🤖 machine learning

MINT: Tensor Decomposition on Stacked Recurrence Matrices for Time Series Data Mining

This paper introduces MINT, a time series data mining method that utilizes tensorized self-similarity matrices derived from stacked recurrence plots to effectively identify co-clustered patterns and motifs across diverse univariate and multivariate datasets.

Original authors: Kaamil Kaka, Audrey Der, Evangelos E. Papalexakis, Zachary Zimmerman, Vikram Jayaram

Published 2026-08-06

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kaamil Kaka, Audrey Der, Evangelos E. Papalexakis, Zachary Zimmerman, Vikram Jayaram

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Hidden Rhythm of the World

Imagine you are a detective trying to solve a mystery, but instead of looking at fingerprints or footprints, you are looking at the heartbeat of the world. This is the job of time series data mining, a branch of science dedicated to finding patterns in data that changes over time. Think of it as listening to a song where the notes are numbers: the temperature rising and falling, the number of people on a subway, or the wind speed at a turbine.

To understand the new idea in this paper, we first need to know about two tools detectives already use. The first is a recurrence plot. Imagine taking a long strip of music and laying it over itself. Where the melody repeats, you get a dot. If you do this for the whole song, you get a map of all the repeating rhythms. The second tool is tensor decomposition. If a regular map is a flat piece of paper, a tensor is like a 3D block of cheese. Decomposition is the act of slicing that block in a very specific way to find the hidden layers inside, revealing the "ingredients" that make up the whole.

Why does this matter? Because in our modern world, we have thousands of sensors watching everything at once. When a subway station gets crowded, or a wind turbine starts acting weird, it's rarely just one sensor screaming. It's a chorus. The big question is: How do we hear the chorus as a single, meaningful song rather than just a bunch of random noise?

The Paper's Big Idea: MINT

The paper you are reading introduces a new detective tool called MINT (which stands for Mplots INto Tensor). The authors, a team of data scientists, are tackling a specific headache: when dozens of sensors record data at the same time, how do we find the patterns that happen across them, not just inside one?

The Problem with Looking at One Sensor at a Time

Imagine you are watching a stadium full of people. If you look at just one person, you might see them clap. If you look at another, you might see them stand up. But if you look at the whole crowd, you might realize they are all doing a "wave."

Previous methods often looked at one sensor at a time. They would find a pattern in Sensor A, then a pattern in Sensor B, and hope they matched up. The paper argues that this is like trying to understand a symphony by listening to one violin at a time. You miss the harmony. The authors show that if you try to find these cross-sensor patterns using standard math on raw data, you often miss the subtle "shape" of the event. You might know when something happened, but not what it looked like across the whole group.

The New Trick: Stacking the Maps

MINT changes the game by changing what it looks at. Instead of staring at the raw numbers, MINT first turns every single sensor's data into a Recurrence Plot (or what the paper calls an Mplot).

Think of an Mplot as a "selfie" of a sensor's history. It's a grid that shows how much the sensor's past looks like its future. If the sensor has a repeating pattern (like a heartbeat), the selfie shows bright, diagonal lines. If it's chaotic, the selfie is a mess of dots.

Here is the magic step: MINT takes these "selfies" from all the sensors and stacks them on top of each other to form a giant 3D block (a tensor). It's like taking a stack of transparent sheets, each with a different drawing, and holding them up to the light to see the combined image.

What They Found

Once they built this 3D block, they used a mathematical technique called tensor decomposition to slice it open. This didn't just give them a list of numbers; it gave them three clear pictures:

Who is involved (which sensors).
When it happens (the time intervals).
What the pattern looks like (the shape of the event).

The authors tested this on real-world data from four very different places:

Subway Stations in Taipei: They found patterns related to the Lunar New Year, where the whole system moved in sync.
Traffic in California: They spotted how traffic jams rippled across thousands of sensors.
Wind Turbines in Portugal: They identified when a turbine was about to break, spotting the "flatline" of a dying machine across multiple sensors.
Electricity Grids in Europe: They separated "Christmas" patterns (a short, specific spike) from "Winter" patterns (a long, seasonal trend), even though both happened at the same time of year.

In a controlled experiment where they mixed real data with random noise, MINT successfully ignored the noise and found the real, shared patterns 92% to 100% of the time. This suggests the method is robust and doesn't just get fooled by random glitches.

Why This Is a Big Deal

The paper doesn't claim to have invented a new way to slice the math (the decomposition part is old news). Instead, they invented a new object to slice: the stacked Mplot.

They argue that looking at raw data is like trying to find a face in a pile of sand. Looking at MINT's stacked Mplots is like sifting the sand to find the gold nuggets first. By turning the data into these "selfie maps" before analyzing them, MINT makes the hidden connections between sensors visible and interpretable.

The authors are careful to say this is a "primitive"—a basic building block for future tools. They aren't saying this solves every problem in the world, but they have shown that it works better than looking at sensors one by one, especially when you need to understand how a whole system moves together. It turns a chaotic chorus of sensors into a clear, readable score.

Technical Summary: MINT – Tensor Decomposition on Stacked Recurrence Matrices for Time Series Data Mining

1. Problem Statement

The paper addresses the challenge of discovering cross-series co-clustering patterns in large collections of univariate time series. While existing methods like the Matrix Profile and Recurrence Plots (RPs) effectively identify motifs and discords within a single time series, they lack a scalable mechanism to analyze structural alignment across multiple series simultaneously.

Current approaches face specific limitations:

Visual Inspection: While Mplots (Matrix Profile plots) allow practitioners to visually identify patterns in individual series, this does not scale to datasets with tens or hundreds of series.
Lossy Compression: Methods like the Matrix Profile compress self-similarity structures into nearest-neighbor distance vectors, discarding the rich pairwise distance information required to detect subtle phase relationships or shape similarities across series.
Lack of Interpretability: Deep learning approaches often rely on learned representations that obscure the specific sensors and time intervals contributing to a pattern.
Subsequence Clustering Limitations: Standard subsequence clustering often yields random results due to the overlapping nature of sliding windows, failing to distinguish genuine cross-series co-occurrence from sampling artifacts.

The core question posed is: When a discordant phenomenon appears across dozens of sensors, is it $N$ anomalies or one shared event? Which sensors share not just the timestamp, but the specific shape of the event?

2. Methodology: The MINT Pipeline

The authors propose MINT (Mplots INto Tensor), a pipeline that treats time series data mining as a tensor decomposition problem. The method constructs a tensorized self-similarity matrix as a new data mining primitive.

Step 1: Mplot Computation
For a dataset $D$ containing $N$ time series of length $n$ , the method computes an Mplot for each series. An Mplot is a matrix of Euclidean distances between all pairs of subsequences of length $m$ .

To manage computational complexity, Mplots are pooled into a smaller $k \times k$ window using max-pooling (for similarity) or min-pooling (for distance), reducing the dimensionality while retaining structural information.

Step 2: Tensor Construction
The $N$ pooled Mplots are stacked along a new axis to form a third-order tensor $\mathcal{T}$ of dimensions $N \times k \times k$ .

Preprocessing: Before stacking, the authors apply Nonnegative Robust PCA to each Mplot to decompose it into a low-rank matrix $L$ and a sparse matrix $S$ .
Mean-Centering: The low-rank matrices are mean-centered. This is a critical step to remove the "global self-similarity" baseline of each sensor, ensuring the tensor decomposition focuses on deviations and co-occurring patterns rather than sensors that are simply more self-similar overall.

Step 3: Tensor Decomposition
The resulting tensor $\mathcal{T}$ is decomposed using Canonical Polyadic Decomposition (CPD) (also known as PARAFAC).

The decomposition yields three factor matrices: $A$ (sensor mode), $B$ (temporal mode 1), and $C$ (temporal mode 2).
The rank $R$ is estimated automatically using the Kneedle algorithm on the reconstruction error curve.
The resulting components represent "archetypal patterns" where specific sensors (rows in $A$ ) exhibit specific temporal structures (columns in $B$ and $C$ ) simultaneously.

3. Key Contributions

New Primitive: The paper introduces the tensorized self-similarity matrix as a novel primitive for time series mining. Unlike previous work that treats tensors as raw data or uses RPs only for classification, this work treats the stacked Mplot as a mineable object for structure discovery.
MINT Pipeline: The authors instantiate this primitive with a concrete pipeline that combines Robust PCA, pooling, and CPD to extract cross-series patterns without requiring labels.
Interpretability: The method provides explicit, interpretable mappings. Each component directly indexes the specific sensors and time intervals participating in a shared pattern, avoiding the "black box" nature of deep learning representations.
Validation of Co-Clustering: The paper provides an existence proof that tensor decomposition on self-similarity matrices can recover genuine periodic structure and phase offsets that are indistinguishable in raw time-space analysis (demonstrated via synthetic data with known phase shifts).

4. Experimental Results

The authors evaluated MINT on four diverse datasets:

Taipei MRT: 108 stations (ridership data).
LargeST: 8,303 traffic sensors (California).
Electricity Load Diagrams: 332 clients (quarter-hourly consumption).
CARE to Compare: 43 sensors from 5 wind turbines.

Validation Protocol:
To test robustness, the authors performed a noise injection experiment. They created synthetic datasets where specific subsets of sensors were transformed to be "Completely Random," "Mostly Random," or "Mostly Normal" around specific disjoint windows. They then tested if MINT could correctly group these transformed sensors into distinct components.

Findings:

Hypothesis Testing: Across 50 trials on each dataset, the method satisfied four distinct validation properties ( $P1$ $P 1$ – $P4$ $P 4$ ) with high frequency (often 100%).
- $P1$ : MINT identified patterns aligned with the injected windows.
- $P2$ : Components correctly grouped sensors transformed by the same noise type.
- $P3$ & $P4$ : The method correctly isolated random noise and identified the dominant "Mostly Normal" patterns.
Statistical Significance: The probability of these results occurring by random chance was extremely low (e.g., $p < 0.001$ for most properties), rejecting the null hypothesis that the co-clustering is random.
Case Study (OPSD): On the Electrical Power Demand dataset, MINT successfully separated distinct cross-sensor patterns, such as Christmas holidays (short-term spikes) and seasonal demand (long-term winter peaks), which would be difficult to disentangle using single-series analysis.
Wind Turbine Data: The method identified time-aligned degenerate flatlines (sensor failures) across multiple turbines, demonstrating utility in anomaly detection.

5. Significance and Claims

The paper claims that MINT offers a scalable, interpretable alternative to both visual inspection and deep learning for cross-series structure discovery.

Modest Claims: The authors explicitly state that MINT does not propose a new decomposition algorithm. Instead, its novelty lies in the object being decomposed (the tensor of self-similarity matrices).
Structure Discovery: The method is framed as a "structure discovery" tool rather than a prediction tool. It recovers latent co-clustering information (e.g., phase offsets, shared motifs) that is not naturally representable in per-series analysis.
Generalizability: The primitive is noted to be naturally extensible to multivariate datasets, though the current work focuses on univariate series treated as independent sensors.

The authors conclude that this approach enables the analysis of cross-series temporal structure at a level of granularity previously limited to what a practitioner could see "by eye" in individual Mplots, but now applicable to large-scale, automated mining.