Statistically validated mobile communication networks: Evolution of motifs in European and Chinese data

Motivation

• Mobile call records provide large-scale information about communication, mobility, communities, and temporal behavior across diverse populations.Their coverage and metadata make them useful proxies for studying social-network structure and dynamics.Source: §I · p. 1

Research questions

• Static motifs in temporal networks change with the aggregation window, motivating questions about which motifs are over-represented, how they emerge, and their characteristic timescales.Time-stamped mobile call data are particularly suited to examining this evolution.Source: §I · p. 2
• Directed 3-motifs are studied because they are socially informative while their reliable estimation requires strict processing conditions for very large samples.The study compares two providers in Europe and China to assess the approach across differing commercial, geographic, and temporal settings.Source: §I · p. 2

Motivation

• Spurious records, commercial activity, wrong numbers, and call centers can create links and hubs that are misread as social relationships, distorting motif statistics.The problem is amplified by fat-tailed communication-network activity and degree distributions.Source: §I · pp. 2–3

Approach

• The Bonferroni network statistically validates directed links, removes extreme non-social hubs, and retains many unidirectional contacts unlike simpler mutual-link filters.Earlier filtering could unintentionally remove socially relevant unidirectional or short-duration links, while local backbone filtering has structural restrictions.Source: §I · p. 3

Contribution

• Filtered networks support motif analysis showing intraday triadic closure and evolution of open triads toward greater reciprocity.The paper argues that unfiltered hubs make 3-motif statistics unreliable.Source: §I · p. 3

Statistical validation

• Mobile-call entries are evaluated as probabilistic proxies for real social interactions before constructing communication networks.The validation problem is framed as multiple hypothesis testing.Source: §II · p. 3

Chinese data

• The Chinese dataset contains 109 recorded days, 4,031,090 subscribers, and 1,091,695,590 calls, with 88.24% involving non-subscribers.The analysis retains 128,410,897 calls among subscribers.Source: A. Data · p. 4

European data

• The European dataset covers 7,387,034 subscribers over 212 days and includes 3,969,043,426 calls, with 81.54% connecting subscribers to non-subscribers.The analysis includes 682,124,009 calls between subscribers.Source: A. Data · p. 4

Scope

• The study analyzes subscriber-only calling networks because calls between two non-subscribers are not recorded, which would alter 3-motif statistics.This defines the dataset boundary for motif analysis.Source: A. Data · p. 4

Statistical validation

• Bonferroni validation tests each directed link against a null model of heterogeneous calling profiles and retains links whose p-values fall below 0.01/NE.The filtered network is called the Bonferroni Network.Source: B. Statistically Validated Networks · pp. 4–5
• Figure 1 compares p-value distributions for Chinese and European links, separating all links and links with weights 1, 3, 5, and 10 against the Bonferroni threshold.Links to the left of the threshold are retained, using the entire observation period and subscriber-to-subscriber links only.Source: Figure 1 · p. 5, B. Statistically Validated Networks · p. 5

Results

• Links with one call are generally excluded from both Bonferroni networks, while links with weights above five are always retained under these analyses.Some links with weights up to five are also filtered out.Source: B. Statistically Validated Networks · p. 6

Statistical validation

• Validation depends on more than average link weight, because links with higher weights can remain unvalidated under the statistical test.The absence of validation is not directly determined by a small average number of calls per link.Source: B. Statistically Validated Networks · p. 6

Threshold choice

• The threshold is set conservatively at 0.01/NE, with NE equal to the number of observed edges, and its choice reflects the intended network comparisons.For short daily and weekly intervals, this choice is described as rather conservative.Source: B. Statistically Validated Networks · p. 6
• Increasing the daily Chinese Bonferroni threshold by one order of magnitude raises the average number of validated links by 4.8%.The authors report that threshold changes within broad limits do not crucially affect network composition across time intervals.Source: B. Statistically Validated Networks · p. 6

Basic metrics

• Bonferroni networks retain roughly one third of the original links, while their daily largest connected component is negligible compared with 12% of nodes in the original European network.For longer aggregation windows, the Bonferroni largest component grows toward the original network values as weaker interconnecting ties are progressively detected.Source: C. Basic metrics and degree distribution · p. 7, Table I · p. 7
• In daily European networks, the Bonferroni network’s 3-motif count fluctuates less than the original network’s and lacks the huge spike at day 79.The authors hypothesize that Bonferroni filtering retains links with more pronounced social motivations and makes the Bonferroni network a better proxy for the underlying social network.Source: C. Basic metrics and degree distribution · p. 7, Figure 2 · p. 8

Degree distribution

• In Chinese data, original subscribers reach out-degree around 3,000 and in-degree around 700, whereas Bonferroni maximum degree is about 200 with more similar tails.The original out-degree tail is substantially more pronounced than the in-degree tail, unlike the filtered network.Source: C. Basic metrics and degree distribution · pp. 7–8
• European degree distributions show the same broad filtering pattern, but the original network has a more pronounced in-degree tail while the Bonferroni network retains a stronger in-degree tail than out-degree.The European Bonferroni network has maximum in-degree near 150 and maximum out-degree near 250, excluding the outlier at 1131.Source: C. Basic metrics and degree distribution · pp. 8–9, Figure 3 · p. 9
• European Bonferroni cumulative degree tails are described by power-law exponents of 3.85 for in-degree and 6.25 for out-degree, although the latter is difficult to distinguish from exponential decay.The large out-degree exponent makes the tail decay rapidly and limits discrimination between the two decay forms.Source: C. Basic metrics and degree distribution · p. 9

III. 3-MOTIFS

• The analysis examines 13 directed 3-motif types as communication motifs in a social network, using a specific mobile-phone operator’s network as a proxy and Milo et al. labeling.Source: §III · p. 10, Figure 4 · p. 10
• Bonferroni networks show more stable daily motif fractions, while original networks exhibit stronger seasonalities and localized spikes.The contrast is especially pronounced for motifs with only two of the three possible pair relationships.Source: §III · p. 10, Figure 5 · p. 11, Figure 6 · p. 12
• Triadic-closure motifs are over-expressed relative to degree-preserving randomized networks across daily, weekly, and monthly scales in both datasets.Motifs with links among all three node pairs exceed random expectations, whereas several motifs involving only two communicating pairs are near or below expectation.Source: §III · p. 10
• Original and Bonferroni networks preserve the qualitative motif pattern but differ substantially in average motif fractions, especially over weekly and monthly windows.The difference is particularly pronounced for the first motif set, including motif 6 in monthly networks.Source: §III · p. 10
• The filtered results are judged more robust and reliable than the original-network results, enabling more detailed analysis of motif formation and disappearance.The subsequent temporal analysis therefore focuses on Bonferroni networks.Source: §III · p. 11

IV. TEMPORAL EVOLUTION OF COMMUNICATION 3-MOTIFS

• Conditional persistence is highest when the motif remains unchanged after a one-day network expansion, with probabilities ranging from 0.395 to 0.916 in China and 0.418 to 0.966 in Europe.Motif 98 is least stable, whereas motif 238 is most stable in both datasets.Source: §IV · p. 13
• Open motifs preferentially evolve into bidirectional open motifs rather than closed motifs, indicating that return calls dominate the observed transitions.The relevant transition probabilities are associated with converting a unidirectional link into a bidirectional one.Source: §IV · p. 13
• Only fully reciprocated open motif 78 tends to evolve toward a closed triad, while most open-to-closed transitions remain limited.This supports an intraday formation timescale for closed communication triads.Source: §IV · p. 13
• More than two-thirds of closed motifs observed on a given day arise from unconnected triples on the preceding day, while about one-quarter form from open motifs.The reported pattern is consistent with many closed motifs appearing through contingent intraday communication.Source: §IV · pp. 13–14
• Extending the detection window can occasionally remove links that were validated in the shorter window, although such disappearances are uncommon.This affects the interpretation of motif evolution during network expansion.Source: §IV · p. 13

• Bonferroni networks provide a more realistic proxy for social ties than original communication networks by removing spurious links and stabilizing motif frequencies over time.This conclusion is supported across two large mobile communication datasets from different countries.Source: §V · p. 14
• Closed communication 3-motifs frequently form on an intraday time scale.Source: §V · p. 14
• Open 3-motifs primarily evolve into other open motifs with more reciprocated calls, and can close when all their calls are reciprocated.Source: §V · p. 14
• The results suggest that communication interactions typically unfold within days, whereas the underlying social structure forms and dissolves over months or years.The authors therefore distinguish the rules governing triadic closure in communication and social networks.Source: §V · p. 14