LiFTER: A Grounded Neuro-Symbolic Microscope for Continuous-Time Dynamic Graph Forecasting

TGN updates temporal state through node memory and message passing, while TGAT uses time encoding and temporal attention. GraphMixer processes neighborhoods with temporal link encoding and MLP mixing, and DyGFormer compresses interaction sequences into patches [8, 13, 3, 17]. CRAFT applies cross-attention to source history with the candidate destination as the query; CRAFT-R further incorporates repeat counts [15]. These approaches offer strong GPU-friendly forecasting with continuous representations. LiFTER is evaluated under the same protocol but constrains the score to a sum of grounded rule executions.

EdgeBank showed that repeated-edge memory can be strong under random negatives [7]. Heuristics based only on recency and popularity also compete with neural models on several temporal benchmarks [4], and representative models deteriorate sharply on the sequential dynamics of TGB-Seq, which reduces repeated edges [16]. These findings establish recurrence and order as central structures that a prediction language should represent, rather than incidental metadata.

T-GNNExplainer searches pre-query events for a prediction-preserving subset, and TempME uses temporal motifs as explanation units [11, 2]. TGIB learns sparse evidence through an information bottleneck, while SIG constructs a self-interpretable temporal subgraph [9, 5]. Their outputs are importance scores, masks, or subgraphs. LiFTER instead returns the historical facts, variable bindings, temporal conditions, and signed contributions actually used in score computation. This output is not a surrogate and accounts for the entire score apart from the prior.

Neural-LP learns differentiable inference over a finite first-order rule space [14]. TLogic converts temporal random walks that preserve timestamp order into rules, and TILP learns temporal intervals and confidence on relation-typed TKGs [6, 12]. CAW also structures temporal walks through recurring entity positions [10]. In this line of work, the designer supplies a language bias that specifies admissible predicates, variable bindings, and maximum body structure, rather than dataset-specific individual rules. LiFTER follows the same principle: it enumerates all admissible combinations and learns their weights and temporal compatibility.

The symbolic primitives differ. A TILP rule is identified by an observed semantic predicate sequence and temporal constraints; in a single-relation CTDG, that sequence collapses to repetitions of one predicate. LiFTER fills this gap with argument roles relative to query endpoints, same-pair recurrence, source-local position, and transitions from recent destination states to a candidate. Preserving a raw interaction as a ground atom does not invent semantic relations; it supplies the representation on which this language executes.

The output contract also differs. In LiFTER, every final logit term other than the prior must be an enumerated execution with cited facts and a signed contribution, and it must be independently replayable from raw history. A direct numerical comparison with TILP would require an artificial conversion from CTDG interactions to a relation-typed interval TKG, changing both the input semantics and prediction protocol. The formulation-level distinction is therefore more faithful than such an adaptation.

An interaction $(u_{i},v_{i},t_{i})$ records entity $u_{i}$ interacting with entity $v_{i}$ at time $t_{i}$. LiFTER represents it as $f_{i}=\operatorname{Link}(u_{i},v_{i},t_{i})$, with source, destination, and timestamp as the first, second, and temporal arguments. For example, user 5619 editing page 949 at timestamp 2352512 becomes the fact $\operatorname{Link}(5619,949,2352512)$.

This representation permits direct substitution of historical entities for rule variables. For query $\operatorname{Link}(X,Y,T_{q})$, the fact $\operatorname{Link}(5619,949,t_{1})$ binds $X=5619$ and $Y=949$. If the same user edited page 103, the fact $\operatorname{Link}(5619,103,t_{2})$ binds $X=5619$ and $Z=103$. The shared variable $X$ receives the same integer identifier 5619 in both facts, preserving the structure that the two interactions share a source.

The local fact database for a query contains the most recent $H$ facts adjacent to its source and candidate endpoints. Thus $H$ bounds the number of entity substitutions examined per query. The selected facts retain their timestamps, from which the executor computes event order, time to query, and recurrence intervals.

LiFTER rules specify which entities must be equal across a query and historical facts, and in what temporal order the events must occur. In the forecasting experiments of Section 5.1, all interactions share one observed predicate $P_{1}$. The typed extension generalizes it to $P_{k}$, $k\in\{1,\ldots,K\}$. Figure 2 illustrates four rule schemas.

Endpoint-bound unary rules. These rules test the six orientations in which one historical fact shares an endpoint with the query source or destination.

Equation (2) distinguishes, for example, a historical $X\rightarrow Y$ interaction from $X\rightarrow Z$ and $Z\rightarrow Y$ roles. Equality with a query argument, rather than the entity identifier itself, determines the role, yielding $6K$ clauses.

Pair-renewal rules. A query pair that occurred twice is represented as

The rule compares the previous recurrence interval $T_{2}-T_{1}$ with the gap from the last interaction to the query, $T_{q}-T_{2}$. The language includes all $K^{2}$ predicate pairs.

Positioned recurrence rules. These rules retain the position of a historical interaction with the same pair in the source history.

In Eq. (4), $j=1$ denotes the most recent fact. Rather than preselecting a position, the language includes every $1\leq j\leq H$, and training determines its signed weight.

Ordered grounded transitions. These rules evaluate the next candidate from the destinations reached by recent outgoing facts of source $X$. The two-event body is

The one-event schema uses only the second fact in Eq. (5). Distinct timestamps determine temporal order; equal timestamps are treated as simultaneous recent facts. The executor first finds actual facts satisfying Eq. (5), and only then computes the compatibility from grounded destinations $Z_{1},Z_{2}$ to candidate $Y$.

A transition score is the learned compatibility between grounded destinations and a candidate. If source $X$ recently interacted with $Z$, the one-event transition asks whether $Y$ follows $Z$. If recent interactions occurred in the order $Z_{1},Z_{2}$, the two-event transition asks whether $Y$ follows that ordered pair. Here $a_{z}$ and $b_{y}$ are $d$-dimensional vectors for a historical destination and candidate, and $p_{k}$ is a predicate factor. The two-event schema multiplies three vectors so that its value is large when both historical destinations support the candidate in the same latent dimension. Equations (6a) and (6b) define these computations.

Here $r_{H}(X,Y)$ indicates whether the query pair appears in the recent source history, and $s_{m,r}=\operatorname{softplus}(\rho_{m})\operatorname{softplus}(\eta_{m,r})/\log 2$ is a learned positive scale. Equation (6a) evaluates a transition from one destination, while Eq. (6b) evaluates a transition from an ordered pair. These vectors are accessed only when a grounding satisfies Eq. (5); neural compatibility alone cannot create an execution. The two potentials contain $5Nd+3Kd+6$ parameters.

For each clause, the executor first finds historical facts that satisfy its body. Because facts satisfying the same rule can have different gaps to the query, Eq. (7) measures how closely each grounding matches the rule’s learned preferred gap.

The kernel lies in $[0,1]$ and approaches one at the preferred gap. Parameter $\mu_{r}$ is rule $r$’s preferred log-time gap, and $\operatorname{softplus}(\lambda_{r})+0.05$ is its tolerance. Renewal rules receive the difference between the previous recurrence interval and the current gap instead of absolute recency. Aggregating all grounding values yields $E_{r}$. Direct-pair unary rules use a maximum to avoid repeatedly counting the same pair; other rules use the capped sum $E_{r}=\min(e^{4},\sum_{g}e_{r}(g,q))$. If no valid grounding exists, $E_{r}=0$.

The candidate logit contains no hidden residual scorer beyond this grounded evidence and is computed as

In Eq. (8), $b$ is the global prior, $w_{r}E_{r}$ represents unary and renewal rules, $u_{j,k}R_{j,k}$ represents positioned recurrence, and $c_{\mathrm{tr},m}$ is an ordered-transition contribution. The weights $w_{r}$ and $u_{j,k}$ and the transition terms may be positive or negative, thereby supporting or inhibiting a candidate. Every nonzero term identifies the facts and conditions that produced it. Summing the prior and all signed contributions exactly reconstructs the logit. Figure 3 follows this accounting from facts to score.

All rule parameters are trained jointly through future-link prediction. The objective raises the positive-candidate logit $s^{+}$ and lowers the sampled-negative logit $s^{-}$:

Rather than assigning every clause a wholly independent weight, clauses with the same structural schema and predicates share the parameters in Eq. (10).

Here $t_{\tau(r)}$ represents a schema, $q_{k_{s}}$ represents a predicate slot in the body, and $\theta_{r}$ is a clause-specific residual. We add $\lambda_{s}\sum_{r}|\theta_{r}|$ to the prediction loss and use a type-usage entropy regularizer only when $K>1$.

Finite histories and batched grounding bound execution cost. After indexing the global event stream once, each batch gathers recent fact tensors once. Unary and renewal clauses use parallel tensor operations, while Eqs. (6a) and (6b) use batched embedding operations. The full stream is never rescanned per query.

The JODIE streams evaluated in Section 5.1 provide one observed interaction type, so LiFTER uses $K=1$. All facts then share a predicate but remain distinct ground atoms through their entity arguments and timestamps.

For datasets in which multiple event semantics are distinguishable from observed context, LiFTER also supports an optional $K>1$ vocabulary. A fact encoder maps the pre-event context of event $i$ to $z_{i}$ and selects a predicate by similarity to learned prototypes $c_{k}$:

Forward execution uses the single discrete atom $P_{k_{i}^{*}}(u_{i},v_{i},t_{i})$ selected by Eq. (11a). Equation (11b) preserves this discrete choice while passing gradients from Eq. (9) to the encoder and prototypes. Section 5.3 and Appendix A establish the conditions under which $K>1$ is useful.

We chronologically split the first 32,768 events of Wikipedia, Reddit, MOOC, and LastFM into 85% training and 15% testing. All forecasting models are evaluated for 10 epochs with seeds 7, 17, and 27. Across all four datasets, LiFTER uses a predicate vocabulary $K=1$, hidden dimension 64, transition rank 32, dropout 0.1, and history length 128. The batch size is 512; AdamW uses a learning rate of 0.004 and weight decay $10^{-5}$. Forecasting, component, and explanation results report the mean and sample standard deviation across three seeds. Deterministic certificate verification covers the complete test split of the seed-7 model for each dataset; hardware latency is reported separately as the mean of repeated runs.

Grounding capacity $H$ is selected by chronological validation disjoint from the held-out test interval. After reserving the final 15% for testing, we use the last 15% of the preceding 85% development prefix for validation. We apply each $H\in\{10,20,40,80\}$ to all datasets and seeds and select the value with the highest macro mean validation Historical AUC, breaking ties in favor of smaller $H$. Table I selects $H=10$, which is fixed for every test evaluation. Although MOOC alone peaks at $H=20$, the difference is only 0.02 percentage points. A shared capacity controls both predicate-position grammar size and per-query grounding cost.

Historical-negative evaluation. For query $(X,Y,T_{q})$, we uniformly sample a historical negative from destinations that $X$ interacted with before $T_{q}$, excluding the current positive $Y$. A random negative is often an entity never observed with the source and is therefore easily separated through source preference, exposure, or direct recurrence. Historical positives and negatives both appeared in the source history; the task asks whether recency, intervals, history position, and transitions provide stronger current evidence for the positive [7].

When no historical destination is available, we draw a random negative from the common destination pool. Historical-candidate coverage is 51.1% on Wikipedia, 48.9% on Reddit, 97.7% on MOOC, and 99.7% on LastFM. Candidates and fallbacks are identical across models and seeds. We report AUC and AP for both historical and random candidates.

Table II compares LiFTER and EdgeBank with six neural CTDG baselines under both negative protocols. LiFTER ranks first on both historical metrics for Reddit and LastFM, second on both for Wikipedia, and second in AUC and first in AP for MOOC. Relative to the strongest neural baseline, its Reddit AUC/AP is 2.05%/2.01% higher and its LastFM AUC/AP is 5.69%/9.22% higher. It trails TGN on Wikipedia by 2.75%/2.34%; on MOOC it trails GraphMixer by 0.43% in AUC but leads by 0.28% in AP.

Under random negatives, LiFTER is within one percentage point of the best AUC and AP on Wikipedia, Reddit, and MOOC. Grounded execution limits paths that reject an irrelevant random destination solely through unconstrained identity or popularity, yet the difference remains small on these datasets. On LastFM, random AUC/AP is 9.87/8.15 points below the best baseline while historical AUC/AP ranks first. This contrast localizes LiFTER’s strength to temporal discrimination among plausible alternatives previously present in the interaction history.

EdgeBank often assigns the same memory score to several candidates because a historical alternative is, by definition, an observed pair. LiFTER exceeds EdgeBank by 20.44–56.30 AUC points across the datasets. Its performance therefore cannot be reduced to the lookup “this pair existed before”; recurrence intervals, source-local position, and candidate transitions determine the ranking through a single grounded program.

LiFTER permits a latent vocabulary with $K>1$ to represent multiple functional roles within one observed relation. We test whether this expressivity aids forecasting by comparing $K=1,2,4,8$ under the same architecture and protocol and by shuffling learned assignments across facts. Grounded binding, recurrence, temporal order, and transition execution remain intact at $K=1$.

The changes from $K>1$ in Table III are small and inconsistent. Wikipedia and Reddit changes are comparable to seed variation, and $K=1$ obtains the best AP on MOOC. LastFM’s 1.08-point AP change nearly survives shuffling the fact–predicate correspondence and is therefore not an effect of functional typing. One predicate is the most concise executable representation for these four JODIE streams. This result, in which recurrence, temporal order, and exposure matter more than finer relation semantics, is consistent with findings on recency/popularity heuristics and TGB-Seq [4, 16].

Typed predicates can nevertheless add expressivity in multi-action streams when event meaning is identifiable from pre-event history and changes the future distribution. The controlled experiment that isolates these two conditions and its complete heatmap appear in the semantic-predicate phase diagram in Appendix A.

The learned $K=1$ LiFTER logit decomposes into seven disjoint components: direct pair, source context, candidate context, pair renewal, positioned recurrence, one-event transition, and two-event transition. If one fact executes multiple rules, each execution contribution is recorded under its component. Let $s_{m}(q)$ be the signed contribution of component $m$; then $s(q)=b+\sum_{m}s_{m}(q)$ holds exactly.

We evaluate all $2^{7}=128$ execution coalitions with the same historical candidates. A component’s Shapley contribution averages the AUC or AP change produced by adding it over every possible addition order. It therefore includes both isolated effects and interactions, and the seven contributions sum to the difference between the complete and empty programs. The exact definition and computation are given in Appendix C; the maximum measured residual of this identity is $1.11\times 10^{-16}$.

Figure 4 assigns the largest predictive value to pair-renewal intervals on Wikipedia and Reddit, the recent two-destination state on MOOC, and the transition from the most recent destination to a candidate on LastFM. On Wikipedia, pair renewal and one-event transition exhibit +4.28 AP points of complementarity. The one-/two-event interaction is $-9.27$ points on MOOC and $-5.10$ on LastFM, indicating competition between the two temporal resolutions on some queries.

Query regimes. We repeat component deletion after partitioning queries by direct recurrence, source activity, destination popularity, and early/late test interval. Activity and popularity are measured from the complete pre-query history, not window-truncated counts, and divided at the dataset median. Table IV localizes Wikipedia renewal to recurrent queries. Reddit uses renewal for high-activity sources and candidate context for low-activity sources. MOOC’s two-event transition contributes 20.91 AP points without recurrence, while LastFM’s one-event transition dominates in both regimes.

Grounded-fact intervention. We select 512 queries distributed throughout each dataset’s test interval and each seed. The fact with the largest absolute contribution in the dominant component is removed from the available prefix, and all rules are grounded again. The control removes a random fact matched on query, evidence bank, and recency quartile.

In Table V, top-fact deletion changes the MOOC and LastFM logits 3.63 and 9.29 times more than matched-random deletion and reduces AP by 13.38 and 9.61 points. The top Wikipedia fact also yields a larger logit change and AP drop. On Reddit, individual renewal executions have larger local logit effects without producing a dataset-level AP drop, separating local execution influence from global ranking responsibility.

We also verify signed roles. A supporting fact has $\Delta_{f}=c_{f}(q^{+})-c_{f}(q^{-})>0$ and should reduce the margin when deleted; an opposing fact has $\Delta_{f}<0$ and should increase it. Direction agreement for the top supporting facts is 99.4%/99.5% on Wikipedia/Reddit and 99.5%/93.9% for opposing facts. On MOOC/LastFM, supporting-fact deletion decreases the mean margin 5.22/8.22 times more than matched-random deletion.

We retrain variants that alter exact binding, binding schema, renewal, and transition, and compare a matched-capacity MLP receiving the same input.

In Table VI, the full program on LastFM outperforms every restricted variant and exceeds the matched MLP by 6.60 AP points. MOOC’s signal is concentrated in short ordered transitions, so transition-only execution reproduces full forecasting. This component is not an arbitrary low-rank recommender: removing exact source binding lowers AP by 1.91 points, and randomizing the prescribed binding schema lowers it by 2.23 points. The symbolic executor decides which facts satisfy binding and order and thus instantiate an execution; the neural potential determines candidate compatibility within that valid execution.

Following TempME and TGIB, we retain only selected events at explanation ratios $r\in\{.05,.10,\ldots,.30\}$ and measure agreement with the original decision. ACC-AUC is the normalized area under this accuracy–ratio curve. Following SIG, Deletion AUFSC integrates the AP reduction caused by removing selected events at the same ratios. ACC-AUC measures explanation-only sufficiency, whereas Deletion AUFSC measures ranking responsibility. Exact definitions appear in Appendix B.

For each dataset, we use the same 256 queries and three seeds. Every method receives the same evidence bank formed by the recent 10 events of the query source and candidate, with identical selection counts at each ratio. T-GNNExplainer runs explorer–navigator MCTS for 40 rollouts using the coalition reward of a frozen TGN. TempME uses the same TGN. Search-budget sensitivity at 100 and 200 rollouts and fixed budgets $k\in\{1,2,3,5,10\}$ are reported in Appendix B.

Figure 5 and Table VII show that LiFTER ranks first in both macro ACC-AUC (0.8671) and AUFSC (0.0552). It also achieves the best ACC and deletion fidelity for every fixed budget $k\leq 5$. Evidence with high ACC-AUC need not have high deletion fidelity: the former measures decision reproduction by a small set, while the latter measures disruption of the global ranking after deletion. LiFTER’s signed execution and re-grounding expose these properties separately.

Independent execution verification. Explanation metrics quantify how selected evidence affects a prediction. A separate verifier tests whether a LiFTER trace faithfully describes the claimed grounded execution. For all 19,664 test predictions across four datasets, it checks raw-fact existence, integer equality bindings, $T<T_{q}$, outgoing-fact selection for Q1/Q2, and temporal order. It then recomputes groundings and signed contributions from raw history and learned parameters and compares them with the trace and candidate logit. All predictions are reproduced within numerical tolerance $2\times 10^{-5}$. Deliberate changes to a historical fact, timestamp, bound entity, execution, or signed contribution are all detected under the five corruption types in Table VIII.

As Reddit grows from 32,768 to 672,447 events, indexing sustains 31.1–31.5k events/s and payload grows linearly from 7.5 to 153.9 MiB. The full index takes 21.63 s to construct. The query executor operates within fixed $H$ without scanning the global stream. At $H=10,K=1$, it takes 0.041 ms/query with batch 128 and reaches 101.6k queries/s with batch 2,048. Complete re-execution after a fact edit takes 0.138 ms/query; a human-readable JSON certificate takes 1.185 ms/query and averages 6.1 KiB.

Under architecture-level timing with the same recent-10 tensors and batch 512, neural predictors reach 234.0k–487.0k queries/s and LiFTER reaches 66.9k. Grounded execution incurs a cost but retains practical full-stream throughput. Intrinsic tracing takes 0.024 ms/query, approximately 3,800 times faster than T-GNNExplainer’s 91.533-ms search. Feed-forward attribution by TempME, TGIB, and SIG is faster at 0.006, 0.005, and 0.008 ms/query. Figure 6 therefore does not claim the fastest attribution; it shows that LiFTER emits an exact executable trace with its prediction and without a separate search.