Exploring the Temporal Degradation and Drift of AS Path Inference

1  Introduction

The Internet is composed of numerous Autonomous Systems (ASes), and the Border Gateway Protocol (BGP) is the de facto standard for inter-domain routing. The AS path information is important for network management [1,2], traffic engineering [3], and security [4,5]. It is also closely related to routing incidents, anomalous path propagation, and inter-domain attack analysis [6–8]. However, BGP path information is not directly available to the public and can only be obtained through measurement or path inference. Measurement-based approaches include traceroute and collecting BGP data from routers. Due to limited measurement resources, however, these measurement-based approaches are not scalable. Therefore, path inference approaches have been proposed to broaden the availability of path information.

Internet inter-domain path inference has been studied for nearly two decades; approaches can be broadly categorized into heuristic policy-aware inference and data-driven inference. The classic heuristic policy-aware inference approaches are based on the valley-free principle and AS relationships, such as KnownPath [9], iPlane [10], and iNano [11]. Data-driven inference approaches, on the other hand, leverage machine learning techniques to infer paths based on observed data, such as Sibyl [12], ProbInfer [13], and the recent PathRadar [14]. More recently, our prior work proposed a generative and measurable process for fine-grained AS path inference (GMPI) [15], and further extended this line toward personalized and adaptive inference (PA-GMPI) [16]. We have also systematically reviewed the methodological landscape, applications, and open challenges of Internet inter-domain path inference [17].

These studies have substantially improved the ability to infer unmeasured AS paths, primarily by constructing more accurate inference models under limited or incomplete observations. However, they typically evaluate inferred paths against ground-truth data collected at the same time as the inference inputs or models, and therefore mainly emphasize instantaneous inference accuracy. In practical deployment, path inference services are often used at a later time, when routing conditions may already have changed and the previously inferred paths may no longer remain fully representative of the current routing state.

On the other hand, prior studies on routing stability do not present a fully uniform picture. Green et al. [18] reported substantial persistence of primary inter-domain paths over multi-month observations, whereas Bakhshaliyev et al. [19] showed that stable-path ratios decrease as the observation horizon expands from hours to months. Comarela et al. [20] further demonstrated that routing states evolve measurably over long timescales rather than remaining static. Taken together, these studies suggest that inter-domain paths may appear highly stable over certain observation horizons, yet exhibit noticeable temporal variation over others. However, these studies characterize routing behavior itself rather than the usable lifetime of path inference outputs. As a result, the literature still lacks a direct evaluation of what these mixed stability findings imply for the validity and usable lifetime of AS path inference results.

Motivated by this gap between AS path inference and routing-stability research, this paper aims to systematically answer the following question: How do the performance and temporal drift of AS path inference approaches evolve over time? In this sense, our work is positioned not as another inference method, but as an evaluation study on the temporal degradation and temporal drift of inferred AS paths. Answering this question is essential for understanding the reliability and practical applicability of path inference results in dynamic inter-domain routing environments.

In practical applications, users expect path inference services to return paths that are representative of the current network state, even though the underlying routing data may have been collected earlier. Therefore, some degree of performance degradation over time is inevitable, and understanding its extent is crucial because it informs two practical decisions: (1) the backward time range of the data that can be used for path inference and (2) the update frequency of path inference services. A fundamental limitation inherent in path inference methods is the lack of measurement data. Therefore, the temporal degradation pattern of path inference methods can help indicate how far back in time the data can be used for path inference. Path inference services are computationally intensive, so a trade-off between update frequency and computational resources is often necessary.

Despite the practical importance of temporal degradation in path inference, there is still limited understanding of how quickly stale inference results become outdated. Thus, in this paper, we conduct a quantitative validation study and a temporal drift analysis to examine the evolution of path inference performance over time. The quantitative validation study uses real-world data from a specific day and evaluates the inferred paths over the following eight weeks to quantify the performance degradation of path inference approaches over time. Thus, it helps to determine how far back in time the data can be used for current path inference. The temporal drift analysis further identifies when the inference results derived from the baseline snapshot are no longer representative of the current network state, thereby informing refresh scheduling decisions under the evaluated setting.

The contributions of this paper are as follows:

•   The validation study reveals how representative path inference approaches degrade over time. The results show that the minimal performance degradation of the selected path inference approaches is only 2.09% in nearly two months. The maximum observed degradation is 7.44% in nearly two months, which occurs for ProbInfer in our evaluation, a data-driven path inference approach.

•   The temporal drift analysis provides an operational comparison of refresh horizons across methods. Under the current decision rule, material drift is first detected on day 20 for KnownPath, on day 10 for GMPI, and on day 2 for ProbInfer. This comparison indicates that the three evaluated methods have different refresh horizons and temporal sensitivities under the evaluated setting.

•   This research provides preliminary evidence on how historical data can be leveraged despite limited measurement resources and can inform refresh-frequency decisions for path inference services under computational constraints.

2  Related Work

2.1 Traceroute Methods

Traceroute is a commonly used tool for measuring the path between two hosts. The tool sends a series of packets to the destination, with each packet containing an increasing Time-to-Live (TTL) value. When a packet reaches an intermediate router, the router decrements the TTL value by one. When the TTL value reaches zero, the router discards the packet and sends an ICMP Time Exceeded message back to the source. The source then records the IP address of the router and the round-trip time of the packet. This process is repeated with increasing TTL values until the packet reaches the destination. The recorded IP addresses are used to obtain the path between the source and the destination.

In [21], the authors summarized the limitations (including “loops”, “cycles”, and “diamonds”) of traditional traceroute and proposed Paris Traceroute to provide more accurate path information. To extend path measurement to the reverse direction, revtr 2.0 [22] combines novel measurement techniques with a large-scale deployment and significantly improves throughput, accuracy, and coverage for Internet-scale reverse-path exploration. Recent studies continue to improve traceroute-based topology discovery and active probing efficiency, especially in IPv6 networks. 6Search [23] proposes a reinforcement learning-based traceroute approach for efficient IPv6 topology discovery, while TNet [24] further improves the efficiency of active IPv6 network discovery.

Though multiple traceroute methods have been proposed, these methods share the same limitation: the measurement resources are too limited to obtain the path information between any two ASes. On the other hand, mapping IPs to ASes is also a challenging task [25–28]. As a result, path inference approaches are proposed to address these challenges.

2.2 AS Path Inference

Heuristic-based path inference approaches. KnownPath [9] improves AS path inference by incorporating the valley-free principle [29] to constrain feasible routing paths. In iNano [11], in addition to constructing an Internet topology and adhering to the valley-free principle, the authors proposed the creation of a routing preference database. This database stores AS-level routing preferences in 3-tuples, such as (AS1, AS2 > AS3), which indicates that AS1 favors a path through AS2 over AS3 when both paths are of equal length. When inferring a path, iNano initially searches for the shortest paths that comply with the valley-free principle. It then selects paths that align with the routing preferences. In [30], a method called PredictRoute was introduced. This technique utilizes Traceroute data acquired through active detection to infer paths. The researchers developed distinct directed acyclic graphs for each destination prefix and subsequently employed a destination-specific probabilistic Markov model to derive the inferred paths. In iPlane [10], segments of the measured Internet paths that converged at a common point (which could be an AS, a PoP, or a router) were stitched to infer paths between two target points. The HyperPath proposed in [31] also obtains paths between two ASes by concatenating segments of measured paths. This paper initially establishes that the structure of the AS-level network topology can be conceptualized as a tree, providing a theoretical foundation for stitching path segments to infer paths. Then, a heuristic algorithm is proposed to stitch path segments to infer paths between two ASes.

Data-driven path inference approaches. Due to the intricate nature of the Internet routing system, refined heuristic rules derived from collected measurement data often lack the accuracy needed to model it effectively. Heuristic methods frequently encounter issues with redundant paths. To tackle this challenge, researchers have put forth data-driven inference techniques. Sibyl [12] and ProbInfer [13] both involve stitching path segments to deduce inferred paths, although they employ distinct approaches to minimize redundancy. Sibyl employs a supervised machine learning model called RuleFit [32] to select the most optimal path. In contrast, ProbInfer introduces a probability model (decision tree) to mitigate redundancy. More recently, PathRadar [14] proposed a fine-grained AS path inference framework with a progressive learning process that applies different learning models to different AS categories. The process of Generative and Measurable Path Inference (GMPI) [15] utilizes heuristic algorithms to generate paths for an AP. Subsequently, a dual attention neural network is utilized to extract features from AS paths and estimate the likelihood of these generated paths. In [33], the authors proposed a method called RouteInfer, which infers AS paths by inferring the routing policies of ASes. First, a three-layer policy model is proposed to extract the routing policies for ASes, but it cannot obtain policies for all ASes. Hence, a learning-based approach is proposed to mitigate this limitation. Recent studies have also extended this line toward personalized and adaptive inference settings [16] and survey-style syntheses of inter-domain path inference methods and applications [17]. Taken together, these studies mainly focus on how to construct more accurate or more flexible inference models. In contrast, the present work focuses on how the validity of inferred paths evolves as the routing system changes over time.

2.3 AS Relationship Inference

AS relationship inference is closely related to AS path inference because inferred business relationships provide key constraints and signals for path generation, valley-free filtering, and routing-preference modeling. Gao [29] established the classic relationship model underlying many later inference and routing studies. More recent work has focused on improving relationship inference under incomplete and biased observations. ProbLink [34] improves inference stability and practicality on hard links, while TopoScope [35] recovers AS relationships from fragmentary observations. Prehn and Feldmann [36] further show that available validation data can be substantially biased, which directly affects the evaluation of relationship inference algorithms. Recent learning-based extensions include multiclass relationship inference with graph convolutional networks [37] and HELA [38], which combines empirical and learning-based components to improve accuracy and stability. These studies provide important support for path inference, but they mainly aim to infer the relationships themselves rather than evaluate how path inference outputs degrade over time.

2.4 Mixed Evidence on Internet AS-Level Path Stability

Prior studies on Internet AS-level path stability do not present a fully uniform picture. Green et al. [18] analyze multi-month BGP observations and show that primary inter-domain paths can remain highly persistent over time. In contrast, Bakhshaliyev et al. [19] report that the stable-path ratio decreases from 97% over one hour to 89% over one day, 76% over one week, 54% over one month, and 44% over two months. Comarela et al. [20] further show that routing states evolve measurably over long timescales rather than remaining static. Taken together, these studies suggest that Internet AS-level paths may appear highly stable over some observation horizons, yet exhibit noticeable temporal variation over others. However, these studies characterize routing behavior itself rather than the usable lifetime of path inference outputs. Therefore, the literature still lacks a direct evaluation of how temporal routing dynamics affect the validity of inferred AS paths. This mixed picture is exactly why it is necessary to conduct a validation study on the temporal degradation and temporal drift of path inference approaches.

3  Validation Methodology

To quantify the temporal degradation and drift of path inference approaches, we conduct a validation study using real-world data. The validation process uses data from a specific day to evaluate inferred paths over the following eight weeks. The temporal behavior of path inference approaches is characterized by both performance degradation over time and a temporal drift analysis of stale inference results.

3.1 Preliminaries

As defined in [15], AS path inference focuses on identifying a feasible AS path (s→d) satisfying underlying and unrevealed routing policies for any unmeasured AS pair (AP, two different ASes) Q=(s,d). For ease of understanding, we first introduce some basic concepts and notations in path inference.

•   AS Pair (AP): An AP is a pair of two different ASes, denoted as Q=(s,d), where s is the traffic source AS and d is the destination. If the path between s and d is known, it is called a measured AP; otherwise, it is called an unmeasured AP. The AP of an AS path refers to the AP composed of its first and last ASes.

•   Starting and ending AS: The starting AS of an AP (AS path) is the first AS of the AP (AS path), and the ending AS is the last AS of the AP (AS path).

3.2 Input Data

Three types of data are used in the validation study: AS paths derived from BGP routing tables, AS relationship data, and AS Rank data.

3.2.1 AS Paths Derived from BGP Routing Tables

The routing tables are collected from the Route Views project [39] and the RIPE RIS project [40], two widely used public platforms for BGP data collection. These platforms collect routing information through peering sessions with external networks and provide routing-table snapshots for research use. Route Views offers data from multiple route collectors, while RIPE RIS operates a globally distributed set of Remote Route Collectors, many of which are located at Internet Exchange Points.

In this paper, we use 20 routing-table snapshots collected at 8 a.m. on selected dates from October to November 2023, as summarized in Table 1. The fixed collection time is used to reduce diurnal effects and ensure comparability across snapshots. From these routing tables, we extract AS paths and use them to construct the validation dataset for each snapshot. Following common practice in prior work, the extracted AS paths are treated as ground-truth routing outcomes observed at the control plane.

At the same time, BGP routing tables on any given day may still contain localized abnormal events or temporary policy shifts, and 1 October 2023 is not necessarily an exception. For this reason, the present study should be understood as a large-scale aggregated analysis rather than as evidence that a single snapshot is universally representative. When the evaluation covers a very large AP set and multiple subsequent snapshots, limited local fluctuations are more likely to introduce noise or mild bias than to invalidate the overall temporal trend, unless they affect a large proportion of the observed routing state. In our manual review of publicly reported Internet operational conditions around 1 October 2023, we did not identify evidence of a large-scale Internet-wide disruption on that date.

To assess path inference approaches, both unmeasured APs and their corresponding ground-truth paths are required. In the snapshot of 20231001, the collected AS paths are divided into training and testing sets using a 70%–30% split. The split is performed randomly at the AP level to avoid bias toward specific AS pairs. Seventy percent of the APs and their paths are used as input to the path inference approaches, while the remaining 30% are used for evaluation.

As this study focuses on temporal behavior across multiple snapshots, we further construct a common evaluation set that remains observable throughout the full observation window. The evaluation data is obtained as follows:

•   The APs corresponding to the testing set in snapshot 20231001 are treated as initially unmeasured APs, denoted as U′.

•   The AP sets of the remaining 19 snapshots are denoted as U1 to U19. We compute the intersection U=U′∩U1∩…∩U19 to obtain the final set of unmeasured APs that remain observable across all snapshots, thereby ensuring temporal consistency of the evaluation set.

•   For each snapshot, the AS paths corresponding to these APs are used as ground-truth data for both quantitative validation and temporal drift analysis.

As shown in Table 1, the number of starting ASes is much smaller than the number of ending ASes. This is a direct consequence of the collector-based observation model: the starting ASes are constrained by the locations and peering relationships of the collectors and their connected vantage points, whereas the ending ASes span a much broader destination space.

3.2.2 AS Relationship Data

We use the CAIDA AS relationship dataset dated 1 October 2023 [41,42]. According to CAIDA, this dataset is inferred primarily from publicly available BGP data and captures inter-AS business relationships such as customer–provider and peer–peer links.

AS relationships are commonly classified into customer-to-provider, peer-to-peer, and sibling-to-sibling types [43]. Based on these relationships, Gao proposed the valley-free principle, which states that AS-level routing paths typically follow policy-compliant patterns [29]. In this study, the inferred AS relationships provide policy-related constraints for path inference, particularly for methods that rely on valley-free assumptions or relationship-aware path construction.

3.2.3 AS Rank Data

We use CAIDA’s AS Rank dataset dated 1 October 2023 [44]. AS Rank orders ASes primarily based on customer-cone size, with additional attributes such as AS degree, country, and organization information. In this paper, these attributes are used in GMPI to characterize the relative position and structural importance of ASes, supporting both path generation and AS representation learning.

3.3 Selected Path Inference Approaches

Three approaches are selected to conduct the validation study: GMPI, KnownPath, and ProbInfer. GMPI, proposed in [15], represents the state-of-the-art path inference approach. On the other hand, KnownPath, introduced in [9], is a classic path inference approach.

GMPI: GMPI is the state-of-the-art path inference approach, which involves a generative and measurable path inference process. GMPI first generates paths for an AP using a heuristic algorithm. Then, some AS paths related to the AP are obtained from the collected data to provide latent routing preferences of the AP. Finally, a dual attention neural network is utilized to extract features from AS paths and estimate the likelihood of these generated paths. After that, the path with the highest likelihood is selected as the inferred path.

GMPI uses the BGP routing tables, the AS relationship data, and the AS Rank data to accomplish the path inference. In this paper, the parameters of GMPI are set to the same values as in the previous work [15].

KnownPath: It is a classic and widely used path inference approach, which is derived from the Bellman-Ford algorithm. For a destination prefix p and a set of ASes A={a1,…,an}, suppose that the path between ai and p is known. Let A′={a1′,…,am′} be the set of ASes for which the path between ai′ and p is unknown. The KnownPath approach first constructs an AS-level Internet topology using BGP data. Then, it expands the paths between ASes in A and p to obtain the paths between ASes in A′ and p based on the constructed topology and the AS relationships.

KnownPath uses the BGP routing tables and the AS relationship data to accomplish the path inference.

ProbInfer: ProbInfer is a data-driven path inference approach, which obtains inferred paths by stitching path segments and reducing redundancy with a decision-tree model. ProbInfer uses the BGP routing tables and the AS relationship data to accomplish the path inference.

3.4 Metrics

3.4.1 Metrics for Quantitative Validation

For an AP, the inference accuracy of its inferred path is defined as the Jaccard similarity between the inferred path Pinf and the ground-truth path Pgt. The Jaccard similarity is calculated as follows:

Jaccard(Pinf,Pgt)=|Pinf∩Pgt||Pinf∪Pgt|(1)

where Pinf and Pgt denote the sets of ASes in the inferred and ground-truth paths, respectively.

Three metrics are used to measure the performance of path inference approaches in quantitative validation: upper bound accuracy (UBA), average accuracy (AA), and exact same ratio (ESR).

Upper Bound Accuracy (UBA) measures the best achievable accuracy when multiple candidate paths are generated. For an AP, given k candidate paths {Pi,1inf,…,Pi,kinf}, UBA is defined as the maximum Jaccard similarity among these candidates. This metric reflects the upper bound performance of the path generation process.

Average Accuracy (AA) is defined as the average Jaccard similarity across all APs in a snapshot:

AA=1|U|∑i∈UJaccard(Piinf,Pigt)(2)

where U denotes the set of APs in the evaluation dataset.

Exact Same Ratio (ESR) is defined as the fraction of APs whose inferred paths exactly match the ground-truth paths:

ESR=|{i∈U∣Piinf=Pigt}||U|(3)

3.4.2 Metrics for Temporal Drift Analysis

To characterize how inference results deviate over time, we analyze temporal drift relative to a baseline snapshot t0. For each subsequent snapshot tx, we compute the following three complementary metrics based on the distribution of Jaccard similarities.

Mean Shift measures the directional change in average similarity relative to the baseline:

Δμ(tx)=S¯tx−S¯t0(4)

Cohen’s d measures the standardized effect size of this change:

d(tx)=S¯tx−S¯t0sp(tx),sp(tx)=(n−1)σt02+(n−1)σtx22n−2(5)

First-order Wasserstein Distance measures the overall distributional difference between similarity distributions at t0 and tx.

These three metrics are reported together because they capture complementary aspects of temporal drift: mean shift reflects absolute change, Cohen’s d captures standardized effect size, and Wasserstein distance characterizes global distributional deviation.

3.5 Quantitative Analysis of Temporal Degradation

To observe the daily and weekly performance degradation of the path inference approaches, a quantitative analysis is conducted. With the data from 1 October 2023, the paths between unmeasured APs are inferred for all 20 snapshots and validated using the ground-truth paths from each snapshot. The AA and ESR of the path inference approaches are calculated for each snapshot. For GMPI, the UBA of the path generation process is also observed.

We intentionally use a fixed baseline snapshot in this analysis. This design isolates how inference results derived from one reference state degrade as the network evolves over time, without introducing additional variation from changing the training snapshot itself. It also reflects a practical deployment scenario in which path inference results or models trained at time t0 are reused for some period before being refreshed.

Accordingly, the reported degradation curves should be interpreted as fixed-baseline temporal behavior under the selected baseline setting, rather than as a claim that the same degradation pattern must hold for every possible starting date. The purpose of this section is to characterize how quickly performance decays when the baseline is held constant, whereas the question of whether the same trend is invariant across multiple baseline dates is a separate robustness issue discussed in Section 5.

3.6 Temporal Drift Analysis of Path Inference

To understand how long path inference results derived from a baseline snapshot remain valid, we analyze the temporal drift of path similarity over time. Unlike conventional hypothesis testing that focuses on whether two samples are statistically different, our goal is to determine after how many days the inference results generated from a baseline snapshot t0 become no longer representative of the current network state.

This formulation aligns with the notion of concept drift in data-driven systems, where previously learned knowledge may become outdated due to distributional changes in the underlying system.

For each AP i, let pi(t0) denote the path inferred using data collected at time t0, and let gi(tx) denote the ground-truth path observed at time tx. The inference accuracy at time tx is defined as:

Si(tx)=J(pi(t0),gi(tx)),(6)

where J(⋅,⋅) denotes the Jaccard similarity defined in Section 3.4.1. The distribution of {Si(tx)} characterizes the validity of stale inference results at time tx.

Drift quantification. To quantify the difference between the baseline distribution S(t0) and the distribution at time tx, we use three complementary metrics defined in Section 3.4.2: mean shift, Cohen’s d, and the first-order Wasserstein distance.

Let S(t0)={Si(t0)}i=1n and S(tx)={Si(tx)}i=1n be the sampled similarity sets for a repeated run, where n is the sample size. Let S¯t and σt denote the sample mean and standard deviation at time t, respectively.

These metrics capture complementary aspects of temporal drift: mean shift reflects absolute directional change, Cohen’s d captures standardized effect size, and Wasserstein distance quantifies global distributional differences beyond mean-based statistics.

From statistical difference to material drift. Given the large number of APs, even small differences may appear consistently across repeated samples. Therefore, instead of relying solely on statistical significance, we jointly consider both the uncertainty of the mean shift and the magnitude of distributional drift.

We say that the stale inference results at time tx become materially different from those at t0 when both:

•   the 95% empirical interval of the mean shift over repeated samples excludes zero, and

•   the Wasserstein distance exceeds a practical threshold.

Expiration time. We define the expiration time T∗ of inference results derived from t0 as:

T∗=min{tx:S(tx) is materially different from S(t0)}.(7)

This time point provides an operational estimate, under the current decision rule, of when stale path inference results are no longer representative of the current network state.

Implementation details. In our experiments, each snapshot contains more than 106 APs. For computational efficiency and to reflect realistic query scenarios, we repeatedly sample 5000 APs from the common AP set between the baseline and target snapshots. This sampling process is repeated 50 times for each target snapshot.

We report the mean and 95% empirical interval of the mean shift over repeated samples, together with the average Cohen’s d and Wasserstein distance across repeated samples. We use 0.01 as a practical Wasserstein cutoff to distinguish very small distributional changes from operationally meaningful drift in this study, rather than as a universal statistical constant. A snapshot is considered to exhibit material drift when the 95% empirical interval of the mean shift excludes zero and the average Wasserstein distance exceeds 0.01. While different practical cutoffs may change the exact estimated expiration day, they do not change the overall ordering of the three methods in our results.

4  Validation Results

The validation process is as follows: the training dataset (70% of the AS paths, the AS relationship data, and the AS Rank data) obtained from 1 October 2023 is used to infer paths for the unmeasured APs for all 20 snapshots. For GMPI, we observe not only the final path inference performance but also the path generation performance.

We use the same parameter settings for GMPI as in the original paper [15] to ensure that the observed temporal behavior reflects the method’s performance under its intended configuration: the input/output layer size is set to d=512, the learning rate is set to lr=0.00001, dropout is set to 0.1 in the position encoding layer, and the maximum number of observed positive paths is set to m=200. KnownPath and ProbInfer are also configured as in the original paper.

4.1 Quantitative Analysis Results

The performance of the three path inference approaches over time is shown in Tables 2–4. From the tables, ProbInfer shows the largest performance degradation, followed by GMPI, while KnownPath exhibits the smallest degradation over the examined period.

To further observe the temporal behavior of the path generation process in GMPI, the UBA and ESR of the path generation process are calculated, as shown in Figs. 1 and 2. The degradation of the path generation process is relatively small, suggesting that both the path generation and path selection components contribute to the overall temporal degradation of GMPI.

Figure 1: Average UBA of GMPI’s path generation process across snapshots.

Figure 2: ESR of GMPI’s path generation process across snapshots.

KnownPath: The validation results of KnownPath are shown in Table 3. KnownPath has the smallest performance degradation among the three methods, with a decrease of 2.09% and 4.47% in accuracy and ESR over eight weeks, respectively. This relatively small decrease suggests that many AS-level paths remain stable over the examined period.

ProbInfer: The validation results of ProbInfer are shown in Table 4. ProbInfer has the most significant performance degradation among the three methods, with a decrease of 7.44% and 8.23% in accuracy and ESR over eight weeks, respectively. This observation indicates that ProbInfer is less temporally robust than the other two evaluated methods in our dataset.

In the quantitative analysis, several snapshots show slight local rebounds relative to the immediately previous date, such as 20231008, 20231012, and 20231013. Similar local fluctuations also appear for ProbInfer on 20231028 and for KnownPath on 20231104. These observations suggest that the degradation process is not strictly monotonic at every snapshot. However, the current evidence is not sufficient to infer any broader recurring Internet-wide routing pattern. Their underlying drivers require dedicated longitudinal analysis and remain part of our future work.

The above results indicate that the performance of path inference approaches changes over time. In the current evaluation, KnownPath is the most temporally robust of the three methods, GMPI is intermediate, and ProbInfer exhibits the largest degradation.

4.2 Temporal Drift Analysis Result

The temporal drift analysis provides an operational view of when stale inference results cease to be representative of the current network state. The results are shown in Tables 5–7.

Among the three approaches, KnownPath exhibits the slowest temporal drift, GMPI shows a moderate drift rate, and ProbInfer drifts the fastest. Under the current decision rule, material drift is first detected on day 20 for KnownPath, on day 10 for GMPI, and on day 2 for ProbInfer. This ordering is consistent with the quantitative analysis, where KnownPath shows the smallest long-term degradation and ProbInfer shows the largest.

The temporal drift pattern of ProbInfer is also less monotonic than those of GMPI and KnownPath. For example, its drift is material on day 2, not material on day 3 under the current threshold, and becomes material again in subsequent snapshots.

5  Discussion

While this study provides a systematic evaluation of the temporal degradation and drift of path inference approaches, several limitations and potential sources of bias should be considered when interpreting the results.

Limited data and observation bias: Our analysis relies on BGP routing-table data from Route Views and RIPE RIS, which provide control-plane observations from a limited set of vantage points (VPs). As a result, the observed AS paths may not fully reflect data-plane forwarding behavior, and the coverage of routing dynamics depends on the spatial distribution of VPs. This limitation may lead to incomplete visibility of certain routing changes, particularly in less well-observed regions of the Internet.

Evaluation set construction bias: To ensure temporal consistency, we construct a common AP set by intersecting APs across all snapshots. While this design enables controlled temporal comparison, it may introduce a bias toward APs that remain consistently observable over time. Such APs are more likely to correspond to relatively stable routing scenarios, which may lead to a conservative estimation of temporal drift in more dynamic parts of the network.

Single-baseline sensitivity: The current validation uses 1 October 2023 as the only baseline snapshot. In operational Internet routing data, limited abnormal events or temporary policy shifts may occur on many dates, and the selected baseline date is unlikely to be completely free of such local fluctuations. However, because our evaluation is performed on a very large AP set and the degradation curves are observed across multiple subsequent snapshots, limited local anomalies are more likely to act as noise or mild bias than to dominate the aggregate trend, unless they affect a substantial portion of the routing state. Our manual review of publicly reported Internet operational conditions did not identify evidence of a large-scale Internet-wide abnormal event on 1 October 2023. Nevertheless, using a single baseline date still restricts the robustness and generalizability of the conclusions, because it does not test whether the same degradation pattern would be reproduced under other starting dates. A stronger design would repeat the analysis from multiple independent baseline snapshots and compare the resulting curves directly.

Limited method coverage: The evaluation focuses on three representative methods, namely GMPI, KnownPath, and ProbInfer. While these methods span several important inference styles, the conclusions may not fully generalize to approaches that rely heavily on active measurements or other fundamentally different inference paradigms.

Limited mechanism-level interpretation: The current study evaluates temporal degradation and drift at the method level, but it does not include a dedicated structural analysis of failed paths. Therefore, the present results do not support causal attribution of the observed degradation pattern to any specific internal mechanism of ProbInfer. Future work should analyze failed-path structures directly to examine such mechanism-level explanations.

Granularity limitations: This study focuses on AS-level path inference, which provides a coarse-grained abstraction of inter-domain routing. Such abstraction does not capture finer-grained routing dynamics at the prefix level, PoP level, or IP level, where routing decisions may differ due to traffic engineering, load balancing, or intra-AS policies. As a result, temporal variations observed at finer granularities may not be fully reflected in our evaluation.

Despite these limitations, the consistent temporal trends observed across large-scale datasets and multiple complementary metrics suggest that our findings provide a meaningful characterization of the temporal degradation and drift of path inference approaches under the evaluated setting. These limitations also highlight important directions for future work, including incorporating data-plane measurements, expanding method coverage, and exploring finer-grained path representations.

6  Conclusion

In this paper, we explore the temporal degradation and drift of path inference approaches through eight weeks of real Internet path data, three representative path inference methods, and quantitative analysis and temporal drift analysis. This provides preliminary evidence that can inform the deployment and update of path inference services under the evaluated setting.

Our experiments show that path inference methods do degrade over time, but the magnitude of degradation differs across methods and remains measurable rather than severe over the examined eight-week period. The temporal drift analysis further provides an operational reference for refresh scheduling by indicating when stale inference results are no longer sufficiently representative of the current network state under the evaluated setting.

Acknowledgement: The authors would like to express their gratitude to the editors and reviewers for their detailed review and insightful advice.

Funding Statement: This work is supported by the National Natural Science Foundation of China (No. 62472434), the Key Program of NSFC Hunan (2026JJ30028), and the China Postdoctoral Science Foundation (2023TQ0089).

Author Contributions: Xionglve Li conceived the study, designed the methodology, implemented the experiments, analyzed the data, and drafted the manuscript. Changsheng Hou contributed to the study design, supervised the research, and revised the manuscript. Yuzhou Huang and Zhenyu Qiu contributed to data preparation, experiment execution, and result validation. Gang Hu, Bingnan Hou, and Wei Dong contributed to the interpretation of the results and manuscript revision. Zhiping Cai contributed to the overall research design, supervision, and final revision of the manuscript. All authors reviewed and approved the final version of the manuscript.

Availability of Data and Materials: Not applicable.

Ethics Approval: Not applicable.

Conflicts of Interest: The authors declare no conflicts of interest.

References

1. Banerjee R, Razaghpanah A, Chiang L, Mishra A, Sekar V, Choi Y, et al. Internet outages, the eyewitness accounts: analysis of the outages mailing list. In: Passive and active measurement. Cham, Switzerland: Springer International Publishing; 2015. p. 206–19. doi:10.1007/978-3-319-15509-8_16. [Google Scholar] [CrossRef]

2. Lutu A, Bagnulo M, Pelsser C, Maennel O, Cid-Sueiro J. The BGP visibility toolkit: detecting anomalous Internet routing behavior. IEEE/ACM Trans Netw. 2016;24(2):1237–50. doi:10.1109/tnet.2015.2413838. [Google Scholar] [CrossRef]

3. Sundaresan S, Deng X, Feng Y, Lee D, Dhamdhere A. Challenges in inferring Internet congestion using throughput measurements. In: Proceedings of the 2017 Internet Measurement Conference; 2017 Nov 1–3; London, UK. p. 43–56. doi:10.1145/3131365.3131382. [Google Scholar] [CrossRef]

4. Prehn L, Foremski P, Gasser O. Kirin: hitting the Internet with distributed BGP announcements. In: Proceedings of the 19th ACM Asia Conference on Computer and Communications Security; 2024 Jul 1–5; Singapore. p. 19–34. doi:10.1145/3634737.3657000. [Google Scholar] [CrossRef]

5. Stöger F, Birge-Lee H, Giuliari G, Subira-Nieto J, Perrig A. BGP vortex: update message floods can create internet instabilities. In: Proceedings of the 34th USENIX Security Symposium (USENIX Security 25); 2025 Aug 13–15; Seattle, WA, USA. p. 3613–29. [Google Scholar]

6. Testart C, Clark DD. A data-driven approach to understanding the state of internet routing security. In: Proceeding of the Research Conference on Communication, Information and Internet Policy; 2020 Feb 17–19; Washington, DC, USA. doi:10.2139/ssrn.3750155. [Google Scholar] [CrossRef]

7. Jin Y, Renganathan S, Ananthanarayanan G, Jiang J, Padmanabhan VN, Schroder M, et al. Zooming in on wide-area latencies to a global cloud provider. In: Proceedings of the ACM Special Interest Group on Data Communication; 2019 Aug 19–23; Beijing, China. p. 104–16. doi:10.1145/3341302.3342073. [Google Scholar] [CrossRef]

8. Gray C, Mosig C, Bush R, Pelsser C, Roughan M, Schmidt TC, et al. BGP beacons, network tomography, and Bayesian computation to locate route flap damping. In: Proceedings of the ACM Internet Measurement Conference; 2020 Oct 27–29; Virtual. p. 492–505. doi:10.1145/3419394.3423624. [Google Scholar] [CrossRef]

9. Qiu J, Gao L. As path inference by exploiting known as paths. In: Proceedings of the Global Communications Conference; 2005 Nov 28–Dec 2; St. Louis, MO, USA. [Google Scholar]

10. Madhyastha HV, Isdal T, Piatek M, Dixon C, Anderson T, Krishnamurthy A, et al. iPlane: an information plane for distributed services. In: Proceedings of the Symposium on Operating Systems Design and Implementation; 2006 Nov 6–8; Seattle, WA, USA. p. 367–80. [Google Scholar]

11. Madhyastha HV, Katz-Bassett E, Anderson TE, Krishnamurthy A, Venkataramani A. iPlane nano: path prediction for peer-to-peer applications. In: Proceedings of the Symposium on Networked System Design and Implementation; 2009 Apr 22–24; Boston, MA, USA. p. 137–52. [Google Scholar]

12. Cunha Í, Marchetta P, Calder M, Chiu Y-C, Machado BVA, Pescapè A, et al. Sibyl: a practical internet route oracle. In: Proceedings of the Symposium on Networked System Design and Implementation; 2016 Mar 16–18; Santa Clara, CA, USA. p. 325–44. [Google Scholar]

13. Li X, Cai Z, Hou B, Liu N, Liu F, Cheng J. ProbInfer: probability-based AS path inference from multigraph perspective. Comput Netw. 2020;180(6):107377. doi:10.1016/j.comnet.2020.107377. [Google Scholar] [CrossRef]

14. Jin Z, Shi X, Ma Q, Sun L, Wang Z, Yin X, et al. Which way to go? Inferring fine-grained AS paths with PathRadar. In: Proceedings of the IEEE INFOCOM 2025—IEEE Conference on Computer Communications; 2025 May 19–22; London, UK. p. 1–10. doi:10.1109/infocom55648.2025.11044440. [Google Scholar] [CrossRef]

15. Li X, Zhou T, Cai Z, Su J. Realizing fine-grained inference of AS path with a generative measurable process. IEEE/ACM Trans Netw. 2023;31(6):3112–27. doi:10.1109/tnet.2023.3270565. [Google Scholar] [CrossRef]

16. Li X, Wang C, Yang T, Shen A, Qiu Z, Hou B, et al. Realizing personalized and adaptive inference of AS paths with a generative and measurable process. IEEE Trans Netw. 2025;33(2):729–44. doi:10.1109/tnet.2024.3506156. [Google Scholar] [CrossRef]

17. Li X, Wang C, Yang Y, Hou C, Hou B, Cai Z. Internet inter-domain path inferring: methods, applications, and future directions. Comput Mater Contin. 2024;81(1):53–78. doi:10.32604/cmc.2024.055186. [Google Scholar] [CrossRef]

18. Green T, Lambert A, Pelsser C, Rossi D. Leveraging inter-domain stability for BGP dynamics analysis. In: Passive and active measurement. Cham, Switzerland: Springer International Publishing; 2018. p. 203–15. doi:10.1007/978-3-319-76481-8_15. [Google Scholar] [CrossRef]

19. Bakhshaliyev K, Canbaz MA, Gunes MH. Investigating characteristics of Internet paths. ACM Trans Model Perform Eval Comput Syst. 2019;4(3):1–24. doi:10.1145/3342286. [Google Scholar] [CrossRef]

20. Comarela G, Gürsun G, Crovella M. Studying interdomain routing over long timescales. In: Proceedings of the 2013 Conference on Internet Measurement Conference; 2013 Oct 23–25; Barcelona, Spain. p. 227–34. doi:10.1145/2504730.2504771. [Google Scholar] [CrossRef]

21. Augustin B, Cuvellier X, Orgogozo B, Viger F, Friedman T, Latapy M, et al. Avoiding traceroute anomalies with Paris traceroute. In: Proceedings of the 6th ACM SIGCOMM Conference on Internet Measurement; 2006 Oct 25–27; Brazil: Rio de Janeiro. p. 153–8. doi:10.1145/1177080.1177100. [Google Scholar] [CrossRef]

22. Vermeulen K, Gurmericliler E, Cunha I, Choffnes D, Katz-Bassett E. Internet scale reverse traceroute. In: Proceedings of the 22nd ACM Internet Measurement Conference; 2022 Oct 25–27; Nice, France. p. 694–715. doi:10.1145/3517745.3561422. [Google Scholar] [CrossRef]

23. Liu N, Jia C, Hou B, Hou C, Chen Y, Cai Z. 6Search: a reinforcement learning-based traceroute approach for efficient IPv6 topology discovery. Comput Netw. 2023;235(5):109987. doi:10.1016/j.comnet.2023.109987. [Google Scholar] [CrossRef]

24. Zhao J, Shi F, Xu C, Peng J, Ge M, Xue P, et al. TNet: efficient IPv6 active network discovery. Comput Netw. 2026;280(14s):112170. doi:10.1016/j.comnet.2026.112170. [Google Scholar] [CrossRef]

25. Zhang B, Bi J, Wang Y, Zhang Y, Wu J. Refining IP-to-AS mappings for AS-level traceroute. In: Proceedings of the 2013 22nd International Conference on Computer Communication and Networks (ICCCN); 2013 Jul 30–Aug 2; Nassau, Bahamas. p. 1–7. doi:10.1109/icccn.2013.6614180. [Google Scholar] [CrossRef]

26. Luckie M, Dhamdhere A, Huffaker B, Clark D, Claffy K. Bdrmap: inference of borders between IP networks. In: Proceedings of the 2016 Internet Measurement Conference; 2016 Oct 26–28; Santa Monica, CA, USA. p. 381–96. doi:10.1145/2987443.2987467. [Google Scholar] [CrossRef]

27. Marder A, Smith JM. MAP-IT: multipass accurate passive inferences from traceroute. In: Proceedings of the 2016 Internet Measurement Conference; 2016 Oct 26–28; Santa Monica, CA, USA. p. 397–411. doi:10.1145/2987443.2987468. [Google Scholar] [CrossRef]

28. Marder A, Luckie M, Dhamdhere A, Huffaker B, Claffy K, Smith JM. Pushing the boundaries with bdrmapIT: mapping router ownership at Internet scale. In: Proceedings of the Internet Measurement Conference 2018; 2018 Oct 31–Nov 2; Boston, MA, USA. p. 56–69. doi:10.1145/3278532.3278538. [Google Scholar] [CrossRef]

29. Gao L. On inferring autonomous system relationships in the Internet. IEEE/ACM Trans Netw. 2001;9(6):733–45. doi:10.1109/90.974527. [Google Scholar] [CrossRef]

30. Singh R, Tench D, Gill P, McGregor A. PredictRoute: a network path prediction toolkit. In: Proceedings of the Abstract Proceedings of the 2021 ACM SIGMETRICS/International Conference on Measurement and Modeling of Computer Systems; 2021 Jun 14–18; Virtual. p. 21–2. doi:10.1145/3410220.3460107. [Google Scholar] [CrossRef]

31. Tao N, Chen X, Fu X. AS path inference: from complex network perspective. In: Proceedings of the 2015 IFIP Networking Conference (IFIP Networking); 2015 May 20–22; Toulouse, France. p. 1–9. doi:10.1109/ifipnetworking.2015.7145303. [Google Scholar] [CrossRef]

32. Friedman JH, Popescu BE. Predictive learning via rule ensembles. Ann Appl Stat. 2008;2(3):916–54. doi:10.1214/07-aoas148. [Google Scholar] [CrossRef]

33. Wu T, Wang JH, Wang J, Zhuang S. RouteInfer: inferring interdomain paths by capturing ISP routing behavior diversity and generality. In: Passive and active measurement. Cham, Switzerland: Springer International Publishing; 2022. p. 216–44. doi:10.1007/978-3-030-98785-5_10. [Google Scholar] [CrossRef]

34. Jin Y, Scott C, Dhamdhere A, Giotsas V, Krishnamurthy A, Shenker S. Stable and practical AS relationship inference with ProbLink. In: Proceedings of the 16th USENIX Symposium on Networked Systems Design and Implementation (NSDI 19); 2019 Feb 26–28; Boston, MA, USA. p. 581–98. [Google Scholar]

35. Jin Z, Shi X, Yang Y, Yin X, Wang Z, Wu J. TopoScope: recover AS relationships from fragmentary observations. In: Proceedings of the ACM Internet Measurement Conference; 2020 Oct 27–29; Virtual. p. 266–80. doi:10.1145/3419394.3423627. [Google Scholar] [CrossRef]

36. Prehn L, Feldmann A. How biased is our validation (data) for AS relationships? In: Proceedings of the 21st ACM Internet Measurement Conference; 2021 Nov 2–4; Virtual. p. 612–20. doi:10.1145/3487552.3487825. [Google Scholar] [CrossRef]

37. Peng S, Shu X, Ruan Z, Huang Z, Xuan Q. Classifying multiclass relationships between ASes using graph convolutional network. Front Eng Manag. 2022;9(4):653–67. doi:10.1007/s42524-022-0217-1. [Google Scholar] [CrossRef]

38. Shi X, Jin Z, Xiong B, Huang X, Xi X, Li D, et al. HELA: inferring AS relationships with a hybrid of empirical and learning algorithms. IEEE Trans Netw. 2025;33(5):2648–63. doi:10.1109/ton.2025.3572148. [Google Scholar] [CrossRef]

39. RouteViews. Oregon route views project. 2026 [cited 2026 Jan 1]. Available from: http://www.routeviews.org. [Google Scholar]

40. RIPENCC. RIPE RIS. 2026 [cited 2026 Jan 1]. Available from: http://www.ripe.net/ris/. [Google Scholar]

41. AS relationships (serial-1). 2026 [cited 2026 Jan 1]. Available from: https://catalog.caida.org/dataset/as∖_relationships∖_serial∖_1. [Google Scholar]

42. Center for applied internet data analysis. 2026 [cited 2026 Jan 1]. Available from: https://www.caida.org/. [Google Scholar]

43. Huston G. Interconnection, peering and settlements. Internet Protoc J. 1999;2(1):1–29. [Google Scholar]

44. CAIDA. AS rank. 2026 [cited 2026 Jan 1]. Available from: http://as-rank.caida.org/. [Google Scholar]