Game event prediction method and device, equipment and storage medium

Abstract

The invention relates to the technical field of video processing, and discloses a game event prediction method and device, equipment and a storage medium. The game event prediction method comprises the following steps: receiving a game video to be predicted, wherein the game video comprises a game live video or a game video; extracting a plurality of pieces of battle state data corresponding to each time slice in the game video, and splicing the battle state data after vectorization processing into a battle state vector; constructing a battle state vector sequence corresponding to each time point based on the battle state vectors of the plurality of continuous time slices; and inputting each battle state vector sequence into a pre-trained game event prediction model for processing, and outputting an outbreak probability vector including types and positions of various game events in the future. The resource consumption of game event prediction is reduced, and the game event prediction efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of video processing technology, and in particular to a method, apparatus, device, and storage medium for predicting game events. Background Technology

[0002] With the rise of online games, their competitiveness and strategic elements have increasingly attracted the attention of players and the esports industry. In games, "key events" such as team fights, ganking, and tower pushing are core factors that directly influence the course of the battle and determine victory or defeat. Currently, the industry's analysis of key game events mainly relies on two technical approaches. The first is a game log-based analysis method. This method analyzes massive amounts of log data recorded by the game server, including player behavior, unit positions, skill usage, and economic changes. After an event occurs, data mining and pattern matching are performed to summarize the type and process of the key event. The second is a computer vision-based recognition method. This method performs image recognition on game screens or live video streams, using technologies such as object detection and behavior recognition to detect visual elements such as hero positions, health changes, and skill effects, thereby determining the occurrence of key events.

[0003] However, the aforementioned existing technologies share a common core technical challenge: their analytical model is essentially "post-hoc induction," failing to achieve effective "pre-hoc early warning." Specifically, the existing technologies have the following shortcomings: First, it consumes a lot of resources, impacting game performance: Both log analysis and visual recognition require processing and analyzing massive amounts of data (such as thousands of lines of log files or high-frame-rate video streams), resulting in extremely high computational overhead and severely consuming the CPU, GPU, and memory resources of game devices or analysis servers. Especially in scenarios where real-time analysis is performed on the player's client, this high load directly affects the smoothness of the game, causing stuttering or frame drops, severely damaging the user experience.

[0004] Secondly, the analysis is inefficient and lacks real-time capability: Due to the massive amount of data processed and the complex analysis process, existing technologies suffer from low analysis and identification efficiency and significant delays. This delay means that analysis results are often only available after the key event has occurred or even ended, completely negating the timeliness of early warnings and failing to provide effective references for players in the moment of decision-making.

[0005] Third, the prediction accuracy is low: Due to the complexity of the game environment and the unpredictability of player behavior, log-based or vision-based models often only capture the surface phenomena of events, making it difficult to deeply understand the underlying logic of events and the potential intentions of various characters. For example, the model may identify multiple heroes gathering, but cannot accurately determine whether this is an impending team fight or simply a routine tactical lane swap or roaming. This results in limited prediction accuracy and a tendency for misjudgments or omissions. Summary of the Invention

[0006] The main objective of this invention is to provide a method, apparatus, device, and storage medium for predicting game events, aiming to solve the technical problems of high resource consumption, low analysis efficiency, and low prediction accuracy in existing game key event analysis.

[0007] The first aspect of this invention provides a game event prediction method, the game event prediction method comprising: Receive game videos to be predicted, including live game videos or game recordings; Extract multiple battle state data corresponding to each time slice in the game video, and concatenate the vectorized battle state data into a battle state vector. Based on the battle state vectors of multiple consecutive time slices, construct the battle state vector sequence corresponding to each time point; The sequence of battle state vectors is input into a pre-trained game event prediction model for processing, and the output is a probability vector containing the type and location of future game events.

[0008] Optionally, in a first implementation of the first aspect of the present invention, the battle state data includes high-frequency structured data and low-frequency unstructured data, and the step of extracting multiple battle state data corresponding to each time slice in the game video and concatenating the vectorized battle state data into a battle state vector includes: Within each time slice, the high-frequency structured data and the low-frequency unstructured data are extracted from the game screen and bullet screen text of the game video, respectively. The high-frequency structured data and the low-frequency unstructured data are vectorized respectively to obtain the first extracted feature vector and the second extracted feature vector. The first extracted feature vector and the second extracted feature vector corresponding to each time slice are concatenated in a fixed order to obtain the battle state vector.

[0009] Optionally, in a second implementation of the first aspect of the present invention, the step of extracting the high-frequency structured data and the low-frequency unstructured data from the game screen and bullet screen text of the game video in each time slice includes: Within each time slice, the high-frequency structured data is extracted from the game screen of the game video using a trained interface parsing model. The high-frequency structured data includes one or more of the following: the position of the game character, key states and core attributes, and the global state of the game. Within each time slice, the bullet screen text of the game video is classified using a pre-trained bullet screen text classification model to obtain the low-frequency unstructured data, which includes the probabilities of various emotional intent categories.

[0010] Optionally, in a third implementation of the first aspect of the present invention, the training method used for the game event prediction model includes: Acquire game recordings of multi-player battles; Extract multiple battle state data corresponding to each time slice in the game recording, and concatenate the vectorized battle state data into a battle state vector. Based on the battle state vectors of multiple consecutive time slices, a battle state vector sequence corresponding to each time point is constructed, and each of the battle state vector sequences is used as training samples. Locate the game events that occur in the game recording, and based on each game event, construct a sample label vector corresponding to each time point, which has the type of event and the location of the event. The model is trained based on the training samples and sample label vectors corresponding to each time point to obtain a trained game event prediction model.

[0011] Optionally, in a fourth implementation of the first aspect of the present invention, the step of locating game events occurring in the game recording and constructing a sample label vector corresponding to each time point with the type and location of the event based on each game event includes: Obtain the game battle logs corresponding to the game recordings; Based on the game battle log, the time and location of all game events are located in the game recording; Based on the occurrence time of each game event, an event prediction time window is set for each game event before the occurrence time. For each time point within the event prediction time window, generate a positive sample label vector corresponding to multiple event types and location pointers; For each time point outside the event prediction time window, a negative sample label vector corresponding to multiple event types and location pointers is generated.

[0012] Optionally, in a fifth implementation of the first aspect of the present invention, the step of training the model based on the training samples and the sample label vectors corresponding to each time point to obtain a trained game event prediction model includes: The training samples corresponding to each time point are sequentially input into the pre-trained model; The pre-trained model is used to predict game events, and the output is a probability vector containing the type and location of various game events in the future. Based on the sample label vector and the outbreak probability vector corresponding to each time point, a target loss function is constructed; Based on the target loss function, the model parameters of the pre-trained model are updated to obtain a trained game event prediction model.

[0013] A second aspect of the present invention provides a game event prediction device, the game event prediction device comprising: A receiving module is used to receive game videos to be predicted, including live game videos or game recordings. The extraction module is used to extract multiple battle state data corresponding to each time slice in the game video, and to concatenate the vectorized battle state data into a battle state vector. The input building module is used to construct a sequence of battle state vectors for each time point based on battle state vectors from multiple consecutive time slices. The prediction module is used to input the sequence of battle state vectors into a pre-trained game event prediction model for processing, and output a probability vector containing the type and location of future game events.

[0014] Optionally, in a first implementation of the second aspect of the present invention, the battle situation data includes high-frequency structured data and low-frequency unstructured data, and the extraction module is specifically used for: Within each time slice, the high-frequency structured data and the low-frequency unstructured data are extracted from the game screen and bullet screen text of the game video, respectively. The high-frequency structured data and the low-frequency unstructured data are vectorized respectively to obtain the first extracted feature vector and the second extracted feature vector. The first extracted feature vector and the second extracted feature vector corresponding to each time slice are concatenated in a fixed order to obtain the battle state vector.

[0015] Optionally, in a second implementation of the second aspect of the present invention, the extraction module is further configured to: Within each time slice, the high-frequency structured data is extracted from the game screen of the game video using a trained interface parsing model. The high-frequency structured data includes one or more of the following: the position of the game character, key states and core attributes, and the global state of the game. Within each time slice, the bullet screen text of the game video is classified using a pre-trained bullet screen text classification model to obtain the low-frequency unstructured data, which includes the probabilities of various emotional intent categories.

[0016] Optionally, in a third implementation of the second aspect of the present invention, the game event prediction device further includes: The training module is used to acquire game recordings of multi-player game battles; extract multiple battle state data corresponding to each time slice in the game recordings, and concatenate the vectorized battle state data into a battle state vector; based on the battle state vectors of multiple consecutive time slices, construct a battle state vector sequence corresponding to each time point, and use each battle state vector sequence as a training sample; locate game events occurring in the game recordings, and based on each game event, construct a sample label vector corresponding to each time point with the event type and location; train the model based on the training samples and sample label vectors corresponding to each time point to obtain a trained game event prediction model.

[0017] Optionally, in a fourth implementation of the second aspect of the present invention, the training module is further configured to: Obtain the game battle logs corresponding to the game recordings; Based on the game battle log, the time and location of all game events are located in the game recording; Based on the occurrence time of each game event, an event prediction time window is set for each game event before the occurrence time. For each time point within the event prediction time window, generate a positive sample label vector corresponding to multiple event types and location pointers; For each time point outside the event prediction time window, a negative sample label vector corresponding to multiple event types and location pointers is generated.

[0018] Optionally, in a fifth implementation of the second aspect of the present invention, the training module is further configured to: The training samples corresponding to each time point are sequentially input into the pre-trained model; The pre-trained model is used to predict game events, and the output is a probability vector containing the type and location of various game events in the future. Based on the sample label vector and the outbreak probability vector corresponding to each time point, a target loss function is constructed; Based on the target loss function, the model parameters of the pre-trained model are updated to obtain a trained game event prediction model.

[0019] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the computer device to execute the above-described game event prediction method.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described game event prediction method.

[0021] The technical solution provided by this invention transforms the complex game analysis problem into an efficient, computable, time-series-based supervised learning problem through three core technologies: data abstraction, analysis temporalization, and training supervision. This ultimately achieves low resource consumption, high efficiency, and high accuracy in predicting game events.

[0022] (1) Based on multiple game tactical analysis dimensions (such as character dimension, environment dimension, and macro-battle situation dimension), corresponding battle situation data is extracted from game recordings. All redundant information unrelated to tactical decisions (such as character skin textures and ground grass details) is discarded, retaining only the core elements that can quantify the battle situation. At the same time, the above battle situation data is vectorized and concatenated, thereby compressing a complex game situation into a floating-point vector that may only have a few hundred dimensions. The exponential compression of data volume significantly reduces the resource requirements for both storage used for model training and memory and computation used for real-time prediction. In actual prediction, the system does not need to perform complex image rendering or log parsing. It only needs to directly obtain these core values ​​through the game API or memory reading, generate a lightweight vector, and then input it into the trained model. This process has extremely low CPU usage, thus completely solving the problem of lag caused by the analysis and recognition process occupying game device resources.

[0023] (2) This invention adopts a technical architecture that separates offline training and online inference, completing the most time-consuming and complex model training process offline. Once the model training is complete, when applied in real-time games, only the inference task needs to be performed, which can usually be completed within a few milliseconds. It can continuously make predictions in the seconds before key events occur in the game, thus giving players or the system sufficient reaction time. This invention uses a temporal modeling approach for model training, so that the model learns not a static snapshot of the battle situation, but the complete trend of changes in the battle situation in multiple dimensions such as space, state, economy, and emotion over a period of time.

[0024] (3) The multi-dimensional battle state vector adopted in this invention integrates information from multiple dimensions such as position, economy, skills, and status, enabling the model to make judgments that are far more intelligent and in line with game logic than visual recognition. Through this multi-dimensional quantitative comparison, the model can accurately identify potential signs before key events occur in the game, avoiding misjudgments caused by superficial phenomena. This invention adopts a supervised learning paradigm, which provides the model with a massive amount of accurate standard answers. The model no longer learns fuzzy correlational knowledge, but direct causal knowledge. Through extensive learning, the model can establish an extremely complex and accurate mapping relationship between battle trends and key events. The massive amount of video data contains all tactical patterns from novices to professional players. During the learning process, the model can not only master conventional tactics, but also discover many subtle tactical combinations and intentions that game players have not summarized. This understanding of the game's meta-level is a powerful guarantee for high-accuracy prediction of key game events.

[0025] This invention, through data abstraction, maps the game world from a high-dimensional sensory space to a low-dimensional tactical space, fundamentally avoiding the high resource consumption of analyzing and identifying key events. Simultaneously, it solves efficiency bottlenecks through offline / online separation and enables the model to predict the future from trends through temporal modeling, thereby achieving efficient pre-emptive warnings. Multi-dimensional information fusion provides rich judgment criteria, and precise supervised learning allows the model to directly grasp the deep logic from battle trends to event outcomes, thus significantly improving prediction accuracy. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of an embodiment of the game event prediction model training method in this invention. Figure 2 This is a schematic diagram of one embodiment of the game event prediction method in this invention; Figure 3 This is a schematic diagram of one embodiment of the game event prediction device in this invention; Figure 4 This is a schematic diagram of one embodiment of the computer device in this invention. Detailed Implementation

[0027] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention 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 described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a 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.

[0028] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the game event prediction model training method in this invention includes: 101. Obtain game recordings of multi-player battles; In this embodiment, multi-player game battles refer to multiple players controlling their own game characters (such as heroes, soldiers, etc.) to battle against game characters controlled by opposing players. Game recordings refer to recordings of the entire game process from the start to the end.

[0029] In this embodiment, the game recording includes not only a video stream but also a live chat text stream. The video stream contains all the information from the game screen, such as hero positions, skill effects, minimap status, hero health and mana bars, and economic and equipment information. These are the most direct and objective basis for judging the battle situation. The live chat text stream represents the real-time feedback from different viewers watching the game at the same moment. The viewers' excitement, tension, and worry, as well as the live chat text such as "Watch out!" and "Charge!", often foreshadow changes in the battle situation earlier than the game data itself.

[0030] 102. Extract multiple battle state data corresponding to each time slice in the game recording, and concatenate the vectorized battle state data into a battle state vector. In this embodiment, battle status data refers to data used to describe the state of the game battle. Battle status data changes dynamically in real time. Therefore, battle status data needs to be extracted for each time slice of the game recording, including high-frequency structured data extraction for the video stream and low-frequency unstructured data extraction for the bullet screen text stream.

[0031] In an optional embodiment, step 102 above further includes: 1021. Within each time slice, extract the high-frequency structured data and the low-frequency unstructured data from the game screen and bullet screen text of the game recording, respectively. In this optional embodiment, high-frequency structured data refers to data in the game interface that is well-formatted, easily quantifiable, and dynamically changing in real time, such as hero positions, skill effects, minimap status, hero health and mana bars, and economic and equipment information. Low-frequency unstructured data refers to the text content in the bullet screen messages, which contains high-value information for predicting the battle situation by reflecting the emotions and intentions of the audience.

[0032] In this embodiment, a fixed time step Δt (e.g., 0.5 seconds) is preferably used as the unit, and high-frequency structured data extraction and low-frequency unstructured data extraction are performed in parallel at each time slice t.

[0033] (1) Extraction of high-frequency structured data In one embodiment, preferably, high-frequency structured data is extracted from the game footage of the game recording within each time slice using a trained interface parsing model.

[0034] In one embodiment, the high-frequency structured data includes one or more of the following: the position of a game character, key states and core attributes, and the global state of the game.

[0035] Game character positions: The preferred method is to use the hero positions on the minimap, specifically extracting the coordinates [x, y] of all heroes on the minimap. If a hero is outside the field of view, a special value is used, such as [-1, -1].

[0036] Key states of game characters: It is preferable to use the key states of heroes, specifically extracting all heroes' [health, mana, whether the ultimate skill is ready (0 or 1)].

[0037] The core attributes of game characters: The core attributes of heroes are preferred, specifically the [level and current economy] of all heroes.

[0038] Global game status: Specifically extracts [game time (seconds), friendly kills, enemy kills, countdown to the next neutral creep spawn, and economic difference between the two sides]. These are key indicators of the overall battle situation and are crucial for predicting resource contention and large-scale team battles.

[0039] The interface parsing model used to extract high-frequency structured data is, for example, a real-time object detection model. In game scenarios, the object detection model can be pre-trained to identify and locate UI elements such as hero icons on the minimap, health / mana bars in the status bar, economic figures, and skill cooldown indicators. Then, OCR (Optical Character Recognition) is used to extract text information from video images. For example, specific level numbers and economic values ​​can be identified from the hero status bar, or kill counts and neutral resource countdowns can be identified from the global game status bar.

[0040] (2) Extraction of low-frequency unstructured data In one embodiment, preferably, within each time slice, the bullet screen text of the game recording is classified using a pre-trained bullet screen text classification model to obtain low-frequency unstructured data.

[0041] In one embodiment, the low-frequency unstructured data includes probabilities of multiple emotional intent categories.

[0042] 1) The emotional categories of bullet screen text mainly include: Excitement / praise (e.g., "Nice!", "That was amazing!", "That was incredible!") Disappointment / blame (e.g., "This is really appetizing," "How can this be empty?") Anxiety / worry (e.g., "Be careful!", "Don't go!", "You're going to get ambushed") 2) The main categories of intent in bullet screen text include: Offensive commands (such as "Go in", "Charge!", "Kill him in no time") Defense / retreat commands (such as "run fast", "don't engage", "defend the high ground") Information inquiry / discussion (e.g., "What equipment does he have?", "Why doesn't he zoom in?") Low-frequency unstructured data is extracted and used for deep semantic understanding using a Transformer-based model. To ensure efficiency in real-time inference, a lightweight version that has been distilled and quantized is specifically selected. A single bullet comment may simultaneously contain multiple emotions and intentions. For example, "Run, don't engage the enemy!" contains both the emotion of "tension / worry" and the intention of "defense / retreat." Therefore, this is designed as a multi-label classification task, where the model outputs not a single category, but rather the probability of each category.

[0043] 1022. The high-frequency structured data and the low-frequency unstructured data are vectorized respectively to obtain the first extracted feature vector and the second extracted feature vector. In this optional embodiment, the extracted high-frequency structured data is vectorized using normalization to obtain the first extracted feature vector. For example, in each time slice t (e.g., within 1 second), for the hero's position on the minimap, the x and y coordinates of the hero on the minimap are detected, and then the coordinates are normalized by dividing them by the length and width of the minimap respectively to obtain the quantized hero position on the minimap. The hero's health and mana are vectorized as percentages of maximum health and mana, with 1 representing the ultimate skill being ready and 0 representing it not being ready. The hero's level and current economy are vectorized as percentages of the highest level and maximum economy. Game time, friendly kills, enemy kills, countdown to the next neutral creep spawn, and the economic difference between the two sides are also vectorized using normalization.

[0044] For the extracted low-frequency unstructured data, aggregated weighted vectorization is used to obtain the second extracted feature vector. In each time slice t (e.g., within 0.5 seconds), the game application receives a large number of bullet screen texts. Therefore, each bullet screen text is first classified, and then the classification results of all bullet screens within that time slice are aggregated. The aggregation method can be simple averaging or weighted averaging. Finally, a bullet screen feature vector is output, such as [excitement percentage, disappointment percentage, tension percentage, offensive intent percentage, defensive intent percentage]. For example, the bullet screen feature vector for a certain time slice might be [0.65, 0.10, 0.20, 0.55, 0.15]. This vector clearly indicates that the current audience's emotions are mainly "excitement" and "offensive intent," a strong signal that battle is about to break out.

[0045] 1023. The first extracted feature vector and the second extracted feature vector corresponding to each time slice are concatenated in a fixed order to obtain the battle state vector.

[0046] To facilitate subsequent calculations, the first and second extracted feature vectors corresponding to each time slice need to be concatenated in a fixed order to form a battle state vector.

[0047] For example, all the extracted numerical features—minimap position, hero status, core attributes, global status, and bullet screen features—are concatenated into a single-row, multi-column vector, namely the battle state vector V_t, according to a preset, fixed order. The column dimension N of this vector is the sum of the dimensions of all features. This fixed order is crucial, ensuring that the meaning of each vector element remains consistent during model training and inference. Through this step, at each time point t, a complex, multi-dimensional snapshot of the battle situation is condensed into a high-dimensional numerical vector V_t. This vector can be seen as the model's understanding of the battle situation.

[0048] 103. Based on the battle state vectors of multiple consecutive time slices, construct the battle state vector sequence corresponding to each time point, and use each of the battle state vector sequences as training samples. A single time point t's battle state vector V_t is merely a static snapshot and cannot capture the dynamic evolution of the game. Events in multiplayer online games, such as roaming and ganking, are often caused by a series of continuous actions and state changes (e.g., the jungler starts roaming, the target hero pushes too far into the lane, vision disappears). Therefore, temporal information must be introduced. For example, the instantaneous snapshots of the states "peaceful standoff" and "imminent conflict" are very similar (e.g., both sides' heroes are near the dragon pit). If the model only looks at the current instant, it is difficult to distinguish between these two states. However, by observing the trajectories of the past 10 seconds, the model can detect that one side is gathering while the other is dispersing, thus making an accurate prediction.

[0049] In this embodiment, the training samples corresponding to time point t are selected from the battle state vectors of the n time slices closest to the current time point t, and stacked in chronological order to form an input matrix M_t of shape n x d, where n represents the number of time slices and d represents the dimension of the battle state vector corresponding to a single time slice. Specific parameter settings are as follows: Time step Δt: The interval used to perform feature extraction, preferably 0.5 seconds; Vector sequence length n: the number of time slices. Taking n=40 means taking 40 time slices.

[0050] Historical window: n * Δt = 40 * 0.5 seconds = 20 seconds, meaning that at any prediction time point t, the model input includes the changing trends of all battle states in the past 20 seconds.

[0051] This embodiment preferably uses a sliding window mechanism to construct a sequence of battle state vectors.

[0052] For example, at time point t, the input matrix M_t is composed of 40 vectors stacked together: [V_{t-39}, V_{t-38}, ..., V_{t-1}, V_t]. At the next time point t+1, the input matrix M_{t+1} slides forward one step, becoming [V_{t-38}, V_{t-37}, ..., V_t, V_{t+1}]. This sliding window mechanism ensures the continuity of the data flow, allowing the model to perform uninterrupted real-time analysis of the game.

[0053] The input matrix M_t is an n x d matrix, representing the final input to the model at time t, and thus a training sample corresponding to time t. This training sample can be viewed as a "short video," containing a complete history of the battle situation's changes over the past 20 seconds across multiple dimensions, including space, state, economy, and sentiment. This input matrix contains rich contextual information, forming the foundation for accurate predictions by the subsequent deep temporal model. Through this step, static feature vectors are transformed into dynamic temporal sequences, enabling the model to learn the deep-seated patterns of battle evolution.

[0054] 104. Locate the game events that occur in the game recording, and based on each game event, construct a sample label vector corresponding to each time point, which has the type and location of the event; In this embodiment, to address the problems of high cost, low efficiency, and strong subjectivity of traditional manual sample annotation methods, a weakly supervised or self-supervised mechanism is employed. First, all game events occurring in the game recording are located. Then, based on each game event, a sample label vector corresponding to each time point is constructed, indicating the event type and location. This achieves automated, low-cost, and large-scale label generation.

[0055] In an optional embodiment, step 104 above further includes: 1041. Obtain the game battle log corresponding to the game recording; The core idea of ​​this optional embodiment is to use reliable and easily accessible game logs as anchor points to deduce the time and location of key events in the game. Preferably, a pre-written automated scanning script is used to scan massive amounts of game recordings and their corresponding game logs.

[0056] 1042. Based on the game battle log, locate the time and location of all game events in the game recording; The game battle log records various game data, such as battle routes and the status of both sides. Therefore, various game events can be predefined based on the data in the game battle log. In an optional embodiment, the following event definition rules are preferably used to define game events: Roaming gank events: For example, "Within 8 seconds, 1-2 kills occur on a certain lane (top, mid, or bottom), and the killers include a jungler or a roaming hero." When scanning the game log, if this condition is met, the time point T_gank and location L_gank of a roaming gank event are recorded (the route is determined by the location of the kill).

[0057] Teamfight Events: For example, "Three or more kills occur within a small area (such as near the Dragon Pit) within 15 seconds." When scanning the game logs, if a series of consecutive kills are detected, it is marked as a teamfight event, and the time T_teamfight and location L_teamfight of the event are recorded.

[0058] Resource contest events: For example, "Two or more kills occurred near a neutral resource within 30 seconds before it was killed." When scanning the game log, the kill records within a 30-second window are checked, starting from the neutral resource's death time. If the above definition is met, it is marked as a resource contest event, and the event's time point T_objective and location L_objective are recorded.

[0059] Using the game event definition rules described above, all real game events that have actually occurred can be automatically and accurately located from the game recording, along with their time points and geographical locations accurate to the second.

[0060] 1043. Based on the occurrence time of each game event, set an event prediction time window located before the occurrence time for each game event; After locating game events in the game recording, a corresponding label vector Y_t needs to be generated for each time point t in the training data. To facilitate sample generation, for each located game event in the game recording, an event prediction time window is set based on the event's occurrence time, located before that time point.

[0061] For example, consider a top lane gank event occurring in T_gank. A prediction time window W is set, such as 5 to 15 seconds, representing 5 to 15 seconds before the top lane gank event. The significance of setting this prediction time window is that, within this window period, an ideal model should be able to predict the upcoming top lane gank event.

[0062] 1044. For each time point within the event prediction time window, generate a positive sample label vector corresponding to multiple event types and location pointers; For all time points t within the event prediction time window (i.e., T_gank - 15s <= t <= T_gank - 5s), they are marked as positive samples, and the corresponding label vector is Y_t.

[0063] The dimension of the label vector Y_t is exactly the same as the probability vector P_t output by the model. Assuming the game event is a gank event in the top lane, then for all Y_t within the prediction time window of this event, the dimension element corresponding to the gank event in the top lane has a value of 1, and all other dimensions have a value of 0.

[0064] Taking League of Legends as an example, suppose Y_t = [y1, y2, y3, y4, y5, y6, y7], where y1 represents the probability of a top lane gank, y2 represents the probability of a mid lane gank, y3 represents the probability of a bottom lane gank, y4 represents the probability of a river team fight, y5 represents the probability of a base team fight, y6 represents the probability of an elemental dragon contest, and y7 represents the probability of a Baron Nashor contest. If the identified game event is a top lane gank, then all Y_t values ​​within the corresponding prediction time window are Y_t = [1, 0, 0, 0, 0, 0, 0]. The label vector used in this embodiment not only tells the model that something is about to happen, but also precisely tells the model "what happened" and "where" it happens.

[0065] 1045. For each time point outside the event prediction time window, generate a negative sample label vector corresponding to multiple event types and location pointers.

[0066] For each time point in the game replay that is not covered by any positive event sample window, its corresponding label vector Y_t is a vector of all zeros, such as [0, 0, 0, 0, 0, 0, 0]. These vectors of all zeros constitute a massive amount of negative samples, which are used to teach the model to identify stable laning phases, development phases, and other game states where "nothing happens".

[0067] This embodiment assigns a precise label vector Y_t with spatial and type orientation to each time point t in the training data. The entire process is fully automated, requiring no manual intervention, thus realizing the transformation from unlabeled video recordings to massive, high-quality training data pairs.

[0068] 105. Based on the training samples and sample label vectors corresponding to each time point, the model is trained to obtain a trained game event prediction model.

[0069] This step utilizes the massive dataset (M_t, Y_t) constructed in the previous step to train the selected deep learning model.

[0070] Step 103 constructs the input matrix M_t for each time point t, and step 104 constructs the label vector for each time point. Then, the (M_t, Y_t) pairs for each time point t are combined into a complete training data pair. This automated process can generate hundreds of millions of training data pairs from tens of thousands of hours of game recordings, requiring no manual annotation.

[0071] After the model training data is constructed, model training can be performed. In an optional embodiment, step 105 above further includes: 1051. Input the training samples corresponding to each time point into the pre-trained model in sequence; The pre-trained model is the core of game event prediction, responsible for inferring the probability of future events from the input matrix. Pre-trained models preferably employ an attention-based deep temporal prediction network architecture (such as the Transformer). This architecture is particularly well-suited for processing long sequence data; its self-attention mechanism can capture the complex dependencies between any two time points within the sequence, perfectly meeting the needs of game battle analysis.

[0072] A pre-trained model based on the Transformer Encoder architecture includes the following components: Input layer: Used to receive an n x d dimensional input matrix M_t, with multiple time points forming a sequence of input matrices.

[0073] Positional encoding: used to add time dimension information to the input matrix sequence, enabling the model to understand the order of features.

[0074] Multi-head self-attention layer: Contains multiple stacked encoder layers. Each layer allows the model to dynamically compute the association weights between the battle state vector V_t at each time point t and the battle state vector V_k at all other times t in the input matrix sequence when processing the battle state vector V_t at each time point t. This enables the model to learn complex cross-time and cross-feature dependencies (e.g., associating "jungler disappearance" with subsequent "side lane hero health decrease").

[0075] Output layer: This layer performs global average pooling on the final output sequence of the Transformer, obtaining a vector representing the essence of the entire time series. This vector is then fed into one or more fully connected layers for non-linear transformation. Finally, the output layer passes through k neurons, with the sigmoid activation function applied independently to each neuron. The value of k is the same as the number of game event types to be predicted. For example, if there are 7 types of game events to predict, then k should be 7.

[0076] 1052. Predict game events using the pre-trained model and output an outbreak probability vector containing the type and location of all future game events; In this optional embodiment, the final model outputs a k-dimensional probability vector P_t. Each element P_t[i] is a probability value between (0, 1), corresponding to a specific game event prediction task. For example, a 7-dimensional output vector P_t might look like this: P_t = [ P_gank_top, P_gank_mid, P_gank_bot, P_teamfight_river, P_teamfight_base, P_objective_dragon, P_objective_baron ] For example, taking a 15-second event prediction time window as an example: P_gank_*: represents the probability of a gank occurring on the corresponding route within the next 15 seconds. P_teamfight_*: represents the probability of a large-scale team fight breaking out in the corresponding area within the next 15 seconds. P_objective_*: represents the probability of both sides fighting for the corresponding neutral resource within the next 15 seconds.

[0077] This embodiment can accurately and quantitatively describe the probability of various future game events by using the output vector P_t.

[0078] 1053. Construct a target loss function based on the sample label vector and the outbreak probability vector corresponding to each time point; 1054. Based on the target loss function, update the model parameters of the pre-trained model to obtain a trained game event prediction model.

[0079] For example, in a 30-minute game, the actual time spent ganking, resource contention, or team fights might only be a few minutes. That is, in all training data pairs (M_t, Y_t), positive samples with label vectors Y_t not all zero are very few, their proportion being far smaller than that of negative samples. Therefore, to address the problem of extreme imbalance between positive and negative samples, this embodiment preferably uses a weighted binary cross-entropy loss function as the target loss function for model training.

[0080] Binary cross-entropy is a commonly used loss function in multi-label classification tasks, used to measure the difference between the model's predicted probability and the true label. Weighting is key to solving the imbalanced sample problem. That is, instead of treating every prediction error equally, different "penalty weights" are assigned to different types of errors.

[0081] Higher weights are applied to the loss for positive samples (labeled 1). For example, if the true situation is "an impending team fight" (an element in Y_t is 1), but the model predicts "nothing happening" (the corresponding element in P_t is close to 0), this is a serious false negative. The system will multiply the incorrectly calculated loss value by a large weight (e.g., 50 or 100). Lower weights are applied to the loss for negative samples (labeled 0). For example, if the true situation is "nothing happening," but the model predicts "a fight is about to break out," this is a false positive. The system will multiply the incorrect loss value by a small weight (e.g., 1).

[0082] The model training process uses an optimizer that iteratively adjusts the network weights through backpropagation to minimize the weighted total loss. Under the intense pressure of severely penalizing missed detections, the model learns to focus on subtle signals that foreshadow impending battles, ultimately resulting in a well-trained deep learning model capable of accurately predicting future key events—a game event prediction model.

[0083] This embodiment uses data abstraction to map the game world from a high-dimensional sensory space to a low-dimensional tactical space, fundamentally avoiding the high resource consumption of analyzing and identifying key events. Simultaneously, offline / online separation solves efficiency bottlenecks, and temporal modeling enables the model to predict the future from trends, thus achieving efficient pre-emptive warnings. Multi-dimensional information fusion provides rich judgment criteria, and precise supervised learning allows the model to directly grasp the deep logic from battle trends to event outcomes, thereby significantly improving prediction accuracy.

[0084] See Figure 2 One embodiment of the game event prediction method of the present invention includes: 201. Receive the game video to be predicted, wherein the game video includes live game video or game recording; In this embodiment, the game recording refers to the recording of a live game video, and the live game video supports bullet comments. Event prediction can be performed on either the game recording or the live game video.

[0085] 202. Extract multiple battle state data corresponding to each time slice in the game video, and concatenate the vectorized battle state data into a battle state vector. In this embodiment, battle status data refers to data used to describe the state of the game battle. Battle status data changes dynamically in real time. Therefore, battle status data needs to be extracted for each time slice of the game video, including high-frequency structured data extraction for the video stream and low-frequency unstructured data extraction for the bullet screen text stream.

[0086] In this selected embodiment, the extracted high-frequency structured data is vectorized using normalization. For example, in each time slice t (e.g., within 1 second), for the hero's position on the minimap, the x and y coordinates of the hero on the minimap are detected, and then the coordinates are normalized by dividing them by the length and width of the minimap, respectively, to obtain the quantified hero position on the minimap. The hero's health and mana are vectorized as percentages of maximum health and mana, with 1 representing the ultimate ability being ready and 0 representing it not being ready. The hero's level and current economy are vectorized as percentages of the highest level and maximum economy, respectively. Game time, friendly kills, enemy kills, countdown to the next neutral creep spawn, and the economic difference between the two sides are also vectorized using normalization.

[0087] For the extracted low-frequency unstructured data, aggregated weighted vectorization is used. In each time slice t (e.g., within 0.5 seconds), viewers send a large number of bullet screen texts. Therefore, each bullet screen text is first classified, and then the classification results of all bullet screens within that time slice are aggregated. The aggregation method can be simple averaging or weighted averaging. Finally, a bullet screen feature vector is output, such as [excitement percentage, disappointment percentage, tension percentage, offensive intent percentage, defensive intent percentage]. For example, the bullet screen feature vector for a certain time slice might be [0.65, 0.10, 0.20, 0.55, 0.15]. This vector clearly indicates that the current viewer's emotions are mainly "excitement" and "offensive intent," a strong signal that a battle is imminent. To facilitate subsequent calculations, the feature vectors extracted for each time slice need to be concatenated in a fixed order to form a battle state vector.

[0088] 203. Based on the battle state vectors of multiple consecutive time slices, construct the battle state vector sequence corresponding to each time point; In this embodiment, the training samples corresponding to time point t are selected from the battle state vectors of the n time slices closest to the current time point t, and stacked in chronological order to form an input matrix M_t of shape nxd, where n represents the number of time slices and d represents the dimension of the battle state vector corresponding to a single time slice.

[0089] This embodiment preferably employs a sliding window mechanism to construct a sequence of battle state vectors. For example, at time point t, the input matrix M_t is composed of 40 vectors stacked together: [V_{t-39}, V_{t-38}, ..., V_{t-1}, V_t]. At the next time point t+1, the input matrix M_{t+1} slides forward one step, becoming [V_{t-38}, V_{t-37}, ..., V_t, V_{t+1}]. This sliding window mechanism ensures the continuity of the data flow, enabling the model to perform uninterrupted real-time analysis of the game.

[0090] 204. Input the sequence of battle state vectors into a pre-trained game event prediction model for processing, and output a probability vector containing the type and location of future game events. The game event prediction model is obtained based on the game event prediction model training method.

[0091] Suppose a game event prediction model outputs a 7-dimensional probability vector P_t. Each element P_t[i] is a probability value between (0, 1), corresponding to a specific game event prediction task. For example, a 7-dimensional output vector P_t might look like this: P_t = [ P_gank_top, P_gank_mid, P_gank_bot, P_teamfight_river, P_teamfight_base, P_objective_dragon, P_objective_baron ] For example, using a 15-second event prediction time window: P_gank_top=0.8: indicates an 80% probability of a gank occurring in the top lane within the next 15 seconds. P_teamfight__river=0.9: indicates a 90% probability of a large-scale team fight breaking out in the river area within the next 15 seconds. P_objective_dragon=0.1: indicates a 10% probability of both sides fighting for the corresponding element, the dragon, within the next 15 seconds.

[0092] This embodiment transforms the complex problem of multiplayer online game analysis into an efficient, computable, time-series-based supervised learning problem by employing three core technologies: data abstraction, analysis temporalization, and supervised training. Ultimately, it achieves low resource consumption, high efficiency, and high accuracy in predicting game events.

[0093] Please see Figure 3 One embodiment of the game event prediction device in this invention includes: The receiving module 301 is used to receive the game video to be predicted, the game video including live game video or game recording; The second extraction module 302 is used to extract multiple battle state data corresponding to each time slice in the game video, and to concatenate the vectorized battle state data into a battle state vector. Input construction module 303 is used to construct a sequence of battle state vectors corresponding to each time point based on battle state vectors of multiple consecutive time slices; The prediction module 304 is used to input the sequence of battle state vectors into the game event prediction model for processing, and output the outbreak probability vector containing the type and location of various future game events, wherein the game event prediction model is obtained based on the game event prediction model training method described above.

[0094] Optionally, in one embodiment, the battle situation data includes high-frequency structured data and low-frequency unstructured data, and the extraction module 304 is specifically used for: Within each time slice, the high-frequency structured data and the low-frequency unstructured data are extracted from the game screen and bullet screen text of the game video, respectively. The high-frequency structured data and the low-frequency unstructured data are vectorized respectively to obtain the first extracted feature vector and the second extracted feature vector. The first extracted feature vector and the second extracted feature vector corresponding to each time slice are concatenated in a fixed order to obtain the battle state vector.

[0095] Optionally, in one embodiment, the extraction module 304 is further configured to: Within each time slice, the high-frequency structured data is extracted from the game screen of the game video using a trained interface parsing model. The high-frequency structured data includes one or more of the following: the position of the game character, key states and core attributes, and the global state of the game. Within each time slice, the bullet screen text of the game video is classified using a pre-trained bullet screen text classification model to obtain the low-frequency unstructured data, which includes the probabilities of various emotional intent categories.

[0096] Optionally, in one embodiment, the game event prediction device further includes: Training module 305 is used to acquire game recordings of multi-player game battles; extract multiple battle state data corresponding to each time slice in the game recordings, and concatenate the vectorized battle state data into a battle state vector; construct a battle state vector sequence corresponding to each time point based on the battle state vectors of multiple consecutive time slices, and use each battle state vector sequence as a training sample; locate game events occurring in the game recordings, and construct a sample label vector corresponding to each time point with event type and location based on each game event; train the model based on the training samples and sample label vectors corresponding to each time point to obtain a trained game event prediction model.

[0097] Optionally, in one embodiment, the training module 305 is further configured to: Obtain the game battle logs corresponding to the game recordings; Based on the game battle log, the time and location of all game events are located in the game recording; Based on the occurrence time of each game event, an event prediction time window is set for each game event before the occurrence time. For each time point within the event prediction time window, generate a positive sample label vector corresponding to multiple event types and location pointers; For each time point outside the event prediction time window, a negative sample label vector corresponding to multiple event types and location pointers is generated.

[0098] Optionally, in one embodiment, the training module 305 is further configured to: The training samples corresponding to each time point are sequentially input into the pre-trained model; The pre-trained model is used to predict game events, and the output is a probability vector containing the type and location of various game events in the future. Based on the sample label vector and the outbreak probability vector corresponding to each time point, a target loss function is constructed; Based on the target loss function, the model parameters of the pre-trained model are updated to obtain a trained game event prediction model.

[0099] Since the embodiments of the device part correspond to the embodiments of the above method, please refer to the above method embodiments for the introduction of the game event prediction device provided by the present invention. The present invention will not be described again here, but it has the same beneficial effects as the above game event prediction method.

[0100] Figure 4This is a schematic diagram of the structure of a computer device 500 provided in an embodiment of the present invention. The computer device 500 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 510 (e.g., one or more processors) and a memory 520, and one or more storage media 530 (e.g., one or more mass storage devices) for storing application programs 533 or data 532. The memory 520 and storage media 530 can be temporary or persistent storage. The program stored in the storage media 530 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the computer device 500. Furthermore, the processor 510 may be configured to communicate with the storage media 530 and execute the series of instruction operations in the storage media 530 on the computer device 500.

[0101] Computer device 500 may also include one or more power supplies 540, one or more wired or wireless network interfaces 550, one or more input / output interfaces 560, and / or one or more operating systems 531, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 4 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0102] The present invention also provides a computer device, the computer device including a memory and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the processor performs the steps of the game event prediction method described in the above embodiments. The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions, and when the instructions are executed on a computer, the computer performs the steps of the game event prediction method described in the above embodiments.

[0103] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0104] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0105] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting game events, characterized in that, The game event prediction method includes: Receive game videos to be predicted, including live game videos or game recordings; Extract multiple battle state data corresponding to each time slice in the game video, and concatenate the vectorized battle state data into a battle state vector. Based on the battle state vectors of multiple consecutive time slices, construct the battle state vector sequence corresponding to each time point; The sequence of battle state vectors is input into a pre-trained game event prediction model for processing, and the output is a probability vector containing the type and location of future game events.

2. The game event prediction method according to claim 1, characterized in that, The battle situation data includes high-frequency structured data and low-frequency unstructured data. The step of extracting multiple battle situation data points corresponding to each time slice in the game video and concatenating the vectorized battle situation data points into a battle situation vector includes: Within each time slice, the high-frequency structured data and the low-frequency unstructured data are extracted from the game screen and bullet screen text of the game video, respectively. The high-frequency structured data and the low-frequency unstructured data are vectorized respectively to obtain the first extracted feature vector and the second extracted feature vector. The first extracted feature vector and the second extracted feature vector corresponding to each time slice are concatenated in a fixed order to obtain the battle state vector.

3. The game event prediction method according to claim 2, characterized in that, The step of extracting the high-frequency structured data and the low-frequency unstructured data from the game screen and bullet screen text of the game video within each time slice includes: Within each time slice, the high-frequency structured data is extracted from the game screen of the game video using a trained interface parsing model. The high-frequency structured data includes one or more of the following: the position of the game character, key states and core attributes, and the global state of the game. Within each time slice, the bullet screen text of the game video is classified using a pre-trained bullet screen text classification model to obtain the low-frequency unstructured data, which includes the probabilities of various emotional intent categories.

4. The game event prediction method according to any one of claims 1-3, characterized in that, The training methods used in the game event prediction model include: Acquire game recordings of multiplayer battles; Extract multiple battle state data corresponding to each time slice in the game recording, and concatenate the vectorized battle state data into a battle state vector. Based on the battle state vectors of multiple consecutive time slices, a battle state vector sequence corresponding to each time point is constructed, and each of the battle state vector sequences is used as training samples. Locate the game events that occur in the game recording, and based on each game event, construct a sample label vector corresponding to each time point, which has the type of event and the location of the event. The model is trained based on the training samples and sample label vectors corresponding to each time point to obtain a trained game event prediction model.

5. The game event prediction method according to claim 4, characterized in that, The process of locating game events occurring in the game recording and constructing a sample label vector corresponding to each time point, with the event type and location pointer, based on each game event includes: Obtain the game battle logs corresponding to the game recordings; Based on the game battle log, the time and location of all game events are located in the game recording; Based on the occurrence time of each game event, an event prediction time window is set for each game event before the occurrence time. For each time point within the event prediction time window, generate a positive sample label vector corresponding to multiple event types and location pointers; For each time point outside the event prediction time window, a negative sample label vector corresponding to multiple event types and location pointers is generated.

6. The game event prediction method according to claim 4, characterized in that, The process of training the model based on the training samples and sample label vectors corresponding to each time point to obtain the trained game event prediction model includes: The training samples corresponding to each time point are sequentially input into the pre-trained model; The pre-trained model is used to predict game events, and the output is a probability vector containing the type and location of various game events in the future. Based on the sample label vector and the outbreak probability vector corresponding to each time point, a target loss function is constructed; Based on the target loss function, the model parameters of the pre-trained model are updated to obtain a trained game event prediction model.

7. A game event prediction device, characterized in that, The game event prediction device includes: A receiving module is used to receive game videos to be predicted, including live game videos or game recordings. The extraction module is used to extract multiple battle state data corresponding to each time slice in the game video, and to concatenate the vectorized battle state data into a battle state vector. The input building module is used to construct a sequence of battle state vectors for each time point based on battle state vectors from multiple consecutive time slices. The prediction module is used to input the sequence of battle state vectors into a pre-trained game event prediction model for processing, and output a probability vector containing the type and location of future game events.

8. The game event prediction device according to claim 7, characterized in that, The game event prediction device also includes: The training module is used to acquire game recordings of multi-player game battles; extract multiple battle state data corresponding to each time slice in the game recordings, and concatenate the vectorized battle state data into a battle state vector; based on the battle state vectors of multiple consecutive time slices, construct a battle state vector sequence corresponding to each time point, and use each battle state vector sequence as a training sample; locate game events occurring in the game recordings, and based on each game event, construct a sample label vector corresponding to each time point with the event type and location; train the model based on the training samples and sample label vectors corresponding to each time point to obtain a trained game event prediction model.

9. A computer device, characterized in that, The computer device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the computer device to execute the game event prediction method as described in any one of claims 1-6.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements the game event prediction method as described in any one of claims 1-6.

Images