Temporal Contrastive Link Prediction

• TCLP is a framework that integrates future link prediction with contrastive objectives to learn robust, time-sensitive graph representations.
• It employs diverse mechanisms, including local/global contrast, hierarchical multi-view alignment, and counterfactual augmentation to regularize predictions.
• Empirical results demonstrate significant gains in AUC, AP, and robustness across various temporal regimes and network types.

Searching arXiv for the cited TCLP-related papers to ground the synthesis. Temporal Contrastive Link Prediction (TCLP) is best understood as a family of temporal link prediction methods in which future edge forecasting is coupled to a contrastive or contrastive-like representation objective over time, structure, or multiple views of the same node, edge, or graph state. In current arXiv literature, direct instantiations include teneNCE, which combines next-step link prediction with graph reconstruction and local/global contrastive predictive coding on dynamic snapshot sequences; CLP, which performs hierarchical contrast over spatial and temporal views in temporal heterogeneous networks; and CoDCL, which augments continuous-time dynamic link prediction with counterfactual pair construction and an InfoNCE-based edge-level objective (Nouranizadeh et al., 2024, Tai et al., 2024, Feng et al., 30 Jan 2026).

1. Scope and problem formulations

TCLP addresses the general temporal link prediction problem: infer whether a link will occur in the future, conditioned on the observed graph history. The literature spans two main temporal regimes. In snapshot-based formulations, the input is a sequence of graphs such as $$\mathcal{G}=\{\mathcal{G}^1,\mathcal{G}^2,\cdots,\mathcal{G}^T\}$$, and the target is link existence in the next snapshot $\mathcal{G}^{T+1}$. CLP follows this formulation for temporal heterogeneous networks, while teneNCE discretizes a continuous-time temporal network into snapshots $G=\{G_1,\dots,G_N\}$ and predicts $\mathbf{A}_{k+1}$ from $G_1,\dots,G_k$ (Tai et al., 2024, Nouranizadeh et al., 2024).

A second regime is continuous-time event modeling. CoDCL defines a dynamic network $G=(\mathcal{V},\mathcal{E},\mathcal{T})$ with timestamped interactions $(u,v,t)\in\mathcal{E}$, historical prefix $\mathcal{H}_t$, and target probability

$P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$

Its prediction target is explicitly a future binary link indicator rather than event-time density estimation or survival modeling (Feng et al., 30 Jan 2026).

The exact phrase “Temporal Contrastive Link Prediction” is not uniformly used as a field label in these papers. This suggests that TCLP is more accurately a technical category than a standardized nomenclature. In substance, however, the category is coherent: temporal graph history is encoded, future links remain the downstream target, and contrastive supervision is used to enforce temporal predictiveness, view consistency, or robustness to structural change (Nouranizadeh et al., 2024, Tai et al., 2024, Feng et al., 30 Jan 2026).

2. Contrastive objectives and learning principles

The defining feature of TCLP is that supervised edge prediction is not the only learning signal. The representation is additionally regularized by a contrastive objective whose positives and negatives are temporally or structurally meaningful.

In teneNCE, the total objective is

$$L = L_{\textrm{pred}} + \alpha L_{\textrm{recon}} + \beta L_{\textrm{cpc}},$$

where $\mathcal{G}^{T+1}$0 is next-step binary cross-entropy on $\mathcal{G}^{T+1}$1, $\mathcal{G}^{T+1}$2 reconstructs current snapshots, and $\mathcal{G}^{T+1}$3 is a Contrastive Predictive Coding term accumulated over all future times $\mathcal{G}^{T+1}$4. The contrastive signal is future-oriented: a current temporal state should discriminate the correct future latent structural embedding from temporally mismatched alternatives, at both node and graph scales (Nouranizadeh et al., 2024).

In CLP, the intended total loss is

$\mathcal{G}^{T+1}$5

with supervised link loss $\mathcal{G}^{T+1}$6, node-level heterogeneity differentiation $\mathcal{G}^{T+1}$7, edge-level heterogeneity differentiation $\mathcal{G}^{T+1}$8, and time-level heterogeneity differentiation $\mathcal{G}^{T+1}$9. The contrast is hierarchical rather than predictive-coding-based: it aligns two spatial encoders at node and edge granularity and aligns long-term and short-term temporal channels (Tai et al., 2024).

In CoDCL, the total loss is

$G=\{G_1,\dots,G_N\}$0

where $G=\{G_1,\dots,G_N\}$1 is the factual link prediction loss and $G=\{G_1,\dots,G_N\}$2 is an InfoNCE-based contrastive loss over factual positive edges, counterfactual edges, and negative sampled edges. The paper’s distinctive claim is that contrastive learning becomes more useful when the augmented views are counterfactually meaningful, rather than arbitrary perturbations (Feng et al., 30 Jan 2026).

A common misconception is to treat any temporal link predictor with negative edge sampling as “contrastive.” The surveyed literature does not support that equivalence. Standard positive/negative edge classification in supervised link prediction is repeatedly distinguished from contrastive learning proper; this distinction is explicit in discussions of non-contrastive temporal baselines and transfer models (Fard et al., 2024, Chatterjee et al., 15 Apr 2025).

3. View construction, positive pairs, and negatives

The technical character of a TCLP method is largely determined by how it defines multiple views of the same temporal entity and how it constructs positives and negatives.

CLP uses architectural multi-view contrast rather than stochastic graph augmentation. At node level, the positive pair is $G=\{G_1,\dots,G_N\}$3: the same node, same relation type, same time, but two spatial encoders. At edge level, the positive pair is $G=\{G_1,\dots,G_N\}$4: the same node and timestamp under two cross-relation fusion views. At time level, the positive pair is $G=\{G_1,\dots,G_N\}$5: the same node under long-term LSTM and short-term GRU channels. Negatives come from other nodes in the corresponding typed neighborhood, global neighborhood, or final snapshot. The similarity function is dot product, and the temperature parameter is $G=\{G_1,\dots,G_N\}$6, with best default $G=\{G_1,\dots,G_N\}$7 (Tai et al., 2024).

teneNCE defines positives and negatives through future latent prediction. At local scale, for node $G=\{G_1,\dots,G_N\}$8, context time $G=\{G_1,\dots,G_N\}$9, and future time $\mathbf{A}_{k+1}$0, the positive pair is $\mathbf{A}_{k+1}$1. Negatives are deliberately heterogeneous: the same node at different times, different nodes at the same time, and different nodes at different times. At global scale, $\mathbf{A}_{k+1}$2 is positive, while summaries from other times are negatives. This makes the contrastive task jointly temporal and identity-sensitive (Nouranizadeh et al., 2024).

CoDCL constructs views through counterfactual matching. It first defines a binary treatment

$\mathbf{A}_{k+1}$3

where $\mathbf{A}_{k+1}$4 and $\mathbf{A}_{k+1}$5 is a percentile threshold over interaction intensities. For a factual pair $\mathbf{A}_{k+1}$6, a counterfactual pair $\mathbf{A}_{k+1}$7 must have opposite treatment, lie within explored $\mathbf{A}_{k+1}$8-hop neighborhoods, and maximize average cosine similarity of contextual node representations. The resulting edge representations $\mathbf{A}_{k+1}$9, $G_1,\dots,G_k$0, and $G_1,\dots,G_k$1 enter an InfoNCE-based contrastive loss. This is not merely temporal augmentation; it is matched augmentation under similar context but opposite temporal treatment (Feng et al., 30 Jan 2026).

These mechanisms reveal three broad TCLP design patterns. One is future-state discrimination, exemplified by teneNCE. Another is multi-view heterogeneity alignment, exemplified by CLP. The third is counterfactual contrast under temporal structural interventions, exemplified by CoDCL. The underlying principle is shared: representations should preserve what remains predictive across time while separating temporally or structurally incompatible alternatives (Nouranizadeh et al., 2024, Tai et al., 2024, Feng et al., 30 Jan 2026).

4. Architectural regimes

Direct TCLP methods differ not only in loss design but also in temporal regime, graph type, and encoder architecture.

Method	Temporal regime	Contrastive mechanism
teneNCE	Discrete-time snapshots	Local/global CPC over future latent states
CLP	Snapshot-based temporal heterogeneous network	Hierarchical node-, edge-, and time-level contrast
CoDCL	Continuous-time dynamic network	Counterfactual edge-level InfoNCE augmentation

teneNCE uses a recurrent message-passing architecture. Each snapshot is encoded by a 3-layer GCN into $G_1,\dots,G_k$2, then updated through a Graphical GRU: $G_1,\dots,G_k$3 with fixed time encoding $G_1,\dots,G_k$4. Current graph reconstruction and next-step link prediction both use linear projection, inner product, and sigmoid; the contrastive part predicts future latent node embeddings and future graph summaries from the current recurrent state (Nouranizadeh et al., 2024).

CLP is specialized to temporal heterogeneous networks. For each relation-specific snapshot subgraph $G_1,\dots,G_k$5, it builds attentive typed embeddings $G_1,\dots,G_k$6 and unified pooled embeddings $G_1,\dots,G_k$7, then fuses relation types with attention to obtain $G_1,\dots,G_k$8 and $G_1,\dots,G_k$9. Temporal evolution is modeled through dual channels: $G=(\mathcal{V},\mathcal{E},\mathcal{T})$0 and the final link score is

$G=(\mathcal{V},\mathcal{E},\mathcal{T})$1

Its contrastive design is therefore deeply tied to heterogeneity modeling rather than being an external regularizer (Tai et al., 2024).

CoDCL is presented as a plug-and-play module that can be integrated into temporal backbones such as TGAT, GraphMixer, and DyGFormer, and in experiments is attached to DyGFormer, FreeDyG, and CorDGT. The backbone computes temporal node embeddings $G=(\mathcal{V},\mathcal{E},\mathcal{T})$2; CoDCL adds time-constrained neighborhoods, treatment computation, structural neighborhood exploration, counterfactual pair matching, and edge-level contrastive training, without changing the internal memory or attention equations of the backbone (Feng et al., 30 Jan 2026).

5. Empirical profile

The direct empirical case for TCLP rests on three distinct result patterns: gains from auxiliary temporal self-supervision, gains from hierarchical heterogeneity-aware contrast, and gains from counterfactual augmentation.

teneNCE reports best results across Enron, COLAB, and Facebook on the standard Rand-Pos/Rand-Neg protocol. On Enron it reports AUC $G=(\mathcal{V},\mathcal{E},\mathcal{T})$3, AP $G=(\mathcal{V},\mathcal{E},\mathcal{T})$4, and MRR $G=(\mathcal{V},\mathcal{E},\mathcal{T})$5; on COLAB, AUC $G=(\mathcal{V},\mathcal{E},\mathcal{T})$6, AP $G=(\mathcal{V},\mathcal{E},\mathcal{T})$7, and MRR $G=(\mathcal{V},\mathcal{E},\mathcal{T})$8; on Facebook, AUC $G=(\mathcal{V},\mathcal{E},\mathcal{T})$9, AP $(u,v,t)\in\mathcal{E}$0, and MRR $(u,v,t)\in\mathcal{E}$1. Its ablation sequence is especially diagnostic: link prediction only, then $(u,v,t)\in\mathcal{E}$2 reconstruction, then $(u,v,t)\in\mathcal{E}$3 localNCE, then $(u,v,t)\in\mathcal{E}$4 globalNCE yields monotonic improvement, with local contrastive learning producing the largest boost and global contrastive learning adding further gains (Nouranizadeh et al., 2024).

CLP reports the best AUC and AP on four temporal heterogeneous datasets: Math-overflow, Taobao, OGBN-MAG, and COVID-19. The reported scores are AUC/AP $(u,v,t)\in\mathcal{E}$5, $(u,v,t)\in\mathcal{E}$6, $(u,v,t)\in\mathcal{E}$7, and $(u,v,t)\in\mathcal{E}$8, respectively, with an average improvement of $(u,v,t)\in\mathcal{E}$9 in AUC and $\mathcal{H}_t$0 in AP over the second-best model. Its ablations show that removing edge-level loss produces average degradation of AUC $\mathcal{H}_t$1, AP $\mathcal{H}_t$2; removing time-level loss yields AUC $\mathcal{H}_t$3, AP $\mathcal{H}_t$4; removing node-level loss yields AUC $\mathcal{H}_t$5, AP $\mathcal{H}_t$6. These results support the claim that the contrastive terms are not decorative regularizers but materially shape predictive performance (Tai et al., 2024).

CoDCL reports mostly consistent AP improvements when attached to strong continuous-time backbones in both transductive and inductive settings over Wikipedia, UCI, Enron, MOOC, Reddit, LastFM, and CanParl. The paper summarizes gains as ranging from $\mathcal{H}_t$7 to $\mathcal{H}_t$8 across some settings and emphasizes stronger benefits on communication or irregular temporal networks. The ablation study identifies counterfactual learning as the dominant contributor, with removals causing the largest drops on datasets such as UCI and CanParl; it also shows that temporal-aware treatments such as Dynamic Interaction, Temporal Proximity, and Activity Synchrony are often strongest (Feng et al., 30 Jan 2026).

Taken together, these results indicate that TCLP gains do not arise from a single recipe. Improvement can come from making current states predictive of future latent structure, from contrasting heterogeneous multi-granular views, or from contrasting factual and counterfactual temporal edge contexts. The empirical commonality is that auxiliary temporal structure learning improves downstream future-edge prediction beyond purely supervised training (Nouranizadeh et al., 2024, Tai et al., 2024, Feng et al., 30 Jan 2026).

6. Relation to non-contrastive temporal link prediction

A substantial portion of the literature most useful to TCLP is not itself contrastive. These papers define stronger encoders, better supervision targets, or more rigorous evaluation regimes.

“Temporal Link Prediction Using Graph Embedding Dynamics” models node embeddings as Newtonian positions and defines per-node velocity

$\mathcal{H}_t$9

then predicts future velocity with an LSTM, aggregates recent velocities, smooths with one-hop neighborhoods, and scores future links by Euclidean similarity. The method is explicitly non-contrastive, but its node-wise temporal evolution and interpretability layer are natural precursors for trajectory-aware TCLP (Fard et al., 2024). “Improving Temporal Link Prediction via Temporal Walk Matrix Projection” is also non-contrastive, yet it unifies relative encodings through temporal walk matrices, introduces time-decayed walk scores

$P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$0

and maintains projected temporal walk matrices efficiently via random feature propagation, with a reported maximum speedup of $P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$1 over the SOTA baseline and $P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$2 over CAWN on LastFM. This provides a principled pair-aware encoder that could be reused in TCLP objectives (Lu et al., 2024).

Other non-contrastive papers illuminate adjacent axes of the problem. “Transfer Learning for Temporal Link Prediction” isolates the graph-specificity of memory-based TLP and learns a 3-layer MLP mapping from structural features to TGN memory embeddings for zero-shot memory initialization on unseen graphs; although not contrastive, it frames structure-to-memory alignment as the central transfer bottleneck (Chatterjee et al., 15 Apr 2025). “Link-aware link prediction over temporal graph by pattern recognition” is query-aware rather than node-centric: it constructs a query-conditioned sequence of historical links, transforms it into an attention tensor over link-link relations, and classifies temporal patterns ending in the query link, achieving AUC $P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$3, $P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$4, $P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$5, $P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$6, $P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$7, and $P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$8 on UCI, Social Evolution, Enron, Wikipedia, Lastfm, and MOOC (Liu et al., 2024). “HGWaveNet” contributes a hyperbolic spatio-temporal encoder with Hyperbolic Diffusion Graph Convolution, Hyperbolic Dilated Causal Convolution, and a Fermi–Dirac decoder, reporting relative AUC improvement by up to $P(y_{uv}(t+\Delta t)=1\mid \mathcal{H}_t).$9 on temporal new link prediction over SOTA methods (Bai et al., 2023).

Two further works expand the conceptual scope of TCLP. “NPGLM: A Non-Parametric Method for Temporal Link Prediction” treats temporal link prediction as time-to-event modeling with censoring, learning a conditional arrival-time distribution $$L = L_{\textrm{pred}} + \alpha L_{\textrm{recon}} + \beta L_{\textrm{cpc}},$$0 rather than only a binary horizon label; its main value is the survival-analysis formulation and the treatment of unformed links as censored rather than naive negatives (Sajadmanesh et al., 2017). “Generating Causal Temporal Interaction Graphs for Counterfactual Validation of Temporal Link Prediction” is not a learning method but a causality-aware evaluation framework that generates continuous-time temporal interaction graphs with known ground-truth causal structure and defines a distance $$L = L_{\textrm{pred}} + \alpha L_{\textrm{recon}} + \beta L_{\textrm{cpc}},$$1 between generating models, enabling counterfactual validation under controlled causal shifts and timestamp shuffling (Rahman et al., 2 Feb 2026).

7. Limitations, misconceptions, and open directions

The present TCLP literature is heterogeneous, and its limitations are correspondingly varied. A first limitation is temporal regime fragmentation. CLP and teneNCE are snapshot-based and thus trade event-level precision for tractable sequence modeling; CoDCL operates in continuous time but introduces additional counterfactual search overhead. These are not interchangeable settings, and conclusions about one do not automatically transfer to the other (Tai et al., 2024, Nouranizadeh et al., 2024, Feng et al., 30 Jan 2026).

A second limitation concerns generalization. Snapshot-based methods discussed here are primarily transductive over fixed node universes, and CLP explicitly removes newly appearing nodes in $$L = L_{\textrm{pred}} + \alpha L_{\textrm{recon}} + \beta L_{\textrm{cpc}},$$2-th snapshots. The transfer paper shows that in memory-based continuous-time models, the central failure mode is not only encoder generalization but graph-specific hidden state, since unseen nodes otherwise start from zero memory (Tai et al., 2024, Chatterjee et al., 15 Apr 2025).

A third limitation concerns computational and methodological stability. teneNCE accumulates CPC over all $$L = L_{\textrm{pred}} + \alpha L_{\textrm{recon}} + \beta L_{\textrm{cpc}},$$3, so contrastive cost grows with sequence length; CoDCL states that the dominant cost is counterfactual pair generation with complexity $$L = L_{\textrm{pred}} + \alpha L_{\textrm{recon}} + \beta L_{\textrm{cpc}},$$4; and the transfer paper reports divergent total loss and strong random-seed sensitivity for structure-to-memory mapping (Nouranizadeh et al., 2024, Feng et al., 30 Jan 2026, Chatterjee et al., 15 Apr 2025).

A fourth limitation concerns the semantics of augmentations and negatives. CoDCL’s treatment construction is heuristic rather than a formal identified causal model, while causality-aware benchmarking shows that timestamp shuffling can be interpreted as a measurable causal distortion rather than an innocuous corruption. This suggests that TCLP should not assume all invariances are desirable; invariance to causally consequential temporal changes may be harmful, not beneficial (Feng et al., 30 Jan 2026, Rahman et al., 2 Feb 2026).

Several open directions follow directly from these observations. A plausible implication is that TCLP will benefit from integrating stronger temporal state variables, such as node-wise latent velocity, into the contrastive target or projection head (Fard et al., 2024). Another plausible implication is that direct regression between structural and temporal states, as in structure-to-memory mapping, could be replaced or augmented by explicit contrastive alignment for cross-graph transfer (Chatterjee et al., 15 Apr 2025). A further implication is that binary future-edge supervision may be too narrow for some applications, and that survival-style heads over contrastively learned pair representations could extend TCLP from “whether” to “when” an edge appears (Sajadmanesh et al., 2017). Finally, causality-aware benchmarks indicate that future TCLP work may need to evaluate not only AUC or AP, but also whether the learned invariances survive counterfactual shifts in temporal interaction mechanisms (Rahman et al., 2 Feb 2026).