Resource-Centric Next-Activity Prediction

• Resource-centric next-activity prediction is defined as forecasting the next activity based on a resource’s temporally ordered execution history rather than a case instance.
• It employs encoding strategies such as 2-gram transitions and repetition features to capture work behavior from event logs and assess model performance via accuracy metrics.
• Empirical evaluations on BPIC logs show that models like Transformer with repetition encoding yield high predictive accuracies and improved generalization over traditional case-centric methods.

Resource-centric next-activity prediction is the formulation of next-activity prediction from the viewpoint of an individual resource rather than a case: instead of asking which activity will occur next in a process instance, it asks which activity a worker, employee, agent, or other resource will perform next on the basis of that resource’s temporally ordered execution history. In predictive process monitoring, this perspective complements the conventional case-centric or control-flow perspective and is motivated by improved work organization, workload balancing, capacity forecasting, smarter resource allocation, strategic workforce planning, and personalized employee support (Kurowski et al., 26 Aug 2025). Related work places it at the intersection of context-enriched sequence modeling, predictive resource allocation, and prescriptive process monitoring (Park et al., 2019).

1. Conceptual scope and formalization

In the conventional case-centric setting, a case trace $\sigma$ is a sequence of events with associated activities $\langle a_1, a_2, \dots, a_T \rangle$, and next-activity prediction estimates

$$\hat{a}_{T+1} = \arg\max_{a \in A} P(a \mid a_1, \dots, a_T).$$

A resource-centric formulation shifts the unit of analysis from the case to the resource. The corresponding prediction target is the next activity executed by a specific resource $r$:

$$\hat{a}_{t+1}^{(r)} = \arg\max_{a \in A} P\big(a \mid a_1^{(r)}, a_2^{(r)}, \dots, a_t^{(r)}, r\big).$$

This formulation was stated explicitly in a 2025 study that framed the topic as a distinct line of predictive process monitoring and argued that, although resource information had already been shown to enhance process performance analysis, its role in next-activity prediction had remained unexplored (Kurowski et al., 26 Aug 2025).

A broader formal vocabulary for this setting already existed in context-enriched predictive business process monitoring. Events can be modeled as tuples $e = (a, c, t, D)$ over activities, case identifiers, timestamps, and context attributes, with categorical context attributes naturally accommodating resource identifiers, roles, or organizational units (Weinzierl et al., 2020). This established a technical basis for treating resources not as auxiliary metadata but as predictive covariates. Resource-centric prediction departs from that tradition by reindexing the sequence itself around the resource.

The perspective change is substantive rather than cosmetic. Case-centric models primarily support process-state forecasting, whereas resource-centric models target work behavior. This suggests a different operational semantics: the model is less about the next control-flow step in a case and more about the near-future workload profile of a human or machine actor.

2. Resource-centric event-log construction

The standard starting point is a case-centric event log

$$L = \{ \sigma_1, \sigma_2, \dots, \sigma_n \}, \quad \sigma_i = \langle e_{i1}, e_{i2}, \dots, e_{iT_i} \rangle,$$

where each event carries at least a case identifier, an activity label, a resource identifier, and a timestamp. Resource-centric prediction transforms this log by grouping events by resource:

$L^R = \{ \tau_r \mid r \in R \},$

with

$$\tau_r = \langle e_1^{(r)}, e_2^{(r)}, \dots, e_{T_r}^{(r)} \rangle,$$

ordered by timestamp and satisfying $\mathit{res}(e_i^{(r)}) = r$ for all events in the sequence (Kurowski et al., 26 Aug 2025).

The resulting activity sequence for resource $\langle a_1, a_2, \dots, a_T \rangle$0 is

$\langle a_1, a_2, \dots, a_T \rangle$1

Training instances are then generated from prefixes: for a chosen prefix length $\langle a_1, a_2, \dots, a_T \rangle$2, each resource with $\langle a_1, a_2, \dots, a_T \rangle$3 contributes an input prefix $\langle a_1, a_2, \dots, a_T \rangle$4 and label $\langle a_1, a_2, \dots, a_T \rangle$5 (Kurowski et al., 26 Aug 2025). Because cases can overlap in time, the resource sequence interleaves work from many cases; the ordering is purely temporal at the resource level.

The 2025 benchmark used four real-life BPIC logs selected to ensure at least 100 unique resources, which in turn guaranteed at least 20 test resources. Their reported profiles make clear that resource-centric traces differ sharply across domains (Kurowski et al., 26 Aug 2025).

Log	Unique resources	Avg seq len / specialization
BPIC2013 Incidents	1,440	40.51 / 0.34
BPIC2017	149	8,068.91 / 0.31
BPIC2018	165	15,237.97 / 0.39
BPIC2019	628	2,541.28 / 0.78

The same study defined resource specialization through Shannon entropy over the activity distribution of each resource, normalized and inverted to obtain a score in $\langle a_1, a_2, \dots, a_T \rangle$6, where $\langle a_1, a_2, \dots, a_T \rangle$7 denotes a highly specialized resource (Kurowski et al., 26 Aug 2025). This creates a direct empirical bridge between behavioral regularity and predictability.

3. Encoding strategies and predictive models

The central methodological question is how to encode a resource’s behavioral history. The resource-centric benchmark compared a baseline sequence encoding with three explicit feature strategies: capability, 2-gram transitions, and repetition (Kurowski et al., 26 Aug 2025).

The baseline uses only the activity sequence prefix. Capability encoding appends a binary “can do” vector defined for each activity $\langle a_1, a_2, \dots, a_T \rangle$8 by

$\langle a_1, a_2, \dots, a_T \rangle$9

The authors note that, from a strict machine-learning standpoint, this can be viewed as information leakage because capabilities are computed on the full log, but argue that it is realistic operationally because organizations know what tasks an employee can perform from job descriptions (Kurowski et al., 26 Aug 2025).

The 2-gram encoding introduces explicit transition counts. For each ordered pair $$\hat{a}_{T+1} = \arg\max_{a \in A} P(a \mid a_1, \dots, a_T).$$0,

$$\hat{a}_{T+1} = \arg\max_{a \in A} P(a \mid a_1, \dots, a_T).$$1

Because the raw feature space is of size $$\hat{a}_{T+1} = \arg\max_{a \in A} P(a \mid a_1, \dots, a_T).$$2, the study computed mutual information between each transition feature and the next-activity class, then retained the top 20 features (Kurowski et al., 26 Aug 2025).

The repetition encoding extends the 2-gram representation with run-length structure. For the set of maximal runs $$\hat{a}_{T+1} = \arg\max_{a \in A} P(a \mid a_1, \dots, a_T).$$3 in a prefix, the reported feature set includes the average run length

$$\hat{a}_{T+1} = \arg\max_{a \in A} P(a \mid a_1, \dots, a_T).$$4

along with an additional run-related feature such as the number of runs in the prefix (Kurowski et al., 26 Aug 2025). This makes repetition encoding explicitly sensitive to activity “stickiness,” that is, whether a resource tends to remain on the same activity or switch frequently.

The benchmark evaluated four predictive models. Random Forest and LightGBM consumed the fixed-length encoded feature vectors. The LSTM architecture used two stacked LSTM layers, each with 50 units, followed by dropout of 0.2. The Transformer used 128-dimensional activity embeddings, positional encoding, two encoder layers, four attention heads, hidden size 128, dropout 0.1, and a final fully connected classification layer (Kurowski et al., 26 Aug 2025). Accuracy was the main evaluation metric.

A related comparison of context-enriched next-activity prediction in case-centric logs reported that encoding choice affected predictive quality more than architecture type, with hash encoding best overall and LSTM best among the tested architectures (Weinzierl et al., 2020). This suggests that, even in the resource-centric setting, representation engineering is likely to dominate incremental architecture changes unless the model is explicitly redesigned around resource structure.

4. Empirical performance and evaluation characteristics

The most direct empirical result is that LightGBM and Transformer performed best with an encoding based on 2-gram activity transitions, while Random Forest benefited most from the encoding that combined 2-gram transitions and activity repetition features; the combined encoding also achieved the highest average accuracy overall (Kurowski et al., 26 Aug 2025). Capability encoding generally tracked the baseline and sometimes underperformed it.

The reported averages show that repetition-based encoding was the strongest overall representation, outperforming 2-gram encoding in 9 dataset-model settings, with 2-gram outperforming repetition in 4 settings and one tie (Kurowski et al., 26 Aug 2025). This suggests that local transition structure and explicit run-length behavior are especially informative when the predictive unit is the resource rather than the case.

Model-wise, Transformer achieved the highest peak accuracies on complex datasets but was sensitive to prefix length and encoding, whereas Random Forest was described as the most stable and robust across datasets and prefixes. LSTM underperformed overall in this benchmark despite its standard suitability for sequential prediction (Kurowski et al., 26 Aug 2025). The strongest absolute results concentrated on BPIC2019, where most models were in the 0.85–0.95 range and the Transformer with repetition reached $$\hat{a}_{T+1} = \arg\max_{a \in A} P(a \mid a_1, \dots, a_T).$$5; BPIC2017 was more difficult, with Random Forest plus repetition reaching $$\hat{a}_{T+1} = \arg\max_{a \in A} P(a \mid a_1, \dots, a_T).$$6 (Kurowski et al., 26 Aug 2025).

The same study also compared resource-centric and case-centric evaluation characteristics. Example leakage, measured following the concern that commonly used event logs can exhibit substantial overlap between train and test prefixes, was 3%, 0.3%, 3%, and 14% in the resource-centric setting for BPIC2013 Incidents, BPIC2017, BPIC2018, and BPIC2019, compared with 83%, 80%, 12%, and 85% in the corresponding case-centric setting (Kurowski et al., 26 Aug 2025). Earlier work had already argued that commonly used case-centric benchmarks suffer from “an enormous amount of example leakage,” and that trivial prediction approaches can perform almost as well as deep models under such conditions (Abb et al., 2023). The contrast indicates that resource-centric prediction is empirically less redundant and therefore a more demanding test of generalization.

The benchmark further proposed that average specialization and sequence length per resource can serve as indicators of predictive performance: BPIC2019 combined the highest specialization score, 0.78, with the best overall accuracies, whereas BPIC2013, BPIC2017, and BPIC2018 showed lower specialization and more variable performance (Kurowski et al., 26 Aug 2025). This is an interpretive claim in the source material, but it aligns with the broader observation that resource-level regularity drives forecastability.

5. From forecasting to resource allocation and prescriptive monitoring

Resource-centric next-activity prediction is closely related to prescriptive process monitoring because predictions about what a resource will do next can be fed into allocation and control mechanisms. A 2019 study on prediction-based resource allocation explicitly proposed a method that predicts both processing time and next activity of an ongoing instance using Bayesian Neural Networks and then performs online resource allocation in a non-clairvoyant environment with an extension of the minimum cost and maximum flow algorithm (Park et al., 2019). The reported setting remained case-centric in its prediction target, but it already connected next-activity forecasting to concrete online scheduling decisions.

Subsequent work in prescriptive monitoring shifted from forecasting what will happen next to recommending what should happen next. One reinforcement-learning approach learned an optimal policy from past executions and recommended the best activities to carry on for optimizing a KPI from the perspective of one process actor (Branchi et al., 2022). Another introduced goal-oriented next best activity recommendation, using a deep learning model to predict the next best activity and an estimated value of a goal, and an RL method to explore conformant sequences likely to meet one or more goals (Agarwal et al., 2022). These formulations are not resource-centric in the narrow sense of predicting a resource’s own next activity history, but they show how predictive signals can be embedded into prescriptive control.

Graph-based next-activity prediction in service-oriented architectures has also made the resource link explicit. RLHGNN defines an event as $$\hat{a}_{T+1} = \arg\max_{a \in A} P(a \mid a_1, \dots, a_T).$$7, where the additional attributes explicitly include the executing resource, and motivates next-activity prediction as enabling proactive resource allocation and dynamic service composition (Wang et al., 3 Jul 2025). In the current RLHGNN formulation, resources are attributes rather than graph nodes. This suggests a natural extension of resource-centric prediction toward heterogeneous graphs in which resources become first-class entities.

Taken together, these strands define a pipeline: resource-centric prediction forecasts the immediate behavioral horizon of actors; prescriptive monitoring turns that forecast into a recommendation or allocation decision; and optimization machinery enforces capacity, conformance, and cost constraints.

6. Explainability, generalization, and research directions

Explainability remains underdeveloped in explicitly resource-centric settings. XNAP introduced an explainable, LSTM-based next-activity predictor using layer-wise relevance propagation, but encoded only activities, not resources or other attributes. Its authors explicitly identified future work to adapt the propagation rules so as to determine relevance values of context attributes, which includes resources (Weinzierl et al., 2020). For resource-centric prediction, this implies that attribution should eventually be able to answer not only which past activities drove the forecast, but also which past resource states or organizational attributes did so.

Knowledge-guided modeling addresses a different weakness: under-sampling and drift. A Symbolic[Neuro] approach combined an attention-based neural predictor with background process knowledge expressed as a procedural process model, using compliance scores to modulate predictions and improve performance on exceptional executions or when the process changes due to concept drift (Donadello et al., 2023). That system operated on the control-flow perspective only. This suggests a parallel resource-centric direction in which organizational rules, role constraints, or resource-capability models modulate predictions about future resource behavior.

Several neighboring research lines show how the topic may broaden. Object-centric predictive process monitoring with graph embeddings focuses on next activity and next event time in object-centric event logs and has been described as naturally adaptable to settings where resources are treated as first-class objects (Gherissi et al., 21 Jul 2025). Semantic language-model approaches build textual “stories” from event logs for next-activity prediction and already incorporate role information where available (Oved et al., 2024). Entropy-guided model selection with attribute-aware Transformers explicitly lists resource as a context-specific feature and argues for selecting between complex and simpler models based on process entropy (Zare et al., 14 Feb 2025). In a different application domain, LLM-based temporal reasoning in smart environments showed that activity prediction with duration estimation can be achieved from compact contextual cues, with strong zero-shot behavior and refinement from one or two demonstrations (Doctorarastoo et al., 20 Jan 2026). These are not direct resource-centric BPM results, but they indicate that resource-aware forecasting can plausibly benefit from semantic context, object-centric structure, and few-shot temporal priors.

A stable research agenda therefore emerges. One line concerns representation: whether resource histories are best modeled with transition counts, repetition features, sequence encoders, heterogeneous graphs, or semantic prompts. A second concerns supervision: whether the target should remain the next activity of a resource, or be extended to joint prediction of next activity and next resource, or further to prescriptive recommendations. A third concerns evaluation: the low example leakage reported for resource-centric logs indicates that the setting is better suited than many case-centric benchmarks for studying actual generalization (Kurowski et al., 26 Aug 2025). A plausible implication is that resource-centric next-activity prediction will become an important testbed not only for operational workforce analytics but also for methodological progress in predictive process monitoring more broadly.