Social robot recognition method based on multi-scale entropy feature, terminal and medium

Abstract

The application relates to the technical field of information processing, and discloses a social robot identification method based on multi-scale entropy features, a terminal and a medium. The method acquires social network behavior data of a user to be detected, constructs initial binary time series for each predefined behavior type, sets a plurality of time windows with different granularities, maps the initial binary time series into coarse-grained behavior sequences with different time scales, calculates the behavior information entropy of each coarse-grained behavior sequence with a time scale to form a behavior entropy set, calculates the variation coefficient of the entropy values between different time scales based on the behavior entropy set to quantize the dynamic fluctuation features of the user behavior, fuses the behavior entropy set and the variation coefficient to construct a comprehensive feature vector, inputs the comprehensive feature vector into a trained supervised classification model, and outputs the classification result of the user as a real user or a social robot. The application improves the identification ability for the complex camouflage behavior of social network robots.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, specifically to a social robot identification method, terminal, and medium based on multi-scale entropy features. Background Technology

[0002] With the rapid development of social media and content-sharing platforms, the way and scale of information dissemination have undergone profound changes. Online social networks have become an important channel for the spread of ideas and opinions, and have had a far-reaching impact on billions of users worldwide. Social bots refer to algorithmic agents deployed on social media platforms that can automatically perform specific behaviors based on preset human goals. Researchers have proposed various detection frameworks aimed at accurately identifying social bots from real users.

[0003] Existing social bot detection methods can be broadly categorized into three main types: content-based, network structure-based, and behavior-based. Content-based methods primarily rely on user-generated text or multimedia information. However, with the rapid improvement in the content generation capabilities of large language models, these methods are finding it increasingly difficult to effectively identify bot-generated information, and their discriminative ability is constantly weakening. Network-based methods mainly utilize the connection structure between users and social topology features for identification; however, due to the highly dynamic and massive scale of social networks, these methods face significant challenges in terms of efficiency and scalability. In contrast, behavior-based methods have received increasing attention in recent years. These methods analyze the temporal distribution and interaction patterns of user behavior sequences, exhibiting strong interpretability and adaptability, and can achieve more robust detection results without relying on content or network structure.

[0004] However, current mainstream behavioral modeling methods generally rely on manually predefined statistical features or fixed-granularity time windows. This static approach is not only highly dependent on specific scenarios but also almost completely ineffective in environments with complex user behavior structures or where deception exists. Especially on social media platforms, attackers constantly evolve their behavioral patterns, making traditional methods prone to "feature collapse," causing models to fail to identify highly deceptive social bots. Furthermore, existing methods typically model behavior on a single time scale, failing to capture the dynamic changes between short-term bursts and long-term patterns. This coarse-grained representation ignores the "multi-layered structure" of behavioral evolution, causing a sharp decline in classification performance when faced with accounts exhibiting irregular rhythms or mixed behaviors, becoming a major performance bottleneck for detection systems. Meanwhile, although some methods incorporate deep learning frameworks, their internal features lack transparency, making the output difficult for humans to understand or trace. This "uninterpretability" is particularly fatal in security-sensitive scenarios; errors not only make troubleshooting difficult but can also lead to serious misjudgments. As social bot attacks become increasingly complex, single behavioral feature construction strategies are no longer sufficient to meet practical detection needs. Existing methods generally lack a unified modeling mechanism for different behavior types (such as posting, forwarding, and commenting), resulting in poor model generalization ability and being almost powerless to detect bots across platforms and contexts. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a social robot identification method, terminal, and medium based on multi-scale entropy features. By combining independent modeling of multiple behavior types with temporal fluctuation analysis, it significantly improves the ability to identify complex camouflage behaviors of social network robots.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] This invention discloses a social robot recognition method based on multi-scale entropy features, comprising:

[0008] S1. Obtain the social network behavior data of the user to be detected during their life cycle, and construct initial binary time series for each predefined behavior type;

[0009] S2. Set multiple time windows with different granularities to map the initial binary time series into multiple coarse-grained behavior sequences with different time scales;

[0010] S3. For each coarse-grained behavior sequence at each time scale, calculate its corresponding behavior information entropy to form a set of behavior entropies;

[0011] S4. Based on the set of behavioral entropies, calculate the coefficient of variation of the entropy values ​​at different time scales to quantify the dynamic fluctuation characteristics of user behavior;

[0012] S5. The behavioral entropy set is fused with the coefficient of variation to construct a comprehensive feature vector. The comprehensive feature vector is input into the trained supervised classification model, and the classification result of whether the user is a real user or a social robot is output.

[0013] As a further improvement to the above scheme, in step S1, the predefined behavior types include posting, forwarding, and replying; the initial binary time series is constructed as follows: if a user performs a specified type of behavior at a specific time point, the value of that time point is recorded as 1, otherwise it is recorded as 0; the lifecycle is defined as the time interval from the start of user account registration to the last recorded behavior timestamp.

[0014] As a further improvement to the above scheme, in step S2, the multiple time windows of different granularities are non-overlapping time windows; the calculation formula for mapping the initial binary time series into multiple coarse-grained behavioral sequences of different time scales is as follows:

[0015] ;

[0016] In the formula, Indicates the first k Coarse-grained behavioral states within a time window; Time scale; For the first k The starting reference point for each time window; Given an initial binary time series, Indicates behavior In time occur; , A global set of behaviors for each user.

[0017] As a further improvement to the above scheme, in step S3, the formula for calculating the behavioral information entropy is:

[0018] ;

[0019] In the formula, Indicates behavior In time scale Information entropy below; Represents a binary state in a coarse-grained behavioral sequence. The empirical probability of occurrence;

[0020] The calculated behavioral information entropy is normalized to limit its value range to [0,1]; by traversing... K Different time scales , construct behavior exist K A set of entropy values ​​at different time scales That is, the set of behavioral entropy.

[0021] As a further improvement to the above scheme, in step S4, the formula for calculating the coefficient of variation is:

[0022] ;

[0023] In the formula, For behavior The corresponding coefficient of variation; For set standard deviation For set The mean.

[0024] As a further improvement to the above scheme, the specific process of constructing the comprehensive feature vector in step S5 is as follows:

[0025] Regarding posting ,Forward and reply Three behavior types were identified, and their behavioral information entropy at different time scales was extracted to construct a user profile. u Multi-scale information entropy feature matrix :

[0026] ;

[0027] In the formula, , and These represent the three behavior types: posting, forwarding, and replying. m Time scale The behavioral information entropy, ;

[0028] Get Posting ,Forward and reply Coefficients of variation for the three behaviors , and , build users u eigenvectors of coefficient of variation :

[0029] ;

[0030] Multi-scale information entropy feature matrix With coefficient of variation eigenvector The features are then combined and aggregated to form a comprehensive feature vector that represents the user's dynamic behavior patterns.

[0031] As a further improvement to the above scheme, in step S5, the supervised classification model is a random forest classifier; before inputting the comprehensive feature vector into the supervised classification model, the historical user data is also processed using downsampling technology to balance the distribution of category data between social robots and real users.

[0032] As a further improvement to the above scheme, in steps S1 to S4, each behavior type does not overlap in the process of constructing the initial binary time series, performing multi-scale time series mapping, extracting multi-scale behavior information entropy, and calculating the coefficient of variation, so as to preserve the inherent temporal evolution characteristics of a single behavior type; until step S5, the independently extracted behavior entropy sets and the coefficient of variation are concatenated to achieve feature-level fusion of multiple behavior dimensions.

[0033] The present invention also discloses a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the social robot recognition method based on multi-scale entropy features as described above.

[0034] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the social robot recognition method based on multi-scale entropy features as described above.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] 1. The social robot recognition method based on multi-scale entropy features disclosed in this invention can effectively capture the temporal consistency and diversity differences of user behavior by constructing multi-scale time series and calculating the entropy value of behavior series, significantly improving the accuracy and robustness of social robot recognition, outperforming traditional behavior feature-based methods, and showing excellent performance on multiple public datasets.

[0037] 2. The entropy-based framework proposed in this invention can quantify the dynamic changes in user behavior, reveal the inherent differences in behavioral patterns between real users and social robots, provide theoretical support and interpretability indicators for behavior modeling, and break through the limitations of existing single time scales or static features.

[0038] 3. This invention relies solely on behavioral entropy and its derived dynamic features, reducing the dependence of complex models on large amounts of training data and computing resources. It possesses strong generalization ability and applicability, making it easy to deploy and apply in large-scale social network environments.

[0039] 4. This invention comprehensively analyzes various behavioral categories such as posting, forwarding, and commenting, as well as their contributions at different time scales, to determine short-term posting dynamics as the core identification signal, with forwarding and commenting serving as auxiliary information, effectively improving detection results and system stability. Attached Figure Description

[0040] Figure 1 This is a flowchart of the social robot recognition method based on multi-scale entropy features in Embodiment 1 of the present invention.

[0041] Figure 2 This is a schematic diagram of the data preprocessing process in Embodiment 1 of the present invention.

[0042] Figure 3 This is a schematic diagram of the feature extraction process in Embodiment 1 of the present invention.

[0043] Figure 4 This is a schematic diagram of the classification process in Embodiment 1 of the present invention.

[0044] Figure 5 This is a schematic diagram of the structure of the computer terminal in Embodiment 2 of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Example 1

[0047] Please see Figures 1 to 4 This embodiment provides a social robot identification method based on multi-scale entropy features, including steps S1 to S5.

[0048] S1. Obtain social network behavior data of the user to be detected during their life cycle, and construct initial binary time series for each predefined behavior type.

[0049] In this embodiment, the predefined behavior types include posting (P). t (posting), forwarding R t (retweeting) and reply R p (replying); The initial binary time series is constructed as follows: if a user performs a specified type of behavior at a specific time point, the value of that time point is recorded as 1, otherwise it is recorded as 0; The lifecycle is defined as the time interval from the start of user account registration to the last recorded behavior timestamp.

[0050] To effectively capture the temporal dynamics of user behavior and use it for social bot detection, this embodiment targets three main interaction behavior types: posting, forwarding, and replying. Each behavior type is independently modeled to retain unique temporal characteristics, which are then used to build the detection model. Formally, users... u The original behavioral sequence is represented as And defined as:

[0051] ;

[0052] in, , and These represent the timestamp binary sequences corresponding to the posting, forwarding, and replying behavior types, respectively.

[0053] In existing social bot identification technologies, content-based detection methods are increasingly struggling to handle the anthropomorphic and highly deceptive text generated by large language models. To overcome this deficiency, this embodiment abandons the reliance on the semantic content of the text and focuses only on the time-point feature of whether an action occurs, abstracting the action into a time signal. Posting, forwarding, and replying cover the three most core and representative interaction patterns in online social networks. Using binary sequences (0 and 1) to record whether an action occurs at a specific time point effectively transforms the user's complex lifecycle trajectory into a highly structured time series model, providing a data foundation for subsequently discovering fine-grained temporal pattern differences between social bots and real users.

[0054] S2. Set multiple time windows with different granularities to map the initial binary time series into multiple coarse-grained behavior sequences with different time scales.

[0055] Multiple time windows of different granularities are non-overlapping; the calculation formula for mapping the initial binary time series into multiple coarse-grained behavioral sequences of different time scales is as follows:

[0056] ;

[0057] In the formula, Indicates the first k Coarse-grained behavioral states within a time window; Time scale; For the first k The starting reference point for each time window; Given an initial binary time series, Indicates behavior In time occur; , A global set of behaviors for each user.

[0058] In social networks, user behavior dynamically changes between short-term bursts (such as continuous posting on a trending event) and long-term patterns (such as daily active periods). A single time scale often cannot fully capture this "multi-layered structure" characteristic. Therefore, this embodiment preferably sets four types ( K 4) Representative time window scales: 1 hour (1H), 1 day (1D), 1 week (1W), and 1 month (1M). These scales are highly consistent with the typical rhythm of human activities. The mapping formula employs a "binary state aggregation" strategy, meaning that if the behavior occurs within a given time window (regardless of the frequency), the coarse-grained state of that window is 1. It should be understood that this approach intentionally ignores the absolute frequency of the behavior, effectively filtering out the large amount of noise interference generated by social bots sending spam messages at high frequencies within a short period, allowing the model to focus on extracting the underlying rhythmic features of behavior.

[0059] The resulting coarse-grained behavior sequence Represented as:

[0060] ;

[0061] in, , and These are coarse-grained behavioral sequences corresponding to posting, forwarding, and replying. The selection of these behavioral categories and time resolutions is intended to align with typical human activity rhythms, enabling the analysis of short-term fluctuations and long-term behavioral trends.

[0062] S3. For each coarse-grained behavior sequence at each time scale, calculate its corresponding behavior information entropy to form a set of behavior entropies.

[0063] The formula for calculating behavioral information entropy is:

[0064] ;

[0065] In the formula, Indicates behavior In time scale Information entropy below; Represents a binary state in a coarse-grained behavioral sequence. The empirical probability of occurrence;

[0066] The calculated behavioral information entropy is normalized to limit its value range to [0,1]; by traversing... K Different time scales , construct behavior exist KA set of entropy values ​​at different time scales That is, the set of behavioral entropy.

[0067] Entropy is used to measure the uncertainty of stochastic systems. In this embodiment, information entropy provides a principled quantitative method for quantifying the complexity, regularity, and unpredictability of user behavior sequences. Since the absolute frequency distribution of different types of behavior (such as posting and forwarding) in users' daily activities inherently differs, normalization is crucial. A normalized entropy value of 0 corresponds to behavior that is completely deterministic or extremely rigid at a given time scale (a characteristic commonly seen in pre-programmed social bots); while a value of 1 indicates the greatest uncertainty, reflecting highly random or extremely spontaneous human behavior patterns. In this way, the smooth and bounded entropy distribution characteristics of real users form a clear contrast with the restricted and rigid entropy patterns of social bots.

[0068] S4. Based on the set of behavioral entropies, calculate the coefficient of variation of the entropy values ​​at different time scales to quantify the dynamic fluctuation characteristics of user behavior.

[0069] The formula for calculating the coefficient of variation is:

[0070] ;

[0071] In the formula, For behavior The corresponding coefficient of variation; For set standard deviation For set The mean.

[0072] A single time scale merely reflects local behavioral characteristics. To characterize the dynamic evolution of behavior over time, this embodiment introduces the coefficient of variation (CV). The coefficient of variation provides a scale-invariant measure of the dispersion of information entropy across time scales. A higher coefficient of variation reflects greater variability in the temporal irregularity of the corresponding behavior, indicating that its activity pattern is more dynamic and fluctuating over time. This can effectively identify complex machine accounts that appear well-disguised at fine-grained time scales but reveal automation flaws (such as a lack of long-term adaptability) at coarse-grained time scales.

[0073] S5. The behavioral entropy set is fused with the coefficient of variation to construct a comprehensive feature vector. The comprehensive feature vector is input into the trained supervised classification model, and the classification result of whether the user is a real user or a social robot is output.

[0074] Based on multi-scale behavior sequences, this embodiment quantifies the inherent randomness of each behavior type by calculating the behavioral information entropy at each time scale. For each behavior and its corresponding time scale, an entropy value is calculated and organized into a feature vector. By aggregating these vectors across all behavior types and time scales, a feature vector is generated for the user. u Constructing a multi-scale information entropy feature matrix :

[0075] ;

[0076] In the formula, , and These represent the behavioral information entropy of posting, forwarding, and replying over one hour, respectively. The definitions of the remaining parameters follow the same pattern as 1 day (1D), 1 week (1W), and 1 month (1M), and will not be elaborated further.

[0077] Get Posting ,Forward and reply Coefficients of variation for the three behaviors , and , build users u eigenvectors of coefficient of variation :

[0078] ;

[0079] Multi-scale information entropy feature matrix With coefficient of variation eigenvector The features are concatenated and aggregated to form a comprehensive feature vector that represents the user's dynamic behavior patterns. This comprehensive representation is then used as input to a supervised classification model to detect social bots.

[0080] In this embodiment, the supervised classification model can be a random forest classifier. Before inputting the comprehensive feature vector into the supervised classification model, the historical user data is processed using downsampling techniques to balance the distribution of category data between social bots and real users. Random forest classifiers are preferred because they exhibit strong robustness and excellent interpretability when processing structured features. Downsampling techniques are introduced to mitigate the model evaluation bias caused by the extreme imbalance in the number of samples from normal users and various types of bots in real-world social networks.

[0081] In steps S1 to S4, each behavior type does not overlap in the process of constructing the initial binary time series, performing multi-scale time series mapping, extracting multi-scale behavior information entropy, and calculating the coefficient of variation, so as to preserve the inherent temporal evolution characteristics of a single behavior type; until step S5, the independently extracted behavior entropy sets and the coefficient of variation are concatenated to achieve feature-level fusion of multiple behavior dimensions.

[0082] To evaluate the effectiveness of the method proposed in this invention, the following experiments are provided in this embodiment:

[0083] 1. Dataset Description

[0084] This example uses the Cresci-2017 dataset, which contains four different types of Twitter accounts (now X): real users, app promotion bots, product promotion bots, and traditional bots. A brief description of each account category is provided below:

[0085] Real users: This category includes real platform accounts that are randomly selected and verified through a hybrid crowdsourcing method.

[0086] App promotion bots: This category includes commercial promotion bots associated with a particular app, which primarily share positive content and occasionally mention users to promote their VIP version.

[0087] Product promotion bots: This category includes automated accounts that advertise for shopping platforms by embedding promotional links for products on the platform within seemingly ordinary posts.

[0088] Traditional bots: This category contains some benchmark spam bots that were originally created to train and evaluate machine learning models for traditional bot detection.

[0089] Table 1: Statistics of different account types in the Cresci-2017 dataset

[0090]

[0091] As shown in Table 1, the Cresci-2017 dataset contains comprehensive statistics for each account type, including the number of users, posts, replies, and retweets. This information reflects the different activity patterns between real users and different types of bots.

[0092] 2. Evaluation Indicators

[0093] To evaluate the performance of the classification model, this embodiment uses the following metrics:

[0094] Precision: Measures the proportion of correctly predicted positive instances out of all instances predicted as positive. It is defined as follows:

[0095] ;

[0096] Where TP represents the number of true positives and FP represents the number of false positives.

[0097] Recall: Measures the proportion of all samples that were actually positive that were correctly predicted as positive. It is defined as follows:

[0098] ;

[0099] FN represents the number of false positives.

[0100] F1 score: The harmonic mean of precision and recall, providing a balanced single metric. Its calculation formula is:

[0101] ;

[0102] AUC (Area Under the ROC Curve): This curve plots the relationship between the true positive rate (TPR) and the false positive rate (FPR) at different thresholds. The definitions of TPR and FPR are:

[0103] ;

[0104] ;

[0105] Here, TN represents the true negative number. It measures the model's ability to distinguish between different classes, with 0.5 indicating random guessing and 1.0 indicating perfect classification.

[0106] These evaluation metrics collectively provide a comprehensive assessment of the classification model's performance. Precision and recall focus on the accuracy of positive class predictions, while the F1 score combines these two aspects into a single metric. The AUC metric provides a measure of the model's overall discriminative ability across different decision thresholds.

[0107] 3. Benchmark Method

[0108] The method proposed in this invention is compared with a set of representative baseline methods that rely solely on account behavior information. These baseline methods are mainly divided into two categories:

[0109] (1) Behavioral pattern-based methods, which identify salient features between real users and bots by analyzing time-series patterns of user behavior, including:

[0110] 1) Seasonality: These features capture periodic behavioral patterns (e.g., daily or weekly cycles) and are designed to distinguish regular human activities from overly regular robotic activities.

[0111] 2) Shapelet-based representation: Shapelets are discriminative subsequences in a time series, capable of capturing unique patterns. Shapelets are extracted from labeled sequences by using the tslean library to identify segments with the highest information gain.

[0112] 3) Features based on temporal variance: These features are used to measure the stability of behavior. The temporal variance of the posting interval (Vati) captures temporal regularity, while the activity variance within each time window (Vatw) reflects the stability of breeding. The ratio of Vati to Vatw (Titw) characterizes sporadic occurrence, indicating that irregular time intervals coexist with stable activity levels.

[0113] (2) Entropy-based methods, which utilize entropy measures to quantify the complexity and predictability of behavioral signals, specifically include:

[0114] 1) Time Interval Entropy: This metric applies Shannon entropy to the posting interval to quantify the randomness of user activity time. The lower the entropy value, the stronger the regularity of user activity time, which is usually a characteristic of bot behavior.

[0115] 2) Sample entropy: Used to measure the irregularity of a time series. A user's timeline is divided into fixed-length intervals, and the number of tweets in each interval forms a sequence. The sample entropy value reflects the spontaneity of the sequence.

[0116] These baseline methods collectively cover a wide range of temporal behavior modeling strategies and can serve as a strong reference for evaluating the performance and generalization ability of the methods of this invention.

[0117] 4. Evaluate single-feature temporal entropy to detect robots

[0118] To rigorously evaluate the discriminative power of entropy-based features, we conducted single-feature classification experiments. The entropy values ​​for posting, forwarding, and replying behaviors were calculated separately. Furthermore, we introduced the coefficient of variation of entropy to quantify temporal fluctuations. Due to its robustness and interpretability, this embodiment employs a random forest classifier. To mitigate class imbalance, downsampling techniques were applied to ensure fair evaluation across all robot categories.

[0119] This embodiment uses an application promotion robot as an example. The detection results are shown in Table 2.

[0120] Table 2: Performance metrics of single-feature classification for application promotion robots

[0121]

[0122] It can be seen that the entropy feature based on posting exhibits strong discriminative ability, and the detection effect improves as the time window becomes coarser, highlighting the importance of the temporal variability of posting patterns. The robot achieves a high F1 score at a coarser time granularity, demonstrating strong separability.

[0123] In summary, entropy features based on posting behavior are the most informative and versatile single feature metric for detecting bots, especially at a coarser temporal granularity. Reply behavior adds complementary value to conversational bot detection, while entropy features based on forwarding behavior remain insufficient in terms of expressiveness and discriminability. Computer vision (CV) has improved detection performance for social bots with unstable or sporadic activity patterns. These findings underscore the importance of tailoring behavioral features to behavior type and temporal resolution in social bot detection.

[0124] 5. Detection performance of the present invention (EntroBot entropy ensemble model)

[0125] As can be seen from Section 4 above, a single feature is insufficient to characterize the diversity and multi-scale behavior of social robots. This invention addresses this problem by integrating entropy measures across multiple behavior types and time scales. To verify the effectiveness of this invention, seven baseline methods relying solely on behavioral time series were selected. These baseline methods include Shapelet and Seasonality for capturing periodic behavioral patterns; Vatw, Titw, and Vati for characterizing behavioral frequency and burstiness; and entropy-based methods such as sample entropy and interval entropy. Consistent with previous experiments, random forest was used as the classifier, and downsampling was applied to maintain class distribution balance.

[0126] The performance comparison results are shown in Table 3:

[0127] Table 3: Performance Comparison of Different Methods in Social Robot Detection

[0128]

[0129] As shown in Table 3, the method of this invention (EntroBot) consistently outperforms a range of benchmark methods, demonstrating the effectiveness of combining multiple behavioral dimensions and temporal granularity for social robot detection. By fusing multiple behavior types, EntroBot achieves near-perfect detection performance across all four robot categories, highlighting its strong generalization ability. These results underscore the importance of multi-feature fusion strategies in overcoming the limitations of single-feature models and effectively capturing the complex multi-scale features of various robot behaviors.

[0130] Shape sequence-based methods rely on matching fixed-length subsequences and perform poorly across most robot categories due to their limited generalization ability to rigid or irregular behavioral patterns. In contrast, methods modeling periodic and sporadic activities, such as Seasonality, Vati, Vatw, and Titw, achieve better results by capturing temporal regularity, but their effectiveness decreases when robots mimic human-like randomness, resulting in only moderate F1 scores. Entropy-based methods, by quantifying behavioral uncertainty, generally outperform shape matching methods, especially for robots with repetitive or structured activities. However, these methods are sensitive to data sparsity and user inactivity, leading to unstable estimates at fine-grained temporal resolutions and reduced robustness across various behavioral conditions.

[0131] In summary, while current methods can capture certain aspects of social bot behavior, such as repetitiveness, sudden dynamics, or temporal structure, they fail to provide a unified, multi-scale feature description of user activity. EntroBot addresses this limitation by integrating these complementary behavioral signals into a unified entropy-based feature space, thereby enhancing its ability to detect highly automated accounts as well as those designed to mimic human behavioral patterns.

[0132] 6. Performance comparison with deep learning methods

[0133] To evaluate the effectiveness of deep learning models in capturing temporal behavioral dynamics and to compare their performance and interpretability with proposed entropy-based methods, a diverse set of representative architectures was implemented. These architectures include Long Short-Term Memory (LSTM) networks, Convolutional Neural Networks (CNNs), Transformer Encoders (Transformers), Temporal Convolutional Networks (TCNs), Residual Networks (ResNets), and Attention-Based Recurrent Neural Networks (Att-RNNs). Each model processes user behavior sequences represented in time-series form. Table 4 compares the performance of various deep learning architectures with that of the present invention.

[0134] Table 4: Performance Comparison of the Invention and Deep Learning Methods

[0135]

[0136] The observed performance differences among deep learning models can be attributed to inductive bias and representational power in their architectures. LSTM and TCN outperform other models due to their inherent suitability for modeling temporal dependencies, effectively capturing sequential behavioral patterns. CNNs, with their strong locality bias, excel in categories dominated by short-term behavioral patterns but lag slightly behind in more complex dynamics. Transformer-based models are the weakest, likely due to their heavy reliance on data and lack of local bias, making them ineffective on limited or noisy datasets. ResNet, originally designed for image data, performs poorly in adapting to sequential inputs, resulting in significant performance drops in some categories. Att-RNN achieves moderate results by combining recurrent structures with attention mechanisms, enhancing its ability to aggregate long-range cues, but it still lags behind more specialized temporal models.

[0137] In contrast, the method proposed in this invention demonstrates strong robustness and generalization ability across all robot categories. While it does not always outperform deep learning models on certain metrics, its performance remains highly competitive. By relying on entropy-based behavioral features rather than simply deep feature hierarchies, this invention effectively captures key variations in robot activity while avoiding overfitting to specific patterns. Even when benchmark deep learning methods experience significant performance degradation, the model maintains stable performance. Its architecture-independent design ensures its resilience to various behavioral noises and data sparsity, highlighting its practicality in real-world deployments, especially when robot features vary over time and across different platforms.

[0138] Example 2

[0139] This embodiment provides a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the social robot recognition method based on multi-scale entropy features as described in Embodiment 1.

[0140] like Figure 5 As shown, the computer terminal provided in this embodiment includes: at least one processor 101, and a memory 102 connected to at least one processor 101. This embodiment does not limit the specific connection medium between the processor 101 and the memory 102. Figure 5 The example shown is the connection between processor 101 and memory 102 via bus 100. Bus 100 is... Figure 5 The connections between other components are shown in bold lines and are for illustrative purposes only, not as limiting information. Bus 100 can be divided into address bus, data bus, control bus, etc., for ease of representation. Figure 5The bus is represented by a single thick line, but this does not indicate that there is only one bus or one type of bus. Alternatively, the processor 101 may also be called a controller; there is no restriction on the name.

[0141] In this embodiment, the memory 102 stores instructions that can be executed by at least one processor 101. The at least one processor 101 can execute the aforementioned method by executing the instructions stored in the memory 102.

[0142] The processor 101 is the control center of the device. It can connect to various parts of the control device through various interfaces and lines. By running or executing instructions stored in memory 102 and calling data stored in memory 102, the processor can perform various functions and process data, thereby monitoring the device as a whole.

[0143] In one possible design, processor 101 may include one or more processing units. Processor 101 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into processor 101. In some embodiments, processor 101 and memory 102 may be implemented on the same chip; in some embodiments, they may also be implemented on separate chips.

[0144] Processor 101 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit, field-programmable gate array or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the social robot recognition method based on multi-scale entropy features disclosed in Embodiment 1 can be directly implemented by the hardware processor, or implemented by a combination of hardware and software modules in processor 101.

[0145] Memory 102, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 102 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic storage, magnetic disk, optical disk, etc. Memory 102 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. In this embodiment, memory 102 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0146] By designing and programming the processor 101, the code corresponding to the social robot recognition method based on multi-scale entropy features described in the aforementioned embodiments can be embedded into the chip, thereby enabling the chip to execute the code during runtime. Figure 1 The steps of the social robot recognition method based on multi-scale entropy features are shown. How to design and program the processor 101 is a technique well-known to those skilled in the art and will not be described further here.

[0147] Example 3

[0148] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the steps of the social robot recognition method based on multi-scale entropy features as described in Embodiment 1.

[0149] The computer-readable storage medium may include flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the storage medium may be an internal storage unit of a computer device, such as the hard disk or memory of the computer device. In other embodiments, the storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc., provided on the computer device. Of course, the storage medium may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is typically used to store the operating system and various application software installed on the computer device. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0150] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A social robot recognition method based on multi-scale entropy features, characterized in that, include: S1. Obtain social network behavior data of the user to be detected during their lifecycle, and construct initial binary time series for each predefined behavior type; the predefined behavior types include posting, forwarding, and replying; the initial binary time series are constructed as follows: if the user performs a specified type of behavior at a specific time point, the value of that time point is recorded as 1, otherwise it is recorded as 0; the lifecycle is defined as the time interval from the start of user account registration to the last recorded behavior timestamp; S2. Set multiple time windows with different granularities to map the initial binary time series into multiple coarse-grained behavior sequences with different time scales; S3. For each coarse-grained behavior sequence at each time scale, calculate its corresponding behavior information entropy to form a behavior entropy set; the formula for calculating the behavior information entropy is: In the formula, Indicates behavior In time scale Information entropy below; Represents a binary state in a coarse-grained behavioral sequence. The empirical probability of occurrence; , A global set of behaviors for each user; The calculated behavioral information entropy is normalized to limit its value range to [0,1]; by traversing... K Different time scales , construct behavior exist K A set of entropy values ​​at different time scales That is, the set of behavioral entropy; S4. Based on the set of behavioral entropies, calculate the coefficient of variation of the entropy values ​​at different time scales to quantify the dynamic fluctuation characteristics of user behavior; The formula for calculating the coefficient of variation is: In the formula, For behavior The corresponding coefficient of variation; For set standard deviation For set The mean; S5. The behavioral entropy set is fused with the coefficient of variation to construct a comprehensive feature vector. The comprehensive feature vector is input into the trained supervised classification model, and the classification result of whether the user is a real user or a social robot is output.

2. The social robot recognition method based on multi-scale entropy features according to claim 1, characterized in that, In step S2, the multiple time windows of different granularities are non-overlapping time windows; the calculation formula for mapping the initial binary time series into multiple coarse-grained behavioral sequences of different time scales is as follows: In the formula, Indicates the first k Coarse-grained behavioral states within a time window; Time scale; For the first k The starting reference point for each time window; Given an initial binary time series, Indicates behavior In time occur.

3. The social robot recognition method based on multi-scale entropy features according to claim 2, characterized in that, In step S5, the specific process of constructing the comprehensive feature vector is as follows: Regarding posting ,Forward and reply Three behavior types were identified, and their behavioral information entropy at different time scales was extracted to construct a user profile. u Multi-scale information entropy feature matrix : In the formula, , and These represent the three behavior types: posting, forwarding, and replying. m Time scale The behavioral information entropy, ; Get Posting ,Forward and reply Coefficients of variation for the three behaviors , and , build users u eigenvectors of coefficient of variation : Multi-scale information entropy feature matrix With coefficient of variation eigenvector The features are then combined and aggregated to form a comprehensive feature vector that represents the user's dynamic behavior patterns.

4. The social robot recognition method based on multi-scale entropy features according to claim 1, characterized in that, In step S5, the supervised classification model is a random forest classifier; before inputting the comprehensive feature vector into the supervised classification model, the historical user data is also processed using downsampling technology to balance the distribution of category data between social robots and real users.

5. The social robot recognition method based on multi-scale entropy features according to claim 1, characterized in that, In steps S1 to S4, each behavior type does not overlap in the process of constructing the initial binary time series, performing multi-scale time series mapping, extracting multi-scale behavior information entropy, and calculating the coefficient of variation, so as to preserve the inherent temporal evolution characteristics of a single behavior type. Until step S5, the independently extracted behavioral entropy sets and the coefficient of variation are concatenated to achieve feature-level fusion of multiple behavioral dimensions.

6. A computer terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the social robot recognition method based on multi-scale entropy features as described in any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the social robot recognition method based on multi-scale entropy features as described in any one of claims 1 to 5.

Images