Evolution of Communities with Focus on Stability (extended abstract)

The detection of communities is an important tool used to analyze the social graph of mobile phone users. Within each community, customers are susceptible of attracting new ones, retaining old ones and/or accepting new products or services through the leverage of mutual influences. The communities of users are smaller units, easier to grasp, and allow for example the computation of role analysis – based on the centrality of an actor within his community. The problem of finding communities in static graphs has been widely studied. However, from the point of view of a telecom analyst, to be really useful, the detected communities must evolve as the social graph of communications changes over time – for example, in order to perform marketing actions on communities and track the results of those actions over time. Additionally the behaviors of communities of users over time can be used to predict future activity that interests the telecom operators, such as subscriber churn or handset adoption. Similary group evolution can provide insights for designing strategies, such as the early warning of group churn. Stability is a crucial issue: the analysis performed on a given community will be lost, if the analyst cannot keep track of this community in the following time steps. This is the particular use case that we tackle in this paper: tracking the evolution of communities in dynamic scenarios with focus on stability. We propose two modifications to a widely used static community detection algorithm. We then describe experiments to study the stability and quality of the resulting partitions on real-world social networks, represented by monthly call graphs for millions of subscribers.

💡 Research Summary

The paper addresses a practical problem faced by telecom analysts: how to track communities of mobile‑phone users over time while preserving the stability of the partitions. Community detection on static graphs is well studied, but when the underlying call‑graph evolves month by month, the communities produced by standard algorithms such as Louvain can change dramatically, making it impossible to follow the same group of users across successive periods. This instability hampers downstream tasks such as role analysis, targeted marketing, churn prediction, or handset‑adoption studies.

To tackle this, the authors use a large, anonymized call‑and‑SMS dataset from a mobile operator covering January 2012 to January 2013. For each month T they build a graph G_T by aggregating traffic from the three months T, T‑1, and T‑2. After symmetrizing the edges (keeping only bidirectional interactions) and discarding nodes with degree > 200 (to remove call‑centers and service numbers), each monthly snapshot contains roughly 56 million nodes and 133 million undirected edges.

The baseline method is the Louvain community detection algorithm, optionally enhanced by the Aynaud‑Guillaume dynamic extension. The authors find that even with this extension the partitions between consecutive months are highly unstable. They therefore propose two simple modifications that give the previous month’s partition a “momentum” when processing the next snapshot.

1. Fixed‑node probability p – A fraction p of the nodes that exist in both G_T and G_{T+1} is randomly selected and forced to retain the community label they had at time T. Formally, for the set F ⊆ R (where R = N_T ∩ N_{T+1}), |F| = p·|R|, and for every x ∈ F, γ_{T+1}(x) = γ_T(x). This creates a “hard” anchor for a portion of the network.

2. Preferential‑attachment probability q – A fraction q of the nodes (including newcomers) is allowed to preferentially attach to an existing community from time T if it has at least one neighbor already belonging to that community. During the first pass of Louvain’s node‑moving phase, such a node x ∈ P (|P| = q·|N_{T+1}|) is assigned to the neighboring community that yields the highest modularity gain, provided the set A(x) of eligible neighboring communities is non‑empty.

The authors evaluate the impact of p and q by varying each from 0 to 1 in steps of 0.25. For each pair (p, q) they compute three metrics on the partitions of two consecutive months: (i) Normalized Mutual Information (NMI) between the two partitions (restricted to the intersecting node set), (ii) the number of “matching” communities, defined as pairs of communities whose overlap exceeds 51 % of both community sizes, and (iii) the modularity of the new partition.

Results:

• NMI rises sharply with increasing p, reaching a plateau near p = 1. This indicates that fixing more nodes dramatically improves the similarity between successive partitions. The effect of q is noisy; it neither consistently raises nor lowers NMI.
• Matching communities follow a similar pattern: the count grows dramatically as p approaches 1, confirming that many communities can be followed across months. Small values of q provide a modest boost when p is low, but the overall influence remains weak.
• Modularity shows a modest decline as p increases (typically a 1–2 % drop), reflecting the trade‑off between stability and the quality of the partition. Higher q values introduce additional fluctuations and a slight downward trend, but the effect is less systematic than that of p.

The authors conclude that the fixed‑node mechanism (parameter p) is an effective and low‑cost way to enforce temporal stability, with only a minor penalty in modularity. The preferential‑attachment mechanism (parameter q) does not provide a clear benefit and may be omitted in practice.

In the discussion, the authors note that setting p = 1 essentially forces the new partition to be identical to the previous one, which could mask genuine structural changes. Future work is proposed along three lines: (i) using finer temporal granularity (e.g., weekly snapshots) to capture more subtle dynamics, (ii) validating the approach on publicly available dynamic network benchmarks to enable reproducibility, and (iii) refining the matching criteria and explicitly modeling community events such as birth, death, merge, split, expansion, and contraction.

Overall, the paper contributes a pragmatic solution for telecom analysts who need to monitor evolving user groups over time. By introducing a simple “momentum” concept into the widely used Louvain algorithm, it demonstrates that stability can be substantially improved without sacrificing much community quality, thereby supporting downstream marketing, churn‑prediction, and product‑adoption analyses in real‑world, large‑scale communication networks.