CENDRe Synthetic Benchmarks for Time-Series Concept Extraction This collection contains 47 synthetic binary-classification time-series datasets designed to evaluate concept extraction (CE) methods for 1D CNNs against known ground truth. Each variant provides per-sample binary masks of every class-discriminative primitive (and its complementary "absence"), enabling quantitative assessment of representation correctness and importance correctness in both the time and frequency domains. The datasets are organized into four families. Shape ch1 (1 channel, 15 variants) Localized shape primitives (square, triangle, half-circle), each of length 40 timesteps, injected at random non-overlapping positions into a noisy sinusoidal base. - Single-shape vs. nothing (3 variants). One class contains a single shape, the other only the noisy base.- Paired (3 variants). Each class is characterized by a different shape.- Disjunctive vs. nothing (3 variants). One class contains either of two shapes (chosen at random per sample), the other only the noisy base.- Paired with confounder (6 variants). The two paired bases are augmented with the third (unused) shape as a confounder at three class-confounder ratios (40/60, 50/50, 60/40). Shape ch2 (2 channels, 9 variants) Two-channel version of Shape ch1 for testing per-channel attribution. - Single-shape on all channels (3 variants). The shape is present on both channels of the positive class.- Single-shape on random channel (3 variants). The shape is injected into a uniformly chosen channel; tests channel-agnostic detection.- Specific-channel paired (3 variants). Each of two class-discriminative shapes is assigned to a designated channel, one shape per channel. Frequency ch1 (1 channel, 14 variants) Class-discriminative spectral peaks in three bands - mid (70-75 Hz), mid-high (125-130 Hz), high (170-175 Hz) - injected into the rFFT of a frequency-domain base with random phases, then transformed back to the time domain. - Pure paired (3 variants). Each class is characterized by a different band.- Paired with confounder (5 variants). Paired bases augmented with the third (unused) band; the midFreq-highFreq base sweeps three ratios (40/60, 50/50, 60/40), the other two paired bases use only the 50/50 ratio.- Single-band vs. nothing (3 variants). One class contains a band, the other only the noisy base.- Single-band with confounder (3 variants). midFreq-nothing augmented with highFreq as a confounder at three ratios. Frequency ch2 (2 channels, 9 variants) Two-channel extension of Frequency ch1. - Specific-channel paired (3 variants) and paired with 50/50 confounder (3 variants). Each of two class-discriminative bands is assigned to a designated channel.- Single-band vs. nothing (3 variants). Single band injected on a designated channel. Shared generation parameters - Sample length T: 400- Samples per variant: 1,000- Class balance: 50/50- Sampling frequency: 1,000 Hz for Shape families, 400 Hz for Frequency families- Primitive overlap within a sample: not allowed Ground truth Binary masks are provided per primitive in both time-domain (length T = 400) and frequency-domain (length F = floor(T/2) + 1 = 201) localizations, plus complementary "absence" masks that treat the absence of every involved primitive as a primitive of its own, since a model can learn "absence of X" as a discriminative feature just as readily as its presence. Format All datasets are stored as HuggingFace Dataset objects. The mapping from human-readable variant name to dataset path is provided in INDEX.md accompanying this upload.
Anonymous (Wed,) studied this question.