1 of 8

RocketLens

Interactive Visual Analytics for Explaining MiniROCKET Time Series Classifiers

1

Gonzalo Méndez (UPV), Luis Galárraga (Inria), Rodne Quijije (ESPOL), Paul Viallard (Inria)

​

Demo Track at ECML/PKDD 2026

Naples, Italy

2 of 8

Time Series Classification

  • Time series are time-indexed sequences of measurements
  • Many applications rely on categorizing such sequences

​

2

3 of 8

Time Series Classification – Quick Survey

  • Distance-based, e.g., k-NN
  • Kernel methods, e.g., SVM
  • Shapelet-based classifiers
  • Dictionary-based approaches
  • Deep Learning: LSTM, GRU, etc.
  • Convolutional features, e.g., ROCKET
  • Imaging time series, e.g., VLAIR
  • Ensemble methods, e.g., HIVE-COTE

3

4 of 8

The ROCKET Family of Methods

  • Map time series to a latent space using many 1D convolutions
    • Mask weights, dilations, and padding are randomly chosen
    • Compatible with any SOTA classifier

4

  • Derivatives: MiniROCKET, MultiROCKET, etc.​

​

5 of 8

ROCKET classification is a black box

  • Classifier's response relies on many non-interpretable signals
  • What if we look at inside the rocket?
    • The classifier uses some of the ROCKET features
    • Those depend on the time series observations
    • We can trace those features back to their observations

5

MiniROCKET

transformation

6 of 8

RocketLens: Inspecting the Rocket

6

Video Demo

​

7 of 8

RocketLens: Inspecting the Rocket

7

8 of 8

Come to see our demo if you want to see inside the rocket!

8

https://rocketlens.webs.upv.es/

​

Explanation Video

​