Method and system for recommending synthetic game orders based on player behavior profile

Abstract

This invention provides a method and system for recommending game orders based on player behavior profiles. It constructs a comprehensive profile system to fully reconstruct a player's game status, gameplay preferences, unmet needs, and spending potential. By integrating an improved collaborative filtering algorithm and an optimized deep learning model, it performs multi-dimensional optimizations tailored to the characteristics of game scenarios. A dual recall mechanism addresses the cold start problem for new / low-activity players. The Lambda MART algorithm achieves multi-objective optimized order ranking, ensuring both paid conversion and revenue while also considering player retention and experience. Furthermore, a trigger-based, contextualized push mechanism is designed to precisely target players during peak periods of their willingness to pay, while controlling push frequency to avoid excessive player disturbance. This achieves a balance between commercial effectiveness and player lifecycle, addressing the industry-wide problem of prioritizing monetization over user experience in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of digital technology in gaming and entertainment, and in particular to a method and system for recommending game orders based on player behavior profiles, as well as an electronic device and a computer-readable storage medium. Background Technology

[0002] With the rapid development of the digital entertainment industry, online games have become one of the core tracks of global cultural consumption, and the in-app purchase model is currently the core profit-making method for online game commercialization. According to industry statistics, more than 85% of mobile online games adopt the "free download + in-app purchase" commercialization model. The recommendation effect of game item packs, recharge packages, and other combined orders directly determines the game's payment conversion rate, average order value, and overall revenue capability.

[0003] The current order recommendation technology in the gaming industry mainly suffers from the following core pain points and technical shortcomings: 1. Player profiles are too narrow in scope and suffer from significant static lag: Most existing technologies only build user profiles based on players' historical spending behavior, ignoring key data on non-paying dimensions such as core gameplay behavior, game progress, social interaction, and operating habits, thus failing to reflect players' true gaming needs. At the same time, profiles often adopt an offline periodic update mode, which cannot respond in real time to changes in players' game progress, shifts in gameplay preferences, and fluctuations in spending ability, resulting in a serious disconnect between recommended content and players' current status.

[0004] 2. The recommended content is highly homogenized and lacks personalized adaptation capabilities: The industry generally uses fixed templates for bulk push of game gift packs / orders, which can only be divided into simple tiers based on player level and VIP level. It cannot dynamically synthesize personalized orders based on players' gameplay preferences, level obstacles, item shortages, and payment habits. This results in mismatches such as "casual players receiving high-intensity PVP item packs" and "free-to-play players receiving high-tier recharge packages." This not only fails to improve payment conversion rates but also arouses player resentment and may even lead to user churn.

[0005] 3. Recommendation algorithms suffer from technical bottlenecks and poor adaptability: Existing solutions mostly use a single collaborative filtering algorithm for item recommendation, which has serious cold start problems (it cannot effectively recommend items to new players when they have no historical behavior data) and data sparsity problems (there is little behavior data for low-activity / low-paying players, resulting in extremely low recommendation accuracy); some solutions that introduce deep learning do not design model structures that are combined with the business characteristics of the game scenario, and cannot handle the dynamically changing player states and multi-dimensional behavioral features in the game, resulting in poor generalization ability and implementation effect.

[0006] 4. Low matching degree between push scenarios and timing, limiting conversion efficiency: Most existing technologies adopt the mode of timed batch push to all users, without combining the player's game scenario (such as level failure, level breakthrough, event participation, item consumption) for triggered push, which fails to capture the peak of players' willingness to pay; at the same time, order sorting only takes revenue as the sole objective, ignoring players' click intention and retention experience, which will damage the game's life cycle in the long run.

[0007] 5. Lack of a closed-loop feedback and iteration mechanism: Most existing solutions only complete the one-way process of "order push" without collecting and analyzing feedback data such as player clicks, conversions, order cancellations, and retention throughout the entire process. They cannot dynamically optimize profile weights, algorithm parameters, and order synthesis rules based on performance feedback, and the recommendation effect cannot be continuously optimized. Summary of the Invention

[0008] To address the technical problems existing in the prior art, the present invention provides the following technical solution: On the one hand, a method for recommending synthetic game orders based on player behavior profiles is provided, including the following steps: S1 collects multi-dimensional player behavior data across the entire process. The behavior data includes player basic attribute data, core gameplay behavior data, consumption and payment behavior data, payment ability profile data, social interaction behavior data, activity and retention behavior data, and separates real-time behavior data streams from offline behavior datasets. S2 preprocesses and performs feature engineering on the collected raw behavioral data to generate a standardized set of player feature vectors. S3 is based on standardized player feature vectors to build a full-dimensional player behavior profile system that includes six dimensions: basic attributes, game progress, gameplay preferences, spending habits, spending power, and active retention. The profile is updated in real time through a time decay factor. S4 generates a highly matched set of item candidates that match player needs by integrating improved collaborative filtering and Wide & Deep deep learning recommendation models based on a player's full-dimensional behavioral profile. S5 dynamically generates personalized merge game orders based on a highly matched set of item candidates and a full-dimensional profile of player behavior through an order merging rule engine. S6 uses a multi-objective optimized sorting algorithm to intelligently sort synthetic game orders, and combines players' real-time in-game behavior events and activity patterns to complete contextualized and precise push notifications. The S7 collects player feedback data on pushed orders across the entire process, and uses this feedback data to complete closed-loop iterative optimization of profile weights, recommendation models, order synthesis rules, and push mechanisms.

[0009] On the other hand, a synthetic game order recommendation system based on player behavior profiles is provided to implement the above-described method, the system comprising: The player multi-dimensional behavior data collection module is used to collect player behavior data in six dimensions across the entire process, separating real-time data streams from offline datasets; The data preprocessing and feature engineering module is used to clean, standardize, encode, and filter the raw behavioral data to generate a standardized set of player feature vectors. The Player Full-Dimensional Behavior Profile Construction and Dynamic Update Module is used to build a player behavior profile system with six dimensions, and achieves real-time dynamic updates of the profile through a time decay factor. The personalized merge game order generation module is used to generate a highly matching set of item candidates by integrating a recommendation model, and dynamically generate personalized merge game orders through a rule engine. The intelligent order sorting and contextualized push module is used to intelligently sort orders through a multi-objective sorting algorithm and combine player game behavior to achieve contextualized and accurate push notifications. The recommended feedback and model iteration optimization module is used to collect player feedback data and complete the closed-loop iterative optimization of the system.

[0010] On the other hand, an electronic device is provided, comprising: a processor; and a memory storing computer-readable instructions, which, when executed by the processor, implement the method described above.

[0011] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored therein, the at least one instruction being loaded and executed by a processor to implement the above method.

[0012] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. A comprehensive dynamic player behavior profile system has been constructed, addressing the core shortcomings of existing technologies, such as single-dimensional and static profiles that are outdated. This solution breaks through the limitations of existing technologies that only build profiles around payment behavior, constructing a comprehensive profile system covering 6 dimensions to fully restore the player's game status, gameplay preferences, needs gaps, and payment potential. At the same time, through a time decay factor and a dual-mode dynamic update mechanism, the profile is iterated in real time, ensuring that the profile always matches the player's current state, thus guaranteeing the accuracy of recommendations from the decision-making foundation.

[0013] 2. A game-scene-adapted fusion recommendation model was designed to solve the problems of cold start, data sparsity, and poor generalization ability of traditional algorithms. This solution integrates an improved collaborative filtering algorithm and an optimized Wide&Deep deep learning model, and performs multi-dimensional optimizations for the characteristics of game scenes. It solves the cold start problem of new players / low-activity players through a dual recall mechanism, adapts to the sparse characteristics of game payment samples through an improved loss function, and combines the interpretability of linear models with the non-linear fitting ability of deep learning models, which greatly improves the matching degree and generalization ability of item recommendations.

[0014] 3. This solution enables personalized and dynamic synthesis of game orders, completely resolving the pain points of existing fixed gift packs being homogeneous and having low matching rates. It abandons the industry-standard fixed gift pack templates and uses a rule engine to dynamically synthesize game orders. Based on player profile data, it performs comprehensive personalization from dimensions such as item combinations, pricing, discounts, validity periods, and applicable scenarios. Each order is highly matched to the player's current needs and spending power, fundamentally improving players' willingness to pay and conversion efficiency.

[0015] 4. Achieved multi-objective optimized scenario-based precise push notifications, balancing monetization and player experience. This solution uses the LambdaMART algorithm to achieve multi-objective optimized order sorting, ensuring paid conversion and revenue while also considering player retention and experience. It also features a trigger-based scenario-based push notification mechanism that precisely targets players at peak payment intentions, while controlling push frequency to avoid excessive player disturbance. This achieves a balance between commercial effectiveness and player lifecycle, addressing the industry-wide problem of existing technologies prioritizing monetization over player experience.

[0016] 5. A closed-loop feedback iteration mechanism was constructed to achieve continuous self-optimization of recommendation performance. This solution designed a full-link feedback data collection and iterative optimization mechanism, which can continuously optimize profile weights, algorithm models, order synthesis rules and push strategies based on real-time player feedback, forming a complete business closed loop. It can continuously adapt to changes in player behavior and adjustments to the game operation rhythm, ensuring the long-term and stable recommendation performance and business value of the system. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a system architecture provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the data preprocessing and feature engineering process of the present invention; Figure 3 This is a diagram illustrating the working mechanism of the feature selection process in this invention. Figure 4 This is a schematic diagram of the player full-dimensional behavior profile construction mechanism of the present invention; Figure 5 This is a schematic diagram of the operating mechanism of the improved collaborative filtering recall layer of this invention. Figure 6 This is a schematic diagram illustrating the working mechanism of the improved Wide&Deep deep learning ranking layer of this invention; Figure 7 This is a schematic diagram of an intelligent order sorting and scenario-based push mechanism provided by an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0020] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0021] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0022] In this embodiment of the invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0023] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0024] This invention proposes a method and system for recommending game orders based on player behavior profiles. It can solve the core defects of existing technologies by collecting player behavior data in all dimensions, constructing dynamic and real-time behavior profiles, integrating collaborative filtering and deep learning into a personalized recommendation model, dynamically synthesizing game orders, and providing scenario-based precise push and closed-loop feedback iteration. This significantly improves the matching degree of game orders, payment conversion efficiency and player gaming experience.

[0025] The working mechanism of this system and method will be explained in detail below.

[0026] I. System Architecture Introduction like Figure 1 As shown, the synthetic game order recommendation system based on player behavior profiles of the present invention adopts a six-layer closed-loop architecture design of "data layer - feature layer - profile layer - algorithm layer - application layer - feedback layer". The overall architecture has the characteristics of high real-time performance, high scalability and high adaptability, and can be compatible with commercial recommendation scenarios of all types of online games such as MMORPG, card game, casual competitive, and SLG.

[0027] The core components of the system include: a multi-dimensional player behavior data collection module, a data preprocessing and feature engineering module, a player full-dimensional behavior profile construction and dynamic update module, a personalized game order generation module, an intelligent order sorting and contextualized push module, and a recommendation effect feedback and model iteration optimization module. These modules communicate with each other through standardized data interfaces and control processes through a centralized scheduling engine, forming a complete business loop of "data collection - feature processing - profile construction - order generation - push notification - feedback iteration".

[0028] The system adopts a dual-link data communication architecture of "real-time data stream + offline data stream" to ensure the real-time performance and stability of data transmission. 1. Real-time data communication link: Kafka message queue is used as the core message middleware. Real-time behavioral events of player clients and trigger events of the server are transmitted through Kafka to achieve high concurrency and low latency. The data passes through the data acquisition module, real-time preprocessing unit, real-time profile update, and order generation and push module in sequence. The end-to-end latency is <500ms, which meets the requirements of real-time triggered recommendation. The modules are connected through gRPC interface to realize standardized remote calls, ensuring the stability of data interaction.

[0029] 2. Offline Data Communication Link: The Hadoop distributed architecture is used as the core for offline data storage and computing. The offline business data of the game server is synchronized to the Hive data warehouse daily (T+1) through the DataX synchronization tool. The data then undergoes offline preprocessing, offline profile updating, and offline model training to complete the offline iteration and batch recommendation tasks of the system. The Spark distributed computing engine enables efficient processing of massive amounts of data.

[0030] 3. Data storage architecture: A Redis cluster is used to store real-time player profile cache, real-time feature vectors, and order push sequences to ensure high-frequency read and write performance; a MySQL database is used to store offline player profile library, item attribute library, order history, and system configuration parameters; and HDFS is used to store massive amounts of historical behavior data and model training datasets to meet the system's massive data storage needs.

[0031] The following section will detail the functions of the system's core modules and the system's operation process.

[0032] (a) Player multi-dimensional behavior data collection module This module serves as the data source foundation for the entire system, responsible for collecting full-volume behavioral data from players on the game client and game server across the entire chain and multiple dimensions. It provides raw data support for subsequent feature engineering, profile building, and order recommendation, while also enabling the separate collection and unified management of real-time and offline behavioral data.

[0033] This module collects data covering 6 core dimensions, avoiding the shortcomings of existing technologies that only collect paid data. The specific data collected is as follows: The module uses the lightweight event tracking SDK built into the game client to collect behavioral events in real time, and uses Kafka message queues to receive and transmit high-concurrency behavioral data, avoiding impact on the game client's performance. For offline statistical data, it extracts data from the business database on the game server to the system's Hive data warehouse through a T+1 offline synchronization task, achieving unified storage of real-time and offline data.

[0034] At runtime: By inputting user behavior events from the game client, business database data from the game server, and transaction data from the payment system, this module processes the data to generate and output a standardized raw player behavior dataset, which is divided into a real-time behavior data stream (latency < 500ms) and an offline behavior dataset (T+1 update), which are transmitted to the real-time processing unit and the offline processing unit of the data preprocessing and feature engineering module, respectively.

[0035] (II) Data Preprocessing and Feature Engineering Module This module is the core of the system's feature processing. It is responsible for cleaning, transforming, encoding, and filtering the raw behavioral data, converting unstructured behavioral data into structured feature vectors that can be used for profile building and model training, and solving the problems of high noise, inconsistent dimensions, and inability to be directly used for algorithm calculation in the raw data.

[0036] The system takes the collected raw player behavior dataset (real-time data stream + offline dataset) as input, processes it through this module, and outputs a standardized set of player feature vectors. These are divided into an offline feature library (updated T+1, used for offline profile updates and offline model training) and real-time feature vectors (updated by events, used for real-time profile updates and real-time recommendations). These vectors are then transmitted to the player full-dimensional behavior profile construction and dynamic update module.

[0037] like Figure 2 As shown, the module is divided into four core stages: data cleaning, data standardization, feature encoding, and feature selection, as detailed below: 1. Data cleaning process To ensure data quality, missing and outlier values ​​in the original data are handled. Missing value handling: For continuous features of player behavior, a longitudinal time series imputation method is used, filling in the missing values ​​with the historical mean / median of that feature. The formula for determining the fill is: , in, This is a historical valid data sequence corresponding to the player's characteristics. It is a median function. This is a mean function. For categorical features, use "unknown" labels to fill in the blanks and avoid feature loss.

[0038] Outlier Handling: The 3σ principle is used to identify and handle outliers of continuous features. The formula is as follows: , in, For feature sample values, This is the mean of this feature across all players. This represents the standard deviation of the characteristic. For values ​​exceeding... Outliers within a range are truncated using the upper and lower limits of the range to prevent them from interfering with subsequent profiles and models.

[0039] 2. In the data standardization stage, for continuous features with different dimensions, Min-Max standardization is used to map feature values ​​to the [0,1] interval, ensuring the stability of subsequent similarity calculations and model training. The formula is: in, For the original value of the feature, This represents the maximum value of the feature across the entire sample. To be the minimum value, These are the standardized feature values.

[0040] 3. In the feature encoding stage, categorical features (such as gameplay preferences, professions, and payment tiers) are converted into numerical features using different encoding methods: For unordered categorical features (such as occupation and faction), one-hot encoding is used to transform the features of N categories into N-dimensional 0-1 vectors, thus avoiding unordered differences in numerical weights of the features. For ordered categorical features (such as VIP level, payment tier, and churn risk level), label encoding is used to map them into a continuous numerical sequence according to the order of business logic, thus preserving the ordered nature of the features.

[0041] 4. Feature selection process, such as Figure 3 As shown, for the encoded high-dimensional features, a two-step filtering method is used to retain effective features, reduce feature dimensionality, and avoid the curse of dimensionality: Step 1: Variance screening. Remove low-variance features with variance less than a preset threshold (threshold set to 0.01). Such features have no distinguishing effect among all players and do not contribute to user profiles or recommendations. Step 2: Mutual information screening. Calculate the mutual information value between each feature and the player's payment conversion goal. The formula is: in, Features to be selected Pay players to convert tags. The joint probability distribution of features and labels. , The probability distribution is marginal. Weakly correlated features with mutual information values ​​less than 0.05 are removed, and the core effective features are retained.

[0042] (III) Module for Constructing and Dynamically Updating a Full-Dimensional Player Behavioral Profile This module forms the core decision-making foundation of the system. It is responsible for building a comprehensive player behavior profile system with six dimensions based on standardized player feature vectors. At the same time, it realizes real-time dynamic updates of the profile through a time decay mechanism, accurately reflecting the player's current game status, needs and preferences, and spending potential, providing core decision-making basis for subsequent order recommendations.

[0043] 1. Comprehensive Player Behavior Profile System This module constructs a player behavior profile based on a 6-dimensional tagging system. Each dimension includes quantified feature tags and weight values, as detailed below: Basic attribute profile: Includes the player's registration attributes and basic character information. It serves as the basic static tag for the profile and is updated on an event-triggered basis. Game progress profile: Quantifies the player's level, combat power, level progress, and mission completion, accurately pinpointing the player's current game stage and growth bottleneck, with updates triggered in real time; Gameplay Preference Profile: By analyzing players' gameplay participation behavior, we quantify the preference weights of players for different gameplay styles, identify the core enjoyment points of players' games, and update the profile using a combination of real-time and offline methods. Consumer habit profile: Based on players' historical spending behavior, quantify players' spending frequency, spending scenarios, discount sensitivity, and item type preferences to reconstruct players' spending behavior habits. The update cycle is a combination of real-time and offline data. Pay-to-play profile: By analyzing historical recharge data, this feature quantifies players' disposable spending limits, spending tier preferences, and spending potential, providing a core basis for pricing synthetic orders. The update cycle is offline T+1. Activity and retention profiles: quantify player activity patterns, online time periods, and churn risk levels to provide decision-making basis for the timing and scenarios of order push notifications. The update cycle is a combination of real-time and offline data.

[0044] 2. Portrait Construction Process like Figure 4 The build process shown: (1) The improved TF-IDF algorithm is used to calculate the player's preference weight for different gameplay modes, which solves the problem that the traditional algorithm cannot reflect the player's recent behavioral preferences. The algorithm is as follows: in: Let represent the preference weight of player i for gameplay j. The higher the weight value, the stronger the player's preference for that gameplay. For word frequency, it represents the proportion of time player i spends participating in gameplay j out of that player's total game time. The formula is: In the formula: Let n be the total time player i spends participating in gameplay j within the statistical period, and n be the total number of gameplay options in the game. Let be the inverse document frequency, representing the discriminative power of gameplay j, and the formula is: N is the total number of players in the game. To determine the number of players who participated in gameplay j within the statistical period; This is a time decay factor used to increase the weight of a player's recent actions. The formula is explained in the dynamic update mechanism below.

[0045] (2) Determining the weights of the profile dimensions: The Analytic Hierarchy Process (AHP) is used to determine the weights of the six profile dimensions in the order recommendation decision. By constructing pairwise judgment matrices between dimensions, the eigenvector corresponding to the largest eigenvalue is calculated. After passing the consistency test, the final weights of each dimension are obtained. The weight range is [0,1], and the sum of all weights is 1. For different game categories, the judgment matrix can be adjusted to adapt to business characteristics. For example, the game progress profile has a higher weight in MMORPG games, while the gameplay preference profile has a higher weight in casual competitive games.

[0046] (3) Dynamic update mechanism for profiles This module is designed with a dual-mode dynamic update mechanism, which is divided into real-time triggered update and offline periodic update. At the same time, the time decay function of Newton's cooling law is introduced to realize the dynamic adjustment of player behavior weight and solve the problem of static lag in existing profile technology.

[0047] Time decay factor formula: in: The decay weight corresponding to the behavior that occurs at time t, with a value range of (0,1]. The closer the behavior occurs, the higher the weight. This is the attenuation coefficient, which can be adjusted according to the game category. The value range is [0.1, 1]. High values ​​are used for fast-paced casual games, and low values ​​are used for slow-paced MMORPGs. Let t be the current time, and t be the time when the action occurs. It is the ratio of the time interval between the occurrence of the behavior to the statistical period T; T represents the portrait statistics period, which is set to 30 days by default and can be adjusted according to the game's operation rhythm.

[0048] Real-time trigger updates: When a player performs key actions such as leveling up, completing level challenges, consuming items, or making in-game purchases, the corresponding profile tag is immediately updated with an update delay of less than 1 second, ensuring that the profile matches the player's current game status in real time; Offline periodic update: A full profile update is performed every day at midnight (T+1). Based on the player's full behavioral data over the past 30 days, the label weights and feature values ​​of each dimension are recalculated to complete the full iteration of the profile.

[0049] The system saves the player's full-dimensional behavioral profile tag library, which contains each player's 6 major-dimensional quantified tags, weight values ​​and feature vectors. It is divided into real-time profile cache (for real-time recommendation) and offline profile library (for model training and batch recommendation), and is synchronously transmitted to the personalized synthetic game order generation module, the intelligent order sorting and contextual push module.

[0050] (iv) Personalized merge game order generation module This module is the core execution unit of the system. It is responsible for generating a set of item candidates that match the player's needs based on the player's full-dimensional behavioral profile, by integrating collaborative filtering and improving the Wide&Deep recommendation model. Then, through the order synthesis rule engine, it dynamically generates personalized synthesis game orders, solving the core problems of homogenization and low matching degree of existing fixed gift packs.

[0051] The fusion recommendation model designed in this module is divided into an improved collaborative filtering recall layer and an improved Wide&Deep deep learning ranking layer. First, the recall layer filters out the top-N item candidates from the full item library. Then, the ranking layer accurately scores the candidate items and finally determines the list of items to be included in the order synthesis.

[0052] 1. Improved collaborative filtering recall layer, such as Figure 5 As shown, this method integrates user-based collaborative filtering (User-CF) and item-based collaborative filtering (Item-CF), and incorporates a time decay factor to optimize similarity calculation, thus solving the cold start and data sparsity problems of traditional collaborative filtering. The core calculation process is as follows: User-based collaborative filtering: Calculate the Pearson correlation coefficient between the target player and similar players, and optimize it by incorporating a time decay factor. The calculation formula is as follows: in: The similarity between the target player u and the comparison player v is denoted by , and the value ranges from [-1, 1]. The higher the value, the higher the similarity. A collection of items that players u and v have both interacted with; , The scores are for players u and v’s behavior towards item i (based on purchase, use, and click behaviors, with a score of 1-5). , These are the average ratings of players u and v for all items, respectively. This is a time decay factor, consistent with the formula in the profile update module, which increases the weight of similarity in recent behaviors. Based on the similarity calculation results, the Top-50 similar players are selected, and their high-rated and high-purchase-rate items are included in the candidate set to complete User-CF recall.

[0053] Item-based collaborative filtering: Calculates the cosine similarity between items and optimizes it by combining the item's game attribute tags. The formula is: in: The set of players who have acted on both item i and item j simultaneously; This is the item attribute matching coefficient. If items i and j belong to the same type, are applicable to the same gameplay, and are suitable for the same game stage, ,otherwise Increase the similarity weight of items with the same attributes. Based on the similarity calculation results, include the top-30 similar items of the target player's historical preferred items into the candidate set to complete Item-CF recall.

[0054] Cold Start Adaptation: For new players / free-to-play players who have no historical behavior data, based on the player's basic attribute profile and game progress profile, popular items are matched with players of the same level, channel, and profession to complete the cold start recall and solve the cold start problem of traditional collaborative filtering.

[0055] 2. Improve the Wide & Deep learning ranking layer, such as Figure 6 As shown, the ranking layer optimizes the classic Wide & Deep model (refer to its publicly available model structure) for game scene adaptation. The Wide part learns the matching relationship between the player's explicit features and items, while the Deep part learns the implicit non-linear features of the player's behavior. Finally, it outputs a score on the matching degree between items and the player. The core structure and formula are as follows: Wide part: Employs a generalized linear model to learn the explicit matching relationship between player profile tags and item attributes, using the following formula: in: This refers to the output value of the Wide section; The raw features input (player profile tags, item attribute features); For cross-feature transformation, explicit matching relationships between "player gameplay preferences - item type", "player game progress - item level", and "player spending ability - item price" are captured through feature cross-feature transformation; This is the weight matrix for the wide portion. This is a bias term.

[0056] Deep processing: Employs a 3-layer fully connected neural network to learn the implicit non-linear features of player behavior, using the following formula: in: The activation value of the l-th hidden layer is the input layer. High-dimensional feature vectors for player profiles and embedding vectors for items; , These are the weight matrix and bias term of the l-th layer, respectively; ReLU is an activation function used to fit nonlinear features, and its formula is: The final output of the Deep part is: .

[0057] Joint Model Training and Loss Function: The outputs of the Wide and Deep parts are combined, and the final item matching score is output through the Sigmoid function. The formula is as follows: in, For the Sigmoid function, The matching degree between items and players is scored, with values ​​ranging from [0,1]. Higher scores indicate a higher matching degree. The model employs an improved cross-entropy loss function, incorporating the business objective of game monetization conversion and adding positive sample weights to address the sparsity issue of monetization conversion samples. The formula is: in: The total number of training samples, The true labels for the samples are (1 for paid conversions, 0 for unpaid conversions). This is the positive sample weight coefficient, with a value of 5-10, used to increase the loss weight of positive samples in paid conversions, adapting to the sparse characteristics of game paid samples; This is an L2 regularization term to prevent the model from overfitting.

[0058] Model training and inference: An offline batch training + online incremental update mode is adopted. The model is incrementally trained daily based on the full amount of player behavior data from the previous day to ensure the model's adaptability to changes in player behavior. During inference, the item candidate set output by the recall layer is scored item by item matching degree, and the top-10 high matching degree items are selected to enter the subsequent order synthesis stage.

[0059] 3. Dynamic Order Generation Rule Engine for Merge Games Based on the highly matched item list output by the model, and combined with the player's full-dimensional behavioral profile, personalized game orders are dynamically synthesized through a rule engine. The core rules for order synthesis are as follows: (1) Item combination rules: Based on the player's game progress profile and item inventory data, prioritize the combination of core items required for the player's current stuck level / gameplay, and match them with auxiliary items and consumable items to form a combination structure of "core items + auxiliary items + gifts" to avoid item redundancy; (2) Pricing and discount rules: Based on the player's spending ability profile and spending tier preferences, determine the order price tier, and at the same time, combine the player's discount sensitivity to set the matching discount level: for discount sensitive players, set high discount + low threshold orders; for price insensitive paying players, set high value + exclusive gifts orders; (3) Validity period and scenario rules: Based on the player's active profile, set the validity period of the order, and at the same time match the push scenario to set exclusive order benefits, such as "stuck point assistance order" for level failure scenario, "growth assistance order" for level improvement scenario, and "event exclusive order" during the event period; (4) Uniqueness rule: Each synthesis order is exclusive to the player to avoid different players receiving the same fixed gift pack. At the same time, based on the player's historical order records, it avoids repeatedly pushing the same type of order that has been purchased / rejected.

[0060] During runtime, the engine dynamically generates personalized game orders for each player using the input player full-dimensional behavior profile tag library, the game's full item attribute library, player historical order records, and item inventory data. These orders include core information such as order ID, item combination list, pricing, discount, validity period, and applicable scenarios, and are then transmitted to the order intelligent sorting and scenario-based push module.

[0061] (v) Intelligent order sorting and contextualized push module This module serves as the system's outreach and execution unit, responsible for intelligently sorting the generated synthesis orders based on multi-objective optimizations. It also combines players' game scenarios and activity patterns to achieve trigger-based, contextualized, and precise push notifications, thereby improving paid conversion rates while ensuring a smooth gaming experience for players.

[0062] like Figure 7 As shown, the module performs the following main operations: 1. Intelligent order sorting with multi-objective optimization The LambdaMART sorting algorithm is adopted to achieve intelligent order sorting with multi-objective optimization. The optimization objectives include four core dimensions, and the weights of each objective are determined by the Analytic Hierarchy Process (AHP), as follows: The core of the LambdaMART algorithm is to fit the sorted Lambda gradient using a gradient boosting tree (GBDT). The Lambda gradient is a multi-objective weighted sorting gradient calculated as follows: in: Let this be the Lambda gradient value of order i relative to order j; The weights of the k-th optimization objective are 1, and the sum of all weights is 1. The change in the normalized cumulative gain of the k-th optimization objective is used to measure the ranking quality. , Assign scores to the order rankings of orders i and j output by the model; This is the Sigmoid function, used to calculate the ordering probability.

[0063] Through multiple rounds of iterative training, the LambdaMART model outputs the final ranking score for each order, sorts the orders from highest to lowest score, and determines the optimal order sequence for each player.

[0064] 2. Contextualized and precise push mechanism This module is designed with a dual-mode push mechanism: triggered scenario-based push and timed periodic push. The core push scenarios and triggering rules are as follows: (1) Core Trigger Push Scenarios: When a player performs a corresponding action in the game, a matching order push is immediately triggered to capture the peak of the player's willingness to pay. Specific scenarios include: In the event of failing a level / dungeon challenge: a "stuck point assistance order" containing the items needed to clear the level will be sent. Level Up / Combat Power Breakthrough Scenarios: Push "Growth Assistance Orders" matching the new level; In scenarios where item inventory is low or items have been consumed: A "supply order" for the consumed item will be pushed to the user. When a game event starts or a task is completed: a special "limited-time order" will be pushed to the user. After a player completes their first recharge: a "recharge bonus order" matching the first recharge tier will be sent to them.

[0065] (2) Scheduled periodic push scenario: Based on the player's active time profile, push general personalized orders during the player's daily login peak and online peak hours, and push exclusive return gift pack orders to returning players who have not logged in for 7 days.

[0066] (3) Push notification channels: Based on the urgency of the order and the appropriate scenario, select channels such as in-game pop-ups, in-game mail, red dot notifications, and push notifications to avoid excessively disturbing players while ensuring the order reach rate.

[0067] (4) Push frequency control: Set a daily push limit (the default is no more than 3 times per day) to avoid frequent pushes causing players to feel disgusted. At the same time, for orders that players refuse continuously, the push priority of similar orders will be automatically reduced.

[0068] Therefore, by intelligently sorting personalized game orders, player full-dimensional behavior profiles, player in-game real-time behavior events, and game activity operation rules, the sorted order push sequence, push trigger instructions, push channels and timing configurations can be output and pushed to the player's game client through the game server to complete the order delivery.

[0069] (vi) Recommendation effect feedback and model iteration optimization module This module is the system's closed-loop optimization unit, responsible for collecting player feedback data on pushed orders across the entire process, calculating core recommendation performance metrics, and iteratively optimizing user profile weights, algorithm models, and order synthesis rules through offline evaluation and online A / B testing to continuously improve the system's recommendation performance. The specific implementation process includes: 1. Full-link feedback data collection: Collect player behavior data throughout the entire process of pushed orders, including order exposure, clicks, viewing details, adding to cart, payment completion, order cancellation, and continuous rejection, while also linking player activity duration and retention data after the push to form a complete feedback data link.

[0070] 2. Calculation of Core Metrics for Recommendation Effectiveness: Constructing a multi-dimensional effectiveness evaluation indicator system, with core indicators including: Traffic conversion metrics: Click-through rate (CTR), cost-per-click conversion rate (CVR), cost-per-impression conversion rate; Revenue metrics: Paid subscription penetration rate, average revenue per user (ARPPU), average revenue per user (ARPU), and GMV contribution from orders; User experience metrics: order rejection rate, player complaint rate, 7-day retention rate after push notification, and changes in game online time.

[0071] 3. Offline Model Evaluation and Iteration: Based on the feedback data from the previous day, an offline training dataset is constructed daily to incrementally train the fusion recommendation model and the ranking model. At the same time, the model performance is evaluated through offline AUC and NDCG metrics. If the new model's metrics are better than the online model, the model version iteration is completed. Simultaneously, based on the feedback data, the weights of the profile dimension, order synthesis rules, and push scenario triggering conditions are optimized.

[0072] 4. Online A / B Testing: For major model iterations and rule adjustments, online A / B testing is conducted with a control group and an experimental group. The testing period is 7-14 days. The effectiveness of the optimization scheme is verified by the statistical differences of core indicators. Full deployment is only carried out when the experimental group's indicators are significantly better than those of the control group.

[0073] 5. Anomaly Monitoring and Backup Mechanism: Real-time monitoring of order push conversion data and player feedback data. If anomalies such as a sudden increase in order rejection rate or player complaint rate occur, the backup mechanism will be triggered immediately to suspend the corresponding push rules and revert to the stable version of the model to ensure the stable operation of the system.

[0074] The system achieves full-process flow control through a centralized scheduling engine, with core control logic divided into three main lines: 1. Real-time recommendation control main line: After a player's real-time behavior event in the game is triggered, the scheduling engine sequentially calls the data collection, real-time preprocessing, real-time profile update, order generation, and sorting push modules to complete the triggered order recommendation. The entire process adopts an event-driven asynchronous scheduling mode to avoid blocking the game's main thread; 2. Offline Batch Recommendation Control Mainline: Every day at midnight, the scheduling engine executes offline data synchronization, data preprocessing, full profile update, batch order generation, and model offline training tasks in sequence according to the preset workflow. The DAG workflow scheduling mode is adopted to ensure the dependencies and execution order of tasks. 3. System Iteration and Optimization Control Mainline: The scheduling engine collects feedback data from the previous day daily and automatically executes tasks such as performance indicator calculation, model evaluation, and parameter optimization. It also supports manually configured A / B test scheduling, rule adjustment, and model version deployment / rollback operations to achieve controllable system iteration.

[0075] The system application logic will be explained in detail below in conjunction with the methodology and process.

[0076] II. Order Recommendation Method for Merge Games Based on This System The game order recommendation method based on player behavior profiles of the present invention is implemented based on the above system and specifically includes the following steps: Step 1: Full-link collection of multi-dimensional player behavior data This step synchronizes data between the game client's tracking SDK and the server, collecting behavioral data across six dimensions: basic player attributes, gameplay behavior, spending behavior, spending capacity, social interaction, and active retention. It also separates real-time behavioral data streams from offline datasets, providing a complete data source for subsequent processing. This overcomes the limitations of existing technologies that only collect payment data, obtaining comprehensive behavioral data that fully reflects player game status and needs, ensuring the accuracy of subsequent profiling and recommendations from the data source perspective.

[0077] Step 2: Preprocessing and Feature Engineering of Raw Data The collected raw behavioral data undergoes data cleaning to handle missing and outlier values. Min-Max standardization unifies the dimensions of continuous features, and one-hot encoding / label encoding converts categorical features into numerical values. Variance filtering and mutual information methods are then used to filter features, ultimately generating a standardized set of player feature vectors. This process transforms unstructured raw behavioral data into structured feature vectors suitable for profile building and algorithmic computation, removing data noise and invalid features, reducing feature dimensionality, and ensuring the accuracy and efficiency of subsequent algorithmic calculations.

[0078] Step 3: Construction and dynamic updating of player full-dimensional behavioral profiles Based on standardized player feature vectors, a comprehensive player behavior profile system is constructed, encompassing six dimensions: basic attributes, game progress, gameplay preferences, spending habits, spending power, and active retention. An improved TF-IDF algorithm is used to calculate the weight values ​​of each dimension's tags, while a time decay factor is introduced. Through a dual-mode approach of real-time triggered updates and offline periodic updates, the profile undergoes dynamic iteration. This allows for the creation of a precise, dynamic, comprehensive profile that accurately reflects a player's current game state, needs, preferences, and spending potential. This addresses the core issues of existing profiling technologies, such as limited dimensionality and static lag, providing a precise basis for subsequent order recommendations.

[0079] Step 4: Generation of a personalized item candidate set based on a fusion recommendation model By employing an improved collaborative filtering recall layer that integrates User-CF and Item-CF algorithms, initial item recall is achieved, generating a Top-N item candidate set. This addresses the cold start and data sparsity issues inherent in traditional collaborative filtering. An improved Wide&Deep deep learning ranking layer then scores the matching degree of the candidate items, selecting the Top-10 highly matched items for the order synthesis stage. Combining the interpretability of collaborative filtering with the non-linear fitting capabilities of deep learning models, and adapting to specific game scenarios, this significantly improves the matching degree and accuracy of item recommendations, resolving the issues of poor generalization and insufficient cold start adaptation in existing algorithms.

[0080] Step 5: Dynamically generate game orders by matching player profiles. Based on a highly matched item list and combined with a comprehensive player behavior profile, a personalized game order is dynamically generated through an order synthesis rule engine. This engine determines item combinations based on the player's game progress and item shortages, sets pricing tiers based on their spending power profile, adjusts discount levels based on their sensitivity to promotions, and determines order validity and exclusive benefits based on the push notification scenario. Each order is custom-made for each player. This approach breaks through the homogenization limitations of existing fixed gift pack templates, enabling personalized and dynamic synthesis of game orders. The order content is highly matched to the player's current needs, spending power, and game status, fundamentally increasing the player's willingness to pay.

[0081] Step 6: Multi-objective optimized intelligent order sorting and contextualized push notifications Employing the LambdaMART ranking algorithm, with multiple optimization objectives including paid conversion rate, click-through rate, average order value, and player retention, the algorithm intelligently sorts generated synthetic orders to determine the optimal push sequence. Simultaneously, based on players' real-time in-game behavior events, order pushes are triggered in core scenarios such as level failure and level advancement, and scheduled pushes are combined with player activity periods to ensure a consistent player experience. This achieves precise order sorting and contextualized reach, capturing peak player spending intentions and significantly improving order reach, click-through rate, and conversion rate, while balancing monetization and player experience, and avoiding user churn caused by excessive pushes.

[0082] Step 7: Recommendation effect feedback and system closed-loop iterative optimization The system collects player feedback data on pushed orders across the entire process, calculates core metrics for recommendation effectiveness, and iteratively optimizes user profile weights, recommendation models, order composition rules, and push mechanisms through offline evaluation and online A / B testing. Simultaneously, it incorporates anomaly monitoring and fallback mechanisms to ensure stable system operation. Therefore, the system constructs a complete business loop of "data collection - user profile building - order recommendation - feedback iteration," enabling continuous self-optimization of recommendation effectiveness. This allows the system to adapt to changes in player behavior and adjustments to the game's operational rhythm, ensuring long-term recommendation effectiveness and business value.

[0083] Example 1 This embodiment uses a leading domestic mobile game (such as Perfect World, MMORPG, or other similar mobile games) as its application scenario. The monetization model is in-app purchases, with core monetization points being equipment, upgrade items, and cosmetic items. This implementation deploys the system of this invention within the game's monetization recommendation system, while simultaneously setting up a control group (using an industry-standard fixed gift pack + basic collaborative filtering recommendation scheme) for a 14-day online A / B test to verify the technical effectiveness of this invention.

[0084] The specific implementation process is as follows: 1. Data collection and deployment: The lightweight event tracking SDK of this system is built into the game client, deploying 28 core behavior tracking points to collect all player behavior data across 6 dimensions. At the same time, it connects to the business database and payment system of the game server to achieve T+1 synchronization of offline data.

[0085] 2. System Deployment and Model Training: Deploy the full-module system of this invention, and based on the full player behavior data of the game in the first 3 months, complete the pre-training of the integrated recommendation model and ranking model. Use the AHP hierarchical analysis method to determine the profile dimension weights and ranking target weights suitable for the MMORPG category, and configure order synthesis rules and push scene triggering conditions.

[0086] 3. A / B testing grouping: Using stratified random sampling, the 2 million active players in the game were divided into an experimental group and a control group, with 1 million players in each group. There were no statistically significant differences in the level distribution, spending ability distribution, and active time distribution between the two groups, ensuring the fairness of the test.

[0087] Experimental group: The game order recommendation system based on player behavior profiles of this invention was used; Control group: The industry-standard solution was adopted, which pushes fixed template gift packs based on player level and VIP level, and recommends items in combination with basic collaborative filtering algorithm.

[0088] 4. Testing period and data collection: The testing period is 14 days. Core business indicators of two groups of players will be collected daily. After the test, full data statistics and effect comparison will be carried out.

[0089] Example of a specific player order generation: Using two typical players from the experimental group as examples, the order synthesis effect of this system is demonstrated: Player A: Level 35, average monthly recharge of 500 RMB, core gameplay is PVE dungeons, currently stuck on the level 35 elite dungeon, lacking the core item "Enhancement Stones" needed to clear the dungeon, moderate sensitivity to discounts, preferred payment tier is 30-98 RMB. This system creates an exclusive order for him:

Dungeon Clearance Assistance Pack

[0090] 2. Player B: Level 15, a new player who has never spent money on the game. Their core gameplay revolves around casual quests and the main storyline. They have no history of making in-app purchases, are highly sensitive to discounts, and have low to mid-range spending potential. This system provides them with a specially crafted order: the "New Player Growth Welfare Pack," which includes a full set of beginner equipment, 10,000 gold coins, and 3 appearance trial cards. Priced at 6 yuan, it is discounted by 20%, with an additional double gold coin bonus for the first recharge. This pack will be sent via in-game mail after the player completes the level 15 main quest.

[0091] In the control group, both players received the system's fixed "level-up gift packs." Player A received a general level-up gift pack worth 98 yuan, while Player B received a first-time recharge gift pack worth 30 yuan. These gifts were not well-matched to the players' specific needs and spending power.

[0092] After the testing period ended, the core indicators of the experimental group and the control group were compared as shown in the table below: The test results show that, compared with the industry's common solutions, the technical solution of this invention has achieved a comprehensive and significant improvement in core business indicators: 1. The click-through rate and conversion rate of orders have increased several times, proving that the personalized orders generated by this invention match the needs of players much better than traditional fixed gift packs, and players' acceptance and willingness to pay have been significantly improved; 2. The paid penetration rate and ARPU have increased significantly, while ARPPU has remained basically the same, proving that this solution does not achieve revenue growth by increasing the amount recharged per user, but by increasing the scale of paying users. This avoids over-harvesting core paying users and is more in line with the long-term operation needs of the game. 3. The player 7-day retention rate has significantly improved and the order rejection rate has dropped sharply, proving that this solution has not damaged the player experience while improving the commercialization effect. On the contrary, through precise demand matching, it has improved the player's game experience and retention intention, achieving a balance between commercial value and user experience.

[0093] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0094] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0095] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0096] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0097] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0098] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0099] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0100] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0101] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0102] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0103] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for recommending a synthetic game order based on a player behavior profile, characterized in that, Includes the following steps: S1 collects multi-dimensional player behavior data across the entire process. The behavior data includes player basic attribute data, core gameplay behavior data, consumption and payment behavior data, payment ability profile data, social interaction behavior data, activity and retention behavior data, and separates real-time behavior data streams from offline behavior datasets. S2 preprocesses and performs feature engineering on the collected raw behavioral data to generate a standardized set of player feature vectors. S3 is based on standardized player feature vectors to build a full-dimensional player behavior profile system that includes six dimensions: basic attributes, game progress, gameplay preferences, spending habits, spending power, and active retention. The profile is updated in real time through a time decay factor. S4 generates a highly matched set of item candidates that match player needs by integrating improved collaborative filtering and Wide & Deep deep learning recommendation models based on a player's full-dimensional behavioral profile. S5 dynamically generates personalized merge game orders based on a highly matched set of item candidates and a full-dimensional profile of player behavior through an order merging rule engine. S6 uses a multi-objective optimized sorting algorithm to intelligently sort synthetic game orders, and combines players' real-time in-game behavior events and activity patterns to complete contextualized and precise push notifications. The S7 collects player feedback data on pushed orders across the entire process, and uses this feedback data to complete closed-loop iterative optimization of profile weights, recommendation models, order synthesis rules, and push mechanisms.

2. The method of claim 1, wherein, In step S2, the preprocessing and feature engineering construction specifically include: processing missing and outlier values ​​in the original data, identifying and truncating outliers using the 3σ principle, and processing missing values ​​using the longitudinal time series imputation method; mapping continuous features to the [0,1] interval through Min-Max standardization; completing the numerical conversion of categorical features through one-hot encoding and label encoding; and completing the effective feature selection through variance screening and mutual information method to generate a standardized player feature vector set.

3. The method of claim 1, wherein, In step S3, the player's preference weights for different gameplay modes are calculated using the improved TF-IDF algorithm, with the following formula: wherein, is a preference weight of player i for game j, is a proportion of the total game time that player i spent playing game j, is an inverse document frequency of game j, is a time decay factor; The time decay factor is constructed based on Newton's law of cooling, and the formula is: in, The attenuation coefficient is... t represents the current time, t represents the time the action occurred, and T represents the portrait statistics period. The portrait update adopts a dual mode of real-time triggered update and offline periodic update. When a player performs a key action, the portrait is immediately updated in real time, and a full portrait offline update is performed every day on T+1.

4. The method according to claim 1, characterized in that, In step S4, the fusion recommendation model is divided into an improved collaborative filtering recall layer and an improved Wide&Deep deep learning ranking layer. The improved collaborative filtering recall layer integrates user-based collaborative filtering and item-based collaborative filtering. It adds a time decay factor to optimize user similarity calculation and an item attribute matching coefficient to optimize item similarity calculation. It also designs a cold start recall strategy for new players. The improved Wide&Deep deep learning ranking layer learns the explicit matching relationship between player profiles and item attributes through the Wide part, and learns the implicit nonlinear features of player behavior through the Deep part. It is jointly trained using an improved cross-entropy loss function, and positive sample weight enhancement and L2 regularization terms are added. Finally, it outputs the matching degree score between items and players, and selects a candidate set of items with high matching degree.

5. The method according to claim 1, characterized in that, In step S5, the core rules of the order synthesis rule engine include: determining the item combination of "core item + auxiliary item + gift" based on the player's game progress profile and item inventory data; determining the order pricing tier based on the player's spending ability profile and spending tier preference; setting the matching discount level based on the player's discount sensitivity; setting the order validity period and exclusive benefits based on the push scenario, and finally generating a personalized synthesis game order exclusive to the player.

6. The method according to claim 1, characterized in that, In step S6, the LambdaMART sorting algorithm is used to intelligently sort orders. The optimization objectives are order payment conversion rate, click-through rate, average order value, and player 7-day retention rate. The weight of each objective is determined by the analytic hierarchy process, the Lambda gradient of the multi-objective weighted algorithm is fitted, and the final order ranking score is output. The scenario-based push notification adopts a dual mode of triggered scenario push notification and timed periodic push notification. Order push notifications are triggered in core scenarios such as player level failure, level upgrade, item consumption, and event start. Timed push notifications are made based on player activity time profiles, while a daily push frequency limit is set.

7. The method according to claim 1, characterized in that, In step S7, the closed-loop iterative optimization specifically includes: collecting full-process feedback data of player order exposure, clicks, payments, rejections, and order cancellations, and calculating core indicators of recommendation performance; completing incremental training and offline evaluation of the recommendation model daily based on feedback data, verifying the effectiveness of the optimization scheme through online A / B testing, and simultaneously monitoring abnormal data in real time and setting up a fallback mechanism.

8. A synthetic game order recommendation system based on player behavior profiles, used to implement the method described in any one of claims 1-7, characterized in that, The system includes: The player multi-dimensional behavior data collection module is used to collect player behavior data in six dimensions across the entire process, separating real-time data streams from offline datasets; The data preprocessing and feature engineering module is used to clean, standardize, encode, and filter the raw behavioral data to generate a standardized set of player feature vectors. The Player Full-Dimensional Behavior Profile Construction and Dynamic Update Module is used to build a player behavior profile system with six dimensions, and achieves real-time dynamic updates of the profile through a time decay factor. The personalized merge game order generation module is used to generate a highly matching set of item candidates by integrating a recommendation model, and dynamically generate personalized merge game orders through a rule engine. The intelligent order sorting and contextualized push module is used to intelligently sort orders through a multi-objective sorting algorithm and combine player game behavior to achieve contextualized and accurate push notifications. The recommended feedback and model iteration optimization module is used to collect player feedback data and complete the closed-loop iterative optimization of the system.

9. The system according to claim 8, characterized in that, The system adopts a dual-link communication architecture of real-time data stream + offline data stream. Real-time data is transmitted through Kafka message queue, and offline data is stored and computed through Hadoop distributed architecture. The system stores real-time profiles and order data through Redis cluster, offline profiles and business data through MySQL database, and massive historical behavior data through HDFS.

10. The system according to claim 8, characterized in that, The personalized game order generation module has an improved collaborative filtering recall layer and an improved Wide&Deep deep learning sorting layer. It is also configured with a dynamically adjustable order synthesis rule engine, which supports rule adaptation and custom configuration for different game categories.

Images