Overview of streaming-data algorithms

Due to recent advances in data collection techniques, massive amounts of data are being collected at an extremely fast pace. Also, these data are potentially unbounded. Boundless streams of data collected from sensors, equipments, and other data sources are referred to as data streams. Various data mining tasks can be performed on data streams in search of interesting patterns. This paper studies a particular data mining task, clustering, which can be used as the first step in many knowledge discovery processes. By grouping data streams into homogeneous clusters, data miners can learn about data characteristics which can then be developed into classification models for new data or predictive models for unknown events. Recent research addresses the problem of data-stream mining to deal with applications that require processing huge amounts of data such as sensor data analysis and financial applications. For such analysis, single-pass algorithms that consume a small amount of memory are critical.

💡 Research Summary

**
The paper provides a broad overview of clustering algorithms that can be applied to data streams, which are defined as continuously arriving, potentially unbounded sequences of time‑series observations generated by sensors, financial systems, multimedia sources, and other real‑time applications. The authors begin by contrasting static data sets—where data can be stored and queried repeatedly—with streams that must be processed on the fly because storing the entire volume is infeasible. They argue that streaming environments impose two fundamental constraints: algorithms must operate in a single pass over the data and must use only a small, bounded amount of memory.

In the definitions section the paper formalizes a time‑series as a sequence S = {s₁,…,sₙ} and a time‑series dataset D = {S₁,…,Sₘ}. A streaming scenario is modeled as m data sources (e.g., geographic sensors) each emitting n observations over time, where n can increase indefinitely. The authors stress that a streaming clustering system must continuously ingest points from all sources, update its model incrementally, and never rely on the full historical record.

The core of the manuscript reviews several classic clustering techniques and evaluates their suitability for streaming contexts:

1. k‑medoids family (PAM, CLARA, CLARANS) – These partition‑based methods select actual data points as cluster representatives. PAM is exact but requires O(n²) distance calculations, making it unsuitable for large streams. CLARA reduces cost by sampling a subset of the data, applying PAM to the sample, and then assigning the remaining points to the nearest medoid. While this lowers memory usage, the quality of the clustering heavily depends on the representativeness of the sample, and important clusters may be missed when the stream size far exceeds the sample size. CLARANS further randomizes the neighbor search to avoid local minima, yet it still relies on a fixed number of neighbor evaluations (MAX_neigh) and solution attempts (MAX_sol), parameters that are difficult to tune for an unbounded stream.

2. k‑means – The most widely used partitional algorithm, k‑means iteratively updates centroids to minimize within‑cluster sum‑of‑squares. Its simplicity and fast convergence make it attractive, but it assumes a static data set, requires the number of clusters k a priori, and suffers from the “curse of dimensionality” when applied to high‑dimensional time‑series streams. Moreover, the Euclidean distance metric used in k‑means is sensitive to noise and scaling, which are common in streaming sensor data.

3. CLARA (Clustering Large Applications) – Described separately from the k‑medoids family, CLARA also relies on sampling. The algorithm draws a small subset, runs PAM, and then evaluates the clustering cost on the entire data set. The authors note that when the total number of objects n is much larger than the maximum manageable size M, CLARA’s sampling may overlook entire clusters, leading to poor quality results.

4. CLARANS (Clustering Large Applications based upon Randomized Search) – This method searches a random subset of the neighbor space of a current solution, governed by two parameters: MAX_neigh (maximum neighbors examined) and MAX_sol (maximum solutions explored). While it can escape local minima better than CLARA, the need to set these parameters and the lack of incremental update mechanisms limit its applicability to true streaming scenarios.

5. BIRCH (Balanced Iterative Reducing and Clustering using Hierarchies) – BIRCH builds a CF‑Tree, a height‑balanced hierarchical structure that incrementally absorbs incoming points, compresses them into clustering features (CF = (N, LS, SS)), and optionally performs a global clustering step on the leaf entries. Its linear time complexity O(N) and ability to operate with a single pass make it the most promising among the surveyed algorithms for streaming data. However, the performance of BIRCH is highly sensitive to the branching factor B and the threshold T that control tree size; inappropriate settings can cause the tree to overflow memory or to produce overly coarse clusters. The authors also mention an optional refinement step that re‑reads the raw data to improve cluster quality, which may be impractical in strict streaming environments.

6. CURE (Clustering Using REpresentatives) – CURE is a hierarchical algorithm that selects c well‑scattered points from each cluster, shrinks them toward the cluster centroid by a factor α, and uses these shrunken points as representatives. This approach enables detection of non‑spherical clusters and improves robustness to outliers. Nevertheless, the choice of c and α is heuristic, and maintaining and updating the representative set for each cluster in a continuously arriving stream can be computationally expensive.

Throughout the review, the paper emphasizes that most of these classic algorithms were originally designed for static, batch‑processed data sets. To adapt them to streams, researchers typically employ sampling, incremental data structures, or multiple passes, each of which introduces trade‑offs between memory consumption, clustering quality, and processing latency.

The manuscript, however, lacks several critical components. It does not present any experimental evaluation on real or synthetic streams, nor does it compare the surveyed methods against modern streaming‑specific algorithms such as DenStream, CluStream, D‑Stream, or StreamKM++. Consequently, the reader cannot gauge the practical performance of the discussed techniques under realistic constraints like concept drift, variable arrival rates, or distributed execution. Moreover, the paper does not address windowing strategies (sliding or tumbling windows), drift detection mechanisms, or integration with contemporary stream processing frameworks (Apache Flink, Spark Structured Streaming). The omission of these topics limits the paper’s relevance to current practitioners.

In conclusion, the article serves as a high‑level taxonomy of traditional clustering algorithms and highlights the challenges of applying them to data streams—namely, the need for single‑pass processing, bounded memory, and incremental updates. While the overview is useful for newcomers, the lack of recent literature, experimental validation, and discussion of practical deployment issues means that the paper falls short of providing actionable guidance for modern streaming data mining. Future work should incorporate rigorous benchmarking against dedicated stream clustering methods, explore adaptive windowing and drift handling, and consider scalability in distributed environments.