Temporal Tensor Factorization for Evolving Relational Data

45 concepts35 papers1 proposal

When analyzing relational data—such as social networks, transaction logs, or knowledge bases—we often want to understand not just who is interacting with whom, but when those interactions occur and how they change over time.

Traditional approaches collapse these interactions into a static two-dimensional matrix. However, flattening the temporal dimension destroys crucial sequence information and mixes long-term structural patterns with short-term fluctuations. Alternatively, slicing the data into independent time steps prevents us from tracking the global identity of entities across intervals.

To solve this, researchers treat time as an explicit, higher-order dimension within a multi-dimensional array, or tensor. By factorizing this tensor, we can isolate structural relationships from their temporal dynamics, providing a mathematically grounded way to predict future links, detect anomalies, and track community evolution.


Representing Time as a Tensor Dimension

To understand how this works, imagine a dynamic graph of user interactions. Instead of a single adjacency matrix, we stack a sequence of adjacency matrices chronologically to form a three-dimensional tensor of dimensions $\text{Node} \times \text{Node} \times \text{Time}$. As Panisson et al. (2014) and Molaei et al. (2019) demonstrate, this multi-way structure preserves the exact temporal sequence of edges without losing the overarching network topology.

      Static Matrix                     Temporal Tensor
   (Time is collapsed)               (Time as an explicit axis)

       User B                             User B
       ┌───┐                              ┌───┐
User A │ 1 │                       User A │ 1 │ ───► Time Step 3
       └───┘                              └───┼───┐
                                          │ 0 │ 1 │ ───► Time Step 2
                                          └───┼───┼───┐
                                              │ 1 │ 0 │ ───► Time Step 1
                                              └───┴───┘

This concept scales naturally to more complex data structures. For instance, a temporal knowledge graph contains facts represented as triples: $(\text{Subject}, \text{Predicate}, \text{Object})$. As surveyed by Cai et al. (2023), adding time transforms this into a fourth-order tensor ($\text{Subject} \times \text{Predicate} \times \text{Object} \times \text{Time}$). Similarly, Qian et al. (2021) show that adding an extra "update" dimension to a spatio-temporal tensor allows models to explicitly learn systematic reporting delays over time.


Isolating the Temporal Signature

Once data is structured as a tensor, the goal is to decompose it into a set of low-rank latent factors. In a standard CANDECOMP/PARAFAC (CP) decomposition, a three-mode tensor $\mathcal{X}$ is factorized into a sum of vector outer products:

$$\mathcal{X} \approx \sum_{k=1}^{R} \mathbf{u}_k \circ \mathbf{v}_k \circ \mathbf{t}_k$$

Here, $\mathbf{u}_k$ and $\mathbf{v}_k$ represent the latent features of the interacting entities, while $\mathbf{t}_k$ represents the temporal latent factor.

The intuition behind this separation is powerful. Early work by Kolda (2007) highlighted that these higher-order decompositions successfully isolate distinct temporal signatures within communication networks. If $\mathbf{u}_k$ and $\mathbf{v}_k$ map out a community of entities that frequently interact, the corresponding vector $\mathbf{t}_k$ acts as an activity profile, showing when that specific community is active or when its discussion themes shift.

As noted by Cai et al. (2026), this factorization enables the clean separation of short-term, noisy fluctuations from long-term structural trends. It also helps mitigate data sparsity; Tu et al. (2016) demonstrate that by applying non-negativity constraints to the decomposition, the resulting components can be interpreted as global edge-generation probabilities, smoothing out the noise of individual, sparse time snapshots.


Enforcing Temporal Continuity and Smoothness

In a basic tensor decomposition, the time steps along the temporal axis are treated as independent categorical indices. However, physical time has an inherent order and continuity—what happens at time $t$ is highly correlated with what happens at time $t+1$.

To capture this inductive bias, researchers incorporate temporal constraints directly into the factorization framework:

  • Smoothness Priors: The authors of a 2020 study on temporal knowledge bases introduced temporal smoothness regularizers to ensure that latent representations change gradually from one step to the next. This prevents the model from making erratic predictions in data-sparse intervals.
  • Time-Series Constraints: Li et al. (2016) integrated temporal dynamics into a probabilistic tensor framework by applying time-based constraints to the latent factors, significantly boosting predictive accuracy over static baselines.
  • Forecasting Latent Factors: Dunlavy et al. (2010) showed that once the temporal factor vectors ($\mathbf{t}_k$) are isolated, standard time-series forecasting methods (like ARIMA or exponential smoothing) can be applied to project these vectors into the future, allowing for accurate temporal link prediction.

Moving to Continuous Time and Evolving Cores

While CP decomposition assigns a single vector to each mode, Tucker decomposition introduces a "core tensor" that models complex, multi-linear interactions between different latent spaces.

To capture shifting global dynamics, Fang et al. (2022) proposed parameterizing the core tensor itself as a continuous stochastic process (using Gaussian Processes and Stochastic Differential Equations).

   Traditional Tucker Decomposition               Continuous-Time Tucker (BCTT)

       ┌───────────┐                                      ┌───────────┐
       │   Core    │ ──► Static Rules                     │   Core    │ ──► Governed by a
       │  Tensor   │     of Interaction                   │  Tensor   │     Stochastic Process
       └───────────┘                                      └───────────┘     (Changes continuously)
          ▲  ▲  ▲                                            ▲  ▲  ▲
          │  │  │                                            │  │  │
       ┌──┴──┼──┴──┐                                      ┌──┴──┼──┴──┐
       │  Entity   │                                      │  Entity   │
       │  Factors  │                                      │  Factors  │
       └───────────┘                                      └───────────┘

This distinction is crucial for intuition: instead of changing the individual entity features over time, the model changes the global rules of interaction via a time-varying core. This preserves the structural interpretability of the entity factors while allowing the underlying system dynamics to evolve smoothly and continuously.

By treating time as a fundamental axis of multi-linear algebra rather than a sequential index, temporal tensor factorization provides a unified, mathematically rigorous framework for uncovering the hidden, evolving patterns of our interconnected world.

Key Foundational Work

Graph Mining: Patterns, Generators and Tools (2009) — contributed 1 concept elements to this direction.
Data Mining on Attributed Relationship Graphs (ARGs). (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_temporal_tensor_factorization_30d4311f,
  title = {temporal tensor factorization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-tensor-factorization/rdr_30d4311f630545a3b29e786b4e40e09f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_temporal_tensor_factorization_30d4311f_v1,
  title = {temporal tensor factorization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-tensor-factorization/rdr_30d4311f630545a3b29e786b4e40e09f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake