A social game recommendation method and system based on artificial intelligence

Abstract

The present invention discloses a social game recommendation method and system based on artificial intelligence, which relates to the field of game recommendation technology. The method identifies users with low social pressure sensitivity, obtains the virtual image characteristics, skill characteristics, emotional tags and text chat keywords of the users with low social pressure sensitivity on multiple game platforms, and then outputs the social type of the users with low social pressure sensitivity; combines the core recommendation parameters of the games to be recommended and the social type of the users with low social pressure sensitivity to obtain the initial recommendation value of each game to be recommended; recommends and sorts the games according to the size of the initial recommendation value; outputs the adjusted optimization value of each game based on the real-time monitored in-game data and emotional data, and then generates a comprehensive recommendation value for each game. This solves the problem that traditional game recommendation systems lack accurate identification and dynamic adaptation of users' social pressure sensitivity, resulting in mismatch between recommended content and user needs and loss of users due to social pressure.

Description

Technical Field

[0001] The present invention relates to the technical field of game recommendation, and in particular to a social game recommendation method and system based on artificial intelligence. Background Art

[0002] With the rapid development of mobile internet and artificial intelligence technologies, the social gaming market continues to expand, and user demands for social gaming are becoming increasingly diverse and personalized. Currently, the background technologies for social game recommendations primarily include content-based recommendation techniques, collaborative filtering-based recommendation techniques, and social relationship-based recommendation techniques. However, these traditional recommendation technologies are gradually exposing limitations when faced with users' complex and ever-changing social needs and gaming scenarios.

[0003] Existing recommendation technologies often only focus on users' gaming behaviors and preferences, and lack in-depth analysis of users' social characteristics (such as social pressure sensitivity, social type, etc.); traditional recommendation technologies make static recommendations based on users' historical data and cannot promptly respond to users' real-time feedback and changes in demand during the game; and when selecting recommendation parameters, existing recommendation technologies often only consider some of the game's characteristics, such as game type or popularity, and lack comprehensive and integrated consideration of the game's core recommendation parameters. At the same time, they fail to fully integrate users' social behaviors and game characteristics, resulting in the recommended games being difficult to meet user expectations in terms of social interaction.

[0004] Therefore, in response to the above problems, there is an urgent need for a social game recommendation method and system based on artificial intelligence. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides an artificial intelligence-based social game recommendation method and system, which solves the problem that traditional game recommendation systems lack accurate identification and dynamic adaptation of users' sensitivity to social pressure, resulting in mismatch between recommended content and user needs and user loss due to social pressure.

[0006] To achieve the above-mentioned objectives, the present invention is implemented through the following technical solutions: a social game recommendation method based on artificial intelligence, comprising the following steps: S1, identifying users with low social pressure sensitivity based on social avoidance indicators in user historical behavior data; S2, obtaining virtual image characteristics and skill characteristics of users with low social pressure sensitivity across multiple game platforms, and identifying emotional tags and text chat keywords in multiple game platforms of users with low social pressure sensitivity, and then outputting the social type of users with low social pressure sensitivity; S3, extracting core recommendation parameters of the games to be recommended, and then obtaining initial recommendation values ​​for each game to be recommended based on the social type of users with low social pressure sensitivity; S4, recommending and ranking the games according to the size of the initial recommendation values, and simultaneously monitoring in-game data and user emotional data of users with low social pressure sensitivity during the game in real time; S5, outputting adjusted optimization values ​​for each game based on the in-game data and emotional data, generating a comprehensive recommendation value for each game based on the initial recommendation value and the adjusted optimization value, re-ranking the games based on the comprehensive recommendation value, and continuously updating recommended social games for users.

[0007] Furthermore, the specific steps of S1 are: extracting social avoidance indicators, which include the rate of increase in the number of friends, frequency of social interaction, proportion of chat silence, and number of team rejections; taking the social avoidance indicators as input, using support vector machines for training, building a user social pressure sensitivity model, and outputting the user's social pressure sensitivity value; comparing the social pressure sensitivity value with the sensitivity threshold, when the social pressure sensitivity value is greater than or equal to the sensitivity threshold, marking the user as a high social pressure sensitivity user, entering the single-player task game recommendation branch, and recommending single-player task-dominated games; when the social pressure sensitivity value is greater than or equal to the sensitivity threshold, marking the user as a low social pressure sensitivity user, triggering the social multiplayer task game recommendation demand.

[0008] Furthermore, the specific steps of outputting the social type of users with low social pressure sensitivity are: obtaining virtual image characteristics, skill characteristics, emotion labels and text chat keywords of known social types as training samples, and then using the training samples to train a decision tree algorithm to construct a decision forest model composed of multiple decision trees; using the virtual image characteristics, skill characteristics, emotion labels and text chat keywords of users with low social pressure sensitivity as inputs of the decision forest model, and outputting the social type of users with low social pressure sensitivity, wherein the social types include leadership type, assistance type, observation type and follower type.

[0009] Furthermore, the virtual image feature acquisition and analysis is as follows: obtaining multi-source heterogeneous data of the user's virtual image, constructing a virtual image feature vector space matrix, and numerically representing the virtual image using feature encoding and attribute weighting; the skill feature acquisition and analysis is as follows: obtaining skill structured data, extracting skill text features, constructing a skill knowledge graph network, and realizing multi-dimensional quantification of skill features through level mapping, frequency normalization and attribute weighting; the emotion feature acquisition and analysis is as follows: obtaining user multi-platform voice chat records based on user authorization, performing hierarchical emotion annotation based on the voice chat records, and identifying a comprehensive emotion index based on tag frequency and polarity weight; the text chat keyword acquisition and analysis is as follows: keyword screening is performed through the TF-IDF algorithm combined with a domain dictionary, and semantic quantification is realized by dynamically adjusting the weight coefficient according to the scene type.

[0010] Furthermore, the specific steps of S3 are: collecting the core recommendation parameter praise data corresponding to each social type of multi-user games, wherein the core recommendation parameter praise data is specifically the total number of praises from multiple users for a certain core recommendation parameter; by comparing the total number of praises for each core recommendation parameter of users of each social type, using the priority diagram method to assign a proportional coefficient to each core recommendation parameter, and determine the proportional coefficient of each core recommendation parameter of users of each social type; identifying the proportional coefficient of each core recommendation parameter of the social type to which users with low social pressure sensitivity belong, and at the same time combining the core recommendation parameters of the game to be recommended, generating the initial recommendation value of each game to be recommended.

[0011] Furthermore, the core recommendation parameters specifically include social interaction complexity parameters, social role positioning parameters, game atmosphere style parameters and player matching mechanism parameters.

[0012] Furthermore, the specific steps of S5 are: during the game process of users with low social pressure sensitivity, real-time acquisition of in-game data, the in-game data including the participation popularity of activities released in the game and the frequency of player interaction in the community; using emotion recognition technology to obtain user emotion data in real time, the user emotion data including user pleasure, user concentration and user stress level; using in-game data and user emotion data to build an adjustment optimization model, and then output the adjustment optimization value of each game; obtaining a fusion coefficient, combining the initial recommendation value, the adjustment optimization value and the fusion coefficient to generate a comprehensive recommendation value, and then re-ranking the game recommendations.

[0013] A social game recommendation system based on artificial intelligence (AI) is provided, which utilizes the aforementioned AI-based social game recommendation method and comprises: a user type identification module for identifying users with low social stress sensitivity based on social avoidance indicators in historical user behavior data; a social type identification module for obtaining avatar features and skill features of users with low social stress sensitivity across multiple game platforms, identifying emotional tags and text chat keywords across multiple game platforms of users with low social stress sensitivity, and outputting the social type of the users with low social stress sensitivity; an initial recommendation analysis module for extracting core recommendation parameters of the games to be recommended, and then combining the social types of the users with low social stress sensitivity to obtain initial recommendation values ​​for each game to be recommended; a real-time monitoring module for recommending and ranking games based on the initial recommendation values, while simultaneously monitoring in-game data and user emotion data of users with low social stress sensitivity during gaming in real time; and an optimization and adjustment module for outputting an adjusted optimization value for each game based on the in-game data and emotion data, generating a comprehensive recommendation value for each game based on the initial recommendation value and the adjusted optimization value, re-ranking the games based on the comprehensive recommendation value, and continuously updating recommended social games for users.

[0014] The present invention has the following beneficial effects:

[0015] This artificial intelligence-based social game recommendation method and system divides users into two categories: high and low social pressure sensitivity by identifying their social pressure sensitivity. For users with low social pressure sensitivity, their virtual image characteristics, skill characteristics, emotional tags and text chat keywords are further analyzed to determine the user's social type. This multi-dimensional and in-depth user portrait construction enables the recommendation system to accurately capture the user's personalized needs. Compared with traditional recommendation methods, it greatly improves the accuracy and fit of recommendations, and provides users with games that are more in line with their own interests and social preferences. For users with low social pressure sensitivity, the recommended games are matched according to their social type and core game recommendation parameters, and in-game data and user emotional data are monitored in real time during the game, and the recommendation results are dynamically adjusted. It can respond to changes in the user's experience during the game in a timely manner. For example, if the user expresses dissatisfaction with the current game atmosphere or low enthusiasm for participating in activities in the game, the system updates the recommendation by adjusting the optimization value. , ensuring that users always have a good gaming experience and enhancing users' satisfaction and loyalty to the recommendation system and game platform; when recommending games, the core recommendation parameters such as the game's social interaction complexity, social role positioning, game atmosphere style and player matching mechanism are comprehensively considered, and recommendations are made in combination with the user's social type, which helps the game platform to reasonably allocate resources, accurately recommend games to suitable user groups, increase the game's exposure and user participation, and at the same time reduce users' time and energy waste on inappropriate games, and achieve efficient matching of game resources and user needs; in the recommendation process, the user's social behavior and preferences are fully considered, such as identifying user types through social avoidance indicators, and determining social types based on users' social interaction data in the game. The recommended games can better meet users' social needs, promote users' social interaction in the game, enhance the social attributes of the game, form a good game social ecology, and enhance the game's attractiveness and user stickiness.

[0016] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a flow chart of a social game recommendation method based on artificial intelligence in the present invention.

[0018] Figure 2 This is a structural diagram of an artificial intelligence-based social game recommendation system of the present invention. DETAILED DESCRIPTION

[0019] The embodiments of the present application use an artificial intelligence-based social game recommendation method and system to provide users with accurate game recommendations with low social pressure through multimodal data fusion and real-time dynamic optimization, thereby significantly improving user satisfaction, retention rate and platform commercial value.

[0020] The overall idea of ​​the embodiments of this application is:

[0021] Based on the social avoidance indicator in users' historical behavior data, a support vector machine model is used to identify users with low social pressure sensitivity. Then, for these users, their virtual image characteristics and skill characteristics on multiple game platforms are obtained, and emotional tags and text chat keywords are identified. The user's social type is output through a decision forest model, completing deep feature extraction and classification of the user, laying the foundation for subsequent accurate recommendations. The core recommendation parameters of the recommended games are extracted, including social interaction complexity, social role positioning, game atmosphere style, and player matching mechanism. Combined with the social type of users with low social pressure sensitivity, the proportional coefficient of each core recommendation parameter is determined using the priority diagram method, and then the initial recommendation value of each recommended game is calculated to achieve a preliminary match recommendation between the game and the user's social type. After ranking and recommending the games according to the initial recommendation value, the in-game data and user emotion data of the user during the game are monitored in real time. Based on the real-time data, an adjustment optimization model is constructed to calculate the adjustment optimization value of each game, and then merge it with the initial recommendation value to generate a comprehensive recommendation value. The games are re-ranked according to the comprehensive recommendation value, and the recommendations are continuously updated to achieve dynamic and accurate personalized social game recommendations.

[0022] See also Figure 1 , an embodiment of the present invention provides a technical solution: a social game recommendation method based on artificial intelligence, comprising the following steps: S1, identifying users with low social pressure sensitivity based on social avoidance indicators in user historical behavior data; S2, obtaining virtual image characteristics and skill characteristics of users with low social pressure sensitivity across multiple game platforms, and identifying emotional tags and text chat keywords in multiple game platforms of users with low social pressure sensitivity, and then outputting the social type of users with low social pressure sensitivity; S3, extracting core recommendation parameters of the games to be recommended, and then obtaining initial recommendation values ​​for each game to be recommended based on the social type of users with low social pressure sensitivity; S4, recommending and ranking the games according to the size of the initial recommendation values, and simultaneously monitoring in-game data and user emotional data of users with low social pressure sensitivity during the game in real time; S5, outputting adjusted optimization values ​​for each game based on the in-game data and emotional data, generating a comprehensive recommendation value for each game based on the initial recommendation value and the adjusted optimization value, re-ranking the games based on the comprehensive recommendation value, and continuously updating recommended social games for users.

[0023] Specifically, the specific steps of S1 are: extracting social avoidance indicators, which include the rate of increase in the number of friends, frequency of social interactions, proportion of chat silence, and number of team rejections; using the social avoidance indicators as input, training with support vector machines, building a user social pressure sensitivity model, and outputting the user's social pressure sensitivity value; comparing the social pressure sensitivity value with the sensitivity threshold, when the social pressure sensitivity value is greater than or equal to the sensitivity threshold, marking the user as a high social pressure sensitivity user, entering the single-player task game recommendation branch, and recommending single-player task-dominated games; when the social pressure sensitivity value is greater than or equal to the sensitivity threshold, marking the user as a low social pressure sensitivity user, triggering the social multiplayer task game recommendation demand.

[0024] In this implementation plan, the friend increase rate refers to the ratio of the number of new friends of a user within a certain period of time to the number of existing friends, reflecting the user's willingness and behavioral intensity to actively expand social relationships. The higher the ratio, the more actively the user participates in social expansion. Conversely, it means that the user is relatively passive in establishing social relationships. From the user database of the social game platform, the user's friend list data within a specific time period (such as the past week or month) is extracted, and the number of new friends and the number of existing friends are counted. The value is obtained by the ratio of the number of new friends to the number of existing friends.

[0025] The frequency of social interactions is used to measure how frequently users interact with other players within the game, such as the number of times they chat, play in a team, participate in social activities, etc., reflecting the user's level of activity in the game's social scenes. With the help of the game platform's logging system, we collect data on various social interaction behaviors of users during the game, and count the number of social interactions according to set time units (such as per hour or per day) as the frequency of social interactions.

[0026] The chat silence ratio indicates the proportion of time a user is silent in the total time they participate in game social chats, reflecting their participation and enthusiasm in social chats. The game chat system records the start and end time of each chat, as well as the timestamps of when users send messages during the chat, and calculates the total chat time and silence time of the user. The ratio of silence time to total chat time is used to derive the indicator value.

[0027] The number of team rejections indicates the number of times a user chooses to reject a team invitation after receiving one from other players in the game. It directly reflects the user's resistance to social teaming behavior. A counter is set in the game team system. Every time a user rejects a team invitation, the counter automatically increases by 1. The number of team rejections within a certain period of time is regularly counted.

[0028] The specific steps of using support vector machine to build a user social stress sensitivity model are as follows: collect user historical behavior data including social avoidance indicators such as the rate of increase in the number of friends, frequency of social interactions, proportion of chat silence, and number of team rejections from the user behavior database of the social game platform; clean the collected data, remove missing values ​​and outliers, and use the normalization method to map indicator data of different magnitudes to the same numerical range to eliminate the influence of data dimension differences; use the preprocessed social avoidance indicator data as feature vectors, and annotate each feature vector with the corresponding user social stress sensitivity category label (high social stress sensitivity is marked as 1, low social stress sensitivity is marked as 0). is 0), forming a training data set; select a suitable support vector machine kernel function (such as radial basis function RBF, polynomial kernel function, etc.), set the penalty factor C and other hyperparameters; use the training data set to train the support vector machine, and optimize the weight vector and bias term of the support vector machine by minimizing the structural risk, and build a user social pressure sensitivity model; use the reserved validation data set to test the trained user social pressure sensitivity model, and calculate the model's classification accuracy, recall rate, F1 value and other evaluation indicators; according to the evaluation results, adjust the support vector machine's kernel function type and hyperparameter values, and repeat the training and validation process until the model meets the predetermined performance indicator requirements. The support vector machine model expression is: Assume that the training set is ,in For the The social avoidance index feature vector of samples, , which correspond to the rate of increase in the number of friends, frequency of social interactions, percentage of silent chats, and number of team rejections; is the sample label, , 0 represents a user with low social pressure sensitivity, and 1 represents a user with high social pressure sensitivity. The goal of SVM is to find a hyperplane , so that samples of different categories can be separated by the maximum interval. The optimization objective function is: ; ;in, is the weight vector, is the bias term, is a slack variable used to handle abnormal points or linearly inseparable cases in the training set. is the penalty factor, which is used to control the degree of penalty for misclassified samples. The larger it is, the heavier the penalty for misclassification is, and the more the model tends to reduce classification errors.

[0029] The sensitivity value is calculated as follows: For the new user's feature vector , whose social stress sensitivity value Calculated by the decision function of the SVM model: ; where sgn is the sign function, when When, sgn , indicating users with high social pressure sensitivity; when hour, , indicating users with low social stress sensitivity.

[0030] The sensitivity threshold is obtained in the following way: based on the experience of game industry experts, the analysis of user data accumulated over a long period of time by the game platform operation team, and the research results of user behavior of similar social games, an initial sensitivity threshold is set as a reference benchmark; the social pressure sensitivity values ​​and actual game behavior feedback data of a large number of users can also be collected, and indicators such as user recommendation satisfaction and game participation under different threshold settings can be analyzed. Optimization algorithms such as grid search and random search can be used to traverse and search within the threshold range to identify the sensitivity threshold that makes the comprehensive evaluation indicators reach the optimal level; or a real-time monitoring mechanism can be established to continuously track the changes in user behavior and recommendation effect data during the game. When it is found that the behavior pattern of the user group has changed significantly (such as a large number of users have poor feedback on recommended games) or the business needs of the game platform are adjusted, the threshold adjustment program will be automatically triggered to recalculate and update the sensitivity threshold based on the current data.

[0031] The specific steps for recommending single-player task-dominated games are as follows: screen out a collection of games with single-player tasks as the core gameplay from the game resource library, and extract the key characteristic parameters of each game in the collection, including but not limited to game task type (such as adventure, puzzle solving, and cultivation), game difficulty level, plot richness, and task completion time; analyze the historical game behavior data of users with high social pressure sensitivity, and extract the user's preference characteristics for game task type, difficulty, plot, etc.; use similarity calculation algorithms such as cosine similarity and Euclidean distance to calculate the matching score between each single-player task-dominated game and the user's preference characteristics; sort the single-player task-dominated games in descending order according to the matching score, and select several games with high rankings to form a recommendation list. If there are multiple games with the same matching score, further compare the game's user rating, popularity and other additional indicators to determine the final recommendation order, and push the recommendation list to the user client to complete the recommendation process of single-player task-dominated games.

[0032] By quantifying multi-dimensional social avoidance indicators such as the rate of increase in the number of friends and the frequency of social interactions, combined with the powerful classification capabilities of support vector machines, users can be scientifically and accurately divided into two categories: high and low social pressure sensitivity. Compared with traditional user classification methods that only rely on subjective judgment or a single indicator, the accuracy and reliability of classification are improved; based on accurate user classification, differentiated recommendation strategies are formulated for users with different social pressure sensitivities, single-player task-oriented games are recommended for users with high social pressure sensitivity, and social multiplayer task game recommendations are triggered for users with low social pressure sensitivity, meeting users' personalized gaming needs and improving user gaming experience and recommendation satisfaction; after clarifying the user's social pressure sensitivity, the game platform can push games that meet user needs in a targeted manner, avoiding resource waste, improving game promotion efficiency and user participation, achieving efficient matching of game resources and user needs, and promoting the healthy development of the game platform ecosystem.

[0033] Specifically, the specific steps for outputting the social type of users with low social pressure sensitivity are: obtaining the virtual image characteristics, skill characteristics, emotion labels and text chat keywords of known social types as training samples, and then using the training samples to train the decision tree algorithm to construct a decision forest model composed of multiple decision trees; using the virtual image characteristics, skill characteristics, emotion labels and text chat keywords of users with low social pressure sensitivity as input to the decision forest model, and outputting the social type of users with low social pressure sensitivity, which includes leadership type, assistance type, observation type and follower type.

[0034] The analysis of avatar feature acquisition is as follows: obtain multi-source heterogeneous data of user avatar, construct avatar feature vector space matrix, and numerically represent avatar by feature coding and attribute weighting. The quantification of avatar feature is as follows: let avatar feature vector be ,in For the The quantized value of each feature, for appearance features, such as hairstyle, is encoded as , the hairstyle type is converted into a numerical value through the coding table. For clothing, the clothing type quantization value is , the combination complexity weight is , then the quantized value of this part is , and so on, each feature is encoded and weighted, and finally the virtual image feature vector is obtained The skill feature acquisition and analysis is as follows: obtain skill structured data, extract skill text features, build a skill knowledge graph network, and achieve multi-dimensional quantification of skill features through level mapping, frequency normalization and attribute weighting. The skill feature quantification is: let the skill feature vector be ,in For the Skill No. Quantitative value of each dimension, skill level Mapped to , normalized; skill usage frequency After normalization, ; For the key attribute words in the skill effect description, set the word quantization value to , the weight is , then the quantized value of this part is . Combine the quantized values ​​of each dimension to obtain the skill feature vector , and then comprehensively process all skill feature vectors to obtain the user's skill feature representation; the emotional feature acquisition analysis is as follows: obtain the user's multi-platform voice chat records based on user authorization, perform hierarchical emotional annotation based on the voice chat records, and identify the comprehensive emotional index based on the label frequency and polarity weight. The emotional label quantification is: let the emotional label set be , the corresponding quantization value is , such as positive emotions Quantized value , negative emotions Quantized value For voice chat records, count the frequency of each emotion tag , then the comprehensive sentiment index ; The acquisition and analysis of text chat keywords is as follows: keyword screening is performed through the TF-IDF algorithm combined with the domain dictionary, and the weight coefficient is dynamically adjusted according to the scene type to achieve semantic quantification. The quantification of text chat keywords is as follows: for the keywords in the text chat records , whose TF-IDF value is ,in Keywords Frequency of words in the text , is the total number of texts, To contain keywords The number of texts, set the scene type weight to , then the keyword The final quantized value is .

[0035] Decision forest consists of Composition, for the input feature vector (including avatar features, skill features, sentiment index, and text chat keyword quantification values), each decision tree According to the splitting rule of internal nodes Perform classification and output prediction results Leader, Supporter, Observer, Follower , and finally the user's social type is determined by a voting mechanism, that is, the type with the same prediction results from most decision trees is the final social type. If there is a tie, the selection can be made based on the weight of the decision tree or other rules.

[0036] In this implementation scheme, the decision tree algorithm is trained using training samples, and the specific steps for constructing a decision forest model are as follows: the acquired training sample data such as virtual image features, skill features, emotion tags and text chat keywords of known social types are cleaned to remove duplicate, erroneous or incomplete data records; for numerical data such as virtual image features and skill features, normalization or standardization methods are used to map them to specific numerical ranges to eliminate data dimension differences; for text-type emotion tags and text chat keywords, word segmentation and stop word removal are performed, and they are converted into numerical vector form through word vector embedding (such as Word2Vec, BERT-Embedding) and other technologies to form a standardized training sample data set; from the standardized training sample data set, a certain proportion of data is randomly selected as a training subset, and the remaining data is used as a validation subset; for each training subset, an appropriate decision tree splitting criterion (such as information gain, information gain rate, Gini index) is selected, starting from the root node, according to the characteristic attributes of the training sample, the decision tree is split according to the splitting criterion. The dataset is recursively divided according to the criterion to construct a single decision tree. During the construction process, hyperparameters such as the maximum depth of the decision tree, the minimum number of sample splits, and the minimum number of sample leaf nodes are set to prevent overfitting of the decision tree. The constructed decision tree is evaluated using a validation subset to calculate evaluation indicators such as accuracy, recall, and F1 value. The hyperparameters are adjusted based on the evaluation results to optimize the decision tree model. The above decision tree training steps are repeated to generate multiple different decision trees. These decision trees are combined into a decision forest model. During the combination process, a bagging (self-service sampling aggregation) strategy can be adopted, that is, random sampling with replacement is performed on the training samples to generate multiple different training subsets for training each decision tree to increase the differences between the decision trees. Alternatively, a random subspace method can be used to randomly select some feature attributes for splitting when training each decision tree to further improve the generalization ability and classification performance of the decision forest. The final decision forest model can perform comprehensive classification prediction on the input low social pressure sensitivity user feature data and output the user's social type.

[0037] The multi-source heterogeneous data of the user's virtual image refers to data from different channels with different data structures and forms of expression, specifically including: visual appearance data, action posture data, and personalized setting data. The visual appearance data includes the facial features, clothing features, weapons, and tool features of the virtual image. The action posture data includes the various action information of the virtual image in the game, such as the posture parameters of standing, walking, running, combat attack, skill release, etc., as well as dynamic attribute data such as the smoothness and speed of the action; the personalized setting data includes the personalized customization parameters of the user's virtual image, such as unique facial expressions, personalized voice packages, exclusive action special effects and other data. Multi-source heterogeneous data is connected to the official interface of the game platform, and data request instructions are sent according to the interface specifications to obtain the user's virtual image related data stored on the game platform, including static appearance data and dynamic action posture data, etc.; or with the user's authorization, the user's terminal device (such as mobile phone, computer)'s camera, sensor and other hardware devices are used to collect the real-time action data of the user's virtual image during the game; it is also possible to obtain virtual image data published or shared by users from third-party platforms related to the game (such as game communities, virtual image design platforms), such as virtual image design works uploaded by users, virtual image screenshots displayed in the community, etc., to enrich the source of virtual image data.

[0038] The steps for constructing the virtual image feature vector space matrix are as follows: feature analysis is performed on the multi-source heterogeneous data obtained from the user's virtual image. For visual appearance data, facial feature proportion features, hairstyle encoding features, clothing style features, etc. are extracted; for motion posture data, dynamic features such as motion speed, angle, and duration are extracted; from personalized setting data, features such as expression type and voice package keywords are extracted; these extracted features are aggregated to form the original feature set of the virtual image; non-numerical features in the original feature set are encoded and converted using methods such as one-hot encoding and ordinal encoding, and mapped into numerical features. For example, hairstyle styles are encoded as different numerical vectors. For numerical features, normalization or standardization is performed to unify their numerical ranges to obtain the encoded feature vector set; based on the encoded feature vector set, the dimension and elements of the matrix are determined; with the number of feature vectors as the number of columns of the matrix and the dimension of each feature vector as the number of rows of the matrix, each feature vector is arranged in sequence as a column vector of the matrix to construct the virtual image feature vector space matrix; this matrix can comprehensively and systematically represent the various feature information of the user's virtual image, providing a basic data structure for subsequent numerical representation and analysis.

[0039] The steps of constructing the skill knowledge graph network are as follows: collect the structured skill data of users on multiple game platforms from the skill database of the game platform, including skill name, skill level, skill description, skill effect parameters (such as damage value, treatment amount, duration), skill prerequisites, skill-related skills and other information; parse the collected skill data, extract key attributes and relationship information, and convert the unstructured skill description text into a structured data format; use the skill name, skill level, skill effect parameters, etc. as nodes of the knowledge graph, and each node represents a specific skill-related entity; according to the relationship between skills (such as the relationship between the preceding skill and the subsequent skill), the skill description text is transformed into a structured data format. The edges between nodes are defined by setting skill relationships, skill effect complementary relationships, skill combination usage relationships, etc. The edge type and weight are used to represent the nature and strength of the skill relationship; using a graph database (such as Neo4j) or a graph computing framework (such as GraphX), the defined nodes and edges are stored and organized according to a certain topological structure to build a skill knowledge graph network; during the construction process, the nodes and edges are annotated with attributes to supplement the detailed information and relationship descriptions of the skills; by querying, traversing and analyzing the skill knowledge graph network, the complex relationships between skills can be intuitively displayed, providing a structured knowledge foundation for skill feature quantification and recommendation analysis.

[0040] The steps to achieve multi-dimensional quantification of skill characteristics through level mapping, frequency normalization and attribute weighting are as follows: for skill level data, establish a level mapping function to uniformly map the different skill level systems in different games to a standard numerical range; for example, set the standard numerical range to [0, 100]. If the highest skill level in a game is level 10 and the current skill level is level 5, then the mapped level value calculated by the mapping function "mapping value = (current level / highest level) × 100" is 50; map all skill levels in this way to achieve quantitative representation of skill levels at the same scale; count the number of times users use each skill within a certain period of time as the original data of skill usage frequency; use normalization methods (such as minimum-maximum normalization) to convert the original data of skill usage frequency into the original data of skill usage frequency. The initial data is mapped to the interval [0, 1], and the calculation formula is "normalized frequency = (original frequency - minimum frequency) / (maximum frequency - minimum frequency)"; through normalization processing, the impact of the magnitude difference in the frequency of use of different skills is eliminated, making the frequency data comparable and convenient for subsequent analysis; according to the different attributes of the skills (such as offensive, defensive, auxiliary, etc.) and their importance in the game, the level mapping value and frequency normalization value of each skill are assigned corresponding weight coefficients; for example, for offensive core skills, a higher weight is given, while for auxiliary secondary skills, a lower weight is given; through weighted summation, the comprehensive quantitative value of each skill is calculated. The comprehensive quantitative value reflects the characteristics of the skill in multiple dimensions such as level and frequency of use, realizing multi-dimensional quantification of skill characteristics.

[0041] The steps for hierarchical sentiment annotation based on voice chat records are as follows: using speech recognition technology (such as an end-to-end speech recognition model based on deep learning) to convert the acquired multi-platform user voice chat records in real time or offline, converting the voice signal into text content; during the conversion process, performing pre-processing such as noise reduction and enhancement on the speech to improve the accuracy of speech recognition; segmenting the converted text content into sentences based on punctuation or semantics to obtain individual sentences; using sentiment analysis algorithms (such as the BERT-based sentiment classification model) to analyze the sentiment tendency of each sentence, determine the emotion category expressed in the sentence (such as positive, negative, neutral, anger, joy, sadness, etc.), and annotate each sentence with the corresponding sentiment label; based on the sentence-level sentiment labels, perform higher-level sentiment annotation; group sentences according to conversation topic or time period, and divide related sentences into a conversation unit; perform statistical analysis on the sentence sentiment labels within each conversation unit, and determine the sentiment label of the conversation unit using a majority voting method or a weighted average method (assigning different weights based on the importance of sentences); and aggregate and analyze the sentiment labels of multiple conversation units to determine the sentiment annotation results of the entire voice chat record at different levels, forming a hierarchical sentiment annotation system.

[0042] The steps for identifying a comprehensive sentiment index based on label frequency and polarity weight are as follows: for voice chat records with hierarchical sentiment annotation, the frequency of each sentiment label (such as positive, negative, neutral, etc.) at different levels (sentences, dialogue units, and the whole) is counted to form a sentiment label frequency distribution table. For example, in the entire voice chat record, the positive label appears 15 times, the negative label appears 8 times, and the neutral label appears 20 times. Different sentiment labels are assigned corresponding polarity weights to represent the positive and negative tendencies and intensity of the emotions. For example, the weight of the positive label is set to +1, the weight of the negative label is set to -1, and the weight of the neutral label is set to 0. For more detailed sentiment labels (such as anger and joy), the weights are further refined according to their emotional intensity. For example, the weight of the anger label is -0.8, and the weight of the joy label is +0.7. Based on the sentiment label frequency and polarity weight, the comprehensive sentiment index of the voice chat record is calculated. This comprehensive sentiment index can quantitatively reflect the overall emotional tendency and intensity of the user during the voice chat process, providing a quantitative basis for the emotional dimension for subsequent recommendation analysis.

[0043] The steps for keyword screening by combining the TF-IDF algorithm with the domain dictionary are as follows: clean the user's multi-platform text chat records and remove irrelevant content such as special characters, punctuation marks, HTML tags, etc.; convert the text into a unified lowercase format for easy subsequent processing; use word segmentation tools (such as Jieba word segmentation, NLTK word segmentation) to segment the text and split the sentences into individual words; collect professional terms, common words, game-specific words, etc. related to social games to build a domain dictionary; classify and annotate the words in the dictionary, such as game character names, game prop names, game play terms, etc.; continuously update and improve the domain dictionary to ensure that it covers the latest game-related words; for each word after word segmentation, calculate its The word frequency in the previous text and the inverse document frequency of the word are multiplied together to obtain the TF-IDF value of the word. The TF-IDF value reflects the importance of the word in the current text and the distinctiveness of the word in the entire document collection; based on the calculated TF-IDF value, keywords are screened in combination with the domain dictionary; a TF-IDF threshold is set, and words with TF-IDF values ​​higher than the threshold are used as candidate keywords; from the candidate keywords, words that exist in the domain dictionary are screened as the final text chat keywords; for words with similar TF-IDF values, words that are classified as more important or more representative in the domain dictionary are given priority to ensure that the screened keywords can accurately reflect the core information related to the text chat content and the game.

[0044] The steps to achieve semantic quantification by dynamically adjusting the weight coefficient according to the scene type are as follows: define a variety of social game scene types, such as game battle scenes, social interaction scenes, task completion scenes, game store scenes, etc.; establish a scene type recognition model, and judge the scene type to which the text chat record belongs by analyzing the keywords, context, game event triggering information, etc. in the text chat record. For example, when keywords such as "team battle" and "kill" appear in the text, it is identified as a game battle scene; pre-set the initial weight coefficient for the text chat keyword under each scene type, and assign different weight values ​​according to the importance and semantic contribution of the keyword in different scenes. For example, in the game battle scene, the initial weight of the keyword "skill release" is set to 0.8, while in the social interaction scene, the keyword weight is set to 0.3; monitor the user's behavior data and text chat in real time Content changes and scene conversion information; when a change in scene type is detected, the weight coefficient of the keyword is dynamically adjusted according to the characteristics and needs of the new scene. For example, when switching from a game battle scene to a social interaction scene, the weight of social-related keywords (such as "chat" and "friends") is increased, and the weight of battle-related keywords is reduced. At the same time, according to factors such as the frequency of occurrence of keywords in the current scene and the degree of contextual relevance, the weight coefficient is further fine-tuned so that the weight coefficient can more accurately reflect the semantic importance of the keyword in a specific scene; according to the adjusted weight coefficient, the text chat keywords are semantically quantified; for each keyword, its TF-IDF value is multiplied by the corresponding weight coefficient to obtain the semantic quantification value of the keyword; the semantic quantification values ​​of all keywords are summarized, and the entire text chat record is identified by weighted summation or other appropriate aggregation methods.

[0045] Through in-depth mining and analysis of multi-dimensional and multi-source data (virtual image features, skill features, emotional tags and text chat keywords), the social behavior patterns and personality preferences of users with low social pressure sensitivity can be comprehensively and meticulously portrayed. Compared with the analysis method of a single dimension or a small amount of data, the accuracy of user portraits is improved, laying a solid foundation for subsequent personalized recommendations; the decision forest model is used to classify and predict user social types. The decision forest model integrates the advantages of multiple decision trees, has strong generalization ability and robustness, and can effectively handle noise and complex relationships in the data, thereby improving the accuracy and stability of classification. Compared with a single decision tree or other simple classification models, the decision forest model has a strong generalization ability and robustness, and can effectively handle noise and complex relationships in the data, thereby improving the accuracy and stability of classification. type, which can more reliably output the user's social type; based on accurate identification of user social types, the recommendation system can push social games that are highly matched with the characteristics and needs of users of different social types, thereby improving users' acceptance and satisfaction with recommended games, enhancing user stickiness, and also improving the recommendation efficiency and resource utilization efficiency of the game platform; defining multiple social types covers the different behavioral tendencies that users may exhibit in social games, which can meet the diverse social needs of users, so that the recommended social games not only meet user preferences in terms of entertainment, but also match the user's behavioral habits in terms of social interaction mode, and promote users to establish good social relationships within the game.

[0046] Specifically, the specific steps of S3 are: collecting the core recommendation parameter praise data corresponding to each social type of multi-user games, and the core recommendation parameter praise data is specifically the total number of praises from multiple users for a core recommendation parameter; by comparing the total number of praises of each core recommendation parameter of users of each social type, using the priority diagram method to assign a proportional coefficient to each core recommendation parameter, and determine the proportional coefficient of each core recommendation parameter of users of each social type; identifying the proportional coefficient of each core recommendation parameter of the social type to which users with low social pressure sensitivity belong, and at the same time combining the core recommendation parameters of the game to be recommended, respectively generate the initial recommendation value of each game to be recommended.

[0047] The core recommendation parameters specifically include social interaction complexity parameters, social role positioning parameters, game atmosphere style parameters and player matching mechanism parameters.

[0048] In this implementation plan, the method for obtaining positive reviews of core recommendation parameters is as follows: a special review portal is set up in the game platform to encourage users to review the core recommendation parameters of the game. For example, a review button is set up on the game details page, community forum or in the game, and users can score (such as 1-5 points) or provide text reviews on parameters such as the complexity of social interaction and social role positioning of the game. The platform background collects and organizes these review data in real time; questionnaires on the core recommendation parameters of the game can also be sent to users regularly or irregularly. The content of the questionnaire may include questions such as user satisfaction evaluation and importance ranking of the core recommendation parameters of different social types of games. Users are invited to participate in the survey through emails, in-game message push, etc., and the collected questionnaire data is analyzed. Statistical analysis can be performed to obtain positive reviews. Natural language processing and sentiment analysis techniques can also be used to monitor and analyze user discussion content in game communities (such as official forums, social media groups, and player communities), extract topics and comments related to core recommendation parameters, and determine users' emotional tendencies (positive, negative, and neutral) toward each parameter. The number of positive reviews can be counted as positive reviews. Alternatively, in-game behavioral data can be analyzed, such as game duration, frequency of participation in social activities, and number of times game characters are changed, to indirectly infer users' preferences for the core recommended game parameters. For example, if a user frequently participates in high-difficulty social interaction activities and plays for a long time, it can be considered that they have a high evaluation of the social interaction complexity parameter, and positive reviews can be counted accordingly.

[0049] The social interaction complexity parameter indicates the richness, complexity, and depth of in-game social interactions. This includes factors such as the diversity of social gameplay (e.g., the number and types of team dungeons, guild battles, and player trading), the complexity of interaction rules (e.g., the process of task collaboration and the rules for managing social relationships), and the real-time and dynamic nature of interactions (e.g., the fluency of multiplayer real-time battles and instant chat interactions). Information on game-designed social gameplay is obtained from game development documents and official introductions. In-game monitoring tools are used to record the types, frequency, and duration of social activities players actually participate in. Game logs are also analyzed for social interaction-related event data, such as the number of team requests and the number of participants in social activities. A hierarchical scoring method is used to categorize social interaction complexity into multiple levels (e.g., 1-5), with level 1 indicating simple social interaction, including only basic chat functionality, and level 5 indicating very complex social interaction, with multiple in-depth social gameplay and complex rules. Each game can also be assigned a corresponding level score based on the actual social gameplay and interactions. Alternatively, this score can be quantified by calculating the weighted sum of indicators such as the number of social gameplay and the frequency of interaction events, with weights determined based on the importance and complexity of the gameplay.

[0050] Social role positioning parameters refer to the different social roles set for players in the game and their functions and responsibilities, clarifying the player's position and role in social interactions. For example, in team games, there are roles such as commander, attacker, defender, and healer. Each role has specific tasks in the team and has an impact on team collaboration and social relationships. By consulting the game's role setting documents, understanding the game's preset role types and functional descriptions, observing the actual behavior and task allocation of players in the game, analyzing the roles played by players, collecting players' feedback and evaluation of the roles, and understanding the rationality and popularity of the role positioning; constructing a role positioning evaluation index system, evaluating from dimensions such as role functional completeness (such as whether the role has rich skills and whether the responsibilities are clear), role balance (whether the importance and capabilities of different roles in the team are balanced), and role substitutability (whether it is easy to be replaced by other roles). Each dimension uses a 0-10 point system, and the scores of each dimension are combined to obtain the quantitative value of the social role positioning parameters. It can also be weighted calculation based on data such as the role's usage rate in the game and the player's favorable rating of the role.

[0051] The game atmosphere style parameters reflect the overall emotional atmosphere and style characteristics created by the game, including the game's visual style (such as realism, cartoon, magic, ancient style, etc.), music and sound effect style (such as passionate, soothing, tense, cheerful), plot atmosphere (such as relaxed and humorous, suspenseful and tense, touching and inspiring), and social atmosphere (such as friendly and harmonious, fiercely competitive, mutual assistance and cooperation). These aspects have an impact on the player's emotional experience and social willingness. The game's visual style is analyzed through game screenshots and video materials; the game's music and sound effect files are collected, and their style is judged by professionals or through user reviews. The game's plot text can also be read, plot animations can be watched, the plot atmosphere can be evaluated, or the interactive behavior and community atmosphere of players in the game can be observed to understand the social atmosphere; a combination of classification coding and scoring is used. First, the game atmosphere style is classified and coded, such as realistic style is coded as 1, cartoon style is coded as 2, etc.; for each style type, a score of 0-10 is given from multiple sub-dimensions (such as visual beauty, music adaptability, plot appeal, and social activity). The weighted sum of the coding value and the score of each sub-dimension is used as the quantitative value of the game atmosphere style parameter, and the weight is determined according to the importance of each sub-dimension.

[0052] Player matching mechanism parameters refer to the rules and algorithms used by the game to match players for social interaction or competition. These include matching criteria (such as level, skill level, game time, interests, etc.), matching speed (the time from initiating a match request to a successful match), and matching fairness (whether the matched opponents or teammates are of equal strength). These parameters directly impact the player's gaming experience and social outcomes. Obtain technical documentation on the matching mechanism from the game development team to understand the principles and rules of the matching algorithm. Conduct multiple matching experiments in the game, record matching times and information about matched opponents or teammates, collect player feedback on matching results, and analyze matching fairness and satisfaction. Establish matching mechanism evaluation indicators, evaluating the rationality of the matching criteria (such as whether the matching criteria can effectively ensure a good gaming experience), matching speed (quantified using average matching time), and matching fairness (calculating a fairness index by comparing the strength data of the two matching parties, such as level difference and win rate difference). Each dimension uses a standardized score (e.g., 0-1). The scores from each dimension are combined to derive the quantitative value of the player matching mechanism parameters.

[0053] Assume that the core recommendation parameter set is , corresponding to the social interaction complexity parameters, social role positioning parameters, game atmosphere style parameters and player matching mechanism parameters, respectively, to construct The priority diagram matrix Representation parameters With parameters When comparing, relatively The importance score of Compare More importantly, ,otherwise ,and , calculate the sum of the scores of each parameter , then the parameter The proportionality factor .

[0054] Calculation of initial recommendation value: Assume the game to be recommended The core recommended parameter vector is , which correspond to the parameter values ​​of social interaction complexity, social role positioning, game atmosphere style, and player matching mechanism respectively. The core recommended parameter ratio coefficient vector of the social type of users with low social pressure sensitivity is , then the game The initial recommended value of .

[0055] By collecting positive reviews for core game recommendation parameters across different social types, we can understand the preferences of users of different social types for core game elements. For example, leadership-oriented users may place greater emphasis on social interaction complexity and social role positioning. Determining the proportionality coefficients based on these positive reviews allows the recommendation system to accurately match user needs, improving recommendation relevance and user satisfaction. Using the priority graph method to determine the proportionality coefficients for core recommendation parameters systematically quantifies the importance of each parameter to users of different social types. This helps the recommendation system prioritize parameters that are more important to users when recommending games, reducing interference from irrelevant parameters, thereby improving recommendation quality and efficiency and making recommendation results more targeted. Providing game recommendations that match the social type of users with low social pressure sensitivity can meet their personalized needs in social games. When users frequently receive game recommendations that match their preferences, they increase their trust and reliance on the recommendation system and game platform, thereby strengthening user stickiness and promoting the long-term development of the game platform. Based on the preferences of users of different social types for core recommendation parameters, game platforms can optimize game resource allocation in a targeted manner. To satisfy support-oriented users, we focus on optimizing game atmosphere and player matching mechanisms to achieve efficient resource utilization.

[0056] Specifically, the specific steps of S5 are: during the game process of users with low social pressure sensitivity, real-time acquisition of in-game data, the in-game data including the participation rate of activities released in the game and the frequency of player interaction in the community; using emotion recognition technology to obtain user emotion data in real time, the user emotion data including user pleasure, user concentration and user stress level; using in-game data and user emotion data to build an adjustment optimization model, and then output the adjustment optimization value of each game; obtaining the fusion coefficient, combining the initial recommendation value, the adjustment optimization value and the fusion coefficient to generate a comprehensive recommendation value, and then achieve re-ordering of game recommendations.

[0057] In this implementation plan, the participation heat of activities released in the game indicates the degree of attention and participation of players in the activities released in the game, reflecting the popularity and influence of the activities among the player community. It is one of the important indicators for measuring the activity and attractiveness of the game. The game server background records data such as the number of participants, number of participations, completion progress, etc. of each activity, and can also monitor the discussion heat related to the activity, such as the number of posts, comments, and likes about the activity in the game community, or count the changes in the online time and login frequency of players during the activity, which indirectly reflects the participation heat of the activity; a comprehensive scoring method is used to set a basic participation heat score (such as based on the number of participants as the main basis, 1 point for every 100 participants), and then weighted points are added according to factors such as the completion progress of the activity and the heat of community discussion.

[0058] The frequency of player interaction in the community refers to the frequency of communication and interaction between players in the game community (such as official forums, in-game chat channels, player groups, etc.), reflecting the activity of the game community and the social atmosphere among players. The game community's logging system is used to count data such as the number of messages sent by players within a certain period of time, the number of discussions initiated, the number of replies to other people's messages, and the frequency of participation in group activities; the number and frequency of private messages between players in the community are monitored; and the various interactive behaviors of players within a unit of time (such as 1 hour, 1 day) are weighted and summed. For example, sending a normal message counts as 1 point, initiating a discussion counts as 3 points, and replying to other people's messages counts as 2 points. By setting different weights, the quantitative value of the player interaction frequency is calculated.

[0059] User pleasure is used to measure the degree of positive emotions such as pleasure and happiness felt by users during the game. It is a key indicator for evaluating user satisfaction with the gaming experience. Wearable devices (such as smart bracelets and smart glasses) are used to collect users' physiological data (such as heart rate, skin conduction response, and facial expressions), and the user's pleasure is inferred through physiological signal analysis algorithms. A real-time feedback button can also be set in the game client to allow users to actively rate their own pleasure, or analyze the user's voice tone and chat content in the game, and judge their pleasure through sentiment analysis technology; the pleasure is divided into multiple levels (such as 0-10 points), where 0 points means completely unpleasant and 10 points means extremely pleasant. According to the analysis results of physiological data, the user's active scoring or the results of sentiment analysis, the corresponding level score is corresponded to achieve quantification.

[0060] User concentration reflects the degree of user focus during the game, and reflects the user's level of involvement and interest in the game content. By monitoring the user's device operation behavior (such as mouse click frequency, keyboard input frequency, and screen switching times), and analyzing the user's operation rhythm and continuity, concentration can be judged. Eye tracking technology (such as mobile phone cameras or external eye trackers) can also be used to record the user's gaze focus and gaze duration to evaluate the user's attention to the game screen, or analyze the user's stay time and task completion efficiency in the game to indirectly reflect concentration. A concentration evaluation model is constructed to weight indicators such as operation behavior, eye movement data, and stay time. For example, the weight of mouse click frequency and keyboard input frequency is 0.3, the weight of eye movement gaze duration is 0.4, and the weight of stay time is 0.3. Based on the quantitative value and weight of each indicator, a quantitative score of user concentration is calculated.

[0061] The user's stress level indicates the amount of psychological pressure the user endures during the game. It is caused by factors such as game difficulty, competitive pressure, and task urgency, and affects the user's gaming experience and willingness to continue participating. Wearable devices can be used to collect the user's physiological indicators (such as heart rate variability and respiratory rate) and assess the user's stress level through a physiological stress detection algorithm. The user's voice emotions in the game (such as rapid tone and anxious words) and negative emotional expressions in chat content can also be analyzed. Alternatively, a stress self-assessment questionnaire can be set up in the game to regularly invite users to subjectively evaluate their own stress levels. The stress level can be divided into different levels (such as 0-10 points), with 0 points indicating no stress and 10 points indicating extreme stress. The corresponding quantitative score can also be determined based on the results of physiological indicator analysis, voice emotion analysis, and user self-assessment results.

[0062] The steps to build an adjustment optimization model are as follows: clean the in-game data (activity participation enthusiasm, player interaction frequency) and user emotional data (pleasure, concentration, stress level) obtained in real time, remove outliers and missing values, and use normalization or standardization methods to map data of different types and magnitudes to the same numerical range to ensure data consistency and comparability, forming a preprocessed data set; extract key features from the preprocessed data set, such as activity participation enthusiasm data, activity type, participant growth rate and other features; for user emotional data, extract emotional change trends, emotional fluctuation amplitude and other features, and use feature selection algorithms (such as chi-square test, mutual information method) to screen out features that have a significant impact on game recommendation adjustments and construct feature vectors; select machine learning models (such as neural networks, random forests, support vector regression) as adjustment Optimize the basic architecture of the model, use the feature vector and the corresponding game recommendation adjustment target (such as adjusting the recommendation order, increasing or decreasing the recommendation weight) as training data, use the training data to train the model, and adjust the model parameters through optimization algorithms (such as stochastic gradient descent, Adam optimization algorithm) to enable the model to learn the relationship between in-game data, user emotional data and game recommendation adjustments; use the reserved validation data set to evaluate the trained adjustment optimization model, and use evaluation indicators such as mean square error (MSE) and mean absolute error (MAE) to measure the prediction accuracy of the model. According to the evaluation results, adjust the model's hyperparameters (such as the number of layers and nodes of the neural network, the number of trees in the random forest) or reselect features, repeat the training and evaluation process until the model reaches satisfactory performance indicators, and finally build the adjustment optimization model. The expression of the adjustment optimization model is: Let the in-game data vector be , which correspond to activity participation enthusiasm and community player interaction frequency respectively; the user emotion data vector is , corresponding to user pleasure, user concentration, and user stress level, respectively, the adjusted optimization model is: ,in is the weight coefficient, which is set through training or experience. The comprehensive recommendation value is calculated as: Comprehensive recommended value ,in is the fusion coefficient, which is used to balance the importance of the initial recommended value and the adjusted optimized value. .

[0063] The fusion coefficient is obtained in the following way: based on the experience of the game industry and the historical operation data of the recommendation system, the domain experts or developers manually set an initial fusion coefficient as a reference value for the initial operation of the recommendation system. For example, in the early stage of system development, the fusion coefficient is set to 0.6, which means that the initial recommendation value accounts for 60% of the comprehensive recommendation value and the adjusted optimization value accounts for 40%. A real-time monitoring mechanism can also be established to track user feedback data on recommended games (such as game downloads, game time, user reviews) and the performance indicators of the recommendation system (such as recommendation accuracy and recall rate). According to the changes in feedback and performance indicators, an adaptive The fusion coefficient can be dynamically adjusted by an algorithm (such as reinforcement learning algorithm, genetic algorithm). For example, when the recommendation accuracy is found to be declining, the weight of the adjusted optimization value can be appropriately increased, and the weight of the initial recommendation value can be reduced. The optimal fusion coefficient can be found through continuous trial and error and learning. Custom options for the fusion coefficient can also be provided in the user setting interface of the recommendation system, allowing users to manually adjust the proportion of the initial recommendation value and the adjusted optimization value in the comprehensive recommendation value according to their preferences and needs. For example, the user can set the fusion coefficient to 0.8, which is more inclined to the initial recommendation value, or to 0.2, which pays more attention to the adjusted optimization value based on real-time data.

[0064] By acquiring in-game data and user emotional data in real time, it is possible to promptly capture behavioral changes and emotional feedback from users during the game, build an adjustment and optimization model based on this dynamic information, revise the initial recommendation value, and dynamically update game recommendations, so that the recommendation results are more in line with the user's current needs and experience, and improve the accuracy of the recommendation; pay attention to the user's emotional state (pleasure, concentration, stress level). When the user has negative emotions in the game or loses interest in the current game, the system can adjust the recommendation in time according to the adjustment optimization value, provide the user with a more suitable game, avoid user loss due to bad experience, and effectively improve the user's overall experience during the game; based on data such as the popularity of in-game activity participation and the frequency of community player interaction, the recommendation system can understand the real-time popularity and activity of the game, and tilt more resources towards popular and user-friendly games, thereby improving the utilization efficiency of game resources. At the same time, it also helps game developers and platforms understand user needs and optimize game operation strategies.

[0065] See also Figure 2A social game recommendation system based on artificial intelligence (AI) is provided, which applies the above-mentioned social game recommendation method based on AI, and includes: a user type identification module for identifying users with low social pressure sensitivity based on social avoidance indicators in user historical behavior data; a social type identification module for obtaining avatar characteristics and skill characteristics of users with low social pressure sensitivity across multiple game platforms, identifying emotional tags and text chat keywords in multiple game platforms of users with low social pressure sensitivity, and then outputting the social type of users with low social pressure sensitivity; an initial recommendation analysis module for extracting core recommendation parameters of the games to be recommended, and then combining the social types of users with low social pressure sensitivity to obtain initial recommendation values ​​for each game to be recommended; a real-time monitoring module for recommending and ranking games based on the initial recommendation values, while simultaneously monitoring in-game data and user emotional data of users with low social pressure sensitivity during gaming in real time; and an optimization and adjustment module for outputting adjusted optimization values ​​for each game based on in-game data and emotional data, generating a comprehensive recommendation value for each game based on the initial recommendation value and the adjusted optimization value, re-ranking the games based on the comprehensive recommendation value, and continuously updating recommended social games for users.

[0066] In summary, this application has at least the following effects: accurately identifying users with low social pressure sensitivity through social avoidance indicators in user historical behaviors, avoiding misjudgment of highly sensitive users, and reducing invalid recommendations; combining virtual image features, skill features, emotional tags and text chat keywords to more comprehensively characterize user social preferences and output more accurate social types; extracting the core recommendation parameters of the game to be recommended, and calculating the initial recommendation value based on the user's social type to ensure that the recommended game is highly matched with the user's preferences; adjusting the optimization value in real time through in-game data and user emotional data, and dynamically correcting the recommendation value to avoid user loss due to changes in gaming experience; accurately identifying users with low social pressure sensitivity and recommending single-player tasks or light social games to avoid user anxiety and frustration due to forced social interaction, thereby improving the gaming experience; combining multimodal data such as virtual images, skills, emotions, and text, and through real-time monitoring and dynamic adjustment, forming a technical barrier, which is different from the traditional recommendation system based on single behavioral data, and improving user satisfaction and retention rate.

[0067] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods or systems. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0068] The present invention is described with reference to the flowcharts and structure diagrams of the methods and systems according to the embodiments of the present invention. It should be understood that each process and combination of modules in the flowcharts and structure diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate the instructions for implementing the processes in the flowcharts. Figure 1 process or processes and structures Figure 1 A device that specifies functionality within a module or modules.

[0069] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 process or processes and structures Figure 1 Functionality specified in a module or modules.

[0070] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 process or processes and structures Figure 1 Steps for specifying functionality in a module or multiple modules.

[0071] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0072] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A social game recommendation method based on artificial intelligence, characterized in that: The following steps are involved: S1, identifies users with low social stress sensitivity based on the social avoidance index in the user's historical behavior data; S2, obtaining the avatar characteristics and skill characteristics of the low social pressure sensitivity user across multiple game platforms, identifying the emotional tags and text chat keywords in the low social pressure sensitivity user across multiple game platforms, and then outputting the social type of the low social pressure sensitivity user; S3, extracting the core recommendation parameters of the games to be recommended, and then combining them with the social types of users with low social pressure sensitivity to obtain the initial recommendation values ​​of each game to be recommended; S3 specifically comprises the following steps: collecting positive reviews of core recommendation parameters for multi-user games of various social types, wherein the positive reviews of core recommendation parameters are specifically the total number of positive reviews by multiple users for a core recommendation parameter; By comparing the total number of positive reviews for each core recommendation parameter of users of different social types, the proportional coefficient of each core recommendation parameter is assigned using the priority diagram method to determine the proportional coefficient of each core recommendation parameter of users of different social types; Identify the proportional coefficients of the core recommendation parameters of the social type of users with low social pressure sensitivity, and combine them with the core recommendation parameters of the games to be recommended to generate initial recommendation values ​​for each game to be recommended; The core recommendation parameters specifically include social interaction complexity parameters, social role positioning parameters, game atmosphere style parameters and player matching mechanism parameters; S4, recommends and ranks games based on the size of the initial recommendation value, and monitors the in-game data and user emotion data of users with low social pressure sensitivity during the game in real time; S5, based on the in-game data and emotional data output, obtains the adjusted optimization value of each game, combines the initial recommendation value and the adjusted optimization value to generate a comprehensive recommendation value for each game, re-ranks the games according to the comprehensive recommendation value, and continuously updates the recommended social games for users.

2. The method for recommending social games based on artificial intelligence according to claim 1, characterized in that: S1 specifically comprises the following steps: extracting social avoidance indicators, which include the rate of increase in the number of friends, the frequency of social interactions, the percentage of silent chats, and the number of team-up rejections; Taking the social avoidance index as input, the support vector machine is used for training to build a user social stress sensitivity model, and the output is the user's social stress sensitivity value. Compare the social pressure sensitivity value with the sensitivity threshold. When the social pressure sensitivity value is greater than or equal to the sensitivity threshold, mark the user as a high social pressure sensitivity user and enter the single-player task-oriented game recommendation branch, recommending single-player task-oriented games. When the social pressure sensitivity value is greater than or equal to the sensitivity threshold, the user is marked as a low social pressure sensitivity user, triggering the recommendation demand for social multiplayer task games.

3. The method for recommending social games based on artificial intelligence according to claim 1, wherein: The specific steps of outputting the social type of users with low social stress sensitivity are: obtaining virtual image characteristics, skill characteristics, emotion labels, and text chat keywords of known social types as training samples, and then using the training samples to train a decision tree algorithm to construct a decision forest model composed of multiple decision trees; The virtual image characteristics, skill characteristics, emotional tags and text chat keywords of users with low social pressure sensitivity are used as inputs of the decision forest model, and the social types of users with low social pressure sensitivity are obtained as output. The social types include leadership type, assistance type, observation type and follower type.

4. The method for recommending social games based on artificial intelligence according to claim 1, wherein: The virtual image feature acquisition and analysis comprises: acquiring multi-source heterogeneous data of the user's virtual image, constructing a virtual image feature vector space matrix, and numerically representing the virtual image using feature coding and attribute weighting; The skill feature acquisition and analysis includes: acquiring skill structured data, extracting skill text features, building a skill knowledge graph network, and achieving multi-dimensional quantification of skill features through level mapping, frequency normalization, and attribute weighting; The emotional tag acquisition and analysis includes: obtaining user multi-platform voice chat records based on user authorization, performing hierarchical emotional annotation based on the voice chat records, and identifying a comprehensive emotional index based on tag frequency and polarity weight; The text chat keyword acquisition and analysis is as follows: keyword screening is performed using the TF-IDF algorithm in combination with a domain dictionary, and the weight coefficient is dynamically adjusted according to the scenario type to achieve semantic quantification.

5. The method for recommending social games based on artificial intelligence according to claim 1, wherein: S5 specifically comprises the following steps: obtaining in-game data in real time during the game play of the user with low social pressure sensitivity, wherein the in-game data includes participation in activities published in the game and the frequency of player interactions in the community; Using emotion recognition technology to obtain user emotion data in real time, the user emotion data includes user happiness, user concentration and user stress level; Utilize in-game data and user sentiment data to build an optimization model, and then output the optimization value of each game; Obtain the fusion coefficient, combine the initial recommendation value, the adjusted optimization value and the fusion coefficient to generate a comprehensive recommendation value, and then re-rank the game recommendations.

6. An artificial intelligence-based social game recommendation system, applying the artificial intelligence-based social game recommendation method according to any one of claims 1 to 5, characterized in that: include: A user type identification module is used to identify users with low social pressure sensitivity based on the social avoidance index in the user's historical behavior data; A social type recognition module is used to obtain the avatar characteristics and skill characteristics of users with low social pressure sensitivity across multiple game platforms, identify emotional tags and text chat keywords in these users' multiple game platforms, and output the social type of these users with low social pressure sensitivity; The initial recommendation analysis module is used to extract the core recommendation parameters of the games to be recommended, and then combine the social types of users with low social pressure sensitivity to obtain the initial recommendation value of each game to be recommended; A real-time monitoring module is used to recommend and rank games based on the size of the initial recommendation value, while also monitoring in-game data and user emotion data of users with low social stress sensitivity during gaming in real time; The optimization and adjustment module is used to obtain the adjusted optimization value of each game based on the in-game data and emotional data output, combine the initial recommendation value and the adjusted optimization value to generate a comprehensive recommendation value for each game, re-rank the games according to the comprehensive recommendation value, and continuously update the recommended social games for users.

Images