A game strategy recommendation method and a recommendation effect generation method and system thereof

Abstract

The application relates to the technical field of data recommendation, and more particularly to a game strategy recommendation method and a recommendation effect generation method and system thereof. The generation method comprises the following steps: obtaining first technical index data of a target player in a game before the target player clicks a recommended strategy; constructing a personalized historical performance benchmark of the target player according to the first technical index data; obtaining second technical index data of the target player in the game after a first time window after the target player clicks the recommended strategy is determined; obtaining a comparison result of the second technical index data and the personalized historical performance benchmark, and generating effect feedback data of the recommended strategy according to the comparison result. The application can generate more accurate effect feedback data that can reflect individual differences of players, and provide more reliable supervision signals for optimization of a recommendation model of a game strategy.

Description

Technical Field

[0001] This invention relates to the field of data recommendation technology, and more specifically, to a game strategy recommendation method and a method and system for generating recommendation effects. Background Technology

[0002] As a bridge connecting players with a vast amount of strategy content, game strategy recommendation systems directly impact players' gaming experience and skill development. Currently, content platforms both inside and outside the game have accumulated a wealth of game strategy data, including text-based guides, video tutorials, and tactical analyses. How to accurately identify content tailored to a player's current skill gaps from this massive amount of data has become a crucial research direction in the field of game strategy recommendation systems.

[0003] Currently, the main technical solutions for game strategy recommendations include those based on collaborative filtering or keyword matching, and those based on manual tagging or search enhancement. Collaborative filtering or keyword matching solutions recommend based on analyzing player historical behavior or matching query keywords, but they struggle to understand the deep semantics of game content and cannot identify content that players actually need but haven't actively searched for, resulting in cold start issues and insufficient personalization. Manual tagging or search enhancement solutions assist recommendations by manually labeling or retrieving relevant examples, but manual labeling is costly, has limited coverage, and players are unlikely to actively initiate precise search queries. Therefore, most existing recommendation systems optimize based on click-through rate (CTR), but this only reflects players' immediate click behavior and cannot measure the effectiveness of recommended content in improving players' actual gameplay skills. Summary of the Invention

[0004] This invention provides a method and system for recommending game guides, which can generate more accurate feedback data on recommendation effects that better reflects individual player differences, and provide more reliable monitoring signals for optimizing game guide recommendation models.

[0005] According to a first aspect of this application, a method for generating game strategy recommendation effects is provided, the method comprising: Obtain the target player's primary technical metrics in the game before they click on the recommended strategy; Based on the first technical indicator data, a personalized historical performance benchmark for the target player is constructed; After determining the first time window after the target player clicked on the recommended strategy, obtain the target player's second technical indicator data in the game; Obtain the comparison results between the second technical indicator data and the personalized historical performance benchmark, and generate the effect feedback data of the recommended strategy based on the comparison results.

[0006] Understandably, this application constructs the personalized historical performance benchmark by obtaining the target player's first technical indicator data before clicking the recommended guide, and obtains the second technical indicator data after the target player clicks the recommended guide within the first time window. By comparing the data before and after the click, effect feedback data is generated. This ensures that the generation of the effect feedback data is based on the target player's own performance baseline rather than a uniform fixed threshold, thus adapting to the individual differences of different players. At the same time, by setting the first time window, it avoids misjudgments caused by premature evaluation before the player has fully digested the guide content, effectively improving the accuracy and reliability of the effect feedback data. This provides a more accurate and reliable monitoring signal for optimizing the game guide recommendation model.

[0007] Optionally, constructing the personalized historical performance benchmark of the target player based on the first technical indicator data includes: Based on the first technical indicator data, obtain the first technical indicator statistical feature quantity of the target player; Based on the statistical features of the first technical indicator, a personalized historical performance benchmark for the target player is constructed.

[0008] Understandably, this application obtains the first technical indicator statistical feature quantity of the target player based on the first technical indicator data, and constructs the personalized historical performance benchmark based on the statistical feature quantity, so that the personalized interval can reflect the fluctuation range of the target player's normal performance level, rather than relying on a fixed threshold, providing an objective and quantifiable comparison benchmark for obtaining subsequent comparison results, and effectively improving the scientificity and reliability of the effect feedback data.

[0009] Optionally, the statistical features of the first technical indicator are the mean and standard deviation of the first technical indicator; The step of constructing the personalized historical performance benchmark of the target player based on the statistical features of the first technical indicator includes: Based on the preset confidence coefficient and the mean and standard deviation of the first technical indicator, the upper boundary and lower boundary of the interval are constructed. A personalized historical performance range is constructed based on the upper and lower boundaries of the range, serving as the personalized historical performance benchmark for the target player.

[0010] Understandably, this application specifies the statistical features of the first technical indicator as mean and standard deviation, and constructs the upper and lower boundaries of the interval based on a preset confidence coefficient, so that the personalized historical performance interval can scientifically reflect the reasonable range of normal level fluctuations of the target player, providing an objective and quantifiable basis for obtaining subsequent comparison results, and avoiding misjudgments caused by single performance fluctuations.

[0011] The step of obtaining the comparison result between the second technical indicator data and the personalized historical performance benchmark, and generating the effect feedback data of the recommended strategy based on the comparison result, includes: Based on the second technical indicator, obtain the statistical feature quantity of the second technical indicator of the target player; If the statistical feature of the second technical indicator is greater than the personalized historical performance benchmark, then positive feedback data is generated as the effect feedback data of the recommended strategy. If the statistical feature of the second technical indicator is within the personalized historical performance benchmark, then the first type of negative feedback data is generated as the effect feedback data of the recommended strategy; If the statistical feature value of the second technical indicator is less than the personalized historical performance benchmark, then a second type of negative feedback data is generated as the effect feedback data of the recommended strategy.

[0012] Understandably, this application compares the second technical indicator data with the personalized historical performance benchmark, and divides the effect feedback data into three categories based on the comparison results: positive feedback data, first-type negative feedback data, and second-type negative feedback data. This achieves a refined hierarchical evaluation of the recommended strategy recommendation effect. The first-type negative feedback data reflects that the target player learned the strategy but had no effect, while the second-type negative feedback data reflects that the target player learned the strategy but it had a negative effect. The distinction between the two types of negative feedback data effectively improves the richness and effectiveness of the effect feedback data, thereby providing a more accurate supervision signal for subsequent model adjustments based on the effect feedback data.

[0013] Optionally, the first time window is obtained in the following way: The time taken for the target player to complete a preset number of games in the game after clicking on the recommended strategy is obtained, and the time taken is used as the first time window.

[0014] Understandably, this application uses the time taken for the target player to complete the first preset number of games after clicking on the recommended strategy as the first time window. This allows the first time window to be dynamically linked to the target player's actual game participation behavior, ensuring that the target player only enters the evaluation stage after fully practicing the strategy content. This effectively avoids the target player being mistakenly judged as having an invalid strategy due to poor performance before digesting the new knowledge in the recommended strategy, thus improving the accuracy and rationality of the effect feedback data.

[0015] Optionally, determining that the target player clicked on the recommended strategy includes: The recommended strategy guide is pushed to the target player. When the target player's click operation information on the recommended strategy guide is obtained within a preset second time window after the recommended strategy guide is pushed, it is determined that the target player has clicked on the recommended strategy guide.

[0016] Understandably, this application determines whether the target player has clicked on the recommended guide by pushing it to the target player and detecting click operation information within a preset second time window. The subsequent effect feedback data generation process is only triggered after the click is confirmed. This avoids including the target player's non-click behavior on the recommended guide in the effect feedback data, thereby effectively eliminating the interference of noise data unrelated to the quality of the guide content on the effect feedback data, and thus improving the accuracy and reliability of the effect feedback data generation.

[0017] Optionally, after determining that the target player clicked on the recommended strategy in the first time window, obtaining the target player's second technical indicator data in the game includes: Within a preset second number of rounds following the first time window, the technical indicator data of the target player is continuously tracked to obtain multiple technical indicator data samples; The second technical indicator data is determined based on the multiple technical indicator data samples.

[0018] Understandably, this application continuously tracks the technical indicator data of the target player within a preset second number of games after the first time window, obtains multiple technical indicator data samples, and determines the second technical indicator data accordingly. This allows the second technical indicator data to comprehensively reflect the overall performance of the target player in multiple games, avoiding the impact of accidental deviations caused by luck factors in a single game on the effect feedback data, and effectively improving the stability and reliability of the effect feedback data.

[0019] According to a second aspect of this application, a method for recommending game strategies is provided, the method comprising: Obtain the target player's player status information; Based on the player status information, multiple candidate strategies are recalled from the strategy library; Construct a hint message that includes the player's status information and the multiple candidate strategies; The prompt information is input into the recommendation model to obtain the target recommended strategy; wherein the recommendation model is fine-tuned based on the effect feedback data, and the effect feedback data is generated using the game strategy recommendation effect generation method described in the first aspect above.

[0020] Understandably, this application obtains the player status information of the target player, recalls candidate guides from the guide library, and uses the prompt information constructed based on the player status information and candidate guides to input into the recommendation model to obtain the target recommended guide. The recommendation model is adjusted based on the generated effect feedback data, forming a closed-loop recommendation of diagnosis, recall, fine ranking, feedback, and self-optimization. This allows the recommendation model to continuously optimize based on the actual performance feedback of the target player after clicking on the recommended guide, and gradually learns which guide to recommend under what player status will bring the most realistic technical improvement. This significantly improves the accuracy and long-term value of game guide recommendations, overcoming the limitation of traditional recommendation systems that only use click-through rate as the optimization goal.

[0021] Optionally, the recommendation model is adjusted based on performance feedback data, including: An unpaired dataset is constructed based on the effect feedback data. Each data point in the unpaired dataset includes the player status information, the recommended strategy, the candidate strategy, and the effect feedback data. The recommendation model is adjusted using the unpaired dataset.

[0022] Understandably, this application constructs an unpaired dataset containing the player state information, the recommended strategy, the candidate strategy, and the effect feedback data based on the effect feedback data, and uses the unpaired dataset to adjust the recommendation model. This allows the effect feedback data to be used for model adjustment and training in a structured form, without the need to construct paired positive and negative sample pairs for each feedback sample, thus reducing the cost of constructing sample data. At the same time, the unpaired dataset can make fuller use of the single-point feedback signals generated in the actual recommendation scenario, improving the efficiency of sample data utilization.

[0023] According to a third aspect of this application, a game strategy recommendation effect generation system is provided, the generation system comprising: The first acquisition module is used to acquire the target player's first technical indicator data in the game before clicking on the recommended strategy; A construction module is used to construct a personalized historical performance benchmark for the target player based on the first technical indicator data; The second acquisition module is used to acquire the second technical indicator data of the target player in the game after determining that the target player clicked on the recommended strategy in the first time window; The generation module is used to obtain the comparison results between the second technical indicator data and the personalized historical performance benchmark, and generate the effect feedback data of the recommended strategy based on the comparison results.

[0024] According to a fourth aspect of this application, a game strategy recommendation system is provided, the recommendation system comprising: The third acquisition module is used by users to obtain the player status information of the target player; The recall module is used to recall multiple candidate strategies from the strategy library based on the player status information. The prompt information construction module is used to construct prompt information that includes the player's status information and the multiple candidate strategies; The recommendation module is used to input the prompt information into the recommendation model to obtain the target recommended strategy; wherein, the recommendation model is obtained by fine-tuning based on the effect feedback data, and the effect feedback data is generated by the game strategy recommendation effect generation system described in the third aspect above.

[0025] According to a fifth aspect of this application, an electronic device is provided, comprising: Memory, used to store one or more computer programs; A processor, when the one or more computer programs are executed by the processor, implements the game strategy recommendation effect generation method described in the first aspect or the game strategy recommendation method described in the second aspect.

[0026] According to a sixth aspect of this application, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions, the computer instructions being configured to cause a processor to execute and implement the game strategy recommendation effect generation method described in the first aspect or the game strategy recommendation method described in the second aspect.

[0027] Based on any of the above aspects, the beneficial effects of the game strategy recommendation method and its recommendation effect generation method, system, electronic device, and storage medium provided in the embodiments of this application include: (1) This application constructs the personalized historical performance benchmark by obtaining the first technical indicator data of the target player before clicking the recommended strategy, and obtains the second technical indicator data after clicking the recommended strategy after determining the first time window. By comparing the data before and after, effect feedback data is generated, so that the generation of the effect feedback data is based on the target player's own performance baseline rather than a uniform fixed threshold, which can adapt to the individual differences of different players. At the same time, by setting the first time window, it avoids misjudgment caused by premature evaluation due to the player not yet digesting the strategy content, effectively improving the accuracy and reliability of the effect feedback data, thereby providing a more accurate and reliable supervision signal for the optimization of the game strategy recommendation model.

[0028] (2) This application obtains the first technical indicator statistical feature quantity of the target player based on the first technical indicator data, and constructs the personalized historical performance benchmark based on the statistical feature quantity, so that the personalized interval can reflect the fluctuation range of the normal performance level of the target player, rather than relying on a fixed threshold, providing an objective and quantifiable comparison benchmark for obtaining subsequent comparison results, and effectively improving the scientificity and reliability of the effect feedback data.

[0029] (3) This application specifies the statistical features of the first technical indicator as mean and standard deviation, and constructs the upper and lower boundaries of the personalized historical performance benchmark based on the preset confidence coefficient, so that the personalized historical performance benchmark can scientifically reflect the reasonable range of the normal level fluctuation of the target player, providing an objective and quantifiable basis for obtaining subsequent comparison results, and avoiding misjudgment caused by single performance fluctuation.

[0030] (4) This application uses the time taken for the target player to complete the first preset number of games after clicking on the recommended strategy as the first time window, so that the first time window is dynamically associated with the target player's actual game participation behavior, ensuring that the target player enters the evaluation stage only after fully practicing the strategy content, thereby effectively avoiding the target player being mistakenly judged as invalid due to poor performance when the target player has not yet digested the new knowledge of the recommended strategy, and improving the accuracy and rationality of the effect feedback data.

[0031] (5) This application compares the second technical indicator data with the personalized historical performance benchmark, and divides the effect feedback data into three categories: positive feedback data, first type negative feedback data, and second type negative feedback data according to the comparison results. This achieves a refined hierarchical evaluation of the recommendation effect of the recommended strategy. The first type negative feedback data reflects that the target player learned it but had no effect, while the second type negative feedback data reflects that the target player learned it but had a negative effect. The distinction between the two types of negative feedback data effectively improves the richness and effectiveness of the effect feedback data, thereby providing a more accurate supervision signal for subsequent model adjustment based on the effect feedback data.

[0032] (6) This application determines whether the target player has clicked the recommended strategy by pushing the recommended strategy to the target player and detecting the click operation information within a preset second time window. The subsequent effect feedback data generation process is only triggered after the click is confirmed, which avoids including the target player's non-click behavior on the recommended strategy in the effect feedback data. This effectively eliminates the interference of noise data unrelated to the quality of the strategy content on the effect feedback data, thereby improving the accuracy and reliability of the effect feedback data generation.

[0033] (7) This application continuously tracks the technical indicator data of the target player within a preset second number of games after the first time window, obtains multiple technical indicator data samples, and determines the second technical indicator data accordingly. This enables the second technical indicator data to comprehensively reflect the overall performance of the target player in multiple games, avoids the influence of accidental deviations caused by luck factors in a single game on the effect feedback data, and effectively improves the stability and reliability of the effect feedback data.

[0034] (8) This application obtains the player status information of the target player, recalls candidate strategies from the strategy library, and uses the prompt information constructed based on the player status information and candidate strategies to input the recommendation model to obtain the target recommended strategy. The recommendation model is adjusted based on the generated effect feedback data, forming a closed-loop recommendation of diagnosis, recall, fine ranking, feedback and self-optimization. This enables the recommendation model to be continuously optimized based on the actual performance feedback after the target player clicks on the recommended strategy. It gradually learns which strategy to recommend under what player status can bring the most realistic technical improvement, thereby significantly improving the accuracy and long-term value of game strategy recommendations and overcoming the limitation of traditional recommendation systems that only use click rate as the optimization target.

[0035] (9) This application constructs an unpaired dataset containing the player status information, the recommended strategy, the candidate strategy and the effect feedback data based on the effect feedback data, and uses the unpaired dataset to adjust the recommendation model, so that the effect feedback data can be used for model adjustment and training in a structured form without having to construct paired positive and negative sample pairs for each feedback sample, reducing the construction cost of sample data. At the same time, the unpaired dataset can make fuller use of the single-point feedback signals generated in the actual recommendation scenario, improving the utilization efficiency of sample data. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is an illustrative application scenario diagram of the game strategy recommendation effect generation method or game strategy recommendation method provided in this embodiment.

[0038] Figure 2 This is a flowchart illustrating the method for generating game strategy recommendation effects provided in this embodiment.

[0039] Figure 3 This is a schematic diagram of the sub-steps of step S200 provided in this embodiment.

[0040] Figure 4 This is a schematic diagram of the sub-steps of step S220 provided in this embodiment.

[0041] Figure 5 This is a schematic diagram of a sub-step of step S300 provided in this embodiment.

[0042] Figure 6 This is a schematic diagram of the sub-steps of step S400 provided in this embodiment.

[0043] Figure 7 This is a flowchart of the game strategy recommendation method provided in this embodiment.

[0044] Figure 8 A schematic diagram illustrating the specific steps for adjusting the recommended model provided in this embodiment.

[0045] Figure 9 This is a schematic diagram of the functional modules of the game strategy recommendation effect generation system provided in this embodiment.

[0046] Figure 10 This is a schematic diagram of the functional modules of the game strategy recommendation system provided in this embodiment.

[0047] Figure 11 This is a schematic diagram of the structure of the electronic device provided in this embodiment. Detailed Implementation

[0048] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this application. To better illustrate the following embodiments, some components in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product; it is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0049] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0050] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0051] The inventors' research revealed shortcomings in existing game strategy recommendation systems, particularly in the individual adaptability of feedback signals, accurate assessment of latency effects, and refined differentiation of negative feedback signals. Most existing systems optimize based on click-through rate, reflecting only the player's immediate click behavior and failing to measure the effectiveness of recommended content in improving the player's actual gameplay. The closest existing solution to this application uses a fixed threshold to determine positive or negative feedback, which cannot adapt to individual player differences; its feedback signals rely on the player's immediate behavior, while the actual effectiveness of strategies often requires time and practice to manifest. Immediate feedback cannot fully reflect the true value of strategies, impacting model optimization.

[0052] This embodiment provides a technical solution that can solve the above problems. The specific implementation of this application will be described in detail below with reference to the accompanying drawings.

[0053] Exemplary, the following is a schematic diagram illustrating an application scenario of a game strategy recommendation effect generation method or a game strategy recommendation method provided in an embodiment of this application. For example... Figure 1 As shown, the application scenario includes at least a server 100 and a terminal 200 that can communicate with the server 100.

[0054] Understandably, the server 100 can be an independent electronic device or a cluster of multiple electronic devices; the terminal 200 can be a smartphone terminal, personal computer, tablet computer, vehicle terminal, etc., but is not limited to these.

[0055] In one feasible approach, server 100 and terminal 200 may respectively execute the game strategy recommendation effect generation method provided in the embodiments of this application, or, optionally, the game strategy recommendation effect generation method provided in the embodiments of this application may be partially executed in server 100 and partially executed in terminal 200.

[0056] In one feasible approach, server 100 and terminal 200 may respectively execute the game strategy recommendation method provided in the embodiments of this application, or, optionally, the game strategy recommendation method provided in the embodiments of this application may be partially executed in server 100 and partially executed in terminal 200.

[0057] Example 1 like Figure 2 As shown in the figure, this embodiment provides a method for generating game strategy recommendation effects, which can be further divided into the following steps: S100: Obtain the target player's primary technical indicator data in the game before clicking on the recommended strategy; In practice, the recommended strategy guide is a guide selected from the strategy guide library and pushed to the target player's terminal device by the game strategy guide recommendation system. Before the recommended strategy guide is pushed, the target player is in a normal game state, and no strategy guide content has been pushed to the target player, nor has the target player clicked on the recommended strategy guide.

[0058] For example, the process of pushing the recommended strategy includes: the game strategy recommendation system recalls multiple candidate strategies from the strategy library based on the target player's player status information; selects a target recommended strategy from the multiple candidate strategies; and pushes the recommended strategy to the target player's terminal device, for example, by presenting it in the target player's game interface in the form of a notification message, a pop-up prompt, or a highlighted card in the strategy recommendation list. Before pushing the recommended strategy, the system has obtained the target player's first technical indicator data, which reflects the target player's normal skill performance level before encountering the recommended strategy.

[0059] Specifically, the first technical indicator data refers to the quantitative indicator data related to the skill level of the target player in the game before clicking the recommended strategy. For example, the first technical indicator data may be one or more of the following: "speech effectiveness score", "voting accuracy rate", "skill release timing score", "operation reaction speed", "teamwork score", "goal achievement rate".

[0060] In practice, the first technical indicator data can be obtained from the game server. For example, the technical indicator score after each game can be extracted from the target player's historical game records. This score can be automatically calculated and generated by the game system based on the target player's performance in various aspects of their historical games. Preferably, the first technical indicator data includes the technical indicator data from the target player's most recent games before clicking the recommended strategy.

[0061] S200. Based on the first technical indicator data, construct a personalized historical performance benchmark for the target player; Specifically, the personalized historical performance benchmark is a personalized assessment reference used to reflect the normal fluctuation range or expected level of the target player. It is constructed based on the target player's own first technical indicator data, rather than using a uniform fixed threshold. By constructing the personalized historical performance benchmark, an assessment standard adapted to the individual differences of the target player can be provided for subsequently determining whether the recommended strategy brings actual technical improvement.

[0062] In the specific implementation process, such as Figure 3 As shown, the specific steps of step S200 may include: S210. Based on the first technical indicator data, obtain the first technical indicator statistical feature quantity of the target player; In one specific implementation, the first technical indicator statistical feature is the mean and standard deviation of the first technical indicator. Specifically, statistical calculations are performed on the technical indicator data of the target player in the most recent few games before clicking the recommended strategy to obtain the mean of the first technical indicator data. and standard deviation The mean The standard deviation reflects the average level of the target player's technical indicators under normal conditions. This reflects the normal fluctuation level of the target player's technical indicators.

[0063] In specific implementation, the first technical indicator data may include multiple types of technical indicators, such as simultaneously including both "speech validity score" and "voting accuracy rate". When the first technical indicator data includes multiple types of technical indicators, for each type of technical indicator, its first technical indicator statistical characteristic is calculated, that is, its mean and standard deviation are calculated separately. Specifically, for the j-th type of technical indicator, based on the j-th type of technical indicator data from the target player's most recent games before clicking the recommended strategy, the mean corresponding to this type of technical indicator is calculated. and standard deviation .

[0064] S220. Based on the statistical features of the first technical indicator, construct the personalized historical performance benchmark of the target player.

[0065] Specifically, due to the inherent differences in skill levels among players, the same technical indicator value is not directly comparable between different players. For example, a "speech effectiveness score" of 80 points might be considered exceptional for a novice player, but an underperformance for a professional player. Therefore, this embodiment does not use a uniform fixed threshold as an evaluation standard. Instead, it independently constructs a personalized historical performance benchmark based on the target player's own statistical characteristics of the first technical indicator, enabling subsequent judgments to adapt to the individual skill differences of the target player.

[0066] In the specific implementation process, such as Figure 4 As shown, the specific steps of step S220 may include: S221. Based on the preset confidence coefficient, the mean of the first technical indicator, and the standard deviation of the first technical indicator, construct the upper boundary and the lower boundary of the interval; Specifically, it can be based on the average of the first technical indicator. The standard deviation of the first technical indicator The upper and lower boundaries of the interval are determined using the preset confidence coefficient k as follows: upper boundary lower boundary .

[0067] The confidence coefficient k controls the width of the personalized historical performance benchmark, thereby adjusting the sensitivity of subsequent effect feedback data generation. Specifically, a larger k value results in a wider interval, indicating a higher tolerance for fluctuations in the target player's normal performance level; a significant change in the second technical indicator data is required to be considered a significant improvement or deterioration. Conversely, a smaller k value results in a narrower interval, indicating a lower tolerance for fluctuations in the target player's normal performance level; a small change in the second technical indicator data is sufficient to be considered statistically significant. For example, k=2, where the interval range is approximately... This can cover approximately 95% of the normal fluctuation range. Those skilled in the art should understand that the value of k can be adjusted according to the game type, technical indicator characteristics, or business needs to adapt to the sensitivity requirements of judgment in different scenarios.

[0068] When multiple technical indicators exist, such as "speech validity score" and "voting accuracy rate" simultaneously, the upper and lower bounds of the intervals for each type of technical indicator are constructed independently. In practice, the confidence coefficients for each type of technical indicator can be the same, or different confidence coefficients can be set according to the importance of each type of technical indicator to reflect the different sensitivities of different technical indicators in assessing the improvement of players' skill levels.

[0069] S222. Construct a personalized historical performance interval based on the upper and lower boundaries of the interval, which serves as the personalized historical performance benchmark for the target player.

[0070] Specifically, the continuous interval defined by the upper boundary and the lower boundary of the interval can be used as the personalized historical performance interval, which is the personalized historical performance benchmark of the target player.

[0071] Accordingly, when multiple types of technical indicators exist, the constructed personalized historical performance benchmark can also be constructed separately according to each type of technical indicator. For example, the personalized historical performance interval corresponding to the j-th type of technical indicator is […]. , ].

[0072] S300: After determining that the target player clicked on the recommended strategy in the first time window, obtain the target player's second technical indicator data in the game; In one specific implementation, determining that the target player clicked on the recommended strategy guide specifically includes: The recommended strategy guide is pushed to the target player. When the target player's click operation information on the recommended strategy guide is obtained within a preset second time window after the recommended strategy guide is pushed, it is determined that the target player has clicked on the recommended strategy guide.

[0073] Specifically, "the target player clicking on the recommended guide" means that within a preset second time window after the recommended guide is pushed to them, the target player performs a click operation on the recommended guide content presented on their terminal device, such as the guide title, summary, or preview card. This click operation is an interactive operation generated through touch screen clicking, mouse clicking, or keyboard confirmation. The click operation indicates that the target player actively viewed the detailed content of the recommended guide, that is, the target player has consumed the content of the recommended guide through the click behavior. The click behavior is an interactive signal from the system confirming that the target player has the intention and behavior to view the guide content.

[0074] After the recommended strategy is pushed out, if no click information is obtained from the target player regarding the recommended strategy within the preset second time window, it is determined that the target player has not clicked the recommended strategy. In this case, the subsequent effect feedback data generation step is not executed to avoid including non-click behavior caused by attention factors, such as the target player not paying attention to the push notification or being too busy in a game to check, in the effect feedback data, thereby eliminating noise interference unrelated to the quality of the strategy content.

[0075] In practice, the preset second time window is used to limit the time range within which the target player clicks on the recommended guide. Only clicks made within a reasonable time after the push notification are considered valid. If no click information is obtained within the preset second time window, it is determined that the target player has not clicked on the recommended guide. At this point, subsequent steps can be terminated, and no effect feedback data is generated. This is to avoid including non-click behavior caused by attention factors in the effect feedback data and to eliminate noise interference unrelated to the quality of the guide content.

[0076] In practice, the first time window refers to the time interval between the target player clicking on the recommended strategy and the start of evaluating its effectiveness. The purpose of setting a first time window is that the target player needs some practice after viewing the recommended strategy, such as completing several games, to internalize the new knowledge and improve their skills. Evaluating too early may lead to misjudgment because the target player has not yet digested the new knowledge.

[0077] In one specific implementation, the first time window is obtained in the following manner: The time taken for the target player to complete a preset number of games in the game after clicking on the recommended strategy is obtained, and the time taken is used as the first time window.

[0078] For example, the preset first number of games can be 3. This application records the time span from when the target player clicks on the recommended strategy to when they complete 3 games, and uses this duration as the first time window. Since different players have different gaming habits and paces, this method dynamically links the first time window to the target player's actual gaming behavior, rather than using a fixed duration, thus more accurately reflecting the actual time required for different players to digest and absorb the strategy content.

[0079] In the specific implementation process, such as Figure 5 As shown, the specific steps of step S300 may include: S310. Within a preset second number of rounds after the first time window, continuously track the technical indicator data of the target player to obtain multiple technical indicator data samples. Specifically, the preset second number of games is used to reflect the short-term verification period of the target player's use of the recommended strategy. After the first time window, the technical indicator data of the target player in the subsequent preset second number of games is continuously tracked, and a sample of the technical indicator data for each game is recorded after the game, resulting in multiple technical indicator data samples.

[0080] S320. Determine the second technical indicator data based on the plurality of technical indicator data samples.

[0081] In one specific implementation, the average value of the plurality of technical indicator data samples is calculated, and the average value is used as the second technical indicator data. For example, if the preset number of games is 10, then the technical indicator data samples (s1, s2, ..., s...) for each of the 10 games are obtained. 10 The average value of these values ​​is calculated as the second technical indicator data. By combining the performance of multiple games to determine the second technical indicator data, biases caused by luck factors in a single game can be effectively avoided, making the evaluation results more reliable.

[0082] In the specific implementation process, when there are multiple types of technical indicators, each game corresponds to a technical indicator data sample, and each sample contains the scores of various technical indicators in that game. Specifically, within a preset second number of games, after the end of each game, the scores of various technical indicators in that game are recorded to form a technical indicator data sample. Thus, multiple games result in multiple technical indicator data samples, and each sample is a vector containing the scores of multiple types of technical indicators.

[0083] For example, if the target player's technical metrics include two categories: "speech validity score" and "voting accuracy," then by continuously tracking over 10 rounds, 10 technical metric data samples are obtained, each of which can be represented as ( , ), where t represents the t-th game; t = 1, 2, ..., 10; This represents the "speech validity score" in game t. This represents the "voting accuracy" in game t.

[0084] When determining the second technical indicator data, for each type of technical indicator, its average value within a preset second number of rounds is calculated. Specifically, for the j-th type of technical indicator, its average value is calculated. Where N is the preset number of the second round, Let be the score of the j-th type of technical indicator in the t-th game. The average values ​​of all technical indicators are combined to form the second technical indicator data; that is, the second technical indicator data is a vector containing the average values ​​of multiple types of technical indicators. For example, the second technical indicator data can be represented as ( , ),in The average of the "speech validity score" over the preset second number of rounds. The "voting accuracy rate" is the average value over a preset second number of games. By calculating the average values ​​of various technical indicators, the second technical indicator data can comprehensively reflect the multi-dimensional technical performance of the target player in multiple games.

[0085] It is understood that the second technical indicator data and the first technical indicator data belong to the same type of technical indicator, that is, both are quantitative values ​​of the target player's technical performance in the same game dimension. For example, if the first technical indicator data is "speech validity score", then the second technical indicator data is also "speech validity score"; if the first technical indicator data includes multiple types of technical indicators, then the second technical indicator data also includes the same multiple types of technical indicators accordingly.

[0086] The first technical indicator data is used to construct a personalized historical performance benchmark for the target player, serving as a reference standard reflecting the target player's normal skill level before encountering the recommended strategy. The second technical indicator data is used to compare with the personalized historical performance benchmark to assess whether the recommended strategy has a significant impact on the target player's skill improvement. Since the first and second technical indicator data belong to the same type of technical indicator, they are comparable. By comparing the changes in the second technical indicator data relative to the personalized historical performance benchmark constructed from the first technical indicator data, the effectiveness of the recommended strategy can be objectively determined. The difference between the first and second technical indicator data in the time dimension is that the first technical indicator data is collected before the target player clicks on the recommended strategy, reflecting the target player's historical skill level; the second technical indicator data is collected after the first time window following the target player clicking on the recommended strategy, reflecting the target player's skill level after practicing the recommended strategy.

[0087] S400: Obtain the comparison result between the second technical indicator data and the personalized historical performance benchmark, and generate the effect feedback data of the recommended strategy based on the comparison result.

[0088] It is understood that the effect feedback data is used to characterize the effect of the recommended strategy on the skill improvement of the target player.

[0089] In the specific implementation process, such as Figure 6 As shown, the specific steps of step S400 may include: S410. Based on the second technical indicator, obtain the statistical feature quantity of the second technical indicator of the target player; In one specific implementation, the statistical characteristic of the second technical indicator is the mean of the second technical indicator, that is, the average value of multiple technical indicator data samples calculated in S320 above.

[0090] S420. If the statistical feature of the second technical indicator is greater than the personalized historical performance benchmark, positive feedback data is generated as the effect feedback data of the recommended strategy. Understandably, the positive feedback data indicates that the recommended strategy has a significant positive effect on the target player's skill improvement. Specifically, after watching and practicing the recommended strategy, the target player's statistical features of technical indicators significantly exceed the upper limit of their personalized historical performance benchmark, demonstrating that the strategy effectively helped the player improve their relevant technical abilities. In subsequent model training, the positive feedback data serves as positive samples to strengthen the recommendation model's strategy of recommending the strategy to similar player states.

[0091] S430. If the statistical feature of the second technical indicator is within the personalized historical performance benchmark, then generate the first type of negative feedback data as the effect feedback data of the recommended strategy. Understandably, the first type of negative feedback data indicates that the recommended strategy did not significantly improve the target player's skills. That is, after watching and practicing the recommended strategy, the target player's technical metric statistics still fluctuated within the range of their personalized historical performance baseline—essentially, "it's as if they didn't learn anything." This first type of negative feedback data demonstrates that the recommended strategy was ineffective in addressing the target player's current skill weaknesses. In subsequent model training, this first type of negative feedback data serves as the first type of negative sample, used to suppress the recommendation model's strategy of recommending the strategy to similar player states.

[0092] S440. If the statistical feature of the second technical indicator is less than the personalized historical performance benchmark, then a second type of negative feedback data is generated as the effect feedback data of the recommended strategy.

[0093] Understandably, the second type of negative feedback data indicates that the recommended strategy had a negative effect on the target player's skill improvement. Specifically, after watching and practicing the recommended strategy, the target player's technical metric statistical features significantly fell below the lower limit of their personalized historical performance benchmark—a case of "learning with the opposite effect," such as the target player learning incorrect techniques leading to a deterioration in performance. In subsequent model training, the second type of negative feedback data, as the second type of negative sample, has a heavier penalty weight than the first type of negative feedback data, so that the recommendation model more strongly avoids recommending the strategy in situations involving similar players.

[0094] In the specific implementation process, when there are multiple types of technical indicators, for each type of technical indicator, comparison and judgment are performed according to the steps described in S420 to S440 to obtain the sub-comparison results corresponding to each type of technical indicator. For example, if there are two types of technical indicators, the sub-comparison results corresponding to "speech validity score" and "voting accuracy rate" are obtained respectively.

[0095] After obtaining the sub-comparison results of various technical indicators, the results are integrated according to a preset fusion strategy to generate the final effect feedback data of the recommended strategy. For example, the fusion strategy can be: when at least one type of technical indicator has a sub-comparison result that is second-type negative feedback data, the final effect feedback data is determined to be second-type negative feedback data; when the sub-comparison results do not have second-type negative feedback data but at least one type of technical indicator has a sub-comparison result that is first-type negative feedback data, the final effect feedback data is determined to be first-type negative feedback data; only when all technical indicators have sub-comparison results that are positive feedback data, the final effect feedback data is determined to be positive feedback data. This strategy prioritizes the worst-case scenario and can effectively identify strategies that negatively impact player skills. The fusion strategy can also be: statistically analyzing the category distribution of the sub-comparison results of various technical indicators, and using the category that appears most frequently as the final effect feedback data. For example, if there are three types of technical indicators, two of which have sub-comparison results that are first-type negative feedback data and one that is positive feedback data, then the final effect feedback data is determined to be first-type negative feedback data.

[0096] In the specific implementation process, when the personalized historical performance benchmark is a personalized historical performance interval, the statistical feature of the second technical indicator is compared with the personalized historical performance interval. If the statistical feature of the second technical indicator is greater than the upper boundary of the personalized historical performance interval, positive feedback data for the recommended strategy is generated; if the statistical feature of the second technical indicator is within the personalized historical performance interval, a first type of negative feedback data for the recommended strategy is generated; if the statistical feature of the second technical indicator is less than the lower boundary of the personalized historical performance interval, a second type of negative feedback data for the recommended strategy is generated.

[0097] Understandably, steps S420 to S440 are mutually exclusive branch decision steps. For the second technical indicator statistical feature obtained in S410, it is compared sequentially with the upper and lower limits of the personalized historical performance benchmark, and the corresponding branch step is executed according to the range in which the comparison result falls.

[0098] For example, if the personalized historical performance benchmark is a personalized historical performance range, it has an upper limit. and lower limit The specific execution order of S420 to S440 is as follows: First, determine whether the statistical feature value of the second technical indicator is greater than the upper limit of the personalized historical performance benchmark. If yes, execute S420 to generate positive feedback data, and the process ends; otherwise, proceed to the next step.

[0099] Next, it is determined whether the statistical feature of the second technical indicator is located between the lower and upper limits of the personalized historical performance benchmark, i.e., whether it is less than or equal to the upper limit and greater than or equal to the lower limit. If yes, then S430 is executed to generate the first type of negative feedback data, and the process ends; if not, then proceed to the next step of judgment.

[0100] Finally, if neither of the above two conditions is met, the statistical feature of the second technical indicator will necessarily be less than the lower limit of the personalized historical performance benchmark. At this time, S440 is executed to generate the second type of negative feedback data.

[0101] By executing the above sequence, it is ensured that each group of second technical indicator statistical features generates only one corresponding effect feedback data, and the three effect feedback data are mutually exclusive and jointly cover all possible comparison results.

[0102] Furthermore, the execution order of S420, S430, and S440 can be interchanged.

[0103] By classifying the effect feedback data into three levels, this embodiment achieves a refined hierarchical evaluation of the effectiveness of the recommended strategies, providing richer and more accurate supervision signals for subsequent model optimization.

[0104] Example 2 Based on Example 1, such as Figure 7 As shown, this embodiment provides a method for recommending game strategies, which can be further divided into the following steps: A100: Obtain the target player's player status information; It is understood that the player status information is a comprehensive set of information used to characterize the target player's current game level and technical shortcomings, and it is structured information.

[0105] In the specific implementation process, the player status information can be generated by integrating and processing the game-related data of the target player. The game-related data can be long-term statistical data and recent game behavior data.

[0106] In the specific implementation process, the player state information can be generated by processing the target player's game-related data using a Large Language Model (LLM). The Large Language Model refers to an artificial intelligence model that has been pre-trained on a large-scale text corpus and possesses natural language understanding and generation capabilities. It can be implemented using a publicly available pre-trained model architecture or other pre-trained language models with similar functions; this application does not limit its implementation in this regard.

[0107] In the specific implementation process, the large language model can process and obtain the player status information of the target player in the following way: Long-term statistical data of the target player, such as win rate, skill utilization rate, average score, and historical averages of various technical indicators within a certain time window, and recent game behavior data, such as technical indicator scores, operation records, and decision-making performance in recent games, are integrated into the input data of the large language model. The large language model is then used to analyze and understand the input data to generate structured player status information. For example, the player status information includes at least one of the following: the target player's role information in the game, such as "Sheriff," "Villager," "Werewolf," etc.; technical weakness information, such as "hesitant to fire," "illogical speech," "low voting accuracy," etc.; and game skill level information, such as "Bronze," "Silver," "Diamond," etc. For example, the player status information can be represented in a structured data form, such as {role: 'Sheriff', weakness: 'hesitant to fire', rank: 'Diamond'}.

[0108] In the specific implementation process, the large language model acquires its ability to process player state information through pre-training. This pre-training can include the following steps: First, a training dataset is constructed. This dataset includes historical game data samples from multiple players and corresponding player state information annotations. The player state information annotations include each player's role information, skill weakness information, and game skill level information. The training dataset can be obtained by collecting a large amount of historical game data and long-term statistical data from players. Game domain experts then annotate each player's role preferences, skill weaknesses, and game skill level based on this data, forming training samples. Preferably, the training dataset contains at least several thousand player samples with different game skill levels and role preferences to ensure the model can learn diverse player state patterns. Then, based on the training dataset, the historical game data of each training sample is converted into natural language descriptions using a pre-constructed input template. These natural language descriptions are input into the large language model, and the corresponding player state information annotations are used as the training objective to fine-tune the large language model. The training objective can be to minimize the difference between the model output and the player state information annotations. Through the aforementioned fine-tuning process, the large language model learns to extract key information from players' historical game data, identify players' role preferences, skill weaknesses, and game skill levels, and output structured player status information in a unified format. After pre-training, in the practical application phase, inputting the target player's historical game data into the large language model will yield the target player's player status information. The pre-training can be completed offline. During actual use, newly accumulated player data can be used to periodically perform incremental updates to the large language model to maintain the accuracy and timeliness of its generated player status information.

[0109] In one alternative implementation, if the large language model itself already possesses strong semantic understanding and information extraction capabilities in the game domain, the historical game-related data of the target player can be directly input into the large language model and corresponding prompts can be used to generate player status information without requiring additional fine-tuning training of the large language model.

[0110] A200: Based on the player status information, recall multiple candidate strategies from the strategy library; Understandably, the strategy guide library is a pre-built searchable strategy guide database. In specific implementation, the strategy guide library is constructed in the following way: acquiring multiple strategy guide data, including but not limited to text and image guides, video explanations, tactical analyses, etc.; tagging the multiple strategy guide data respectively to obtain tagged strategy guide data; vectorizing the tagged strategy guide data to obtain corresponding vectors, storing them in a vector database, and thus obtaining the strategy guide library.

[0111] In one specific implementation, the tagging process includes: predefining a unified tagging system that includes dimensions such as character, strategy direction, difficulty, and game stage; specifically, a large language model can be used to automatically generate accurate tags that conform to the tagging system for massive amounts of unstructured guides. The vectorization process includes: using an embedding model to encode information such as the guide's title, tags, and summary into document vectors.

[0112] In one specific implementation, the step of recalling multiple candidate strategies from the strategy library based on the player state information includes: converting the player state information into a query vector; performing an approximate nearest neighbor (ANN) search in the vector database of the strategy library to obtain multiple candidate strategies with the highest similarity to the query vector. For example, the top-50 candidate strategies can be recalled.

[0113] A300, Construct a prompt message containing the player's status information and the multiple candidate strategies; In practice, the prompt information is text or structured data organized in natural language, which is used to input the recommendation model so that the recommendation model can understand the current player status and available strategy content.

[0114] Understandably, during the retrieval process from the strategy guide library, the multiple candidate strategies are stored in a vector database in vector form. Each candidate strategy's vector carries corresponding strategy metadata, including information such as strategy title, tags, summary, and strategy content identifier. After determining the multiple candidate strategies, the corresponding strategy metadata can be retrieved from the strategy guide library based on the vector identifiers of the multiple candidate strategies for use in constructing subsequent prompt information.

[0115] In one specific embodiment, the prompt information is constructed as follows: the player status information is converted into a first natural language description, the strategy metadata of the multiple candidate strategies is converted into a second natural language description, and the first natural language description and the second natural language description are combined to obtain the prompt information. In another specific embodiment, the prompt information is constructed by: using a preset template, filling the corresponding positions of the player status information and the strategy metadata of the multiple candidate strategies into the template to generate structured prompt information. The template may include a player status field and a candidate strategy list field.

[0116] It is understood that the first natural language description, the second natural language description, and the template can all be set manually or automatically generated by the large language model. This application does not limit this.

[0117] In practice, the number of candidate guides in the prompt information is consistent with the number recalled in A200. For example, if 50 candidate guides are recalled in A200, a candidate guide list can be formed, containing metadata for all 50 candidate guides. After the prompt information is input into the recommendation model, the recommendation model performs inference based on the player's state information and the candidate guide list, selecting the most accurate guide from multiple candidates as the target recommended guide.

[0118] A400. Input the prompt information into the recommendation model to obtain the target recommended strategy; wherein, the recommendation model is obtained by fine-tuning based on the effect feedback data, and the effect feedback data is generated using the game strategy recommendation effect generation method described in Example 1.

[0119] In one specific implementation, the recommendation model is a large language model-based recommendation model, meaning that the recommendation model uses a large language model as its core decision network. The recommendation model intelligently filters and sorts the candidate guides, outputting the most accurate target recommendation guide to the target player from multiple candidate guides. By outputting only a single target recommendation guide instead of multiple options, player choice overload can be avoided. In the specific implementation process, the recommendation model acquires the ability to filter target recommended strategies from the prompt information through the following pre-training method: First, a recommendation training dataset for the recommendation model is constructed. This dataset includes multiple training samples, each comprising: sample player status information, multiple candidate strategies, and a labeled target recommendation strategy. The training dataset can be obtained by collecting historical game data from a large number of players, generating sample player status information for each player using the method described in the A100, recalling multiple candidate strategies for each player status from the strategy library, and having game domain experts or automated annotation methods label the strategy most suitable for the player's status as the target recommendation strategy. Preferably, the training dataset contains at least tens of thousands of training samples to cover different player statuses, roles, skill limitations, and strategy types.

[0120] Then, the recommendation model is fine-tuned based on the recommended training dataset. Specifically, following the method described in A300, the sample player state information and multiple sample candidate strategies in each training sample are used to construct prompt information, which is then used as the input to the recommendation model; the labeled target recommended strategy is used as the expected output of the recommendation model. The training objective is to minimize the difference between the actual output of the recommendation model and the labeled target recommended strategy. For example, the cross-entropy loss function can be used for optimization. During the fine-tuning process, the recommendation model learns the decision pattern of selecting the optimal strategy from the sample state information and the sample candidate strategy list, that is, it learns to understand the matching relationship between the sample player state and the labeled target recommended strategy. After the above fine-tuning, the recommendation model can output the most suitable target recommended strategy for the player based on the given player state information and candidate strategies.

[0121] In one specific implementation, the pre-training of the recommendation model can also be continuously performed during the actual deployment and application of the recommendation model: the recommended strategies generated in actual recommendations and their corresponding effect feedback data are used as new training samples, and the recommendation model is incrementally fine-tuned periodically, so that the recommendation model gradually learns which strategy to recommend under what player state will bring the most realistic skill improvement, thereby continuously optimizing the recommendation strategy. The effect feedback data can be generated by the method described in Embodiment 1 of this application.

[0122] In practice, the target recommended strategy is a strategy selected by the recommendation model from the multiple candidate strategies. For example, if there are 50 candidate strategies, the recommendation model comprehensively evaluates the 50 candidate strategies listed in the prompt information, outputs the best one as the target recommended strategy, and pushes this target recommended strategy to the target player. By outputting only a single target recommended strategy instead of multiple options, player choice overload can be avoided.

[0123] In the specific implementation process, such as Figure 8 As shown, the recommendation model is adjusted based on performance feedback data, including: B410. Construct an unpaired dataset based on the effect feedback data, wherein each data point in the unpaired dataset includes the player status information, the recommended strategy, the candidate strategy, and the effect feedback data; It is understood that the effect feedback data includes the positive feedback data, the first type of negative feedback data, or the second type of negative feedback data generated in Example 1, which are used to indicate the preference type of the recommended strategy.

[0124] Specifically, the unpaired dataset refers to a dataset consisting of multiple independent data points, where each data point corresponds to a complete information record generated during a single recommended strategy interaction. Unlike traditional paired datasets, each data point in the unpaired dataset in this embodiment does not rely on pairwise comparisons with other data points, but independently contains complete information about the player's state, recommended strategy, candidate strategies, and effect feedback data in a single recommended strategy, and can be used independently for model training. In specific implementation, the data points in the unpaired dataset can be constructed according to a preset format. For example, the preset format is (Prompt, Response, Label), where Prompt corresponds to the current player state information and candidate strategy list, Response corresponds to the recommended strategy output by the recommendation model, and Label corresponds to the label type (positive feedback data, first-type negative feedback data, or second-type negative feedback data) corresponding to the effect feedback data. The meanings of each component are as follows: (1) The player state information is part of the Prompt: it corresponds to the state description of the target player when the recommendation occurs, including the target player's role information, technical shortcomings information, game level information, etc., and is used to provide decision context for the recommendation model.

[0125] (2) The candidate strategies are part of the Prompt: strategy metadata corresponding to multiple candidate strategies recalled from the strategy library during recommendation, used to characterize the set of strategies that the recommendation model can choose from when making a decision.

[0126] (3) The recommended strategy is a single recommended strategy that corresponds to the actual output of the recommendation model and is recommended to the target player, i.e., the output result of the recommendation model.

[0127] (4) The effect feedback data serves as a Label: corresponding to the positive feedback data, first-type negative feedback data, or second-type negative feedback data of the recommendation strategy generated using the method described in Example 1, used to indicate the effect category of the recommendation strategy. In the non-paired optimization algorithm, the effect feedback data corresponds to a label, used to guide the optimization direction of the recommendation model: positive feedback data indicates that the recommendation model strengthens the recommendation strategy, first-type negative feedback data indicates that the recommendation model suppresses the recommendation strategy, and second-type negative feedback data indicates that the recommendation model suppresses the recommendation strategy with higher weight.

[0128] It should be noted that data points in the unpaired dataset do not need to appear in pairs. A single data point can independently constitute a training sample, which is fundamentally different from paired datasets that require pairwise comparisons, such as comparing two different strategies for the same player. By using an unpaired dataset, this embodiment can fully utilize the single-point feedback signals generated in actual recommendation scenarios: in actual recommendation scenarios, only one strategy is recommended to the player at a time, and the player generates performance feedback for that single recommendation; therefore, the feedback signal is naturally single-point and unpaired. Directly utilizing these single-point feedback signals for model adjustment eliminates the need to construct additional paired samples, reducing the cost of building training data while improving the efficiency of training data utilization.

[0129] By using the data points of the unpaired dataset as training samples to input into the unpaired optimization algorithm, the recommendation model can update its parameters based on the label categories of the effect feedback data in the data points, and gradually learn which strategy to recommend in which player state can bring the most realistic skill improvement.

[0130] In practice, the effect feedback data can be accumulated after a preset number of recommended guides have been collected. This accumulated feedback data is then used to construct the unpaired dataset to adjust the recommendation model. Specifically, since the effect feedback data for a single recommended guide only reflects its effect in a single recommendation, adjusting the recommendation model based solely on this single data point may lead to overfitting to the single recommendation result and make it difficult to capture the overall performance of the recommendation model across various player states. By accumulating effect feedback data from multiple recommended guides, a dataset covering various player states, guide types, and changes in various technical indicators can be constructed. This allows the adjustment direction of the recommendation model to accommodate various scenarios, improving the model's generalization ability and stability.

[0131] For example, a preset data volume threshold can be set, such as accumulating feedback data corresponding to 100 recommended strategies, or a preset time period can be set, such as every 7 days. When the accumulated feedback data reaches the preset threshold or the current time reaches the preset time period, an adjustment process for the recommendation model is triggered. After the adjustment process is triggered, the unpaired dataset is constructed based on the accumulated feedback data of multiple recommended strategies, and the recommendation model is fine-tuned using the unpaired dataset. After the adjustment is completed, the used feedback data can be cleared, and a new round of data accumulation and adjustment cycle can begin.

[0132] Through the above mechanism, the recommendation model forms a continuous optimization closed loop of target recommendation strategy output, effect feedback data generation, data accumulation, model adjustment, and recommendation strategy update, so that the recommendation model can gradually evolve as player effect feedback data accumulates, and continuously improve recommendation accuracy in the long term.

[0133] B420. Adjust the recommendation model using the unpaired dataset.

[0134] In the specific implementation process, the recommendation model is fine-tuned using the unpaired dataset. The fine-tuning process can be optimized by maximizing the utility corresponding to the positive feedback data and minimizing the loss corresponding to the first type of negative feedback data and the second type of negative feedback data, and the parameters of the recommendation model are updated accordingly.

[0135] In one specific implementation, a foreground utility optimization algorithm can be used to implement the above-mentioned fine-tuning process. Through this fine-tuning process, the recommendation model gradually learns which strategy guide will bring the most realistic skill improvement under what player conditions, thereby significantly improving the accuracy and long-term value of game strategy guide recommendations.

[0136] Example 3 like Figure 9 As shown, this embodiment also provides a game strategy recommendation effect generation system. Optionally, the generation system includes: a first acquisition module 511, a construction module 512, a second acquisition module 513, and a generation module 514, wherein: The first acquisition module 511 is used to acquire the target player's first technical indicator data in the game before clicking on the recommended strategy; In this embodiment, the first acquisition module 511 can be used to perform... Figure 2 For a detailed description of the first acquisition module 511, please refer to the description of step S100 shown.

[0137] The construction module 512 is used to construct a personalized historical performance benchmark for the target player based on the first technical indicator data. In this embodiment, the construction module 512 can be used to execute Figure 2 For a detailed description of the construction module 512, please refer to the description of step S200 shown.

[0138] Specifically, the construction module 512 is further configured to obtain the first technical indicator statistical feature quantity of the target player based on the first technical indicator data, and to construct the personalized historical performance benchmark of the target player based on the first technical indicator statistical feature quantity.

[0139] In this embodiment, the construction module 512 can also be used to execute Figure 3 For a detailed description of the construction module 512, see steps S210-S220 shown below. Further details regarding steps S210-S220 can be found in the description of steps S210-S220.

[0140] The second acquisition module 513 is used to acquire the second technical indicator data of the target player in the game after determining that the target player clicked on the recommended strategy in the first time window; In this embodiment, the second acquisition module 513 can be used to perform... Figure 2 For a detailed description of step S300 shown, and the second acquisition module 513, please refer to the description of step S300.

[0141] Specifically, the second acquisition module 513 is further configured to continuously track the technical indicator data of the target player within a preset second number of rounds after the first time window, obtain multiple technical indicator data samples, and determine the second technical indicator data based on the multiple technical indicator data samples.

[0142] In this embodiment, the second acquisition module 513 can also be used to perform Figure 5 For a detailed description of steps S310-S320 shown, please refer to the description of steps S310-S320 for further details regarding the second acquisition module 513.

[0143] The generation module 514 is used to obtain the comparison result between the second technical indicator data and the personalized historical performance benchmark, and generate the effect feedback data of the recommended strategy based on the comparison result.

[0144] In this embodiment, the generation module 514 can be used to execute Figure 2 For a detailed description of the generation module 514, please refer to the description of step S400 shown.

[0145] Specifically, the generation module 514 is further configured to obtain the second technical indicator statistical feature quantity of the target player according to the second technical indicator, and if the second technical indicator statistical feature quantity is greater than the personalized historical performance benchmark, generate positive feedback data as the effect feedback data of the recommended strategy; if the second technical indicator statistical feature quantity is within the personalized historical performance benchmark, generate a first type of negative feedback data as the effect feedback data of the recommended strategy; and if the second technical indicator statistical feature quantity is less than the personalized historical performance benchmark, generate a second type of negative feedback data as the effect feedback data of the recommended strategy.

[0146] In this embodiment, the generation module 514 can also be used to execute Figure 6For a detailed description of the generation module 514, see steps S410-S440 shown below. Further details regarding steps S410-S440 can be found in the description of steps S410-S440.

[0147] Example 4 like Figure 10 As shown, this embodiment also provides a game strategy recommendation system. Optionally, the recommendation system includes: a third acquisition module 611, a recall module 612, a prompt information construction module 613, and a recommendation module 614, wherein: The third acquisition module 611 is used for users to acquire player status information of the target player; In this embodiment, the third acquisition module 611 can be used to perform... Figure 7 For a detailed description of the third acquisition module 611, please refer to the description of step A100 shown.

[0148] The recall module 612 is used to recall multiple candidate strategies from the strategy library based on the player status information; In this embodiment, the recall module 612 can be used to perform... Figure 7 For a detailed description of the recall module 612, please refer to the description of step A200 shown.

[0149] The prompt information construction module 613 is used to construct prompt information that includes the player status information and the multiple candidate strategies; In this embodiment, the prompt information construction module 613 can be used to execute... Figure 7 For a detailed description of the prompt information construction module 613 shown in step A300, please refer to the description of step A300.

[0150] Recommendation module 614 is used to input the prompt information into the recommendation model to obtain the target recommended strategy; wherein, the recommendation model is obtained by fine-tuning based on the effect feedback data, and the effect feedback data is generated by the game strategy recommendation effect generation system described above. In this embodiment, the recommendation module 614 can be used to perform... Figure 7 For a detailed description of the recommendation module 614, please refer to the description of step A400 shown.

[0151] This application also provides an electronic device, the structure of which is as follows: Figure 11 As shown, the electronic device includes a memory 711, a processor 712, a communication module 713, and an input / output interface 714, etc. Optionally, the memory 711, the processor 712, the communication module 713, and the input / output interface 714 can be connected and communicate with each other through a bus 715.

[0152] The memory 711 is used to store one or more computer programs and to transfer the code of the computer programs to the processor 712; when the one or more computer programs are executed by the processor 712, a method for generating game strategy recommendation effects or a game strategy recommendation method according to an embodiment of this application is implemented.

[0153] Optionally, the electronic device can be connected to a network via the communication module 713 to communicate with other devices, such as terminals or servers, and to interact with data. The electronic device can be various forms of digital computers, exemplarily such as desktop computers, servers, workbenches, mainframes, or other types of computers. The electronic device can also be various forms of mobile terminals, exemplarily such as smartphones, tablets, wearable devices (such as helmets, glasses, watches, etc.), and other similar mobile terminals.

[0154] Optionally, the electronic device can connect to desired input / output devices, such as a keyboard or display device, via the input / output interface 714. The electronic device itself may have a display device, and other display devices can also be connected externally via the input / output interface 714. Optionally, a storage device, such as a hard disk, can also be connected via the input / output interface 714 to store data from the electronic device, read data from the storage device, or store data from the storage device in the memory 711. It is understood that the input / output interface 714 can be a wired interface or a wireless interface. Depending on the actual application scenario, the device connected to the input / output interface 714 can be a component of the electronic device or an external device connected to the electronic device when needed.

[0155] Optionally, the memory 711 may be a volatile memory and / or a non-volatile memory. The volatile memory may be a random access memory, etc., and the non-volatile memory may be a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, or a flash memory, etc.

[0156] Optionally, the computer program stored in the memory 711 can be divided into one or more modules, which are stored in the memory 711 and executed by the processor 712 to perform the method provided in this embodiment. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the electronic device.

[0157] Optionally, the processor 712 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the processor 712 include, but are not limited to, a central processing unit, a graphics processing unit, a digital signal processor, various special-purpose artificial intelligence computing chips, various processors running machine learning model algorithms, and can also be any suitable controller, microcontroller, processor, etc. The processor 712 executes the various methods and processes of this embodiment, exemplarily, such as a game strategy recommendation effect generation method or a game strategy recommendation method according to an embodiment of this application.

[0158] Optionally, the bus 715 may include a path for transmitting information. Depending on its function, the bus 715 may be classified as an address bus, a data bus, a control bus, etc.

[0159] In an optional implementation, this application embodiment also provides a computer storage medium storing a computer program thereon. When the computer program is executed by a computer, it enables the computer to perform the methods described in the above-described method embodiments. Part or all of the computer program can be loaded and / or installed on the memory 711 of an electronic device. When the computer program is executed by the processor 712, it can perform one or more steps of a game strategy recommendation effect generation method or a game strategy recommendation method according to this application embodiment.

[0160] Optionally, the computer-readable storage medium may be a random access memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, etc.

[0161] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for generating game strategy recommendation effects, characterized in that, The generation method includes: Obtain the target player's primary technical metrics in the game before they click on the recommended strategy; Based on the first technical indicator data, a personalized historical performance benchmark for the target player is constructed; After determining the first time window after the target player clicked on the recommended strategy, obtain the target player's second technical indicator data in the game; Obtain the comparison results between the second technical indicator data and the personalized historical performance benchmark, and generate the effect feedback data of the recommended strategy based on the comparison results.

2. The generation method according to claim 1, characterized in that, The step of constructing a personalized historical performance benchmark for the target player based on the first technical indicator data includes: Based on the first technical indicator data, obtain the first technical indicator statistical feature quantity of the target player; Based on the statistical features of the first technical indicator, a personalized historical performance benchmark for the target player is constructed.

3. The generation method according to claim 2, characterized in that, The first technical indicator statistical features are the first technical indicator mean and the first technical indicator standard deviation; The step of constructing the personalized historical performance benchmark of the target player based on the statistical features of the first technical indicator includes: Based on the preset confidence coefficient and the mean and standard deviation of the first technical indicator, the upper boundary and lower boundary of the interval are constructed. A personalized historical performance range is constructed based on the upper and lower boundaries of the range, serving as the personalized historical performance benchmark for the target player.

4. The generation method according to claim 1, characterized in that, The step of obtaining the comparison result between the second technical indicator data and the personalized historical performance benchmark, and generating the effect feedback data of the recommended strategy based on the comparison result, includes: Based on the second technical indicator, obtain the statistical feature quantity of the second technical indicator of the target player; If the statistical feature of the second technical indicator is greater than the personalized historical performance benchmark, then positive feedback data is generated as the effect feedback data of the recommended strategy. If the statistical feature of the second technical indicator is within the personalized historical performance benchmark, then the first type of negative feedback data is generated as the effect feedback data of the recommendation strategy; If the statistical feature of the second technical indicator is less than the personalized historical performance benchmark, then a second type of negative feedback data is generated as the effect feedback data of the recommended strategy.

5. The generation method according to claim 1, characterized in that, The first time window is obtained in the following way: The time taken for the target player to complete a preset number of games in the game after clicking on the recommended strategy is obtained, and the time taken is used as the first time window.

6. The generation method according to any one of claims 1 to 5, characterized in that, The determination that the target player clicked on the recommended strategy includes: The recommended strategy guide is pushed to the target player. When the target player's click operation information on the recommended strategy guide is obtained within a preset second time window after the recommended strategy guide is pushed, it is determined that the target player has clicked on the recommended strategy guide.

7. The generation method according to any one of claims 1 to 5, characterized in that, After determining that the target player clicked on the recommended strategy in the first time window, the process of acquiring the target player's second technical indicator data in the game includes: Within a preset second number of rounds following the first time window, the technical indicator data of the target player is continuously tracked to obtain multiple technical indicator data samples; The second technical indicator data is determined based on the multiple technical indicator data samples.

8. A method for recommending game strategies, characterized in that, The recommendation method includes: Obtain the target player's player status information; Based on the player status information, multiple candidate strategies are recalled from the strategy library; Construct a hint message that includes the player's status information and the multiple candidate strategies; The prompt information is input into the recommendation model to obtain the target recommended strategy; wherein the recommendation model is fine-tuned based on the effect feedback data, and the effect feedback data is generated by the game strategy recommendation effect generation method according to any one of claims 1-7.

9. The recommended method according to claim 8, characterized in that, The recommendation model is adjusted based on performance feedback data, including: An unpaired dataset is constructed based on the effect feedback data. Each data point in the unpaired dataset includes the player status information, the recommended strategy, the candidate strategy, and the effect feedback data. The recommendation model is adjusted using the unpaired dataset.

10. A game strategy recommendation effect generation system, characterized in that, The generation system includes: The first acquisition module is used to acquire the target player's first technical indicator data in the game before clicking on the recommended strategy; A construction module is used to construct a personalized historical performance benchmark for the target player based on the first technical indicator data; The second acquisition module is used to acquire the second technical indicator data of the target player in the game after determining that the target player clicked on the recommended strategy in the first time window; The generation module is used to obtain the comparison results between the second technical indicator data and the personalized historical performance benchmark, and generate the effect feedback data of the recommended strategy based on the comparison results.

11. A game strategy recommendation system, characterized in that, The recommendation system includes: The third acquisition module is used by users to obtain the player status information of the target player; The recall module is used to recall multiple candidate strategies from the strategy library based on the player status information. The prompt information construction module is used to construct prompt information that includes the player's status information and the multiple candidate strategies; The recommendation module is used to input the prompt information into the recommendation model to obtain the target recommended strategy; wherein the recommendation model is obtained by fine-tuning based on the effect feedback data, and the effect feedback data is generated by the game strategy recommendation effect generation system of claim 10.

Images