A method for controlling the push of operational data on an O2O platform

Abstract

The invention discloses a method for controlling the push of O2O platform operation data, relates to the technical field of data push control, and is used to solve the problem of low efficiency of the push control of O2O platform operation data. The method comprises the following steps: collecting user behavior data through multi-source data integration, extracting key context features from the data, fusing the features through a word vector model and vector splicing, then using a hierarchical clustering technology to identify user groups with similar behavior and context features, analyzing the association rules between user behavior and context features, refining user behavior patterns, combining behavior patterns and user preferences, adjusting the timing of content push, and finally, judging the matching between the pushed content and the actual user behavior preference through matching evaluation of the user behavior pattern and preference features, optimizing the push strategy according to the matching result, thereby improving the user experience and enhancing the user participation, and effectively optimizing the resource allocation.

Description

Technical Field

[0001] The present invention relates to the technical field of data push control, and in particular to an O2O platform operation data push control method. Background Art

[0002] O2O platform operation data push control refers to an automated push strategy that effectively manages and controls information, advertisements, recommendations or notifications in an online to offline (O2O) business platform. This push control aims to optimize data collection, processing and analysis so as to push relevant information to users more accurately and improve user experience and business results.

[0003] The existing technology has the following deficiencies: The main deficiencies in the field of data push in O2O platform operations include content relevance, personalization, and push timeliness. The current push system often relies on basic user portraits and historical behavior data, resulting in insufficient personalization and relevance of the pushed content, which easily makes users feel that the information does not match their actual needs, thereby reducing user participation and satisfaction. In addition, the timeliness of push and the grasp of precise timing also affect the efficiency of push. Many pushes cannot accurately predict the best push time, and often push when it is inconvenient for users to receive information, such as working hours or late at night. This inaccurate timeliness may not only reduce the effect of the pushed content, but also have a negative impact on the user experience.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0005] The purpose of the present invention is to provide an O2O platform operation data push control method to solve the problem of inaccurate timeliness of content push by the O2O platform in the background technology.

[0006] In order to achieve the above object, the present invention provides the following technical solution: a method for controlling the push of O2O platform operation data, comprising the following steps:

[0007] Integrate multi-source data, collect user data from the O2O platform, extract contextual features from the collected user data, extract the time and location features of user activities, and use the word vector model for feature fusion and vector splicing;

[0008] The timeliness of content push is analyzed based on user behavior and context patterns, and hierarchical clustering is used to identify user groups with similar behavior and context characteristics. The Apriori algorithm is used to determine the association rules between user behavior patterns and context characteristics.

[0009] Determine user behavior patterns through association rules, adjust content push timing based on user behavior and preference characteristics, and obtain matching evaluation information generated during the matching analysis of user behavior patterns and user behavior and preference characteristics;

[0010] Analyze the matching evaluation information to determine the match between the pushed content and the user's actual behavior preferences, and control the push strategy based on the matching results.

[0011] Preferably, context feature extraction is performed on the collected user data to extract the time and location features of the user activities, and a word vector model is used to perform feature fusion and vector splicing, including:

[0012] Collect users' geographic location data and corresponding activity records, as well as relevant external data, including weather and holidays;

[0013] Perform data preprocessing, standardize location and activity descriptions, and encode weather and holidays as numerical values ​​or one-hot encoded vectors;

[0014] The Skip-gram model is used to maximize the logarithmic probability of context words. The location vector and activity vector of the user data are determined based on the logarithmic probability. Feature fusion and vector concatenation are then performed. The encoding vectors of weather and holidays are concatenated with the location vector and activity vector of the user data to obtain a fused feature vector.

[0015] Preferably, the timeliness analysis of content push is performed based on user behavior and context patterns, and hierarchical clustering is combined to identify user groups with similar behavior and context characteristics. The specific steps include:

[0016] Preprocess multi-source data, clean up missing values, outliers, and incorrectly formatted data, and standardize date and time formats;

[0017] Select variables of user behavior and contextual characteristics, including purchase history, location visit frequency, and weather conditions;

[0018] Conduct exploratory analysis on the distribution of user behavior at different times and locations. The exploratory analysis includes univariate analysis using histograms and box plots to view the distribution and bivariate analysis using scatter plots, line graphs, and heat maps to analyze the relationship between variables. Create new features based on the exploratory analysis results.

[0019] Perform hierarchical clustering analysis on the new features, use Euclidean distance to build a distance matrix, and calculate the distance between all pairs of data points;

[0020] Find the two clusters with the closest distance, merge the two clusters into a new cluster, and after the merger, update the distance matrix to determine the distance between the new cluster and the original cluster;

[0021] Repeat merging clusters and updating the distance matrix to merge all data points into one cluster or reach the required number of clusters as the clustering result;

[0022] Based on the clustering results, the common behaviors and content preferences of each group are analyzed to identify user groups with similar behavior patterns and contextual characteristics.

[0023] Preferably, the Apriori algorithm is used to determine the association rules between the user behavior pattern and the context features, and the specific steps are as follows:

[0024] Step 1: Set the minimum support, that is, the minimum frequency that the feature item set must meet;

[0025] Step 2, generate candidate item sets, that is, start from a single item and gradually expand to a larger candidate item set;

[0026] Step 3: Calculate the support of all candidate feature item sets in the database, and remove the feature item sets whose support is lower than the minimum support threshold;

[0027] Use the remaining item sets to generate candidate item sets, and repeat step 3 until no more frequent item sets can be generated;

[0028] Generate association rules from frequent itemsets and calculate the confidence and lift of each rule.

[0029] Preferably, the user behavior pattern is determined by association rules, and the content push timing is adjusted in combination with the user behavior and preference characteristics, and the matching evaluation information generated in the matching analysis process of the user behavior pattern and the user behavior and preference characteristics is obtained. The specific steps are as follows:

[0030] Obtaining matching evaluation information generated during the matching analysis of user behavior patterns and user behavior and preference characteristics, the matching evaluation information includes user participation information and content adaptation information;

[0031] User participation information includes dynamic user behavior synchronization index, and content adaptation information includes content push optimization index:

[0032] The dynamic user behavior synchronization index and content push optimization index are combined to generate a push control coefficient;

[0033] The dynamic user behavior synchronization index and content push optimization index are directly proportional to the push control coefficient.

[0034] Preferably, the matching evaluation information is analyzed to determine the matching between the pushed content and the user's actual behavior preference, and the push strategy is controlled according to the obtained matching result. The specific steps include:

[0035] Compare the generated push control coefficient with the push control threshold;

[0036] If the push control coefficient is greater than the push control threshold, a push content adaptation signal is generated, the user behavior data involved in generating the push content adaptation signal is annotated, and the annotated data is used as training data for the machine learning model;

[0037] If the push control coefficient is less than or equal to the push control threshold, a push control adjustment signal is generated, the user data is re-analyzed, the user's behavior and preferences are evaluated, and the time of content push is adjusted.

[0038] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0039] The present invention collects user behavior data through multi-source data integration, extracts key contextual features from it, and fuses and vectorizes the features through a word vector model to form a comprehensive user feature representation, and then uses hierarchical clustering technology to identify user groups with similar behavior and contextual features, thereby enhancing the personalization and accuracy of content push. In addition, the association rules between user behavior and contextual features are analyzed through the Apriori algorithm, and user behavior patterns are further refined. In combination with these behavior patterns and the specific preferences of users, the timing of content push is dynamically adjusted to improve its timeliness and relevance. Finally, by evaluating the matching of user behavior patterns and preference features, the match between the pushed content and the user's actual behavior preferences can be accurately judged. Finally, the push strategy is optimized based on this matching result to ensure that each content push efficiently reaches the target user, improves the user experience and enhances user participation, and effectively optimizes resource allocation. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0041] Figure 1 The present invention is a flow chart of an O2O platform operation data push control method. DETAILED DESCRIPTION

[0042] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of the present disclosure will be more comprehensive and complete, and the concept of the example embodiments will be fully conveyed to those skilled in the art.

[0043] Embodiment 1: The present invention provides Figure 1 The method for controlling the push of O2O platform operation data shown comprises the following steps:

[0044] Integrate multi-source data, collect user data from the O2O platform, extract contextual features from the collected user data, extract the time and location features of user activities, and use the word vector model for feature fusion and vector splicing;

[0045] The timeliness of content push is analyzed based on user behavior and context patterns, and hierarchical clustering is used to identify user groups with similar behavior and context characteristics. The Apriori algorithm is used to determine the association rules between user behavior patterns and context characteristics.

[0046] Determine user behavior patterns through association rules, adjust content push timing based on user behavior and preference characteristics, and obtain matching evaluation information generated during the matching analysis of user behavior patterns and user behavior and preference characteristics;

[0047] Analyze the matching evaluation information to determine the match between the pushed content and the user's actual behavior preferences, and control the push strategy based on the matching results.

[0048] Integrate multi-source data to collect user data on the O2O platform, including user behavior data (such as browsing, purchase history, clicks, etc.), social media activities, geographic location data, device usage data, external data, etc., and use sensor data from smartphones or other smart devices (such as accelerometers, GPS) to obtain the activity patterns and context of platform users; collect public data of O2O platform users through APIs of social media and other external services, and deploy SDKs within O2O platform applications to automatically collect device and application usage data;

[0049] Enhance user portraits to more accurately predict user behaviors and preferences through data analysis and feature engineering. The specific implementation methods include data collection expansion, feature engineering optimization, and dynamic portrait updates. The specific steps are as follows:

[0050] Contextual feature extraction is performed on the collected user data to extract the time and location features of user activities. External data (such as weather and holiday information) is combined to enhance context understanding. High-dimensional embedding technology, such as the word vector model (Word2Vec), is used to represent the location and activity type, and the category data is converted into a numerical embedding vector. The specific process is as follows:

[0051] Collect user's location data and corresponding activity records, as well as relevant external data, such as local weather conditions, holidays and special events;

[0052] Perform data preprocessing to encode locations and activity types in preparation for Word2Vec embedding, and clean external data, such as classifying weather conditions (sunny, rainy, snowy, etc.), standardizing location and activity descriptions, such as converting to lowercase, removing special characters, etc., and performing word segmentation (if using Chinese data, word segmentation is required);

[0053] Encode weather and holidays information as numeric values ​​or one-hot encoding vectors. For example, weather can be encoded as: sunny = [1, 0, 0], rainy = [0, 1, 0], snowy = [0, 0, 1], and holidays can be binary flags, such as weekdays = 0, holidays = 1;

[0054] Train the Word2Vec model and set model parameters, such as vector dimension, window size, minimum word frequency, etc. Input the processed "sentence" list into the model for training. There are two architectures of the Word2Vec model. The continuous bag of words model (CBOW) predicts the target word based on the context, and the Skip-gram model uses a word to predict the context. For location and activity type embedding, the Skip-gram model is usually more suitable because the Skip-gram model works better for small data sets and can better handle rare words.

[0055] In the Skip-gram model, given a specific word w t , the goal of the model is to maximize the log probability of the context word, the formula is as follows: Where T is the length of the sentence, c is the context window size, and w t+j is the word w at time t t surrounding context words;

[0056] The probability p(w t+j |w t ) is the approximate calculation process defined by the softmax function, and the expression is:

[0057] Among them, v wi , v′ wo The input and output vectors of word wi (input word) and word wo (output word) respectively, W is the vocabulary size;

[0058] Perform feature fusion and vector concatenation to directly combine the weather and holiday encoding vectors with the user data location vector v location and the activity vector v activity Concatenate to get the fused feature vector: v com =[v location , v activity , v weather, v holiday ], where v weather 、v holiday is the embedding vector obtained by Word2Vec;

[0059] Contextual feature extraction, especially combining the time and location characteristics of user activities and external data (such as weather and holiday information), can enhance the personalization and timeliness of relevant push notifications. Using high-dimensional embedding technologies such as Word2Vec to represent locations and activity types can more effectively capture the complex relationships and intrinsic semantics of these categories of data. Using word vector models for feature fusion and vector splicing can not only enrich the expressiveness of the data, but also greatly enhance the ability of machine learning models to process language and text data.

[0060] Analyze the timeliness of content push, that is, integrate multi-source data and analyze behavior and context patterns. The specific steps are as follows:

[0061] Preprocess the integrated multi-source data to clean up missing values, outliers and data with incorrect formats, and standardize date and time formats;

[0062] Select variables that reflect user behavior and contextual characteristics, such as purchase history, location visit frequency, weather conditions, etc., and encode and standardize the characteristics appropriately to suit subsequent cluster analysis;

[0063] Conduct exploratory analysis, use visualization tools to view the distribution of user behavior at different times and locations, and analyze the relationship between various features and user behavior. Exploratory analysis includes:

[0064] Univariate analysis is to perform descriptive statistics and visualization on each individual variable, such as using histograms and box plots to view distribution;

[0065] Bivariate analysis is to look at the relationship between two variables by analyzing the relationship between variables, such as using scatter plots, line graphs, and heat maps (for correlation analysis);

[0066] The user behavior distribution over time and location uses a time series chart to show the trend of user behavior at different times, and the location data can be displayed through a geographic heat map;

[0067] Based on the results of exploration and analysis, new features are created, such as time feature derivation, where the "time of day" derived from the timestamp is classified into "morning", "noon", "evening", etc., or the day of the week is extracted from the date and converted into a categorical variable "day of the week", such as "weekday" and "weekend", etc.; location feature derivation: If the location information is continuous latitude and longitude, it can be considered to be clustered into geographical areas such as "city center" and "suburbs"; combined feature creation: Combine existing features based on business understanding, such as combining "weather" and "time of day" to predict "user activity type";

[0068] K-means clustering and hierarchical clustering are used to identify user groups with similar behavior and context characteristics. The specific steps are as follows:

[0069] Use the elbow method to determine the number of clusters K, that is, observe the changes in the sum of squares of the total within-cluster (WCSS) under different K values, select the point where the WCSS drops sharply and begins to flatten out as the K value, and calculate the sum of squares of the distance from each point to the cluster center: Among them, ||x-μ i || represents the point x and the cluster center μ i The Euclidean distance between i is the i-th cluster;

[0070] Hierarchical clustering is a method of constructing a clustering hierarchy by gradually merging or splitting clusters. This method does not require the number of clusters to be specified in advance and is more suitable for the analysis of exploratory data, that is, performing hierarchical clustering analysis on new features. The specific steps are as follows:

[0071] Initialize, treating each data point as a separate cluster;

[0072] Construct a distance matrix and calculate the distance between all pairs of data points. Usually, Euclidean distance is used, but Manhattan distance or other distance metrics can also be selected according to the characteristics of the data.

[0073] Merge the nearest clusters, that is, find the two clusters with the closest distance and merge the two clusters into a new cluster;

[0074] Update the distance matrix, that is, after merging, update the distance matrix to reflect the distance between the new cluster and other clusters. The distance update rules include: the distance between clusters is defined as the distance between the two closest points in the cluster as the shortest distance; the distance between clusters is defined as the distance between the two farthest points in the cluster as the longest distance; the distance between clusters is defined as the average distance of all point pairs;

[0075] Repeat merging the nearest clusters and updating the distance matrix until all data points are merged into one cluster or the required number of clusters is reached;

[0076] Based on the clustering results, analyze the common behaviors and content preferences of each group, and identify user groups with similar behavior patterns and contextual characteristics;

[0077] The association rule Apriori algorithm is used to determine the correlation between user behavior patterns and context features, that is, to find frequent feature item sets in the data, and then use these feature item sets to generate strong association rules. For example, users are more likely to buy a certain type of goods on rainy days. The specific steps are as follows:

[0078] Step 1: Set the minimum support, that is, the minimum frequency that the feature item set must meet;

[0079] Step 2, generate candidate item sets, that is, start from a single item and gradually expand to a larger candidate item set;

[0080] Step 3: Calculate the support of all candidate feature item sets in the database, and remove the feature item sets whose support is lower than the minimum support threshold;

[0081] Use the remaining itemsets to generate larger candidate itemsets and repeat step 3 until no larger frequent itemsets can be generated;

[0082] Generate association rules from frequent item sets and calculate the confidence and lift of each rule;

[0083] For example, support indicates the frequency of a certain feature item set (combination of goods) appearing in all transactions;

[0084] Support is an indicator to measure the universality of feature item sets, and the expression is: Among them, X jy is the number of transactions containing feature item set X, X zs is the total number of all transactions;

[0085] Confidence indicates the conditional probability that a transaction containing feature item set X also contains feature item set Y. It is used as an indicator to measure the strength of rule x→Y. The specific formula is: Con(X→Y)=Sup(X∪Y) / Sup(X), where Sup(X∪Y) is the proportion of transactions containing both item sets X and Y to the total number of transactions;

[0086] Lift represents the ratio of the frequency of item sets X and Y appearing together to the frequency of their independent occurrence. A lift greater than 1 means that the occurrence of X and Y has a positive correlation, that is, the occurrence of one increases the probability of the occurrence of the other. The formula is: Lift(X→Y)=Con(X→Y) / Sup(X);

[0087] The above indicators are used in combination to evaluate the quality of association rules. Support helps identify which item sets are common, confidence tells the strength of dependency between feature item sets, and lift provides additional insights into the association between item sets. In practical applications, thresholds are usually set to filter out association rules with both high support, high confidence, and high lift, to ensure that the discovered rules are both statistically significant and practically significant.

[0088] Determine user behavior patterns through association rules and combine user behavior and preference characteristics to accurately adjust the timing of content push, thereby optimizing the push strategy. The specific steps are as follows;

[0089] Obtain the matching evaluation information generated during the matching analysis of user behavior patterns and user behavior and preference characteristics, analyze the matching evaluation information, determine the matching between the pushed content and the user's actual behavior preferences, identify which features or data points most accurately predict user behavior, and be able to foresee and adapt to changes in user preferences, thereby maximizing the long-term effect of the push strategy. The matching evaluation information includes user participation information and content adaptation information;

[0090] User participation information includes the dynamic user behavior synchronization index, and content adaptation information includes the content push optimization index. After collection, the dynamic user behavior synchronization index and content push optimization index are calibrated as DUBSI and CDOI respectively.

[0091] The dynamic user behavior synchronization index is used to measure the degree of synchronization between content push strategies and changes in user behavior. It captures and analyzes time series data of user behavior to evaluate the alignment between push timing and changes in user behavior.

[0092] User behavior change points: Identify the causes of inefficiency. Through time series analysis and variable point detection, the dynamic user behavior synchronization index identifies significant changes in user behavior (such as purchases, browsing, event participation, etc.). These changes may be due to external events, marketing activities, or changes in user preferences;

[0093] Synchronicity between push timing and behavior changes: Analyze the time difference between the time of content push and the time of behavior change to evaluate the timeliness and relevance of content push;

[0094] Optimize push effects: By optimizing push timing to better sync with key changes in user behavior, user engagement and the effectiveness of push content can be improved;

[0095] This in turn improves user engagement and satisfaction. By ensuring that content is pushed when users are most interested or most in need, the dynamic user behavior synchronization index helps increase user engagement and satisfaction, and increases the relevance of content perceived by users, which may increase positive evaluations and loyalty to the brand.

[0096] The logic for obtaining the dynamic user behavior synchronization index is as follows:

[0097] Set the time window W and obtain the collected time series data Y = {y1, y1, ..., y T}, T is the total amount of data, and the posterior probability that any time point t is a change point is calculated: Where P(Y|t) is the probability of the observed data Y appearing under the assumption that t is a change point, P(t) is the prior probability that time point t becomes a change point in the time series, and P(Y) is the probability of observing the entire data sequence Y without a specific change point position; get the number of change points N in the actual window W change , calculate the change point density D, the calculation expression is: D = N change / |W|;

[0098] Determine the change point by setting the change point probability threshold. For each change point t change , determine the most recent content push time t push , and calculate the time difference Δt i , Δt i Indicates the i-th change point, and calculates the synchronization time difference of the change point. The calculation expression is: Where n is the total number of change points. The dynamic user behavior synchronization index is calculated using the following expression:

[0099]

[0100] It should be noted that in time series analysis, change points are obtained through Bayesian change point analysis; a change point refers to the location where the statistical properties of the data (such as mean, variance) change. This change usually represents a structural change in the underlying process, such as a sudden shift in system behavior, a change in environmental conditions, or an adjustment in operating strategy. It is specific to time point t, meaning that the data characteristics before and after t are significantly different. For example, after a website launches a new advertisement, the user click rate may suddenly increase. This point is a change point. By identifying the change point of user behavior (such as a sudden increase in interest in a certain type of content), the content push strategy can be adjusted more accurately to improve user engagement and satisfaction; the time window W is set according to actual needs, such as one day or one week, to observe the distribution of change points; a change point probability threshold is determined to determine whether a certain time point t is considered a change point. This change point probability threshold is usually set according to the needs of the specific application and the tolerance for false positives and false negatives.

[0101] The content push optimization index is used to evaluate and optimize the effectiveness of content push strategies. It determines the content push status by comprehensively considering multiple dimensions such as timeliness, relevance, and user engagement. The timeliness score reflects the degree of match between content push and user demand timing, that is, whether the content is pushed when the user needs it most. By considering the degree of match between content and user preferences and behaviors (relevance score), it determines whether the content meets the user's interests and needs. By evaluating and optimizing various aspects of content push, it ensures that the content can reach the right users at the right time and in the right way, thereby maximizing marketing effects and user satisfaction, and affecting the following aspects:

[0102] Improve user experience: By ensuring the relevance and timeliness of content push, the content push optimization index helps improve user satisfaction with content and overall experience. When content is highly matched with users' real-time needs and preferences, users are more likely to continue to engage and maintain a positive attitude towards the brand;

[0103] Resource allocation efficiency: The comprehensive score provided by the content push optimization index supports data-based decision-making, making content push strategies more scientific and precise. This performance indicator-based approach reduces the impact of guesswork and intuition, and enhances the predictability of strategies and consistency of results.

[0104] The logic for obtaining the content push optimization index is as follows:

[0105] Get the frequency n of user j's interaction with each content type k j，k , get the total number of interactions C of user j j , use entropy to calculate the entropy value of user content rating, the calculation expression is:

[0106] Where K is the total number of content types, and the content weight value is calculated by the entropy value of different user content ratings. The calculation expression is:

[0107] Convert the extracted topic label features of the content pushed to user j into content feature vector r j , the dwell time of user j on the pushed content is used as a feature and converted into a preference vector Use cosine similarity to calculate the matching value. The calculation expression is:

[0108] Get the average time difference Δt from content type generation to user interaction avg , calculate the content push optimization index, the calculation expression is: In the formula, J represents the total number of users.

[0109] It should be noted that content types include videos, articles, pictures, etc.; text analysis technology is used to extract the TF-IDF weight of topic tags, or sentiment analysis tools are used to assess sentiment tendencies and then determine content feature weights.

[0110] The dynamic user behavior synchronization index and content push optimization index are normalized and calculated together to obtain the push control coefficient, which is calibrated as K x , based on the formula:

[0111] In the formula, α and γ are the preset proportional coefficients of the dynamic user behavior synchronization index and the content push optimization index, respectively, and both α and γ are greater than 0;

[0112] It should be noted that the formula is obtained by technicians in this field by collecting multiple groups of sample data and setting corresponding preset proportional coefficients for each group of sample data. The specific calculation process is: substituting the obtained dynamic user behavior synchronization index and content push optimization index into the formula to form an equation group, and then screening and averaging the calculated preset proportional coefficients to obtain the preset proportional coefficients. Since each group of sample data is different, the specific values ​​of the preset proportional coefficients under different sample data are also different.

[0113] From the formula, we can see that the larger the dynamic user behavior synchronization index is, the larger the content push optimization index is, that is, the larger the value of the push control coefficient is, indicating that the content push not only responds to the changes in user behavior in a timely manner, but also highly meets the personalized needs of users. The push strategy attracts the attention of users and stimulates their participation, which usually means higher user satisfaction and better user experience. It can also indicate the efficiency of marketing activities in resource allocation and content delivery, and effectively push content to users who are most likely to respond positively.

[0114] By analyzing the push control coefficient, the marketing team can better evaluate and adjust the push strategy, ensure the effectiveness and timeliness of content push, and provide content that is synchronized with user behavior and highly relevant, which can enhance users' brand loyalty and overall satisfaction. Effective push control coefficient guidance can help maximize marketing investment returns and improve conversion rates and user engagement through precise push.

[0115] The smaller the dynamic user behavior synchronization index and the content push optimization index, that is, the smaller the performance value of the push control coefficient, it indicates that there are efficiency problems in the resource allocation and execution of the content push strategy, resulting in marketing resources being wasted on low-effect pushes;

[0116] A low performance value of the user behavior synchronization index may mean that the pushed content does not match the user's personal preferences and actual needs to a high degree, which may be caused by the inaccuracy of the content recommendation system or the inaccuracy of the user preference model;

[0117] A low performance value of the dynamic user behavior synchronization index may also reflect that users are not very engaged with the pushed content, indicating that the content may not be attractive enough or fail to stimulate users' interest and response.

[0118] By jointly calculating the dynamic user behavior synchronization index and the content push optimization index to obtain the push control coefficient, we can get a comprehensive indicator for push analysis:

[0119] The generated push control coefficient is compared with the push control threshold, and different signals are generated according to the comparison result, and the different signals are respectively generating a push control adjustment signal and a push content adaptation signal;

[0120] After obtaining the generated push control coefficient, the generated push control coefficient is compared with the push control threshold. If the push control coefficient is greater than the push control threshold, a push content adaptation signal is generated, indicating that the current user behavior pattern is more closely matched with the user behavior and preference characteristics, and the selected content push strategy has reached or exceeded the predetermined standards in terms of timeliness, relevance and user engagement. This means that the push content is highly matched with user needs, and the push time is accurate, effectively capturing the key change points of user behavior.

[0121] A high push control coefficient that exceeds the push control threshold and generates a push content adaptation signal is a positive indicator that can optimize current marketing activities and also provide long-term strategic insights.

[0122] If the push control coefficient is less than or equal to the push control threshold, a push control adjustment signal is generated, indicating that the user behavior pattern does not match the user behavior and preference characteristics. The content may not be pushed at the time when the user is most interested or most needed, and the pushed content may not fully consider the user's personal preferences, past behaviors or real-time contextual information, resulting in a disconnect between the content and user needs. Inefficient push strategies may cause marketing resources to fail to produce the expected results, increase costs but fail to bring corresponding returns:

[0123] When push control adjustment signals are generated, re-analyze user data, re-evaluate user behavior and preferences, adjust the customization and personalization level of content based on these insights, re-analyze the time pattern of user behavior, adjust the time of content push to better match the user's real-time needs and behavior patterns, and make corresponding strategy adjustments;

[0124] When generating a push content adaptation signal, collect and annotate the user behavior data involved in generating the push content adaptation signal, including the user's response to the push content (clicks, comments, sharing, etc.), detailed features of the push content (type, release time, content characteristics, etc.), for example, the user behavior pattern determined by the association rules analyzed above and the user behavior and preference features;

[0125] Choose a suitable machine learning model based on the characteristics of the problem. For non-sequential recommendations, you can use a deep neural network (DNN) or a gradient boosting machine (GBM). For those that need to consider time series, you can use an LSTM or GRU model.

[0126] Use the training data of labeled data to train the selected model and adjust the parameters of the model. Deploy the trained model to the actual push system for real-time analysis and prediction of user reactions to the content to be pushed. Update and optimize the system in real time based on newly collected data. Adjust content creation and push strategies based on the analysis results, such as pushing specific types of content to specific user groups.

[0127] Through these labeled training data, the model can better understand which types of content best match the behavior and reaction patterns of specific user groups, making push strategies more refined and personalized, and able to optimize for users' specific needs and preferences, thereby improving user engagement and satisfaction;

[0128] Using behavioral data that has successfully matched user needs to train the model can directly reflect which factors lead to successful user engagement. Such data provides high-quality input, which helps the model learn to more accurately predict user behavior and preferences, thereby improving the accuracy of the overall prediction.

[0129] It should be noted that the relevant threshold information in this embodiment is pre-set by professionals and will not be explained in detail here.

[0130] The present invention integrates multi-source data, collects user behavior data, extracts key contextual features from it, and fuses and vectorizes the features through a word vector model to form a comprehensive user feature representation, and then uses hierarchical clustering technology to identify user groups with similar behavior and contextual features, thereby enhancing the personalization and accuracy of content push. In addition, the association rules between user behavior and contextual features are analyzed through the Apriori algorithm, and user behavior patterns are further refined. In combination with these behavior patterns and the specific preferences of users, the timing of content push is dynamically adjusted to improve its timeliness and relevance. Finally, by evaluating the matching of user behavior patterns and preference features, the match between the pushed content and the user's actual behavior preferences can be accurately judged. Finally, the push strategy is optimized based on this matching result to ensure that each content push efficiently reaches the target user, improves the user experience and enhances user participation, and effectively optimizes resource allocation.

[0131] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0132] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0133] It should be noted that, in this article, if there are relational terms such as first and second, etc., they are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0134] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for controlling the push of O2O platform operation data, characterized in that: The steps include: Integrate multi-source data, collect user data from the O2O platform, extract contextual features from the collected user data, extract the time and location features of user activities, and use the word vector model for feature fusion and vector splicing; The timeliness of content push is analyzed based on user behavior and context patterns, and hierarchical clustering is used to identify user groups with similar behavior and context characteristics. The Apriori algorithm is used to determine the association rules between user behavior patterns and context characteristics. Determine user behavior patterns through association rules, adjust content push timing based on user behavior and preference characteristics, and obtain matching evaluation information generated during the matching analysis of user behavior patterns and user behavior and preference characteristics; Analyze the matching evaluation information to determine the matching between the pushed content and the user's preference characteristics, and control the push strategy based on the matching results; The timeliness of content push is analyzed based on user behavior and context patterns, and hierarchical clustering is used to identify user groups with similar behavior and context characteristics. The specific steps include: Preprocess multi-source data, clean up missing values, outliers, and incorrectly formatted data, and standardize date and time formats; Select variables of user behavior and contextual characteristics, including purchase history, location visit frequency, and weather conditions; Conduct exploratory analysis on the distribution of user behavior at different times and locations. The exploratory analysis includes univariate analysis using histograms and box plots to view the distribution and bivariate analysis using scatter plots, line graphs, and heat maps to analyze the relationship between variables. Create new features based on the exploratory analysis results. Perform hierarchical clustering analysis on the new features, use Euclidean distance to build a distance matrix, and calculate the distance between all pairs of data points; Find the two clusters with the closest distance, merge the two clusters into a new cluster, and after the merger, update the distance matrix to determine the distance between the new cluster and the original cluster; Repeat merging clusters and updating the distance matrix to merge all data points into one cluster or reach the required number of clusters as the clustering result; Based on the clustering results, analyze the common behaviors and content preferences of each group, and identify user groups with similar behavior patterns and contextual characteristics; The Apriori algorithm is used to determine the association rules between user behavior patterns and context features. The specific steps are as follows: Step 1: Set the minimum support, that is, the minimum frequency that the feature item set must meet; Step 2, generate candidate item sets, that is, start from a single item and gradually expand to a larger candidate item set; Step 3: Calculate the support of all candidate feature item sets in the database, and remove the feature item sets whose support is lower than the minimum support threshold; Use the remaining item sets to generate candidate item sets, and repeat step 3 until no more frequent item sets can be generated; Generate association rules from frequent itemsets and calculate the confidence and lift of each rule.

2. The O2O platform operation data push control method according to claim 1, characterized in that: Contextual feature extraction is performed on the collected user data to extract the time and location features of user activities, and word vector models are used for feature fusion and vector splicing, including: Collect users' geographic location data and corresponding activity records, as well as relevant external data, including weather and holidays; Perform data preprocessing, standardize location and activity descriptions, and encode weather and holidays as numerical values ​​or one-hot encoded vectors; The Skip-gram model is used to maximize the logarithmic probability of context words. The location vector and activity vector of the user data are determined based on the logarithmic probability. Feature fusion and vector concatenation are then performed. The encoding vectors of weather and holidays are concatenated with the location vector and activity vector of the user data to obtain a fused feature vector.

3. The O2O platform operation data push control method according to claim 2, characterized in that: Determine the user behavior pattern through association rules, and adjust the content push timing in combination with user behavior and preference characteristics, and obtain the matching evaluation information generated in the matching analysis process between the user behavior pattern and the user behavior and preference characteristics. The specific steps are as follows: Obtaining matching evaluation information generated during the matching analysis of user behavior patterns and user behavior and preference characteristics, the matching evaluation information includes user participation information and content adaptation information; User participation information includes dynamic user behavior synchronization index, and content adaptation information includes content push optimization index: The dynamic user behavior synchronization index and content push optimization index are combined to generate a push control coefficient; The dynamic user behavior synchronization index and content push optimization index are directly proportional to the push control coefficient.

4. The O2O platform operation data push control method according to claim 3, characterized in that: Analyze the matching evaluation information to determine the matching between the pushed content and the user's preference characteristics, and control the push strategy based on the matching results. The specific steps include: Compare the generated push control coefficient with the push control threshold; If the push control coefficient is greater than the push control threshold, a push content adaptation signal is generated, the user behavior data involved in generating the push content adaptation signal is annotated, and the annotated data is used as training data for the machine learning model; If the push control coefficient is less than or equal to the push control threshold, a push control adjustment signal is generated, the user data is re-analyzed, the user's behavior and preferences are evaluated, and the time of content push is adjusted.

Images