Multi-Strategy Improvement and Comparative Research on Data-Driven Social Network Construction in Edge-Deficient Scenarios for Social Bot Account Detection

Abstract

Accurate social bot detection relies on simulated data to alleviate the scarcity of labeled real-world datasets. Synthetic graph data serves as the core training resource for detection models within simulated data; nevertheless, edge deficiency in real social networks (induced by privacy constraints and data collection limitations) gives rise to “pseudo-isolated nodes” and distorts the quality of synthetic graph data. Furthermore, mainstream data-driven synthetic graph generation methods lack systematic and credible comparative analyses. To tackle these problems, this study optimizes two representative synthetic graph generation approaches (the Chung-Lu model and the Random Classifier-based Multi-Hop (RCMH) sampling + diffusion model) and puts forward an edge completion strategy grounded in sociological theories. Multiple groups of comparative experiments are conducted to assess the performance of the improved methods and the edge completion strategy. Experimental results demonstrate that the “interest + social association” edge completion strategy achieves an F1-score (F1) of 0.7051, and the improved sampling + diffusion model integrated with edge completion reaches an F1-score of 0.7071, which performs better than traditional and unmodified methods to a certain extent. This work preliminarily enhances the reliability of synthetic graph generation methods and provides relatively high-quality synthetic social graph data for social bot detection. It should be noted that the proposed methods are validated solely on Twitter-derived datasets, and their effectiveness remains to be verified in cross-platform adaptation and dynamic social network scenarios.

1. Introduction

Social bots pose severe threats to the information ecosystem security of social networks, and accurate social bot detection is an urgent research need for network security governance [1]. However, the increasingly stringent privacy protection policies of social platforms have led to the scarcity of high-quality real labeled data for bot detection, making simulated social network data the primary training source for bot detection models. Privacy constraints and malicious behavior threats are core challenges in social network security. In the field of general network security, researchers have explored rule-based defense against collusion attacks in cloud environments [2] and machine learning-based secure data transmission for IoT [3], which provide security-oriented research ideas for network data processing. For social bot detection, the privacy-induced edge deficiency in social networks further distorts topological structure, which is the key problem addressed in this study.

Existing social network simulation, data synthesis and augmentation methods—including diffusion-model-based synthetic graph generation, GNN/rule-based topology replication, and variational generative adversarial network-assisted minority sample synthesis—all aim to enhance the consistency between simulated social network data and real-world data. Yet these methods share an unvalidated core assumption: they treat real data as a complete and reliable reference benchmark. For instance, diffusion-generated social graphs [4] train models on real datasets (e.g., TwiBot-22) without considering topological distortion caused by edge deficiency; comparative studies of GNN and rule-based methods [5] assume sampled subgraphs preserve original network patterns while ignoring the pre-existing community fragmentation; Topology-Aware Gated GNN [6] relies on distorted topological metrics of real nodes for sampling; and HMM-VGAN [7] only removes superficial feature noise while neglecting deep flaws such as missing social relations and label-related noise in real data.

The construction of realistic synthetic graph data must rely on the topological characteristics and interaction preference of real social network data as the core reference [8]. Regrettably, the limited real social network data available for reference is widely plagued by the problem of edge deficiency, which is caused by two core factors: on the one hand, users’ active privacy settings make part of the social connection edges invisible; on the other hand, the restricted scope of data collection makes it impossible to fully tally all the real connection targets of each node [9,10], thus generating a large number of “pseudo-isolated nodes” in the real network. Based on the following relationship data of 1 million nodes in the TwiBot-22 dataset, our experimental results demonstrated that the proportion of isolated nodes in the real social network constructed from this data is as high as 30.62%. As the proportion of isolated nodes further increased, the downstream detection performance declined most sharply before it reached 35%. This observation indirectly confirms that if the real reference data serving as the benchmark for synthetic social graph data construction suffers from increasingly severe edge deficiency, the quality of the resulting synthetic social graph data will inevitably be compromised, and such structural deficiencies of real data directly lead to serious topological distortions and interaction preference deviation of synthetic social graph data constructed with it—meaning that the quality of such data is far from meeting the training requirements of bot detection models [11].

To address the critical challenges caused by edge deficiency in real reference data, this study is dedicated to improving the quality of synthetic social graph data for bot detection. We systematically uncover the structural causes of edge deficiency in real social networks and its distortion mechanism for synthetic social graph data and propose a real data processing approach of “completion first, reference later”. On this basis, we optimize two mainstream data-driven synthetic social graph generation methods (the Chung-Lu model and the RCMH sampling + diffusion model) to enhance their performance in generating high-quality task-adaptive synthetic social graph data (see Figure 1). Furthermore, we design a node label-independent edge completion strategy that integrates interest identification and social association mechanisms and validate its effectiveness using the improved synthetic social graph generation methods as a consistent experimental setup for edge-deficient scenarios—by evaluating the topological quality enhancement of completed real data and the performance improvement of downstream bot detection (refer to Appendix A for relevant experimental methods).

Therefore, the main contributions of this paper are threefold, as follows:

(a)

It systematically reveals the structural causes and synthetic social graph data distortion mechanisms of edge deficiency in real social networks and proposes a “complete first, then reference” research approach for real reference data, which mitigates the negative impact of incomplete real data on high-quality synthetic social graph construction.

(b)

It systematically compares and improves two mainstream data-driven synthetic social graph generation strategies: node degree-driven models and real incomplete network-driven methods, enhancing their ability to generate high-quality synthetic social graph data for the social bot detection task in edge-deficient scenarios.

(c)

We design and verify a node label-independent edge completion strategy for real edge-deficient social network data, which integrates interest identification and social association mechanisms based on user behavior logic, helps infer and supplement potential topological connections in real data, improves the structural authenticity of real reference data, and supports the construction of high-quality synthetic social graph data.

2. Improved Strategies for Social Network Construction in Edge-Deficient Scenarios

2.1. Principle, Defect Analysis, and Improvement of Node Degree-Driven Strategy

The Chung-Lu model is a classic node degree-driven directed network generation method, which is widely used in synthetic social graph generation for its advantage of retaining the real node degree distribution. Aiming to address its adaptability defects in edge-deficient scenarios for social bot detection, this section clarifies the model’s core principle, quantifies its key defects and their impacts on detection tasks, and proposes targeted improvement strategies with clear design motivations, while explaining the physical meaning of key formula and design choices.

2.1.1. Core Principle of the Original Chung-Lu Model

The model takes the out-degree/in-degree sequence of real social network nodes as the core input and must satisfy the hard constraint for directed network topology rationality: ${\sum }_{i=1}^n{d}_i^{out}={\sum }_{i=1}^n{d}_i^{in}=M$ (where $d_i^{out}/d_i^{in}$ is the out-degree/in-degree of node $i$, $n$ is the total number of nodes, and $M$ is the total number of edges).

For any non-self-loop node pair ($i$,$j$), the edge connection probability is calculated as follows: $p_{ij}={\displaystyle \frac{d_i^{out}·d_j^{in}}{M}}$. This formula is based on the basic social network feature that higher-degree nodes have higher connection probability—the product of $d_i^{out}$ and $d_j^{in}$ reflects the basic connection potential of the node pair, and $M$ normalizes the probability to the $[0, 1]$ interval. The model generates the synthetic social graph by calibrating the degree sequence to meet the constraint, calculating the connection probability of all node pairs, and sampling source/target nodes by remaining degree weighting and normalized probability, respectively, while retaining the human–bot labels of real nodes (refer to Appendix B for the relevant pseudocode) [12].

2.1.2. Core Defects and Impacts on Bot Detection

Experimental results indicate that the original model has two core defects in edge-deficient scenarios, which lead to notable topological distortions of the synthetic social graph and further reduce downstream bot detection performance:

• The randomization of edge connections leads to distortions in high-order topology and associations. Without reference to real networks, edge connections lack constraints from association logic. Data such as the human–bot interaction ratio, community structure, and node attribute associations in real networks can only be extracted with real networks as references, which the node degree-driven strategy lacks. This inevitably results in topological and associative distortions, making this flaw inherently unavoidable [13].
• There is a mismatch between a node’s labeled degree and its actual degree in the network, which stems from the lack of complete network topology. While degree information such as in-degrees and out-degrees of social network nodes is easy to collect, the target network fails to cover all associated nodes due to limitations in data collection scope and node selection rules. This leads to significant differences between the global association degree labeled in node attributes and the local actual degree within the network. This has been verified on the TwiBot-22 dataset [14]: the global total number of follows and being followed for 1 million nodes in the dataset reaches 43.7 billion, while the number of actual effective association edges in the target network constructed based on it is only 3.74 million, showing an enormous scale gap.

2.1.3. Targeted Improvement Strategies

In the absence of full and complete data, should we construct the network based on the global real degrees of nodes or restore the topology using available local node degrees? Both approaches have their disadvantages. Aiming to address the above two defects, this study proposes targeted improvement strategies based on Group A (70% of real data, reference set), with all optimizations retaining the model’s core advantage of degree distribution preservation.

• Proportional scaling and fine-tuning of the real degree sequence

Design motivation: Direct use of global/local degree alone either aggravates distortion or loses the real degree distribution difference (humans are mostly high-degree nodes, bots are mostly low-degree). This method balances the two by Gini coefficient matching + local statistical feature alignment.

Implementation: Extract the out-degree/in-degree distribution, Gini coefficient, total edges and average degree of human/bot nodes in Group A; proportionally scale the global degree sequence to align its total edges and average degree with Group A, while keeping the Gini coefficient consistent to retain the relative difference in the original degree distribution; and, finally, calibrate the scaled sequence to meet the model’s hard constraint [15]. This method reduces pseudo-isolated nodes by making the degree sequence consistent with the actual data collection range and real distribution characteristics.

2.

Integration of human–bot interaction preference weights into connection probability

Design motivation: Introduce real human–bot interaction preference as a soft constraint to solve the randomness of edge connection, and restore the differential interaction characteristics between human and bot nodes. Implementation: First count the edge numbers of four interaction types in Group A (${count}_{BB}$, ${count}_{BH}$, ${count}_{HB}$, ${count}_{HH}$), and normalize to obtain the interaction preference weight:

$${\omega }_{XY}={\displaystyle \frac{{count}_{XY}}{{\sum }_{X,Y\in \left\{B,H\right\}}{count}_{XY}}}(X,Y\in \{B,H\left\}\right)$$

(1)

${\omega }_{XY}$ reflects the proportion of connection probability that X-type nodes initiate to Y-type nodes and satisfies ${\omega }_{BB}+{\omega }_{BH}+{\omega }_{HB}+{\omega }_{HH}=1$. On the premise of meeting the model’s hard constraint, integrate ${\omega }_{XY}$ into the original formula to obtain the optimized connection probability [16]:

$$p_{ij}={\displaystyle \frac{d_i^{out}·d_j^{in}}{M}}{\times \omega }_{t\left(i\right)t\left(j\right)}$$

(2)

where $t\left(i\right)/t\left(j\right)$ is the type (human/bot) of node $i/j$. In edge generation, the target node sampling probability is subject to the dual constraint of degree product and interaction preference (source node still by remaining degree weighting), ensuring the synthetic social graph aligns with real-world interaction preference.

Physical meaning: Formula (2) combines the degree distribution feature (first part) and real human–bot interaction preference (second part), making edge connection no longer random.

2.2. Research and Improvement of Real Incomplete Network-Driven Strategy

To alleviate labeled data scarcity for bot detection, the subgraph sampling + diffusion model is a mainstream data-driven synthetic social graph generation method. It uses Rejection-Controlled Metropolis–Hastings (RCMH) sampling [17] to extract representative nodes from incomplete real networks and then generates synthetic social graphs via a diffusion model [4,18]. However, the original method is designed for general social network topology restoration, which is ill-suited to the core goal of bot cluster feature capture. This section first clarifies the original logic and task inadaptability and then proposes targeted improvements with clear design motivations.

2.2.1. Original RCMH Sampling + Diffusion Model: Core Logic

The original RCMH sampling performs node extraction via degree-biased random walk, aiming to retain the global topology of social networks. For any node v in the real network, its total degree is defined as the sum of the out-degree and in-degree (Formula (3)):

$$deg\left(v\right)= ou{t}_{deg\left(v\right)}+ i{n}_{deg\left(v\right)}={\sum }_{u\in V} A_{v,u}^{\left(0\right)}+{\sum }_{u\in V} A_{u,v}^{\left(0\right)}$$

(3)

Equation (3) is the basic degree definition of directed social networks, where $A^{\left(0\right)}$ is the $0-1$ adjacency matrix ($1$ means a follow relationship exists).

The random walk transition acceptance probability (Formula (4)) controls the walk tendency to high-degree nodes:

$$\alpha (u\to v)=\mathrm{m}\mathrm{i}\mathrm{n}(1,{\displaystyle \frac{{deg\left(v\right)}^{\alpha }}{{deg\left(u\right)}^{\alpha }}})$$

(4)

The larger $\alpha \in [0, 1]$ is, the more the walk prefers high-degree nodes, which facilitates rapid capture of the network’s core skeleton. To avoid the walk stagnating without new nodes, a restart mechanism is set (Formula (5)):

$$u_{new}=\left\{\begin{array}{c}v if \alpha (u\to v)\ge ϵ (ϵ\sim Uniform(0,1\left)\right)\\ \\ Random\left(V\backslash S\right) if idle steps\ge \max \_idle\end{array}\right.$$

(5)

When the walk is stuck for too long, randomly select an unsampled node to restart, ensuring full coverage of the network.

After sampling, the original diffusion model generates synthetic graphs based on the entire graph density, without considering the human–bot interaction characteristics required for bot detection (refer to Appendix B for the relevant pseudocode).

2.2.2. Core Defects and Impacts on Bot Detection

The original method has two defects for bot detection, directly verified by the TwiBot-22 dataset (

Table 1. Human–Bot Proportion of High-Degree Nodes and Their Interaction Objects.

Statistical Dimension	Human Proportion (%)	Bot Proportion (%)
Composition of high-degree nodes (deg > 105)	97.69	2.31
Composition of interaction objects of high-degree human nodes	95.05	4.95

):

Sampling bias misses bot nodes: Social network degrees follow a power-law distribution; $97.69\%$ of high-degree nodes ($deg>{10}^5$) are humans, while $96.09\%$ of bots concentrate in medium and low-degree nodes. The original degree-biased RCMH oversamples high-degree humans and ignores low-degree bots, leading to the loss of bot cluster features.

Diffusion distorts topology and interaction ratios: The original diffusion uses the entire graph density to calculate synthetic edges, causing inconsistency between the synthetic graph’s sparsity and the real subgraph. It also does not retain the real human–bot interaction ratio, diluting bot-specific interaction preference (e.g., high-density bot–bot connections).

2.2.3. Improved RCMH Sampling: Node Importance Weighting and Human–Bot Balance

Improvement motivation: Eliminate sampling bias, increase the sampling probability of medium and low-degree bots, and keep the human–bot ratio of the sampled set balanced.

• Step 1: Degree interval division

Based on the power-law degree distribution of social networks and the human–bot degree statistics in Table 1, nodes are divided into 5 intervals: (0, 10), (10, 100), (100, 1000), (1000, 10,000), (10,000,$\mathrm{\infty }$). Bots are mainly concentrated in 0–100, while humans dominate intervals above 10,000. This division enables precise node-type differentiation.

• Step 2: Node importance weight assignment

A differentiated weighting rule is designed based on node type (bot/human) and degree interval (grounded in Table 1):

Bot nodes: Apply positive enhancement weighting to low- and medium-degree intervals $(0-100)$ to raise the sampling probability; apply negative weakening weighting to high-degree intervals (>10,000) to reduce redundant sampling of human-dominated nodes [19].

Human nodes: Adopt the reverse weighting strategy—enhance high-degree intervals and weaken low/medium-degree intervals to balance the human–bot ratio. All weights $\omega \left(v\right)$ are uniformly calculated and normalized by the code in Appendix H, serving as the sole weight source in the model.

• Step 3: Improved transition acceptance probability

The node importance weight is integrated into the original transition probability to retain topological rationality while correcting sampling bias (Formula (6)):

$${\alpha }^{\prime }\left(u\to v\right)=\min \left(1,{\left({\displaystyle \frac{\mathrm{deg}\left(v\right)}{\mathrm{deg}\left(u\right)}}\right)}^{\alpha }\right)\times \omega \left(v\right)(\alpha =0.95)$$

(6)

The first half preserves the degree-biased topological logic of the original RCMH; the second half $\omega \left(v\right)$ (automatically computed weight) helps boost sampling chances for medium- and low-degree bots, alleviating the under-sampling of bot nodes in traditional sampling. Weighted sampling is conducted after normalization to better retain the key bot cluster features in the sampled subgraph.

2.2.4. Improved Diffusion Model: Subgraph Density Constraint and Human–Bot Interaction Preservation

Improvement motivation: Solve topological sparsity distortion and bot feature dilution caused by the original diffusion model.

• Step 1: Subgraph density constraint (Formula (7))

The original diffusion uses the entire graph density, leading to structural distortion. We use the density of the optimized RCMH sampled subgraph to calculate the target number of synthetic edges:

$$M_{synth}=sample\_density\times N_{synth}^2$$

(7)

Ensure the synthetic graph has the same sparsity as the real sampled subgraph, avoiding over-dense/over-sparse topology distortion.

• Step 2: Human–bot interaction weight preservation (Formula (8))

To retain real bot interaction preference, we calculate the weight of four interaction types (bot → bot, bot → human, human → bot, human → human) from real data:

$${\omega }_{XY}={\displaystyle \frac{{count}_{XY}}{{\sum }_{X,Y\in \{B,H\}}{count}_{XY}}}$$

(8)

Allocate synthetic edges according to this weight to avoid diluting bot cluster features (e.g., high-probability bot→bot connections).

Finally, the GraphMaker diffusion model is used: SVD dimensionality reduction ($256$ dimensions) reconstructs node association features [20], reverse diffusion generates a node similarity matrix, and the top $M_{\mathrm{s}\mathrm{y}\mathrm{n}\mathrm{t}\mathrm{h}}$ non-self-loop edges are selected to generate a synthetic graph that meets bot detection requirements.

2.3. Research on Edge Completion Strategies for Social Networks with Edge Deficiency

To address the prevalent edge deficiency in real social networks, this section proposes a data-driven and node label-independent edge completion strategy. Supported by the sociological “common identity and common bond theory” as its core theoretical foundation [21,22], this strategy aims to use publicly observable user behavior data in Group A to capture two types of potential connection patterns [23], “interest identification” and “social association”. On the premise of protecting user privacy, it constructs a more complete and structurally authentic social network topology, providing high-quality data support for downstream bot detection tasks.

The formation of connections in social networks is essentially driven by two core mechanisms, which are highly consistent with the core connotation of the common identity and common bond theory: first, the common identity mechanism—users with similar interests form identity recognition due to shared topic preferences, leading to a significantly higher connection probability; second, the common bond mechanism—frequently interacting users form emotional or behavioral bonds through direct social associations, making it easier to generate explicit connections. Therefore, inferring potential follow relationships based on users’ public behavior trajectories (e.g., posted content and mention relationships) is theoretically logical and practically grounded.

This strategy adheres to a key principle: using the behavioral statistical laws of public nodes to infer the potential connection patterns of hidden nodes. All completion rules are learned from the visible public follow relationship network in Group A and applied equally to all nodes (whether human or bot), thereby minimizing the introduction of human bias or label information leakage. Combining the above theories and principles, the completion strategy is materialized into the following three sequential and complementary logical modules—with the first two corresponding to the two core connection mechanisms, respectively, and the third serving as supplementary enhancement [5].

2.3.1. Interest Identification: Edge Completion Based on Topic Participation

The topic categories that users participate in are the core representation of their interests. All topics in the dataset are classified into 17 macro categories based on IPTC Media Topics [24] by invoking the Gemma3-12b model. First, the average in-degree and out-degree of public nodes under each topic category are calculated to establish a mapping model of “topic category and connection strength”. For a hidden node, its potential interest community can be inferred by analyzing the topic categories of its historical posts. Subsequently, based on the prior connection strength of its main topic category and combined with the number of follows and followers indicated in its profile for weighted allocation, directed edges pointing to other nodes in the same topic community are generated for it. This method aims to restore social bonds based on shared interests (refer to Appendix D for detailed methods).

2.3.2. Social Association: Edge Completion Based on Mention Relationships

Mutual mentions between users are core evidence of strong social interaction. The relevant edge completion rules and extension strategies are as follows:

If user A and user B have had direct mentions in any post (including A actively @B or B actively @A), both parties are confirmed to have explicit interaction intentions. Regardless of how many times such mention behaviors occur, a directed edge is added between these two nodes with high confidence. This module can directly capture strong social interaction signals and restore explicit social connections that were broken due to data deficiency.

For users with no direct mentions, the “similarity of common mention preferences among user groups” (i.e., the overlapping characteristics of the groups frequently mentioned by different users) is combined, and the verified association data distribution rules in real social networks are referenced to add high-probability relationship edges for such users with potential associations (refer to Appendix D for detailed methods).

2.3.3. Edge Completion Based on Link Prediction Technology

Link prediction technology can infer potential high-probability connections between nodes under the condition of known partial network topology structure. Most existing link prediction methods (such as those based on common neighbors, Jaccard coefficient, Adamic–Adar index, node embedding, etc.) mainly rely on local topological similarity for statistical inference [25,26]. Whether such methods can reveal the deep-seated differences between humans and bots in network structures needs to be verified. To this end, this study systematically tests a variety of link prediction methods and their combination strategies, constructs the completed network based on these methods, and then comprehensively evaluates the impact of different completed networks on model training and testing performance through bot detection tasks (among the following four types of methods, each method supplements the top 1 million relational edges with the highest probability that do not exist). The specific methods are divided into the following four categories:

• Link prediction method combining traditional topological features+ network embedding+ tree-model supervised learning [27,28] (basic supervised paradigm);
• Link prediction method combining enhanced topological features + community-aware embedding [29] + ensemble tree model (enhanced supervised paradigm);
• End-to-end graph neural network + hard negative sampling link prediction method [30,31,32] (representation learning paradigm);
• Link prediction method combining multi-embedding fusion + multi-feature integration + meta-learning [33] (hybrid integration paradigm).

3. Results

To systematically evaluate the effectiveness of different social network construction strategies for social bot detection tasks in edge-deficient scenarios, this section designs comparative experiments with raw real social network data as the unified benchmark, aiming to comprehensively compare the performance of various methods in topological restoration and detection performance. The experiments are mainly divided into two groups of controlled tests.

First, we compared six basic and improved network construction methods under the baseline condition without introducing the edge completion strategy.

Second, we integrated the edge completion strategy proposed in this paper (as detailed in Section 2.3) with each of these network construction methods to form another five groups of experimental conditions. Therefore, this section conducts performance comparison and in-depth analysis of social networks under 11 different construction conditions in total.

In addition, to clarify the contribution of each improved module, this section also sets up corresponding ablation experiments. All experiments adopt a unified feature set (9-dimensional-node features) and classification model (random forest) and use accuracy, precision, recall, and F1-score as the core evaluation metrics on the same real test set (Group B data). The specific experimental group design is shown in

Table 2. Comparison of Social Bot Detection Performance of Different Network Construction Methods.

Dataset	Method Name	ACC (95% CI)	Pre (95% CI)	Rec (95% CI)	F1 (95% CI)
TwiBot-22	Real Data (train:val:test = 7:2:1)	0.6170 [0.6115, 0.6224]	0.6006 [0.5928, 0.6076]	0.6997 [0.6929, 0.7064]	0.6463 [0.6404, 0.6521]
	Attribute-Based Degree	0.4999 [0.4947, 0.5052]	0.5000 [0.4947, 0.5048]	1.0000 [1.0000, 1.0000]	0.6666 [0.6620, 0.6710]
	Network-Based Degree	0.5009 [0.4957, 0.5062]	0.7734 [0.6557, 0.8724]	0.0024 [0.0017, 0.0031]	0.0047 [0.0033, 0.0061]
	Network-Based Degree (Non-Isolated)	0.5088 [0.5034, 0.5139]	0.5054 [0.4998, 0.5105]	0.8353 [0.8297, 0.8409]	0.6297 [0.6246, 0.6343]
	Subgraph Sampling + Diffusion Model	0.6061 [0.6014, 0.6111]	0.5760 [0.5703, 0.5821]	0.8041 [0.7984, 0.8093]	0.6712 [0.6664, 0.6762]
	Improved Degree-Based Method	0.5788 [0.5732, 0.5839]	0.5529 [0.5468, 0.5595]	0.8235 [0.8179, 0.8295]	0.6615 [0.6564, 0.6666]
	Improved Subgraph Sampling + Diffusion Model	0.6190 [0.6142, 0.6241]	0.6042 [0.5975, 0.6108]	0.6902 [0.6833, 0.6970]	0.6443 [0.6386, 0.6501]
TwiBot-20	Real Data (train:val:test = 7:2:1)	0.5750 [0.5532, 0.5978]	0.5951 [0.5597, 0.6347]	0.4752 [0.4424, 0.5087]	0.5286 [0.4974, 0.5565]
	Attribute-Based Degree	0.4023 [0.3822, 0.4223]	0.5077 [0.3709, 0.6305]	0.0218 [0.0143, 0.0297]	0.0417 [0.0274, 0.0576]
	Network-Based Degree	0.4320 [0.4111, 0.4530]	0.5542 [0.5160, 0.5929]	0.2540 [0.2308, 0.2761]	0.3490 [0.3224, 0.3784]
	Network-Based Degree (Non-Isolated)	0.4423 [0.4209, 0.4642]	0.5438 [0.5123, 0.5743]	0.4144 [0.3876, 0.4407]	0.4702 [0.4446, 0.4945]
	Subgraph Sampling + Diffusion Model	0.5750 [0.5535, 0.5917]	0.6596 [0.6384, 0.6902]	0.6058 [0.5804, 0.6273]	0.6297 [0.6127, 0.6528]
	Improved Degree-Based Method	0.5656 [0.5447, 0.5861]	0.6425 [0.6146, 0.6686]	0.6153 [0.5875, 0.6426]	0.6287 [0.6059, 0.6512]
	Improved Subgraph Sampling + Diffusion Model	0.5722 [0.5530, 0.5915]	0.6497 [0.6237, 0.6742]	0.6282 [0.6050, 0.6514]	0.6365 [0.6153, 0.6566]

Bold values indicate the highest performance of the corresponding metric among all methods under the same dataset (This explanation applies to all figures and tables in this paper).

and

Table 3. Bot Detection Performance of Edge Completion Strategies and Improved Methods After Edge Completion.

Method Name	ACC (95% CI)	Pre (95% CI)	Rec (95% CI)	F1 (95% CI)
Real Data (train:val:test = 7:2:1)	0.6170 [0.6115, 0.6224]	0.6006 [0.5928, 0.6076]	0.6997 [0.6929, 0.7064]	0.6463 [0.6404, 0.6521]
Real Data (Interest-Based Edge Completion)	0.6833 [0.6784, 0.6881]	0.6771 [0.6704, 0.6840]	0.7008 [0.6936, 0.7075]	0.6888 [0.6833, 0.6945]
Real Data (Social Association Edge Completion)	0.6299 [0.6246, 0.6348]	0.6151 [0.6078, 0.6212]	0.6943 [0.6878, 0.7012]	0.6523 [0.6467, 0.6575]
Real Data (Interest + Social Association Edge Completion)	0.6875 [0.6829, 0.6923]	0.6675 [0.6619, 0.6735]	0.7472 [0.7410, 0.7537]	0.7051 [0.7002, 0.7104]
Improved Degree-Based Method (Post Edge Completion)	0.5908 ↑ [0.5861, 0.5956]	0.5587 ↑ [0.5530, 0.5643]	0.8648 ↑ [0.8599, 0.8695]	0.6788 ↑ [0.6741, 0.6834]
Improved Subgraph Sampling + Diffusion (Post Edge Completion)	0.6427 ↑ [0.6383, 0.6474]	0.5991 ↓ [0.5934, 0.6049]	0.8626 ↑ [0.8580, 0.8677]	0.7071 ↑ [0.7025, 0.7117]

Up/down arrows indicate whether the corresponding metric is improved or decreased after edge completion under the same network synthesis model, compared with the baseline without edge completion. Bold values indicate the highest performance of the corresponding metric among all methods under the same dataset (This explanation applies to all figures and tables in this paper).

(refer to Appendix A, Appendix E for specific details of the experimental design; see Appendix F for feature engineering, Appendix G for results of feature importance analysis and ablation experiments, and Appendix I for data preprocessing and partitioning methods).

Comparative experiments were conducted on the TwiBot-22 and TwiBot-20 datasets, which preliminarily verified the effectiveness and stability of the improved network construction methods proposed in this study. With the raw real social network data as the unified benchmark, both the Improved Degree-Based Method and the Improved Subgraph Sampling + Diffusion Model exhibited stable detection performance on the two datasets. In contrast to traditional approaches, our improved methods show no occurrence of extreme metric collapses, and the overall results remain consistent and close to those obtained from real data. On TwiBot-22, the Improved Subgraph Sampling + Diffusion Model achieved an accuracy of 0.6190, close to the detection performance of real data; on TwiBot-20, this method obtained an F1-score of 0.6365, slightly outperforming the traditional Subgraph Sampling + Diffusion Model, which alleviated the severe fluctuation in detection performance seen in traditional degree-based methods.

The two traditional degree-based methods showed extreme detection metrics with distinct failure causes: on TwiBot-22, the recall of the Attribute-Based Degree Method reached 1.0000 because this method only relied on network degree for detection and could not distinguish the differences between humans and bots in the network at all. In addition, the attribute degree of each node in the dataset was higher than the average level of the test data, which further made the model lose its ability to distinguish humans and bots, and this is the key reason for the failure of this method in the detection task. The recall of the Network-Based Degree Method was only 0.0024, mainly due to the extremely high proportion of isolated nodes in the network constructed by the dataset. Even with random 1:1 human–bot sampling, a large number of isolated nodes were likely to be sampled, and such nodes could not be used for network construction, which indirectly broke the actual balance of human–bot samples. When the differences in network features were small, the model tended to predict the category with a larger sample size, ultimately leading to the collapse of detection metrics. Aiming at the experimental bias of the Network-Based Degree Method, this study additionally supplemented the Network-Based Degree (non-isolated) experiment, which first sampled a sufficient number of non-isolated bot nodes and then an equal number of non-isolated human nodes. Consequently, the recall of this method on TwiBot-22 rebounded to 0.8353 with an F1-score of 0.6297, which verifies the necessity of sampling rule optimization for eliminating such experimental biases.

Objectively, the poor performance of traditional degree-based methods in the social bot detection task does not mean their complete ineffectiveness in synthetic social graph generation. These methods can retain basic topological features such as node degree distribution and are applicable to general synthetic social graph generation scenarios. However, they only perform random edge connection based on the product of node degrees without integrating the core logics of social networks (e.g., user interest, social association, and human–bot interaction preference), making them poorly suited to bot detection requirements. By integrating human–bot interaction preference, node importance weight and hard degree constraint, the improved methods in this study are specifically designed to adapt to social bot detection in edge-deficient scenarios, thus exhibiting more stable detection performance on datasets with different characteristics and providing a reference for similar studies.

Taking the detection performance of raw real social network data as the benchmark, this table compares the core performance metrics of social bot detection for different edge completion strategies and improved network construction methods after edge completion, which verifies the significant improvement effect of edge completion strategies on downstream detection tasks.

Among the single edge completion strategies, both interest-based and social association-based edge completion achieve a positive performance improvement, with the interest-based edge completion performing better and reaching an F1-score of 0.6888, which is higher than that of 0.6463 for the raw data. The interest + social association fusion edge completion strategy achieves a further performance breakthrough with its F1-score rising to 0.7051, and all its metrics are superior to those of the raw data and single edge completion strategies, repairing topological structure defects of edge-deficient networks. Meanwhile, edge completion strategies and improved network construction methods have a good synergistic optimization effect: the detection performance of both the Improved Degree-Based Method and the Improved Subgraph Sampling + Diffusion Model is improved to varying degrees after combining with edge completion. Among them, the F1-score of the Improved Subgraph Sampling + Diffusion Model rises to 0.7071 after edge completion, making it the overall optimal solution, and the Improved Degree-Based Method also achieves a comprehensive increase in all core metrics. This indicates that edge completion strategies cannot only directly improve the bot detection performance of real edge-deficient networks but also cooperate with improved network construction methods to optimize the detection performance, which verifies the effectiveness of the proposed strategies in adapting to the downstream task of social bot detection in edge-deficient scenarios.

To further verify the differences between the proposed method in this paper and traditional link prediction methods, we separately adopted four groups of link prediction methods for edge completion and obtained the experimental results in

Table 4. Comparison of Bot Detection Performance (Link Prediction Edge Completion, Edge-Deficient).

Method Name	ACC (95% CI)	Pre (95% CI)	Rec (95% CI)	F1 (95% CI)
Real Data (20% Test Set)	0.6170 [0.6115, 0.6224]	0.6006 [0.5928, 0.6076]	0.6997 [0.6929, 0.7064]	0.6463 [0.6404, 0.6521]
Link Prediction (Basic Supervision)	0.6176 [0.6124, 0.6229]	0.6007 [0.5937, 0.6076]	0.7021 [0.6953, 0.7093]	0.6474 [0.6418, 0.6533]
Link Prediction (Enhanced Supervision)	0.5941 [0.5890, 0.5993]	0.5743 [0.5675, 0.5811]	0.7280 [0.7208, 0.7347]	0.6420 [0.6368, 0.6474]
Link Prediction (Representation Learning)	0.6166 [0.6113, 0.6221]	0.6000 [0.5928, 0.6069]	0.7002 [0.6933, 0.7073]	0.6462 [0.6405, 0.6520]
Link Prediction (Hybrid Integration)	0.6058 [0.6007, 0.6110]	0.5928 [0.5856, 0.5999]	0.6770 [0.6701, 0.6841]	0.6321 [0.6264, 0.6380]

under the same experimental configuration.

Taking the detection performance of raw real social network data as the benchmark, this table presents the core performance of four types of traditional link prediction methods in social bot detection after edge completion. The results show that no effective improvement is achieved in the detection metrics of all methods, some metrics even decline, and the overall performance is basically consistent with that of the raw data. This suggests that the networks after edge completion by traditional link prediction methods have relatively insufficient adaptability in the specific downstream task of social bot detection.

To verify the reliability of the edge completion of the method proposed in this paper, we designed relevant validation experiments, with the detailed contents provided in Appendix C.

4. Discussion

This work provides incremental improvements for social network construction in edge-deficient bot detection, helping mitigate data scarcity and topological distortion. Our task-adaptive optimizations and label-free edge completion offer a practical reference for related research under constrained data conditions.

It should be emphasized that traditional synthetic social graph generation methods have made important contributions to general social network simulation by effectively retaining basic topological features such as node degree distribution and global network structure, which still have high application value in generic network analysis tasks. Their limitations only lie in the mismatch with the specific requirements of social bot detection in edge-deficient scenarios, rather than inherent defects of the methods themselves, yet this mismatch leads to drastic performance fluctuations in bot detection tasks.

Ablation experiments indicate that optimizing only with human–bot interaction preference weights fails to deliver ideal performance, with the single-component model showing a sharp drop in core metrics. In an ideal scenario without data collection limitations and full access to users’ social relationships, the attribute-degree network and network-degree network would be fully consistent, and the interaction preference weight alone would work effectively. However, in real edge-deficient scenarios, such a single constraint can only adjust macro connection ratios between human and bot nodes and cannot resolve the fundamental discrepancy between attribute degree (global public follow/follower counts) and network degree (actual collected valid edges) caused by limited collection scope and user privacy settings. It also fails to mitigate sampling bias toward high-degree human nodes and topological fragmentation from pseudo-isolated nodes. Since social bot detection relies on complete local topology and fine-grained behavioral association features, single-module optimization is far from repairing the inherent structural defects of incomplete real social networks.

In addition, link prediction-based edge completion fails to bring effective performance gains and even degrades partial metrics. On the TwiBot-22 dataset, the basic supervised link prediction only obtains an F1-score of 0.6474 (close to the original data’s 0.6463), and the hybrid integration paradigm drops to 0.6321. Such methods rely solely on topological statistical inference (e.g., common neighbors, node embedding) to generate potential high-probability edges, which cannot correspond to real user behavior-driven social connections and fail to restore human–bot interaction patterns. This only reflects the task misalignment of link prediction methods, not a lack of value in themselves.

The high-recall and low-precision imbalance in some baseline models is also reflected in quantitative results, as the topological discrepancy between synthetic graphs and real networks makes the model misclassify human nodes as bots, which further verifies the necessity of edge completion and task-adaptive synthetic graph optimization.

5. Conclusions

This study focuses on multi-strategy improvement and comparative analysis of data-driven social network construction for social bot detection in edge-deficient scenarios. We optimize two classic synthetic social graph generation methods and propose a node label-independent edge completion strategy combining interest identification and social association. Experimental results on TwiBot-20 and TwiBot-22 datasets validate that the “completion first, reference later” paradigm and multi-module collaborative optimization effectively alleviate topological distortion caused by edge deficiency, and under the specific experimental settings for social bot detection, our improved method exhibits performance metrics that are highly close to and stable with those of real-world data and achieves the optimal F1-score of 0.7071, demonstrating the effectiveness of the proposed approach.

Despite the stable performance improvement, this study has several limitations. First, all experiments are conducted on Twitter-derived datasets, and the generalization ability across multilingual, cross-platform and different topological social networks remains to be verified. Second, the proposed edge completion strategy is static and cannot adapt to the dynamic evolution of real social networks. Third, the balance between data quality and user privacy protection in edge completion and synthetic graph generation needs further exploration. Fourth, the interpretability of topological and behavioral features for bot detection requires enhancement.

Future research will be carried out in four directions: (1) verify the generalization of the proposed methods on more diverse social network datasets; (2) explore dynamic edge completion mechanisms adaptive to network evolution; (3) develop privacy-preserving technologies for edge completion and synthetic graph generation; (4) improve model interpretability to clarify the contribution mechanism of network features to bot detection.