with similar behavior patterns are classified into a class, forming different user groups.

[0285] User interest modeling unit: Based on the clustering results and user behavior characteristics, a long short-term memory network (LSTM) combined with an attention mechanism is used to construct a user interest model in deep learning. LSTM can process time series data and capture the trend of user interest over time; the attention mechanism allows the model to automatically focus on behaviors and features that have a greater impact on user interest, improving the accuracy and generalization ability of the model.

[0286] 4. Content feature extraction module

[0287] Content classification and labeling unit: Marketing content is classified and labeled. A deep learning-based text classification model (such as BERT model) is used to classify text content, and manual annotation and semi-automatic annotation methods are combined to add detailed labels to the content, such as theme labels, style labels, and sentiment labels.

[0288] Content feature vector representation unit: The classified and labeled content is converted into a feature vector. Word embedding algorithms (such as Word2Vec) are used to convert text content into word vectors, and then average pooling or maximum pooling methods are used to combine word vectors into content feature vectors. At the same time, multimedia features such as pictures and videos of the content are combined, and convolutional neural networks (CNN) are used to extract multimedia feature vectors, and text feature vectors and multimedia feature vectors are fused to obtain more comprehensive content feature vectors.

[0289] 5. Precise push module

[0290] User-content matching unit: A method combining collaborative filtering algorithm based on weighted fusion and content-based recommendation algorithm is used for user-content matching. Collaborative filtering algorithm can find marketing content that other users similar to the target user's interest like; content-based recommendation algorithm can directly calculate the similarity between users and content according to user interest model and content feature vector. According to different application scenarios and user needs, the weights of the two algorithms are dynamically adjusted to obtain more accurate matching results.

[0291] Push strategy formulation and execution unit: According to the matching results and user historical behavior data, formulate personalized push strategy. For example, for active users, increase the push frequency and the diversity of push content; for inactive users, use some incentives, such as push coupons, hold activities, etc., to improve user engagement. When executing the push strategy, a combination of real-time push and scheduled push is used to ensure that marketing content can be delivered to target users in a timely and accurate manner.

[0292] 6. Feedback and optimization module

[0293] User feedback collection unit: Collect user feedback information on push content, including user click rate, conversion rate, dwell time, evaluation, etc. Through embedding feedback links in push content or using burying point technology, real-time user feedback data is obtained.

[0294] Model optimization unit: According to user feedback data, optimize user interest model, content feature extraction model and user-content matching algorithm. For example, if it is found that a certain user group has a low click rate on a certain type of content, the interest model parameters of this user group can be adjusted; if it is found that a certain content feature has a greater impact on the matching result, the weight of the content feature vector can be adjusted.

[0295] In this embodiment, the beneficial effects include:

[0296] Improve marketing accuracy: Through deep mining of user behavior data and accurate user interest modeling, the user's needs and interests can be better understood, and the content can be accurately pushed, improving the conversion rate of marketing content.

[0297] Improve user experience: Users will only receive content related to their interests, avoiding the interference of a large amount of irrelevant information, thereby improving user experience and increasing user goodwill and loyalty to the enterprise.

[0298] Optimize marketing resource allocation: Accurate marketing push can avoid waste of marketing resources, concentrating limited resources on the most potential user groups and content, improving marketing efficiency and return on investment.

[0299] System adaptive optimization: Through the feedback and optimization module, the system can continuously optimize various models and algorithms according to user feedback, adapt to market changes and dynamic changes in user needs, and maintain the advancement and effectiveness of the system.

[0300] In a specific implementable embodiment, the processing process of the content precise marketing push system based on deep mining of user behavior is as follows:

[0301] Data collection and preprocessing

[0302] Data collection: Configure the interface programs of the data collection module to connect with different data sources, implement automatic collection of user behavior data, and obtain the raw data. At the same time, set the parameters of real-time data collection to ensure the timely acquisition of important user behavior data.

[0303] Data preprocessing: Write scripts for data cleaning and feature extraction to clean, convert, and extract features from the collected raw data. For example, use Python to write a data cleaning script to remove duplicate records and outliers; use TensorFlow to implement a feature extraction model to convert user behavior data into a feature vector.

[0304] User behavior deep analysis and content feature extraction

[0305] User behavior deep analysis: Use the improved K-means clustering algorithm and the user interest modeling method combining the attention mechanism of LSTM to perform clustering analysis and interest modeling on user behavior data. The clustering algorithm can be implemented using Spark's MLlib library, and the deep learning model can be implemented using TensorFlow.

[0306] Content feature extraction: Use the BERT model to classify text content, use CNN to extract multimedia features, and fuse text features and multimedia features. Specifically, use the Transformers library of Hugging Face to implement the BERT model, and use Keras to implement the CNN model.

[0307] Precise push and feedback optimization

[0308] Precise push: Based on the user-content matching algorithm and the push strategy, realize the precise push of content. The push logic can be written using Python, and the push task can be sent to the push server through the message queue (including RabbitMQ).

[0309] Feedback optimization: Collect user feedback data and use machine learning algorithms to optimize each model and algorithm in the system. For example, use the gradient descent algorithm to optimize the parameters of the user interest model, and use the genetic algorithm to optimize the weights of the user-content matching algorithm.

[0310] In Figure 2 the present application, the embodiments provide a content precise marketing push method based on user behavior deep mining, comprising the following steps:

[0311] Collect user behavior data using the data collection module to obtain raw data;

[0312] Clean, convert, and extract features from the raw data using the data preprocessing module to obtain processed data;

[0313] performing clustering analysis on user behaviors based on the processed data, and constructing a user interest model;

[0314] performing classification, tagging, and feature vector representation on marketing content by using a content feature extraction module to obtain a content feature vector;

[0315] performing matching on users and content according to the user interest model and the content feature vector by using a precise pushing module, and formulating and executing an individualized pushing strategy;

[0316] collecting user feedback data and optimizing the user interest model by using a feedback and optimization module.

[0317] In the above technical solution, the data acquisition module is configured to collect user behavior data to obtain raw data; the data preprocessing module is configured to clean, convert, and extract features from the raw data to obtain processed data; the user behavior deep analysis module is configured to perform clustering analysis on user behaviors based on the processed data, and construct a user interest model; the content feature extraction module is configured to perform classification, tagging, and feature vector representation on marketing content to obtain a content feature vector; the precise pushing module is configured to perform matching on users and content according to the user interest model and the content feature vector, and formulate and execute an individualized pushing strategy; and the feedback and optimization module is configured to collect user feedback data and optimize the user interest model, thereby achieving precise marketing content pushing.

[0318] In one specific implementable embodiment, the method flow includes:

[0319] Data acquisition: The multi-source data access unit and the real-time data acquisition unit of the data acquisition module collect user behavior data from various channels.

[0320] Data preprocessing: The data cleaning unit and the data conversion and feature extraction unit of the data preprocessing module clean, convert, and extract features from the collected raw data.

[0321] User behavior deep analysis: The user behavior clustering unit and the user interest modeling unit of the user behavior deep analysis module perform clustering analysis on user behaviors, and construct a user interest model.

[0322] Content feature extraction: The content classification and tagging unit and the content feature vector representation unit of the content feature extraction module perform classification, tagging, and feature vector representation on marketing content.

[0323] Precise pushing: The user-content matching unit and the pushing strategy formulation and execution unit of the precise pushing module are used to match users with content according to user interest models and content feature vectors, and to formulate and execute personalized pushing strategies.

[0324] Feedback and optimization: The user feedback collection unit and the model optimization unit of the feedback and optimization module are used to collect user feedback data and optimize various models and algorithms in the system.

[0325] In this embodiment, the beneficial effects include:

[0326] Improved marketing accuracy: By deeply mining user behavior data and precisely modeling user interests, the system can more accurately understand user needs and interests, implement precise content pushing, and improve the conversion rate of marketing content.

[0327] Improved user experience: Users only receive content related to their interests, avoiding interference from a large amount of irrelevant information, thereby improving user experience and increasing user goodwill and loyalty to the enterprise.

[0328] Optimized marketing resource allocation: Precise marketing pushing can avoid waste of marketing resources, focusing limited resources on the most promising user groups and content, improving marketing efficiency and return on investment.

[0329] System adaptive optimization: Through the feedback and optimization module, the system can continuously optimize various models and algorithms based on user feedback, adapt to market changes and dynamic changes in user needs, and maintain the advancement and effectiveness of the system.

[0330] Those skilled in the art know that the present application can be implemented as a system, a method or a computer program product.

[0331] Therefore, the present disclosure can be specifically implemented in the following forms: it can be complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, which is generally referred to herein as "circuitry", "module" or "system". In addition, in some embodiments, the present application can also be implemented in the form of a computer program product in one or more computer readable media, which contains computer readable program code.

[0332] Any combination of one or more computer readable medium can be utilized. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. A computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer readable storage medium can be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.

[0333] Although the embodiments of the present application have been shown and described above, it should be understood by those skilled in the art that the above embodiments are exemplary, and cannot be understood as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application. On this basis, various replacements and improvements can be made to the present application, and these all fall within the protection scope of the present application.

Claims

1. A content precision marketing pushing system based on deep mining of user behavior, characterized in that, The application relates to a user behavior analysis and content recommendation system, which comprises the following modules: a data collection module for collecting user behavior data to obtain original data; a data preprocessing module for cleaning, converting and feature extracting the original data to obtain processed data; a user behavior deep analysis module for clustering analysis of user behavior based on the processed data and constructing a user interest model; the user interest model comprises user preference degree information for different types of content, interest persistence information and interest change trend information; a content feature extraction module for classifying, tagging and feature vector representing marketing content to obtain content feature vectors; the marketing content comprises commodity information, service information, article information and video content; the content feature vectors comprise content subject categories, keyword appearance frequencies and emotional tendencies; a precise pushing module for matching users and content according to the user interest model and the content feature vectors and formulating and executing individualized pushing strategies; the matching process comprises: obtaining user interest preference information for different types of content by using the user interest model; at the same time, the similarity or matching degree between the user interest model and the content feature vectors is calculated by a similarity calculation algorithm according to the content feature vectors and the interest preference information; the pushing strategy comprises pushing time, pushing frequency, pushing channel, pushing content and individualized content customization; a feedback and optimization module for collecting user feedback data and optimizing the user interest model. 2.The content precision marketing pushing system based on user behavior deep mining according to claim 1, characterized in that, The data collection module comprises: a multi-source data access unit for collecting user behavior data from multiple channels; the user behavior data comprises user operation records of enterprise self-owned platforms, user interaction data on social media platforms and user portrait data provided by third-party data service providers; a real-time data collection unit for collecting user behavior data with time effectiveness requirements in real time; the user behavior data with time effectiveness requirements comprises real-time click data and purchase operation data of users. 3.The content precision marketing pushing system based on user behavior deep mining according to claim 2, characterized in that, The data preprocessing module comprises: a data cleaning unit for cleaning the original data; a data conversion and feature extraction unit for converting the cleaned original data and extracting key features of user behavior. 4.The content precision marketing pushing system based on user behavior deep mining according to claim 3, characterized in that, The user behavior deep analysis module comprises: a user behavior clustering unit for clustering analysis of user behavior by using an improved K-means clustering algorithm to obtain clustering results; the clustering results comprise classifying users with similar behavior patterns into one category to form different user groups; a user interest modeling unit for constructing the user interest model by using a long short-term memory network combined with an attention mechanism based on the clustering results and the key features of user behavior; the user interest model is used for processing time sequence data, capturing the change trend of user interest with time and automatically focusing on user interest behavior and features. 5.The content precision marketing pushing system based on user behavior deep mining according to claim 4, characterized in that, The content feature extraction module comprises: The content classification and labeling unit is configured to classify and label marketing content; a deep learning-based text classification model is used to classify text content, and a manual labeling and semi-automatic labeling method is used to add detailed labels to the content; the labels include theme labels, style labels, and sentiment labels; The content feature vector representation unit is configured to convert the classified and labeled marketing content into a content feature vector; the content feature vector includes a theme category of the content, a keyword appearance frequency, and an emotional tendency. 6.The content precision marketing pushing system based on user behavior deep mining according to claim 5, characterized in that, The content classification and labeling unit uses a deep learning-based text classification model to classify text content, and a manual labeling and semi-automatic labeling method is used to add labels to the marketing content. 7.The content precision marketing pushing system based on user behavior deep mining according to claim 6, characterized in that, The content feature vector representation unit uses a word embedding algorithm to convert text content in the marketing content into a word vector, and then uses an average pooling or maximum pooling method to combine the word vectors into the content feature vector. 8.The content precision marketing pushing system based on user behavior deep mining according to claim 7, characterized in that, The precise pushing module includes: A user-content matching unit configured to use a method combining a weighted fusion-based collaborative filtering algorithm and a content-based recommendation algorithm to perform user-content matching; including: Using a collaborative filtering algorithm to find marketing content that other users similar to the target user are interested in; using a content-based recommendation algorithm to calculate the similarity between the user and the content according to the user interest model and the content feature vector; and dynamically adjusting the weights of the weighted fusion-based collaborative filtering algorithm and the content-based recommendation algorithm to obtain a user-content matching result; A pushing strategy formulation and execution unit configured to formulate a personalized pushing strategy according to the user-content matching result and historical behavior data of the user; The pushing strategy includes: selecting a time period when the user is most likely to view information for pushing according to the active time regularity of the user; Reasonably controlling the pushing frequency according to the acceptance and feedback of the user to the content; Selecting a suitable pushing channel according to the usage habits and preferences of the user; Combining and pushing multiple related contents according to the user interest model; Customizing the marketing content for individual users according to their interest characteristics before pushing. 9.The content precision marketing pushing system based on user behavior deep mining according to claim 8, characterized in that, The feedback and optimization module includes: A user feedback collection unit configured to collect feedback information of the user to the pushed content; the feedback information includes the click rate, conversion rate, dwell time, and user evaluation of the user; A model optimization unit configured to optimize the user interest model according to the user feedback data, including dynamically adjusting the parameters of the user interest model and the clustering weights of the content feature vector.

10. A content precision marketing pushing method based on deep mining of user behavior, characterized in that, The method includes the following steps: Collecting user behavior data using a data collection module to obtain raw data; Using a data preprocessing module to clean, convert, and extract features from the raw data to obtain processed data; Using a user behavior deep analysis module to perform clustering analysis on user behavior based on the processed data and construct a user interest model; Using a content feature extraction module to classify, label, and represent the features of marketing content to obtain a content feature vector; The precise pushing module is used for matching the user and the content according to the user interest model and the content feature vector, and formulating and executing a personalized pushing strategy; The feedback and optimization module is used for collecting user feedback data and optimizing the user interest model.

Images