Personalized content recommendation and ROI (Region of Interest) improvement method and system based on multi-dimensional user portraits

Abstract

The invention discloses a personalized content recommendation and ROI (Region of Interest) improvement method and system based on a multi-dimensional user portrait, and the method comprises the following steps: collecting user data, and generating the multi-dimensional portrait of a user through feature extraction and weighted fusion; taking the advertisement platform content feature vectors and the multi-dimensional portraits as nodes of a graph structure, taking the interactive behavior data of the users as edges of the graph structure, and constructing an association graph; analyzing the association map and the multi-dimensional portrait by using a causal reasoning model, and identifying invalid correlation variables and key variables which influence a user decision; carrying out recommendation decision making based on a causal reasoning result under a multi-objective optimization framework, and generating a personalized recommendation list of each user; and calculating potential conversion values of different user groups based on a time sequence prediction model, and optimizing an advertisement budget allocation strategy. According to the invention, the personalized degree of advertisement recommendation and the ROI of advertisement putting can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of advertising placement, and particularly relates to a method and system for personalized content recommendation and ROI improvement based on multi-dimensional user portraits. Background Art

[0002] Personalized recommendation technology plays a crucial role in multiple fields such as e-commerce, digital marketing, news feeds, and short video platforms. By analyzing users' historical behaviors and interest preferences, the recommendation system can accurately match users with appropriate content, thereby improving the user experience, increasing the interaction rate, and promoting commercial conversion. With the development of big data and artificial intelligence technologies, personalized recommendation has gradually evolved from traditional methods based on collaborative filtering to an intelligent recommendation system that integrates multi-dimensional data, deep learning, causal reasoning, and other technologies.

[0003] In the fields of advertising placement and commercial marketing, the return on investment (ROI) is a key indicator for measuring the effectiveness of advertising placement. The improvement of ROI not only depends on accurate personalized recommendations but also requires comprehensive consideration of factors such as the diversity of advertising content, click-through rate, and long-term user retention. An efficient recommendation system can not only increase the probability of users clicking on ads or content but also promote user stickiness in the long term and maximize advertising revenue.

[0004] The Chinese patent application with the publication number CN117593053A discloses a big data analysis system for advertising placement in the field of advertising technology, including: a data collection unit responsible for collecting user data; a data storage unit for storing and managing a large amount of user data and performing efficient access and query; a data analysis unit for deeply analyzing user data by applying technologies such as machine learning and data mining to extract user characteristics and behavior patterns; an advertisement matching unit for matching the advertisement information provided by advertisers with user data to determine the best advertising placement plan.

[0005] The above solution realizes the optimization of the advertising placement plan based on big data. However, with the complexity of users' behavior patterns, relying solely on a single data source (such as click behavior) is no longer sufficient to accurately depict users' needs. Therefore, the recommendation technology based on multi-dimensional user portraits has gradually become a research hotspot. In addition, traditional advertising budget allocation usually relies on historical conversion rates or rule settings and is difficult to dynamically adapt to the potential conversion value of different user groups. This solution is prone to being misled by data correlation, which affects the accuracy of advertising placement. Summary of the Invention

[0006] The present invention provides a method and system for personalized content recommendation and ROI improvement based on multi-dimensional user portraits, aiming to solve problems such as insufficient accuracy in the field of advertising placement in the prior art.

[0007] To solve the above technical problems, the personalized content recommendation and ROI improvement method proposed by the present invention includes the following steps: Collect the user's behavioral data, social data, context data, emotional data, and content consumption trend data, and generate a multi-dimensional portrait of the user through feature extraction and weighted fusion; Take the content feature vector of the advertising platform and the multi-dimensional portrait as the nodes of the graph structure, and the user's interaction behavior data as the edges of the graph structure to construct an association graph; Use a causal inference model to analyze the association graph and multi-dimensional portrait, and identify invalid relevant variables and key variables that affect user decisions; Make a recommendation decision based on the causal inference result under a multi-objective optimization framework, and generate a personalized recommendation list for each user; Calculate the potential conversion value of different user groups based on a time series prediction model, and optimize the advertising budget allocation strategy.

[0008] Preferably, the behavioral data includes browsing records, click records, purchase records, stay duration, and favorite records; the social data includes friend relationships, interactions, and shared content; the context data includes device type, geographical location, access time, and network environment; the emotional data is data analyzed based on the user's text, voice, and expressions; the content consumption trend data includes hot topics and data on the change of user interests over time.

[0009] Preferably, the feature extraction includes the following methods: After selecting the feature dimensions for the behavioral data and social data, normalize or standardize the data corresponding to the feature dimensions, and then splice them into a behavioral vector and a social vector respectively; Use one-hot encoding for the context data, and splice the encoded data into a context vector; Use natural language processing for the emotional data for emotional analysis, and convert the emotional analysis score into an emotional vector; The content consumption trend data is converted into a content consumption trend vector through keyword extraction and one-hot encoding.

[0010] Preferably, the specific method for identifying invalid relevant variables and key variables that affect user decisions is: Construct a causal graph, and use a directed acyclic graph to represent the causal relationship between variables; Calculate the impact of each variable on user decisions:

[0011] In the formula, represents the variable The average causal effect on user decisions; represents in the intervention group, the variable The expected value of the overall user decision after the intervention is applied; Indicates that in the control group, the variable The expected value of the overall user decision when no intervention is applied; Divide the variables into ineffective relevant variables and key variables according to the average causal effect according to the set rules.

[0012] Preferably, the specific method for identifying ineffective relevant variables and key variables that affect user decisions further includes calculating individual causal effects for different user groups:

[0013] In the formula, Indicates for a single user For, the variable The individual causal effect of the variable on the user decision, used for decision recommendation; Indicates that when an intervention is applied to a single user The expected value of the user's decision; Indicates that no intervention is applied to a single user The expected value of the user's decision; Divide the variables into ineffective relevant variables and key variables according to the average causal effect and individual causal effect according to the set rules.

[0014] Preferably, the method for making a recommendation decision based on the causal inference result under the multi-objective optimization framework is as follows: Define a multi-objective optimization function:

[0015] In the formula, Is the maximum value of the optimization objective, Is the click-through rate, indicating the degree of interest of the user in the recommended content, Is the return on investment, indicating the purchase conversion rate of the advertisement or product, Is the diversity of the advertisement content, Is the long-term benefit of the advertisement, Is a hyperparameter used to adjust the weights of each objective; Optimize the multi-objective optimization function using the Pareto front; Calculate the final recommendation scores of each user for different contents based on the Pareto front optimization result; Sort the final recommendation scores, select the top one or more contents to generate a personalized recommendation list, and display it to the corresponding users.

[0016] Preferably, the personalized recommendation list dynamically adjusts the recommendation strategy through the multi-armed bandit algorithm to generate an adjusted personalized recommendation list, and displays it to the user according to the adjusted personalized recommendation list; the multi-armed bandit uses the Thompson sampling algorithm to optimize the reward function.

[0017] Preferably, the optimization method of the advertising budget allocation strategy is specifically as follows: Divide the user groups, calculate the multi-dimensional portraits of all individual users in the group, and calculate the corresponding group feature vectors; Construct and train a time series prediction model based on ARIMA or LSTM; Use the trained time series model to predict the advertising conversion rate and click-through rate of different user groups in different future time periods; Calculate the potential conversion value of each group according to the predicted conversion rate and click-through rate; Calculate the budget ratio of each group according to the potential conversion value weights of different groups, and conduct advertising placement according to the budget ratio.

[0018] Preferably, the division of user groups includes the following division methods: Conduct group division based on the demographic characteristics of users; Divide user groups based on the historical behavior data of users; Use the K-means clustering algorithm, density-based clustering algorithm or Gaussian mixture model clustering algorithm to cluster users, and divide users into high-value users, potential conversion users and low-value users.

[0019] Correspondingly, the present invention also proposes a personalized content recommendation and ROI improvement system based on multi-dimensional user portraits. The system is used to implement the above-mentioned personalized content recommendation and ROI improvement method, including: A user data collection module, which is used to collect the user's behavior data, social data, context data, emotional data and content consumption trend data, and store and preprocess the data; A feature extraction and portrait construction module, which is used to extract features from the collected data and generate a multi-dimensional portrait of the user; A graph structure construction module, which is used to use the content feature vector of the advertising platform and the multi-dimensional portrait as the nodes of the graph structure, and the user's interaction behavior data as the edges of the graph structure to construct an association graph; A causal inference analysis module, which is used to analyze the association graph and multi-dimensional portrait based on the causal inference model, and identify invalid correlation variables and key variables that affect user decisions; A recommendation decision optimization module, which is used to make recommendation decisions under a multi-objective optimization framework, and use a multi-armed bandit to achieve dynamic adjustment of the recommendation strategy; An advertising budget optimization module, which is used to calculate the potential conversion value of different user groups based on a time series prediction model and optimize the advertising budget allocation strategy; A recommended content sorting and display module, which is used to calculate the final recommendation scores of each user for different contents based on the Pareto front optimization result and generate a personalized recommendation list for display; A system storage and management module, which is used to store user data, model parameters, optimization results and recommendation records, and support the monitoring and management of the system running status.

[0020] Compared with the prior art, the present invention has the following technical effects: 1. The personalized content recommendation and ROI improvement method proposed by the present invention uses a multi-dimensional user portrait to improve the recommendation accuracy, solves the problem that traditional methods cannot accurately predict user interests, avoids the deviation of recommended content caused by a single user portrait, enables the recommendation system to adapt to the user needs in different scenarios, and improves the personalization degree of recommendations.

[0021] 2. The personalized content recommendation and ROI improvement method proposed by the present invention uses a causal inference model to analyze the user-content association, identify the key variables that really affect user decisions, and eliminate invalid relevant factors. It solves the problem that recommendations based on statistical correlation may lead to misleading recommendations, avoids the influence of pseudo-correlation on recommendations, ensures that the recommended content is more in line with the true interests of users, and improves the conversion rate; makes the advertising placement more targeted, reduces ineffective exposure, lowers the advertising cost, and improves the ROI.

[0022] 3. The personalized content recommendation and ROI improvement method proposed by the present invention uses a multi-armed bandit algorithm for exploration-exploitation balance, can continuously optimize the recommendation list, dynamically adjust the recommendation strategy, and improve the user interaction rate. Combining with the real-time feedback of users, it ensures that the recommended content is continuously optimized to meet the changing interests of users. It solves the problem that a fixed recommendation strategy cannot adapt to the changing interests of users, can enhance the self-adaptability of the system, effectively improve the advertising conversion rate and user click-through rate, and improve the long-term user stickiness.

[0023] 4. The personalized content recommendation and ROI improvement method proposed by the present invention uses an ARIMA / LSTM time series prediction model to predict the potential conversion value of different user groups and optimize the advertising budget allocation. It solves the problem that traditional advertising budget allocation relies on historical data and lacks the ability of dynamic adjustment. According to the predicted future conversion rate and click-through rate, it accurately adjusts the advertising placement strategy and improves the ROI. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 is a schematic flow chart of the personalized content recommendation and ROI improvement method described in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will, in conjunction with specific embodiments of the present application and with reference to the accompanying drawings, clearly and completely describe the technical solutions of the present invention.

[0026] Embodiment 1 This embodiment is a personalized content recommendation and ROI improvement method based on multi-dimensional user portraits, as Figure 1 shown, including the following steps 1 to 5: Step 1: Collect the user's behavioral data, social data, situational data, emotional data, and content consumption trend data, and generate a multi-dimensional portrait of the user through feature extraction and weighted fusion.

[0027] The behavioral data includes browsing records, click records, purchase records, stay duration, and favorite records; the social data includes friend relationships, interactions, and shared content; the situational data includes device type, geographical location, access time, and network environment; the emotional data is the analysis data based on the user's text, voice, and expressions; the content consumption trend data includes hot topics and data on the change of user interests over time.

[0028] The data sources can be the user's own behaviors, social networks, environmental factors, emotional states, and content consumption records, etc. Including but not limited to the following data: The behavioral data comes from the user's browsing, clicking, favoriting, commenting, sharing, purchase records, and stay duration. The data features include click-through rate, stay time, purchase frequency, bounce rate, etc.; The social data comes from the user's friend relationships, interaction records, follow / followed situations, social platform content, and likes / forwards, etc. The data features include friend recommendations, group influence, common interests, etc.; The situational data is the user's device information, geographical location, time, and network environment. The data features include device type, access time period, location preference, etc.; The emotional data can be collected from the user's comments, posts, bullet screens, and voices. The data feature is the emotional tendency of the user's comments (such as positive, neutral, negative); The content consumption trend is manifested as the user's consumption preference for content. The specific data features include preferred content categories, recent popularity trends, etc.

[0029] The feature extraction method of this embodiment uses different methods according to different data: After selecting the feature dimensions for behavioral data and social data, the selectable dimensions include that due to the different value ranges of different features, normalization processing is required. Therefore, the data corresponding to the feature dimensions can be normalized or standardized, and then concatenated into a behavioral vector and a social vector respectively. Specifically, the Min-Max normalization method is used to normalize data with relatively stable value ranges such as click-through rate, the Z-Score standardization method is used to standardize data with a wider data distribution such as the number of purchases, and the Log transformation method is used to standardize data with a long-tail distribution feature such as the number of product purchases. Converting behavioral data into feature vectors generally involves steps such as feature normalization, statistical feature extraction, and time window aggregation. Common methods include One-Hot Encoding, Min-Max normalization, TF-IDF (Term Frequency–Inverse Document Frequency), Embedding representation, etc.

[0030] Taking the behavioral data as an example, the specific process of feature extraction in this embodiment is described. First, a set of key behavioral data is set in the collected behavioral data, for example: click behavior (number of clicks, click frequency), dwell time (page dwell duration, average dwell time), conversion behavior (purchase, favorite, share, etc.), time window (statistics for time periods such as 7 days, 30 days, etc.). Since the value ranges of different features are different and normalization processing is required, those skilled in the art can select a suitable normalization / standardization processing method according to the data characteristics, such as Min-Max normalization, Z-Score standardization, and Log transformation. Then, the behavioral data is converted into a feature vector. Suppose a user's behavioral data is as follows: User ID: 123456; Number of clicks in the past 7 days: 120; Average dwell time in the past 7 days (seconds): 80; Number of purchases in the past 30 days: 5; Number of shares in the past 7 days: 10; Number of favorites in the past 7 days: 20.

[0031] The maximum values of the four indicators of the number of clicks in the past 7 days, the average dwell time in the past 7 days, the number of shares in the past 7 days, and the number of favorites in the past 7 days, which are statistically obtained based on the global data, are 500, 300, 50, and 100 respectively, and the minimum values are all 0. Then, after Min-Max normalization, their values are 0.24, 0.267, 0.2, and 0.2 respectively. For the indicator of the number of purchases in the past 30 days, after standardization using the Log transformation, it is 1.792.

[0032] The above data is concatenated into a behavioral vector: X1=(0.24,0.267,1.792,0.2,0.2).

[0033] The situational data uses one-hot encoding, and the encoded data is concatenated into a situational vector. Among them, key features such as device category (PC, mobile, tablet), operating system (iOS, Android, Windows), active time period in time features (morning / noon / evening), and common IP location (country, province, city) in geographical features can all adopt one-hot encoding.

[0034] The sentiment data uses natural language processing for sentiment analysis, and the sentiment analysis score is converted into a sentiment vector. In this embodiment, BERT, LSTM, and SVM are used for sentiment classification (positive, neutral, negative), and then the sentiment score is calculated (the sentiment score is in the interval [-1, 1]). Then, the emotional fluctuations such as the average emotion in the recent 7 days and the variance of the emotion in the recent 7 days are statistically analyzed. Suppose the recent comment analysis of the user is as follows: text sentiment score: 0.8; average in the recent 7 days: 0.75; variance in the recent 7 days: 0.1, which is converted into a vector: X2=(0.8, 0.75, 0.1).

[0035] The content consumption trend data is converted into a content consumption trend vector through keyword extraction and one-hot encoding. Specifically, after using TF-IDF to extract keywords, the proportion of the keyword in all categories is statistically analyzed as the initial data. Since the proportion data interval is [0, 1], this embodiment does not perform normalization or standardization processing on the proportion data. The content consumption trend data also includes topic popularity data. For the topic popularity, it is also converted into a numerical value by one-hot encoding, and then all content consumption trend data is concatenated into a content consumption trend vector.

[0036] The above concatenated vectors are all vectors corresponding to a single user. Finally, the behavior vector, social vector, situational vector, sentiment vector, and content consumption trend vector of each user are concatenated to generate a multi-dimensional portrait Xu of each user.

[0037] In some other embodiments of the present invention, for the normalized / standardized, encoded, and converted data, principal component analysis (PCA) can also be performed for dimensionality reduction to improve the calculation efficiency; or the Embedding method can be used to represent the categorical features (such as the type of content clicked) in a low-dimensional vector.

[0038] Step 2: Use the content feature vector of the advertising platform and the multi-dimensional portrait as the nodes of the graph structure, and the interaction behavior data of the user as the edges of the graph structure to construct an association graph.

[0039] The portrait vector of a user is only the feature of a single node. However, to form a complete User-Item Graph, interaction behavior data of the user is also required as edges to connect the user and the content. In the graph structure of this embodiment, it includes user nodes and content nodes. In the user nodes, each user u is characterized by its multi-dimensional portrait Xu; in the content nodes, each content c is characterized by its content feature Xc (such as content category, keywords, release time, etc.). The edge (u, c) connecting the user u and the content c represents the user's interaction behavior, such as click, like, purchase, etc. The weight of the edge can be calculated based on the interaction intensity. For example: the number of clicks (the more clicks, the greater the edge weight); the dwell time (the longer the dwell time, the greater the edge weight); the conversion rate (behaviors such as purchase, collection, etc.).

[0040] Suppose there is the following interaction data:

[0041] The edge weight can be defined as:

[0042] Among them, represents the weight of the edge between the user and the content, is a hyperparameter.

[0043] The source of the content nodes mainly depends on the specific scenario of the recommendation system, such as e-commerce, social media, news platforms, etc. Generally speaking, the content nodes come from the recommendable content on the platform, including: E-commerce platforms (Taobao, Amazon, etc.): products (SKUs), product categories, brands; Short video platforms (Douyin, Kuaishou, Bilibili, etc.): video content, authors, tags; News platforms (Toutiao, etc.): news articles, themes, keywords; Social media (Xiaohongshu, Weibo, etc.): posts, topics, tags.

[0044] The features of the content nodes need to be converted into vectors Xc, and the conversion method is the same as that of the multi-dimensional portrait.

[0045] Step 3: Use a causal inference model to analyze the association graph and the multi-dimensional portrait, and identify the ineffective relevant variables and key variables that affect user decisions.

[0046] Through causal inference, this embodiment achieves the following goals: Find key variables (Causal Variables): features that have a direct causal impact on user behaviors (such as clicks, conversions, purchases, etc.); Exclude spurious correlations: variables that only have correlations but no causal relationships, to improve the generalization ability and recommendation accuracy in subsequent steps.

[0047] The specific method for identifying spurious correlated variables and key variables that affect user decisions is as follows: Construct a causal graph, using a directed acyclic graph to represent the causal relationships between variables; Calculate the impact of each variable on user decisions:

[0048] In the formula, represents the average causal effect of variable on user decisions; represents the expected value of the overall user decisions after variable is intervened in the intervention group; represents the expected value of the overall user decisions when variable is not intervened in the control group; Divide the variables into spurious correlated variables and key variables according to the average causal effect according to the set rules.

[0049] In some other embodiments of the present invention, the specific method for identifying spurious correlated variables and key variables that affect user decisions further includes calculating individual causal effects for different user groups:

[0050] In the formula, represents the individual causal effect of variable on user decisions for a single user and is used for decision-making recommendations; represents the expected value of the decisions of a single user when the user is intervened; represents the expected value of the decisions of a single user when the user is not intervened; Divide the variables into spurious correlated variables and key variables according to the average causal effect and individual causal effect according to the set rules. Specifically, based on the foregoing analysis, we can classify the variables:

[0051] For the high, medium, and low classifications of ATE, in this embodiment, ATE represents the increase in the user purchase rate caused by an advertisement. ATE ≥ 10% is considered high, 2% ≤ ATE < 10% is medium, and 0 ≤ ATE < 2% is low. For ITE, in this embodiment, IT represents the purchase promotion rate of an advertisement for specific users. ITE > 5% is high, 1% ≤ ITE ≤ 5% is medium, and ITE < 1% is low. Mark the variables among them as conclusion variables, and reassign their weights during the recommendation decision-making process. For example, the weight of key variables is increased by 60% based on the original weight, the weight of medium variables remains unchanged, and the weight of invalid relevant variables is decreased by 75% based on the original weight. The specific adjustment range can be set as needed during the implementation process.

[0052] Step 4: Based on the causal inference results, perform recommendation decision-making under the multi-objective optimization framework to generate a personalized recommendation list for each user.

[0053] The recommendation decision-making based on the causal inference results under the multi-objective optimization framework specifically includes the following steps S41 - S44: S41: Define a multi-objective optimization function:

[0054] In the formula, is the maximum value of the optimization objective, is the click-through rate, indicating the degree of user interest in the recommended content, is the return on investment, indicating the purchase conversion rate of the advertisement or product, is the diversity of advertisement content, is the long-term benefit of the advertisement, is a hyperparameter used to adjust the weights of each objective. Here, the hyperparameter is reassigned according to the variable marking in Step S3.

[0055] S42: Use the Pareto Front to optimize the multi-objective optimization function. Since the recommendation task involves multiple conflicting optimization objectives (such as CTR and ROI may affect each other), in this embodiment, the Pareto Front optimization is adopted to find the optimal recommendation strategy. Select a set of optimal content from the Pareto Front to participate in the calculation of the recommendation score.

[0056] S43: Calculate the final recommendation score of each user for different content based on the Pareto Front optimization results. In this step, combining the multi-objective optimization results and the user's personalized weights, calculate the recommendation score of each user u for a single piece of content in a set of optimal content:

[0057] Among them, is the content The recommendation score for user u are respectively the click-through rate, return on investment, diversity of advertising content, and long-term benefits of the content determined by the multi-dimensional portrait of the user. For example, users with a higher purchase propensity may increase the ROI weight, and users who like novel content may increase the diversity weight.

[0058] S44: Sort the final recommendation scores, select the top one or more pieces of content to generate a personalized recommendation list, and display it to the corresponding user. In the specific implementation process, only one advertisement can be finally displayed, that is, the number of elements in the personalized recommendation list is 1, or the number of elements in the personalized recommendation list is greater than 1, but only the top-ranked advertisement in the list is displayed to the user; or multiple advertisements can be displayed to the user in a carousel manner, and multiple advertisements are obtained according to the ranking from the personalized recommendation list.

[0059] Sort the calculated recommendation scores, and according to the value, sort the recommended content for user u in descending order, set the number k of recommended content, select the top k recommended content as the final personalized recommendation list, and feedback the generated personalized recommendation list to the front end of the recommendation system.

[0060] In some embodiments of the present invention, the personalized recommendation list dynamically adjusts the recommendation strategy through the multi-armed bandit algorithm to generate an adjusted personalized recommendation list, and displays it to the user according to the adjusted personalized recommendation list; the multi-armed bandit optimizes the reward function using the Thompson sampling algorithm. The multi-armed bandit optimizes the reward function using the Thompson sampling algorithm. Specifically, the data used by the multi-armed bandit includes the user's personalized recommendation list, historical click data, historical conversion data, advertising budget constraints, and user feedback (click / purchase behavior). Add the recommendation score used in step S44 to the exploration revenue score calculated by the multi-armed bandit as the new sorting basis.

[0061] In some other embodiments of the present invention, the ε-Greedy algorithm or the UCB (Upper Confidence Bound) algorithm can be used to optimize the reward function.

[0062] Step five, calculate the potential conversion value of different user groups based on the time series prediction model, and optimize the advertising budget allocation strategy.

[0063] The optimization method of the advertising budget allocation strategy specifically includes the following steps S51 - S55: S51: Divide the user group, calculate the multi-dimensional portraits of all individual users in the group, and calculate the corresponding group feature vectors.

[0064] In an embodiment of the present invention, the division of user groups includes the following division methods: Group division is performed based on the demographic characteristics of users; User groups are divided based on the historical behavior data of users; The K-means clustering algorithm, density-based clustering algorithm, or Gaussian mixture model clustering algorithm is used to cluster users, and users are divided into high-value users, potential conversion users, and low-value users.

[0065] S52: Construct and train a time series prediction model based on ARIMA (Auto Regression Moving Average) or LSTM (Long Short-Term Memory).

[0066] S53: Use the trained time series model to predict the advertisement conversion rate and click-through rate of different user groups in different future time periods.

[0067] The goal of this step is to predict the CTR and CVR of different user groups in the next T time periods. Input the historical CTR and CVR of the recent n days / hours into the trained ARIMA / LSTM model to predict the CTR and CVR in the future time window (such as 7 days, 30 days). Output the prediction results: for high-value users, with an expected CTR≥10% and CVR≥5%; for potential conversion users, corresponding to an expected 2%≤CTR<10% and 1%≤CVR<5%; for low-value users, corresponding to an expected CTR<2% and CVR<1%.

[0068] S54: Calculate the potential conversion value of each group according to the predicted conversion rate and click-through rate. The specific calculation method is as follows:

[0069] In the formula, is the potential conversion value of the user group, is the predicted value of the group click-through rate, is the predicted value of the group conversion rate, is the average revenue per user of the group, which is statistically obtained from historical data.

[0070] S55: Calculate the budget ratio of each group according to the weight of the potential conversion value of different groups, and perform advertisement placement according to the budget ratio.

[0071] Embodiment 2 This embodiment is a personalized content recommendation and ROI improvement system based on multi-dimensional user portraits. The system is used to implement the personalized content recommendation and ROI improvement method as described in Embodiment 1, including a user data collection module, a feature extraction and portrait construction module, a graph structure construction module, a causal inference analysis module, a recommendation decision optimization module, an advertising budget optimization module, a recommended content sorting and display module, and a system storage and management module.

[0072] The user data collection module is used to collect the user's behavioral data, social data, context data, emotional data, and content consumption trend data, and store and preprocess the data. The feature extraction and portrait construction module is used to extract features from the collected data and generate multi-dimensional portraits of users. The graph structure construction module is used to construct an association graph with the content feature vectors of the advertising platform and the multi-dimensional portraits as the nodes of the graph structure, and the user's interaction behavior data as the edges of the graph structure. The causal inference analysis module is used to analyze the association graph and multi-dimensional portraits based on a causal inference model to identify ineffective relevant variables and key variables that affect user decisions. The recommendation decision optimization module is used to make recommendation decisions under a multi-objective optimization framework and adopt a multi-armed bandit to achieve dynamic adjustment of the recommendation strategy. The advertising budget optimization module is used to calculate the potential conversion value of different user groups based on a time series prediction model and optimize the advertising budget allocation strategy. The recommended content sorting and display module is used to calculate the final recommendation scores of each user for different contents based on the Pareto front optimization result and generate a personalized recommendation list for display. The system storage and management module is used to store user data, model parameters, optimization results, and recommendation records, and support the monitoring and management of the system operation status.

[0073] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the inventive concept of the present invention, several modifications and improvements can be made, and these all belong to the protection scope of the present invention.

Claims

1. A personalized content recommendation and ROI improvement method based on multi-dimensional user portraits, characterized in that: The following steps are involved: Collect user behavior data, social data, situational data, emotional data, and content consumption trend data, and generate a multi-dimensional portrait of the user through feature extraction and weighted fusion; The advertising platform content feature vector and the multi-dimensional portrait are used as nodes of the graph structure, and the user's interactive behavior data are used as edges of the graph structure to construct a correlation graph; Use causal reasoning models to analyze the association maps and multi-dimensional portraits to identify invalid related variables and key variables that affect user decisions; Make recommendation decisions based on causal reasoning results in a multi-objective optimization framework to generate a personalized recommendation list for each user; Calculate the potential conversion value of different user groups based on the time series prediction model and optimize the advertising budget allocation strategy.

2. The personalized content recommendation and ROI improvement method based on multi-dimensional user portrait according to claim 1, characterized in that: The behavioral data includes browsing history, click history, purchase history, stay time and collection history; the social data includes friend relationships, interactions and shared content; the contextual data includes device type, geographic location, access time and network environment; the emotional data is based on user text, voice and expression analysis data; the content consumption trend data includes hot topics and data on changes in user interests over time.

3. The personalized content recommendation and ROI improvement method based on multi-dimensional user portrait according to claim 1, characterized in that: The feature extraction includes the following methods: After selecting the characteristic dimensions of the behavioral data and social data, the data of the corresponding characteristic dimensions are normalized or standardized, and then spliced ​​into behavioral vectors and social vectors respectively; The context data is one-hot encoded, and the encoded data is concatenated into a context vector; Sentiment data is analyzed using natural language processing, and sentiment analysis scores are converted into sentiment vectors; Content consumption trend data is converted into content consumption trend vectors through keyword extraction and one-hot encoding.

4. The personalized content recommendation and ROI improvement method based on multi-dimensional user portrait according to claim 1, characterized in that: The specific method for identifying invalid related variables and key variables that affect user decision-making is: Construct a causal graph, using a directed acyclic graph to represent the causal relationship between variables; For each variable, calculate its impact on the user's decision: In the formula, Representation variables The average causal effect on user decisions; In the intervention group, the variable The expected value of the overall user decision after the intervention; In the control group, the variable The expected value of the overall user decision when no intervention is applied; According to the average causal effect, the variables are divided into invalid related variables and key variables according to the set rules.

5. The personalized content recommendation and ROI improvement method based on multi-dimensional user portrait according to claim 4 is characterized in that: The specific method of identifying invalid related variables and key variables that affect user decisions also includes calculating individual causal effects for different user groups: In the formula, For a single user For variables Individual causal effects that influence user decisions, used for decision-making recommendations; Indicates that when a single user The expected value of the user's decision when the intervention is applied; Indicates that no single user The expected value of the user's decision when the intervention is applied; According to the average causal effect and individual causal effect, the variables are divided into invalid related variables and key variables according to the set rules.

6. The personalized content recommendation and ROI improvement method based on multi-dimensional user portrait according to claim 1, characterized in that: The specific method of making recommendation decisions based on causal reasoning results under the multi-objective optimization framework is as follows: Define a multi-objective optimization function: In the formula, To maximize the optimization objective, is the click-through rate, which indicates the user’s interest in the recommended content. Return on investment, which indicates the purchase conversion rate of advertisements or products. To diversify advertising content, For the long-term benefits of advertising, is a hyperparameter used to adjust the weight of each objective; Optimize multi-objective optimization functions using Pareto frontiers; Calculate the final recommendation score of each user for different content based on the Pareto frontier optimization results; The final recommendation scores are sorted, and the previous one or more contents are selected to generate a personalized recommendation list, which is displayed to the corresponding user.

7. The personalized content recommendation and ROI improvement method based on multi-dimensional user portrait according to claim 1, characterized in that: The personalized recommendation list dynamically adjusts the recommendation strategy through a multi-armed bandit algorithm to generate an adjusted personalized recommendation list, and displays the adjusted personalized recommendation list to the user; the multi-armed bandit uses the Thompson sampling algorithm to optimize the reward function.

8. The personalized content recommendation and ROI improvement method based on multi-dimensional user portrait according to claim 1, characterized in that: The optimization method of the advertising budget allocation strategy is specifically as follows: Divide user groups, calculate the multi-dimensional portraits of all individual users in the group and calculate the corresponding group feature vectors; Build and train time series forecasting models based on ARIMA or LSTM; Use the trained time series model to predict the advertising conversion rate and click-through rate of different user groups in different time periods in the future; Calculate the potential conversion value of each group based on predicted conversion rate and click-through rate; The budget ratio for each group is calculated based on the potential conversion value weights of different groups, and advertising is delivered according to the budget ratio.

9. The personalized content recommendation and ROI improvement method based on multi-dimensional user portrait according to claim 8, characterized in that: The division of user groups includes the following division methods: Segment users into groups based on their demographic characteristics; Divide user groups based on historical user behavior data; Use K-means clustering algorithm, density-based clustering algorithm or Gaussian mixture model clustering algorithm to cluster users and divide them into high-value users, potential conversion users and low-value users.

10. A personalized content recommendation and ROI improvement system based on multi-dimensional user portraits, characterized by: The system is used to implement the personalized content recommendation and ROI improvement method according to any one of claims 1 to 9, including: User data collection module, used to collect user behavior data, social data, situational data, emotional data and content consumption trend data, and store and pre-process the data; The feature extraction and portrait construction module is used to extract features from the collected data and generate a multi-dimensional portrait of the user; A graph structure construction module, used to construct a correlation graph by using the advertising platform content feature vector and the multi-dimensional portrait as nodes of the graph structure and the user's interactive behavior data as edges of the graph structure; The causal reasoning analysis module is used to analyze association maps and multi-dimensional portraits based on causal reasoning models to identify invalid related variables and key variables that affect user decisions; The recommendation decision optimization module is used to make recommendation decisions under the multi-objective optimization framework and to use a multi-armed bandit machine to achieve dynamic recommendation strategy adjustment; Ad budget optimization module, which is used to calculate the potential conversion value of different user groups based on the time series prediction model and optimize the advertising budget allocation strategy; The recommended content sorting and display module is used to calculate the final recommendation score of each user for different content based on the Pareto front optimization results, and generate a personalized recommendation list for display; The system storage and management module is used to store user data, model parameters, optimization results and recommendation records, and supports system operation status monitoring and management.

Images