Integrating Forgetting Curves into Sequential Knowledge Tracing

203 concepts134 papers1 proposal

Standard sequential models—such as Recurrent Neural Networks (RNNs), LSTMs, and Transformers—excel at processing sequences where the primary relationship between elements is order. In these architectures, step $t+1$ immediately follows step $t$. This discrete, index-based assumption works well for text or code, where the distance between words is structural rather than temporal.

However, when applied to human learning, this assumption breaks down. Human memory does not operate on discrete sequence indices; it operates in continuous time. If a student practices a math concept, waits five minutes, and practices it again, their retention is vastly different than if they waited five months.

Traditional Knowledge Tracing (KT) models often treat student interaction histories as simple ordered lists. As highlighted by Ghodai Abdelrahman et al. in their comprehensive survey, assuming knowledge is permanent or purely cumulative fails to capture the realistic dynamics of human forgetting. When a sequence model ignores the variable time intervals ($\Delta t$) between interactions, it risks severely overestimating a student's current mastery after long periods of inactivity.

To fix this, a major research direction focuses on embedding cognitive psychology principles—specifically mathematical representations of the Ebbinghaus forgetting curve—directly into the architectures of sequential neural networks.

The Core Mechanism: Time-Interval Decay

To reconcile sequence models with human memory, researchers modify the internal state transitions of neural networks by introducing a decay factor driven by the elapsed time $\Delta t$. Instead of passing an unattenuated hidden state $h_t$ or memory slot to the next step, the network scales the historical knowledge representation before it influences the current prediction.

Standard RNN:      [h_t] ───────────────────────────────> [h_{t+1}]

Time-Decay RNN:    [h_t] ───> [Apply Decay f(Δt)] ──────> [h_{t+1}]

As Quanlong Guan et al. demonstrate in KVFKT, this decay can be explicitly modeled using the Ebbinghaus power law function. The function applies a forgetting rate directly to a stored knowledge state, simulating the universal human trait of rapid initial forgetting followed by gradual stabilization.

The mathematical implementation of this decay typically falls into two categories:

  1. Multiplicative Decay: The historical state is multiplied by a non-linear function of time, such as $e^{-\gamma \Delta t}$ or $t^{-\gamma}$. Effat Farhana et al. note in Predictive Student Modelling in an Online Reading Platform that while additive terms can reduce computational overhead, multiplicative decay functions offer superior biological plausibility and more accurately capture human forgetting patterns.
  2. Continuous-Time States: Rather than applying a step-wise penalty, some architectures allow the hidden state to evolve continuously during periods of inactivity. Haiping Ma et al. pioneer this in CT-NCM by leveraging a neural Hawkes process. This approach models learning as a continuous-time process, capturing both the instantaneous knowledge gains from active study and the gradual, continuous erosion of proficiency between sessions.

Capturing the Spacing Effect via Time Windows

Human memory is also sensitive to how practice is distributed. The spacing effect dictates that review sessions spaced further apart yield more durable long-term retention than "massed" sessions (cramming).

To capture this without relying on raw interaction counts, Benoît Choffin et al. introduced the DAS3H framework (DAS3H: Modeling Student Learning and Forgetting for Optimally Scheduling Distributed Practice of Skills). Instead of treating time as a single scalar, DAS3H tracks a student's past attempts and successes within discrete, expanding time windows (e.g., last 24 hours, last week, last month).

By structuring memory features into temporal bins, the model can explicitly differentiate between a student who completed 10 problems in the last hour versus one who completed 10 problems distributed over a month. This multi-scale temporal granularity allows the system to identify when a student's failure on a complex task is due to the decay of a specific underlying sub-skill.

Similarly, Hongjing Lu demonstrates in Modeling the spacing effect in sequential category learning that incorporating conjugate temporal priors allows cognitive models to simulate how mental representations drift or become less certain during periods of inactivity, providing a elegant explanation for why spaced exposure stabilizes memory.

Why It Works: Aligning Inductive Biases with Cognitive Reality

Integrating forgetting curves works because it injects a powerful, theoretically sound inductive bias into data-driven architectures.

Purely sequential models like standard LSTMs or Transformers have to learn the concept of forgetting entirely from data. If a dataset contains sparse interactions with massive time gaps, a standard network struggles to learn that an interaction from six months ago should be heavily discounted. By explicitly forcing the network's hidden states or attention weights to decay as a function of $\Delta t$, researchers constrain the hypothesis space to match human biology.

This architectural alignment solves several distinct problems across educational data mining:

  • Preventing Mastery Jumps: As shown by JJ Cui et al. in GRKT, combining graph neural networks with explicit forgetting curves prevents deep learning models from making nonsensical "mastery jumps," ensuring that predicted proficiency evolves logically and realistically over time.
  • Disentangling Prerequisite Gaps: Yanxia Pang et al. demonstrate in Prerequisite-related MOOC recommendation that educational systems often falsely assume a student is prepared for advanced material based solely on historical course completion. A time-decay function reveals when foundational knowledge has faded, allowing recommendation engines to prescribe targeted reviews rather than blindly pushing students forward.
  • Balancing Non-Recurrent Frameworks: Even outside of recurrent neural networks, accounting for temporal decay is vital. Hisashi Kashima notes in Knowledge Tracing Machines that static factorization methods which rely on simple success/failure counts fail to capture dynamic learning states, highlighting the universal need for hybrid architectures that marry structural efficiency with sequential, time-aware decay.

Ultimately, by treating time as a continuous variable rather than a sequence index, these models transform from rigid data-loggers into dynamic simulators of human cognition.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 130 papers
  2. Version 1 · 2026-10-04 · 154 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_temporal_knowledge_decay_bdb901be,
  title = {temporal knowledge decay},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-knowledge-decay/rdr_bdb901be38c641fcbf50f9ccc44f877a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 2, frozen text):

@misc{lacuna_temporal_knowledge_decay_bdb901be_v2,
  title = {temporal knowledge decay},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-knowledge-decay/rdr_bdb901be38c641fcbf50f9ccc44f877a/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake