Abstract:Time series data from various domains is continuously growing, and extracting and analyzing temporal patterns within these series can provide valuable insights. Temporal pattern mining (TPM) extends traditional pattern mining by incorporating event time intervals into patterns, making them more expressive but also increasing the computational complexity in terms of time and space. One important type of temporal pattern is the rare temporal pattern (RTP), which occurs infrequently but with high confidence. Mining these rare patterns poses several challenges, for example, the low support threshold can lead to a combinatorial explosion and the generation of many irrelevant patterns. To address this, an efficient approach to mine rare temporal patterns is essential. This paper introduces the Rare Temporal Pattern Mining from Time Series (RTPMfTS) method, designed to discover rare temporal patterns. The key contributions of this work are as follows: (1) An end-to-end RTPMfTS process that takes time series data as input and outputs rare temporal patterns. (2) A highly efficient Rare Temporal Pattern Mining (RTPM) algorithm, which leverages optimized data structures for fast event and pattern retrieval, as well as effective pruning techniques to accelerate the mining process. (3) A comprehensive experimental evaluation of RTPM, demonstrating that it outperforms the baseline in both runtime and memory efficiency.

What problem does this paper attempt to address?

### What problem does this paper attempt to solve? This paper aims to solve the problem of mining rare temporal patterns (RTPs) from time - series data. Specifically, the author focuses on those temporal patterns that have a low occurrence frequency but high confidence in time - series data. Although such patterns are not common, they may be of great value for certain application scenarios (such as traffic management in smart cities, medical health monitoring, etc.). #### Background and Challenges 1. **Growth of Time - Series Data**: The amount of time - series data from Internet of Things (IoT) sensors is constantly increasing. This data contains rich information, but effective tools are required to extract and analyze the temporal patterns within it. 2. **Complexity of Temporal Pattern Mining**: Compared with traditional pattern mining, temporal pattern mining needs to consider not only the occurrence order of events but also the time intervals between events, which increases the computational complexity and difficulty. 3. **Characteristics of Rare Temporal Patterns**: Rare temporal patterns are characterized by low support but high confidence, which means that although they do not occur often, they are often of great significance when they do occur. However, this low support also leads to the combinatorial explosion problem, generating a large number of irrelevant patterns and making the mining process more difficult. #### Solutions To solve the above problems, the author proposes a method named "Rare Temporal Pattern Mining from Time Series (RTPMfTS)", and the main contributions of this method include: 1. **End - to - End RTPMfTS Process**: This process takes time - series data as input and outputs rare temporal patterns. 2. **Efficient RTPM Algorithm**: - Use optimized data structures (such as hierarchical hash tables) to achieve fast event and pattern retrieval. - Utilize the Apriori principle and the transitivity of temporal relationships for effective pruning to accelerate the mining process. 3. **Experimental Evaluation**: Through extensive experimental verification, it is proved that RTPM is superior to the baseline methods in terms of running time and memory usage. #### Formula Representation - **Support**: Used to measure the frequency of a pattern. \[ \text{supp}(P) = |\{S \in DSEQ \mid P \in S\}| \] - **Confidence**: Used to measure the possibility of a pattern. \[ \text{conf}(P) = \frac{\text{supp}(P)}{\max_{1 \leq k \leq |P|} \{\text{supp}(E_k)\}} \] #### Conclusion By proposing the RTPMfTS method, the author has successfully solved the problem of efficiently mining rare temporal patterns from time - series data, providing a valuable tool for fields such as smart cities and medical health.

Efficient Rare Temporal Pattern Mining in Time Series

Efficient Generalized Temporal Pattern Mining in Big Time Series Using Mutual Information

Rare Category Exploration Via Wavelet Analysis: Theory and Applications

NTP-Miner: Nonoverlapping Three-Way Sequential Pattern Mining

Mining frequent temporal duration-based patterns on time interval sequential database

Mining High-utility Temporal Patterns on Time Interval–based Data

OPP-Miner: Order-Preserving Sequential Pattern Mining for Time Series

Sequential Pattern Mining in Databases with Temporal Uncertainty

Mining Scalable Pattern Based on Temporal Logic over Data Streams

Data analysis in real-time supply chain based on frequent pattern mining

A Two-Phase Approach for Unexpected Pattern Mining.

Real-time POI recommendation via modeling long- and short-term user preferences

Repetitive nonoverlapping sequential pattern mining

Addressing Big Data Time Series: Mining Trillions of Time Series Subsequences Under Dynamic Time Warping

Utility-Driven Mining of Trend Information for Intelligent System

OPR-Miner: Order-preserving rule mining for time series

Towards Efficient Sequential Pattern Mining in Temporal Uncertain Databases

Discovering During-Temporal Patterns (dtps) in Large Temporal Databases

Scalable Order-Preserving Pattern Mining

TaSPM: Targeted Sequential Pattern Mining

Efficient Discovery of Periodic Patterns over Event Sequences